Studio Node Reference

Node Reference

Every node in the Flows palette: what it does, what it takes and gives, and every parameter with its exact label, default and allowed values. Use it alongside Analysis Flows, which explains the canvas, and Cleaning and Weighting Data, which puts the preparation nodes in order.

The palette is served by the engine, so this list is the engine's own: 61 nodes in five groups — Sources (4), Prepare (16), Analyze (26), Visualize (7) and Output (8). Within each group the nodes appear here in the palette's order. Each palette item shows the node's title and a short name (the part after the dot below, e.g. crosstab); nodes that need the project database also say · platform.

---

Reading this page

Ports. Every node lists what it takes in and what it gives out, with the port's name and type:

Type	What flows through
SurveyData	the dataset plus its codebook and questionnaire — which is why tables come out labeled
Table	a frequency table, crosstab, group-means table, banner, coefficient table, t-test, correlation matrix, factor loadings, key drivers, a perceptual map's dimensions, rows and columns, price points and curves, a Trend's points…
Chart	a bar chart (a histogram and a donut too), box plot, heatmap, Likert chart, scatter plot or trend, or the Result chart of an analysis
Stat	a test result or a set of statistics (χ², Fisher's p, t, Kruskal-Wallis, Wilcoxon, Cochran's Q, a correlation, CIs, model fit), or what a Tab book wrote
Report	a report section or a whole report
Any	anything — only the Live tile input accepts every type

The Result chart's input takes a Table or a Stat, several of them: an analysis's table and its statistics.

Weights, recodes, flags and derived variables are columns inside SurveyData, not separate wires.

Parameters. "required" in the Default column means the node reports an error until you fill it in. The inspector marks every other parameter optional. Types:

Type	How you set it in the inspector
variable	a list of the variables available at this node, filtered to the scales the node accepts: each by its name, with its label and scale on the line under it (satisfaction over "Overall satisfaction · ordinal"), under headings where they come from different places ("From name" below a Data file). A list of more than ten opens with a filter box ("Filter variables…"); see Lists in the inspector. The empty entry reads — pick a variable — for a required field and — none — for an optional one. A stored variable of another scale stays shown as what the node reads, with the reason after its name in the field: q_md_score_1 "(interval — Rows takes nominal / ordinal)" — an error of the check for a codebook variable, a warning for one a node of the flow makes. A made variable has the scale the nearest node upstream that makes it gives it (a Recode of a derived variable is ratio, like its source), and the list offers it with that scale; a name the codebook has keeps the codebook's. An arm an Assign to a condition script writes is offered as a nominal variable ("assigned by a script") when the codebook does not list it. A field that can read columns the codebook does not describe lists them last, under Beside the answers: a Trend's Time offers the timestamps the survey's responses carry — created_at ("Response date (created_at)"), updated_at ("Last change (updated_at)"), started_at ("Start time (started_at)"). A field leaves out what its node refuses as soon as it is picked: a Trend's Time and Split by and a Bar chart's Split by do not offer multiple-choice questions, rankings or open answers, and a Trend's Measure variable for a mean no nominal or multiple-choice question; a stored one stays shown with the reason after it in the field (aware "(several answers: not one wave or date)")
variables (several)	a checklist of the same variables, filtered the same way; a checked variable the list would not offer stays in it, with its scale or "(not in codebook)", and can be unchecked
choice	a list of the allowed values, as wide as the field. The field shows the code chosen; in the list, where the code is a statistician's shorthand, its name is on the line under it — "Welch's ANOVA" under welch_anova, "Benjamini-Hochberg" under fdr_bh. An optional field's list starts with — default —, which leaves the default, with "the node's default: …" under it. The stored value and the generated script keep the code. A list of more than ten opens with a filter box. How the lists work: Lists in the inspector
answer code	an answer of the variable another parameter names, picked from that variable's value labels: a list of the codes, each with its label under it ("Male" under 1; — pick an answer — when the field is required, — none — otherwise), or a checklist where several answers may be checked. A t-test's Group A and Group B and a Trend's Answer codes leave the codebook's missing codes out — its missing answers and its missing_values alike (both nodes refuse a missing code there). A stored code that is not among them reads 5 "(not an answer of gender)" in a list's field and "5 is not an answer of gender" under a checklist. A code stored as text is read as the node reads it: a t-test and a Trend find their answers by their text, so a Group A of "1" shows as 1, with "Male" under it in the list; Proportion CI's answer, the Counts as yes of McNemar and Cochran's Q, a Perceptual map's Counts as yes (attributes) and Price sensitivity's Counts as would buy compare codes by type, so there text "1" is not the answer 1 and shows as given, "1" "(not an answer of gender)". Where the field reads a list of variables, the answers are those of the first one checked. When that variable has no value labels, the field is a JSON box instead
whole number, number	a number box; the placeholder shows the default
checkbox	checked = on
text	a text box; where the text names a new variable or column, the hint says "names a new variable". A few text boxes suggest values as you type — a Bar chart's Bins (auto, 10, 0, 18, 25, 35, 50, 65) and a Heatmap's Color map (theme, YlOrRd, Blues, viridis, RdBu_r) — and take any other value the node reads
JSON object, JSON	a text box that must contain valid JSON; it is read when you leave the box, and a parse error is shown under it. Where the help gives an example, the empty box shows it ([18, 30, 45, 65, 100])
condition	the Builder's condition editor, over the variables available at this node; in the inspector its variable, operator, value and ALL / ANY are lists (see Filter rows)
formula	a monospaced box, with the variables available at this node listed under it
file name	a file the node writes: a field labeled File name, the name typed between a fixed outputs/ and the file's ending (outputs/ client_q3 .xlsx); a Format list under it where the node writes several. A / in the name makes folders. The lines under it say where a run leaves the file and what is wrong with a name (see Where files go)
codeframe	a list of the project's codeframes by path (analysis/<name>.codeframe.json; — choose a codeframe — when none is chosen), with Edit codeframe…, or New codeframe… when none is chosen (see Code open answers)
upload	a file the node reads: a list, by name, of the uploads under Files the field can read, each with its size and upload date on the line under it (a data file that Files has read with its rows × columns first); a long name wraps in the list and is cut short with "…" in the field. The chosen file's size and date are also under the field, then Upload… and Type a name… (Type another name… once a file is chosen) (see Where files go)

Choices on this page are written as their code, followed by the name the list shows on the line under it: welch_anova ("Welch's ANOVA"). Look for the code in the list, with the name under it; the field shows the code alone.

The variables available at a node are the questionnaire's codebook variables, then those that nodes upstream of it make — a Recode, Derive, Index / scale, Bands, Explode multiple choice, MaxDiff scores, Cluster (k-means), Factor analysis with Add factor scores checked, Response quality, Speeders & partials, a weighting node, or a Code open answers (its Theme variable, or the name its codeframe carries) — labeled "label · made by node" (or "made by node"), then those a table the flow reads brings, labeled "from table table · made by flow". Below a Data file, the file's own columns come first, grouped under "From name" (the file's name), and the questionnaire's variables only when the file is this survey's data (see The file's columns in the flow). A variable made further down the flow is not offered: it does not exist yet when this node runs. See Parameters and variable pickers.

Fields that depend on a choice. Some parameters are read only with some choices of the node — a t-test's Groups only with Design independent, its Second measurement only with paired. The inspector shows the fields the current choices read and hides the others; each node below says which. A hidden field that still holds a value you gave it is named under the others — "Not used with these choices, and kept for when they apply: Groups (with Design = independent)." — with Clear it (Clear them for several); a condition of several parts is bracketed when there is another ("(Show = count and Sort = code and Layout ≠ histogram and Layout ≠ donut) or Layout = grouped"). Its value comes back into use when you switch back; until then it is not checked — a Group A kept after Design went paired, or a Groups naming a variable since removed, is no error, on the canvas or at Save. A field that one of the node's rules makes necessary (below) is marked required rather than optional.

Rules between parameters. A few nodes have rules about how their choices go together, written in the node's own words — for example "Tukey's HSD follows a one-way ANOVA — set Test to anova, or Post-hoc to none." They are checked as you edit, and by the engine at Check and at Save, with the code PARAM_CONFLICT. An error is a combination the node cannot run: the flow cannot run until you change it. A warning names a setting the node would ignore with the other choices, such as a post-hoc test while Significance test is off. Each node's rules are listed with it below.

One-line texts. The texts a node's card summarizes — a Report section's Heading, a Derive formula, a Live tile's Label, a path, a table or variable name — are written into a comment line of the generated script, so they must be one line. A line break (or another control character; a tab is fine) fails the engine check at Check and at Save, on that node, with the error PARAM_LINE_BREAK: "“heading” holds a line break or another control character: it must be one line of text." The flow then has no script until you fix it. A flow saved before this check with such a text is left out of a research bundle (see What is inside).

Where files go

No file is typed as a path. The flow still stores the whole path — outputs/client_q3.xlsx, assets/panel_wave2.csv — so flows saved earlier, the engine's check and the generated script read what they always read.

A file a node writes goes in outputs/, the only folder a run keeps: it appears in Files as outputs/<flow>/<file> and on the run's card. The File name field keeps outputs/ and the file's ending fixed, so you type only the name:

• The ending comes from the format. A name typed with an ending the node writes, such as tabs.xlsx, loses it and keeps the format (no tabs.xlsx.xlsx); a pasted outputs/ is dropped, and a space becomes _ as you type. An ending the node cannot write is answered under the field — for a Save report, "Reports are saved as Markdown and HTML — for a PDF, open the HTML and print it." — with a fix such as Use satisfaction.
• A name uses Latin letters (A–Z, a–z), digits, -, _ and ., with / between folders, at most 100 characters. Other characters are named — "A file name here can use letters a–z, digits, -, _ and . — not “(” or “)”." for Q3 (final) — with Use Q3_final. While the name has a problem, the field does not say where the file will be.
• Two nodes writing one file are pointed out: "Another node (tabbook) already writes outputs/tabbook.xlsx — this one would replace it.", with Use tabbook_2. Names that differ only in the case of their letters count as one file, as they do on a Mac or on Windows. When another flow of the project saves a file of the same name, the field says so too: Studio keeps each flow's copy apart, but in a downloaded research bundle one would replace the other.
• Where it ends up. "After a run: Files → outputs/flow/name" (Files becomes a link once a run has made the file), "Beside it: …" for what the node writes next to it, and what Run all leaves out, such as "Run all doesn't keep this file: run this flow on its own to get it." (see What Run all keeps).
• A node added from the palette is named after the flow — outputs/<flow>.md, <flow>_tabbook.xlsx, <flow>_data.csv, <flow>_maxdiff_choices.csv, <flow>_conjoint_choices.csv — numbered when the flow already names that file. An empty field stores nothing and shows, as its placeholder, the name the engine then writes, where the node has one (report for a Save report, tabbook for a Tab book (Excel)).
• A path saved earlier that the field cannot show as a name — outside outputs/, with no ending, or with an ending the node cannot write — is shown as it was written, under "Other location, kept as it was written", with the reason ("This file isn't in outputs/, so a run doesn't keep it under Files.") and a button such as Save it as outputs/tabs.xlsx instead.
• A required name left empty is an error on the node: "No file name yet: type the name to save it under."

A file a node reads is an upload under Files. The field lists, by name, only the uploads it can read — a Data file's File the data files, its Dictionary (JSON) the .json files — each with its size and upload date on the line under it (see Lists in the inspector). The note under the field gives the chosen file's size and upload date (a data file's rows × columns first), then "Each run reads the file as it is in Files at that moment." Under the note:

• Upload… opens Upload a data file (or Upload a dictionary), which says what it takes, refuses a file of another kind, even one dragged in, and selects the new upload once it is stored.
• Type a name… (Type another name… once a file is chosen) names a file that is not uploaded yet, such as one a connector will deliver. You type the name only (assets/ is added), and the field says "name isn't in this project's Files: upload it before the run, or choose another file." A name with characters Files would change is answered with the name Files stores, such as Use Wave_3_final_.csv. Choose from Files goes back to the list.
• With no uploads the field reads "No files uploaded yet. Upload a CSV, Excel, SPSS, Stata or Parquet file (up to 50 MB)."; while the project's files load, "Loading this project's files…"; when they cannot be loaded, it says so, with Try again.
• An upload deleted since stays chosen, with "(not in Files)" after its name, and the same sentence. A chosen upload the field would not offer stays too, marked "(can't be read as data)" or "(not a dictionary (.json))".
• A value saved earlier that a run cannot bring in — a path that is not an upload's name, such as ./assets/panel_wave2.csv — is shown under "Other location, kept as it was written": "A run only brings in files uploaded under Files, so it won't find this one.", with Use the uploaded panel_wave2.csv when Files has it. A Data file's Columns panel then says not read.
• A required file left empty is an error on the node: "No file chosen yet: pick one of the project's Files, or upload one."

Only the uploads a flow names are copied into its run (see Data file).

---

Sources

Nodes with no input that produce SurveyData. A flow needs at least one.

Data file

source.file — reads a data file uploaded under Files as it comes, and brings its own columns into the flow. Its ⓘ describes it in two paragraphs: "A data file from Files: CSV, Excel, SPSS, Stata or Parquet. A CSV may be in any encoding, with a comma (,), a semicolon (;), a tab or a vertical bar (|) between the fields." and "Its columns come into the flow with it: their names, the labels of a label row (a Qualtrics export's question texts), and value labels and missing codes where the format keeps them."

It reads .csv, .tsv and .txt (text tables in any encoding, separated by commas, semicolons, tabs or vertical bars), .xlsx, .xlsm and .xls, .sav (SPSS), .dta (Stata) and .parquet, the ending in any case. The nodes below it offer the file's columns, and the checks know them (see The file's columns in the flow); the questionnaire's variables come with the file only when the file is this survey's data.

In: none → Out: data (SurveyData)

The inspector shows the parameters in this order: File, Codebook, the Reading options, Missing codes, Dictionary (JSON) — then the Columns panel.

Parameter	Type	Default	Allowed	Meaning
File	upload	required	the uploads ending in .csv, .tsv, .txt, .xlsx, .xlsm, .xls, .sav, .dta, .parquet	The data file, chosen by name from the uploads under Files, or uploaded from here with Upload… (stored as assets/<name>). The note under the list gives its rows × columns as the node read it (as Files read it, with the default options, until then; "reading its columns…" while Files reads it), its size and upload date.
Codebook	choice	auto	auto, file, questionnaire	Where the columns' labels, scales and value labels come from. auto: the questionnaire's when the file is this survey's data, else the file's own. file — "the file's own: its names, label row and values": for data from elsewhere whose columns happen to share names with the questionnaire. questionnaire — "the questionnaire's, for the columns it knows". Once the file is read, auto shows what it decided: "auto — the file's own" or "auto — the questionnaire's".
Header rows	choice	auto	auto, 1, 2, 3	Rows at the top that are not answers, the names first: 1 "names only"; 2 "names, then labels (question texts)", whose second row labels the columns; 3 "names, labels, and a row left out (Qualtrics CSV)". auto finds a Qualtrics export (its ImportId row, or StartDate and ResponseId with a row of texts under them), else reads one row.
Skip rows	whole number	—	0 or more	Rows above the names to leave out (a title, a note). Empty finds the names row: the first that fills at least half of the table's width.
Sheet	text	—	—	Excel: the sheet to read, by its name or its number (1 is the first). Empty reads the first sheet that holds a table.
Delimiter	choice	auto	auto, ,, ;, tab, |	Text files: what separates the fields — comma, "semicolon (Excel with Russian or European settings)", "tab (Excel's Unicode text)", vertical bar. auto tries each on the first rows.
Encoding	choice	auto	auto, utf-8, utf-16, cp1251, koi8-r, cp866, cp1252, cp1250, iso-8859-1	Text files: how the letters are written. cp1251 is "Windows-1251, Cyrillic (Russian Excel's CSV)", utf-16 "UTF-16 (Excel's Unicode text)". auto reads a byte-order mark, then UTF-8, then the most likely Windows code page.
Decimal mark	choice	auto	auto, ., ,	How numbers write their fraction, 4.5 or 4,5, in a text file and for numbers an Excel sheet keeps as text. auto reads a decimal comma where the numbers are written with one: in a text file with a semicolon, a tab or a vertical bar between the fields (not a comma), and in a sheet where its text numbers show it. In a text file auto always settles on one mark; only in a sheet, when nothing tells, does it leave values as ambiguous as 1,500 as text, and the Columns panel says so — pick the mark here to read them as numbers.
Missing codes	text	—	—	Codes that mean no answer in this file, per column — q5: -9 and q6_1: -7, -8, one column a line (or ; between them) — or for every column that holds them: -7, -8, -9 (a negative code is then left alone in a column of other negative amounts). Missing values turns them into blanks, and tables leave them out. A box of a few lines; the empty box shows "e.g. q5: -9 (a line per column)".
Dictionary (JSON)	upload	—	the .json uploads	Optional: the file's codebook as a data dictionary (<name>.dictionary.json, as a Siamang export writes it); its labels, scales and missing codes describe the file's columns. <name>.dictionary.json files are listed first, and when the one named after File is in Files the field offers Use name.dictionary.json.

Reading options. Header rows, Skip rows, Sheet, Delimiter, Encoding and Decimal mark sit under one line, Reading options, which says what auto found — "Reading options · all auto (found: XLSX, sheet Sheet1, 1 header row)" — or names the options set ("Reading options · Header rows 2, Delimiter ;"). Click it to open it; it opens by itself on a node where one of them is set. Only the options of the chosen kind of file show: for a text file Header rows, Skip rows, Delimiter, Encoding and Decimal mark; for a workbook Header rows, Skip rows, Sheet and Decimal mark; none for SPSS, Stata and Parquet, which say all of it themselves (an option set for another kind of file still shows, so it can be cleared). An option left on auto says in its field what was found: "auto — found ;" in Delimiter, "found: Sheet1" in an empty Sheet. In each option's list a value has its meaning on the line under it, such as "Windows-1251, Cyrillic (Russian Excel's CSV)" under cp1251; these lists have no — default —, since auto is the default.

Missing codes per file. When Missing codes names columns the file does not have — codes left from a file the node read before, which a run would skip without a word — the field says "Missing codes name columns this file doesn't have: …" with Remove them (Remove it for one).

The Columns panel

Under the parameters, once a File is chosen, the Columns panel shows what the node read from the file with these options — its head says read, reading…, not in Files, not read (a File under "Other location") or cannot read:

• "rows rows × columns columns" and how the file was read: the format, the delimiter, the encoding (or "with a byte-order mark"), a decimal comma, the sheet ("sheet Sheet1 of 3"), rows above the names skipped, or "a Qualtrics export: its question texts label the columns, its ImportId row is left out". A + after the rows means only the first 50,000 rows were read.
• Where the labels and scales come from: "Labels and scales: the questionnaire's — the file is this survey's data.", "… from the file's dictionary.", "… from the file's own SPSS / Stata labels." or "Labels and scales: from the file (column names; scales guessed from the values)." (with "the second header row as labels" when there is one). Notes on how the file's names met the questionnaire's follow, such as "2 of the file's 16 columns share names with questionnaire variables (gender, satisfaction), but satisfaction holds answers that don't fit its variable, so they keep the file's own labels. Set Codebook to questionnaire to use the questionnaire's."
• Codes that look like missing codes in n columns — "-9, -8, -7 (e.g. q5: -9; q6_1: -7). They count as answers until you mark them." — with Mark them as missing in these columns, which writes each code into Missing codes for the columns it was found in only ("Each code only in the columns where it was found: a -7 among a column's amounts, or an age of 99, stays a value."). Once they are, the panel says "Marked as missing in n columns — see Missing codes above." with Undo. When some are marked and others not, the button reads Add this file's missing codes (n columns).
• Looks like personal data, a short warning mark, when some columns' names or values look like an email, an IP address, a location, a name, a phone number, an address or a participant ID. Its ⓘ names those columns, each with its kind ("email (e-mail)"), and says, in a paragraph of its own, "An analysis rarely needs these; leave them out early so no table, report or export carries them." Beside it, Add a Select columns node without them puts a Select columns node labeled "Without personal data", keeping every other column, after the Data file; it takes over what the Data file fed, and is selected. Nothing is dropped until you press it.
• Show the n columns lists each column on two lines: its name, a personal-data mark and its missing codes — declared, and suspected ones with a ? — then its label, its value labels count, its type where it says something (text, date) and its scale, * marking a scale guessed from the values ("* a scale guessed from the values · ? a code that looks like a missing code").
• The last line names the options set ("Read with: Header rows 2, Missing codes.") and says "The nodes below this one can pick these columns. The file is read apart from the flow, and only its column list comes back here, never an answer."

While the file is read the panel says "Reading the file's columns… Nodes below it aren't checked for names until it is read."; a file not read yet with the node's options says "This file has not been read with these options yet; someone who can run previews opens this node to read it." Read again ("Read the file again with these options") reads it once more. A file that cannot be read says why in words: the encoding it looks like and what to set, a line with a different number of fields than the table ("Line 4 has 3 fields where the table has 2 …"), the sheets a workbook has, an empty file, or a format the node does not read. When the chosen file is not in Files, the panel says only not in Files; the note under File says the rest. A File under "Other location" is never read: the panel says not read and "Only a file uploaded under Files is read for its columns. Choose the upload under File above (upload it first if it isn't there), and its columns show here.", with no Read again. The field's Use the uploaded … (for an older path) or Use name (for a typed name), where it offers one, or its list chooses the upload.

The file's columns in the flow

The nodes below a Data file offer the file's columns in their pickers, grouped under "From name" (the file's name as File lists it), then the variables the nodes between make ("Made in this flow", or "From the questionnaire and this flow" when the file is this survey's data). The check on the canvas, at Check, at Save and before a preview knows them:

• A name the file does not have is an error: "param: "name" is not a column of assets/name, nor made by this flow."
• A scale the file's values only suggest is a warning where it does not fit the node, not an error: "param: "name" looks ratio (as guessed from its file's values); this node expects nominal/ordinal." Set Codebook, or recode the column, when the guess is wrong.
• The questionnaire's variables are known below the file only when the file is its data, so a file from elsewhere keeps its own names, labels and scales even where a column shares a name with a survey variable. The survey's response timestamps (created_at and the others) are known only if the file has those columns.
• While the file is not read yet, or cannot be read, the names below it are not checked, and their fields take a typed name ("a column of the file", or "columns of the file, separated by commas").
• A table a Write table node writes from the file brings the file's columns to the flows that read it.

Where the labels come from. With Codebook auto, the node takes the labels, value labels and scales from the dictionary you choose, else the labels inside an SPSS or Stata file, else the questionnaire's codebook when the file is this survey's data — at least half of the file's answer columns are the questionnaire's variables and they are at least half of its variables (or the file's dictionary labels them as the questionnaire does), and their answers fit — else the file's own: its names, the labels of its label row, and scales guessed from its values. Once the file has been read, the flow is checked and run with what auto decided then, so a later edit of the questionnaire does not change it until the flow is checked again.

On the platform:

• The node reads a file you uploaded under Files, chosen by its name in File — for example household_survey.xlsx — or uploaded from the node with Upload… (see Where files go). Files reads each data file once it is uploaded, so the node usually knows its columns before you open it. Runs, Run all and Run to here all get it.
• Only the uploads a flow names are copied into its run. An uploaded dictionary is therefore not found "next to the file" on the platform: choose it in Dictionary (JSON) as well — the field offers Use panel_wave2.dictionary.json when the file is panel_wave2.csv.
• An upload that is missing shows in the run's log before the node fails: "note: assets/panel_wave2.csv is not among this project's Files" (or "… is listed under Files but its content is gone"). Each run reads the file as it is in Files at that moment, and its log says which version it read: "read assets/panel_wave2.csv (sha256 …)", the start of the content's hash, so two runs over a replaced upload can be told apart.
• In a research bundle made with the responses so far, the uploads the flows name are included at the same path; in a bundle without data, put the file there yourself (see Reproducibility).

Responses

source.responses · platform — the answers collected by your survey, from the project's responses table, as SurveyData labeled by the questionnaire's codebook. New flows start with one of these already placed (id src).

In: none → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Table	text	responses	—	Which project table to read. responses is the table your survey writes. A table name has lowercase letters, digits and _, starts with a letter or _, and has at most 63 characters.
Environment	text	main	—	Keep the responses collected by this environment's deployments (plus rows no deployment collected, such as imported or sample data). Typed as text, e.g. main or pilot: lowercase letters, digits and -, starting with a letter, at most 63 characters.
Only completed responses	checkbox	off	—	Drop interviews that were started but not submitted.

What comes in besides the answers: duration_s (interview length in seconds), started_at, captcha, tab_switches, hidden_seconds, pastes and every URL parameter of the survey link (url_…).

How the answers arrive. One column per variable, whichever survey runtime collected the response — responses stored before the current runtime are read into the same layout:

• A question that writes several variables — a matrix, a MaxDiff, a conjoint, a Multiple choice with Data layout wide — arrives as one column per variable. A wide choice is 1 when chosen and 0 when offered and not chosen; with the current survey runtime, an option the respondent never saw (hidden by its own condition) is left empty rather than 0.
• A matrix answer is the column's codebook code: a 0–10 scale is 0–10. Responses collected by an earlier runtime, which stored the column's position (1–11 for a 0–10 scale), are read as the code too — so a matrix whose codes are not 1, 2, 3, … gives different numbers than it did before this was fixed. Check Recode mappings, Filter rows values and weighting targets written against the old values.
• "Other (please specify)" arrives as the question's Other code (-66 unless the question sets another) in the question's variable, and the typed text in its own column, <variable>_other. "None of the above" is the question's None code (-77 unless set), and "Not applicable" the variable's declared not-applicable code — or the text na when the codebook declares none. See Codebook and Variables.
• A variable a custom script writes arrives as its own column; flags an earlier runtime stored together (__flags__) arrive one column each (for example speeder).

Notes:

• Environment keeps the responses of that environment's deployments plus the rows no deployment collected (imported data, a template's sample data).
• Only completed responses has no effect on a table without a partial column. Partial interviews reach the responses table only from surveys published with the current survey runtime: a survey published before partial saves were fixed sends only submitted interviews until you publish it again, and from then on a flow that does not check this box sees its partial interviews too.
• A Table or Environment that no project can have (see the rules above), including a name that ends in a line break, fails the engine check at Save: "“x y” is not a table name: …" or "“Main” is not an environment name: …".
• In a research bundle this node reads a data file instead of the database. A bundle made with data carries one already filtered the way this node filters (data/responses.main.csv, data/responses.main.completed.csv, or data/responses.csv when the filter keeps every row); with a file you supply yourself, the file is read as is. See Reproducibility.

Simulated data

source.simulated — "Synthetic responses generated from the questionnaire's logic — conditions, routing, assigned arms — for building a flow before fieldwork." The same seed gives the same data, including in a research bundle.

In: none → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Respondents	whole number	200	at least 1	How many synthetic respondents to generate.
Seed	whole number	42	—	Random seed: the same seed gives the same data every time.

Each simulated respondent walks the questionnaire as the survey would move them, its scripts included:

• a Scripts → Assign to a condition arm is drawn by the arms' weights and arrives as a column (a nominal variable labeled with the arms), and the pages shown only to one arm are filled for that arm;
• a page shuffle is dealt to each respondent;
• questions hidden by their own, their block's or their page's condition stay empty, and so do answer options hidden by their own condition.

Quotas belong to publishing, not to the questionnaire, so no quota closes here (Test → Simulate in the Builder does apply them). The generated line is n_<id> = simulate_survey(survey, n=…, seed=…). In a research bundle whose engine pin lags behind Studio (the README says so), this node stops with an error — see Reproducibility.

Project table

source.table · platform — a table in the project database, typically one another flow wrote with Write table (for example a cleaned copy of the responses), as SurveyData labeled by the questionnaire's codebook.

In: none → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Table	text	required	—	Name of a project table written by a Write table node (or another table in Data): lowercase letters, digits and _, starting with a letter or _, at most 63 characters.

• Run all runs this flow after the flow whose Write table writes the table named here, whatever the flows are called, and skips it (marked failed) when that flow fails. Running this flow on its own does not run the writer first: it reads the table as it is. See Run all.
• A table no run has written yet stops the node, in a preview and in a run: "The table clean_responses does not exist yet: it is written by 1. Clean raw responses. Run that flow (or Run all) first, then this one." — naming the flow whose Write table writes it, or "…: a flow's Write table node makes it." when none does.
• A multiple-choice question kept as one variable, and a ranking, come back as lists, as from Responses, so Explode multiple choice and a tab book read them.
• The table brings the variables its writer stored with it — labels, scales and value labels of a recode, a derived variable, an index, a cluster, the quality flags — merged with the questionnaire's codebook (the questionnaire's entry wins a name both have). They can be picked in this flow's nodes ("from table table · made by flow"), and naming one passes the engine check. A table last written before tables kept their variables arrives without those labels until its writer runs again. See Cleaning and Weighting Data.
• In a research bundle made with data this node reads data/tables/<table>.csv — the table as it was in the project database when the bundle was made, with its variables in data/tables/<table>.dictionary.json — not a table the writing flow produces there.

---

Prepare

SurveyData in, SurveyData out: the steps a methods section calls "data preparation". The recommended order is described in Cleaning and Weighting Data.

Apply weight

prepare.apply_weight — "Weight the results downstream by a column. Weighted results say so, and a result with no weighted form says it is unweighted." Use it after Cell weights or Rake weights created the column (or when the data already carries one).

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Weight column	text	weight	—	The column that holds each respondent's weight: weight unless you name another. Its ⓘ, a tall tip that opens beside the field where there is room, lists the results that are weighted, those weighted except a part, and those that are unweighted and say so (the tables below).

These nodes use the applied weight, and their output names it (Weight or weight among the statistics). In the weighted tables and charts a weight that is missing or not a number counts as 0.

Node	Weighted
Frequencies	N is the sum of the weights and the percentages follow it, with an Unweighted N column beside it (see Frequencies)
Crosstab	cells and totals are sums of weights and the percentages are taken from them; the chi-square test uses Kish's effective base. Fisher's exact test counts respondents (see Crosstab)
Group means	means, SDs and medians; N, the significance test and the post-hoc pairs stay unweighted (see Group means)
Descriptive statistics	mean, SD, median and quartiles, beside a Weighted N column; N, Missing, skewness and kurtosis are not weighted (see Descriptive statistics)
Correlation, Correlation matrix with Method pearson	the coefficient, with its p-value and CI on Kish's effective base (see Correlation)
Banner table	percentages and counts; tests on Kish's effective base
Net Promoter Score	the shares and the score (its standard error on Kish's effective base); N counts respondents
Regression	linear models become weighted least squares; logistic and ordinal logistic models are weighted too
TURF	the base, each portfolio's reach and its frequency
MaxDiff	every column: Shown, Best, Worst, Score, Utility and Share % (see MaxDiff)
Conjoint	the part-worths and Importance % (see Conjoint)
Share of preference	the shares, from weighted part-worths
Principal components	loadings, eigenvalues and explained variance, from the weighted covariance (or correlation) matrix
Scale reliability	alpha, item means, item–total correlations and alpha-if-deleted
Key drivers	the correlations, betas, R² and every driver's share, from the weighted correlation matrix; the tests on Kish's effective N (see Key drivers)
Perceptual map	every cell of the table it maps is a sum of weights; the chi-square test of a crosstab counts respondents (see Perceptual map)
Price sensitivity	every curve and share; N stays the respondents (see Price sensitivity)
Bar chart	bars are sums of weights or weighted percentages — split into groups too, in a histogram and a donut too — or weighted means with By; its confidence intervals and significance letters are on Kish's effective base (see Bar chart)
Heatmap with By, or without it with Method pearson	weighted means by group; weighted Pearson coefficients
Likert chart	the shares are sums of weights; each item's n counts respondents (see Likert chart)
Trend	each point's percent, mean or count (the sum of weights), its band on Kish's effective base; the bases count respondents (see Trend)
Result chart	as the result it draws is, and its title's second line says which (see Result chart)
Proportion CI	only when its Weighted box is checked (see Proportion CI)
Tab book (Excel)	counts and bases are sums of weights (shown to one decimal) beside the unweighted base, percentages are of those sums, and the letters test on Kish's effective base (see Tab book (Excel))

These have no weighted form. After Apply weight they run on the respondents as they are and say so: "unweighted (the weight 'weight' is not applied)" — as a weight statistic, a Weight line under the table, or a second line of the chart's title (also under a title you set):

Node	Where it says so
Compare groups	statistic weight
Correlation with Method spearman or kendall	statistic weight
Correlation matrix with Method spearman or kendall	Weight under the table
t-test	Weight under the table
Paired tests (Cochran's Q too)	Weight under the table (and under the pairs of Friedman and Cochran's Q)
Factor analysis	Weight under each of its three tables
Group means	the post-hoc table's Weight line (the means above it are weighted)
Crosstab with Test fisher	its Base line: "the test counts respondents (an exact test needs whole counts); weighted counts shown"
Perceptual map's chi-square test	Chi-square counts = "respondents (unweighted), as a test of a table must" (the map itself is weighted)
Cluster (k-means)	statistic weight
Proportion CI with Weighted unchecked	statistic weight
Box plot, Scatter plot, Heatmap without By with Method spearman or kendall	second title line
Response quality, Code open answers, Data check	Weight under their table: they count responses, answers and rows
Bands, MaxDiff scores	Weight in their statistics: they count respondents, or score each one

Describe counts rows and, on weighted data, adds a weighted_n_valid column: the weighted base a table of each variable would report. The data files of Choice data for HB and Conjoint data for HB have no weight column. The Methods draft describes the step as "estimates that support weights were weighted by weight (rank tests, k-means clustering, box and scatter plots and Spearman and Kendall correlation heatmaps stay unweighted, as do t-tests, the tests of group means (ANOVA, Welch's ANOVA) and their post-hoc comparisons, paired tests (Cochran's Q among them), Fisher's exact test, a perceptual map's chi-square test, rank correlations and factor analysis)".

Note. The MaxDiff Utility and Share %, Conjoint, Share of preference, Principal components, Scale reliability, the Bar chart and the Heatmap of means were computed unweighted after Apply weight before this was fixed, and a weighted Proportion CI counted non-respondents in its base. A TURF portfolio's frequency was unweighted too. Run a weighted flow again and these numbers change; reports and tiles from earlier runs keep the old ones.

Bands

prepare.bands — "Cut a number into bands — age into age groups, income into brackets — as a labeled ordinal variable; what falls outside every band is counted."

In: data (SurveyData) → Out: data (SurveyData), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	ordinal / interval / ratio variables	Its missing codes (a 999 "Refused") are left out before cutting, so they never land in a band.
Boundaries	JSON	required	—	Increasing numbers, e.g. [18, 30, 45, 65, 100]. A band includes its lower boundary and runs up to, not including, the next; values outside every band stay blank.
Band labels	JSON	—	—	One fewer than the boundaries, e.g. ["18–29", "30–44", "45–64", "65+"]. Defaults to "18 to under 30" and so on.
New variable	text	required	—	Name of the band variable.
Label	text	—	—	Defaults to the variable's label and "(bands)".
Bands include their upper boundary	checkbox	off	—	Then a band runs from above its lower boundary up to and including the upper one (the first band takes its lower boundary too).

• The new variable is ordinal, coded 1, 2, 3, … in the order of the bands, with the band labels as its value labels — so a Crosstab or Group means by it prints "18 to under 30", not 1. Later nodes offer it in their variable lists ("made by node"), and a t-test's Group A and Group B offer its bands.
• With Bands include their upper boundary checked, the default labels read "18 to 30", "over 30 to 45", ….
• The stat output names the step (Variable = age → age_band), counts each band (Bands = 16 to under 30: 12; 30 to under 45: 15; …) and what fell outside (Outside the bands), and adds Missing codes, Blank and Not numbers when there are any. A value outside every band (below the first boundary, or at or above the last) is blank in the new variable, never forced into the nearest band.
• Boundaries that do not increase, or labels whose number does not match, stop the run with the reason: "Band boundaries must increase, got [18, 18, 30].", "5 boundaries make 4 bands, but 3 labels were given."
• The counts are of respondents; on weighted data the statistics add Weight = "unweighted (the weight 'weight' is not applied)". Tables of the band variable after Apply weight are weighted as usual.

Cell weights

prepare.cell_weights — post-stratification on one variable: after weighting, that variable's distribution matches your targets. It adds the weight column but does not apply it — follow it with Apply weight.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	nominal / ordinal variables	The single-answer variable to weight on.
Targets (code → share)	JSON object	required	—	Target share per code, e.g. {"1": 0.45, "2": 0.55}. Shares or counts both work; they are normalized.
Weight column	text	weight	—	Name of the new weight column.
Cap	number	—	at least 1	Upper bound for weights of mean 1 (must be greater than 1). Capping makes the match to the targets approximate.

• Targets may be shares or counts; they are normalized, so {"1": 45, "2": 55} and {"1": 0.45, "2": 0.55} mean the same.
• A category you leave out of the targets keeps its current share.
• Weights are scaled to a mean of 1, so the weighted N equals the number of rows. Cap must be greater than 1.
• A multiple-answer variable is refused: a respondent in two categories has no single cell. Weight on a single-answer variable, or on the 0/1 columns Explode multiple choice creates.

Dedup respondents

prepare.dedup — one row per respondent. A respondent who resumed the survey or submitted twice is kept once: the latest submission by default.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Respondent id column	text	respondent_id	—	Column that identifies a respondent.
Order by	text	submitted_at	—	Column that orders a respondent's submissions. If it is missing, Studio falls back to submitted_at, started_at, updated_at or created_at, whichever exists.
Keep	choice	last	first, last	last keeps the most recent submission, first the earliest.

Rows with a blank respondent id are always kept, each as its own respondent. Data without the id column passes through unchanged.

Derive

prepare.derive — a new variable computed from others with a small formula language: arithmetic, functions and if … then … else. The formula is stored as text and parsed, never executed as code, so what you write is exactly what a reviewer reads in the generated script.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
New variable	text	required	—	Name of the new variable.
Formula	formula	required	—	e.g. round(spend_year / 12, 2), or: if age < 30 then 1 else 2. Functions: mean, sum, min, max, abs, round, log, coalesce. A missing answer stays missing; dividing by zero gives missing, not infinity.
Label	text	—	—	Defaults to the formula itself — the most honest description of the number.
Scale	choice	ratio	ratio, interval, ordinal, nominal	Measurement scale of the new variable. Choose nominal or ordinal when the result is a code rather than an amount.
Value labels	JSON object	—	—	For a formula that yields codes: code → label, e.g. {"1": "Under 30", "2": "30 or over"}. Set the scale to ordinal or nominal with them.

Value labels are typed as JSON, codes as keys: {"1": "Under 30", "2": "30 or over"} for if age < 30 then 1 else 2 — the example the empty box shows. Tables of the new variable then print the labels, not 1 and 2, and a t-test's Group A and Group B downstream offer 1 ("Under 30") and 2 ("30 or over"). To cut a number into ranges, Bands writes the labels for you.

The formula language.

Element	Syntax
Numbers	12, 0.5, .5
Variables	codebook names, e.g. spend_year, q1
Arithmetic	+, -, *, /, unary -, parentheses
Comparisons	=, !=, >, >=, <, <=
Logic	and, or, not
Conditional	if <condition> then <value> else <value>
Functions	mean(…), sum(…), min(…), max(…) — across their arguments, row by row; abs(x), round(x, digits), log(x), coalesce(x, y, …)

Examples: round(spend_year / 12, 2) · (q1 + q2 + q3) / 3 · mean(q1, q2, q3) · if age < 30 then 1 else 2 · if income > 0 then round(income / 12, 2) else 0 · coalesce(children, 0).

Precedence, loosest first: if … then … else, or, and, not, comparisons, + -, * /, unary minus. The comparison and logic words mean exactly what they mean in a skip-logic condition. contains (chose) is not available: a formula works on numbers — run Explode multiple choice first and use the 0/1 column.

Two behaviors to know before you read a result:

• A blank stays blank. A respondent who skipped a question has no value, and arithmetic on it has none either. Dividing by zero also gives a blank rather than infinity — an infinity looks like a number all the way into the report. Write coalesce(x, 0) where you mean "treat a blank as zero".
• A text variable is refused by name, rather than turning into a column of blanks. Recode it first.

A formula is checked when you press Check and when you Save: an unreadable one says where reading stopped ("… (at character 14)"), and a variable that is neither in the codebook nor made earlier in the flow is named. Write it on one line: the box wraps long formulas, but a line break in it fails the check (see One-line texts).

Explode multiple choice

prepare.explode — one 0/1 column per option of a multiple-answer question (brand_1, brand_2, …), so that weights, regression, clustering and TURF can read it.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Multiple-choice variable	variable	required	nominal variables	A question whose answers are lists of codes.
Column prefix	text	—	—	Defaults to the variable's own name and an underscore, so option 1 becomes brand_1.

Each column is named with the option's code as the engine writes it: a code the codebook's {code: label} form writes as 1.0 makes brand_1.0, and the variable pickers downstream offer that name. | Drop the original column | checkbox | off | — | Off by default — the list column still answers questions the indicators cannot, such as how many options each respondent picked. |

Someone who did not answer the question gets a blank in every indicator, not a row of zeroes — "chose nothing" and "was never asked" are different facts, and only the first belongs in a base. The indicator names follow the codebook's codes (<prefix><code>), and later nodes offer them in their variable lists ("made by node") — for example as a TURF's Options. A Multiple choice question with Data layout wide already has one 0/1 variable per choice and needs no Explode.

Filter rows

prepare.filter — keeps only the respondents who satisfy a condition, built with the same visual editor as the questionnaire's logic.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Condition	condition	required	—	Rows that satisfy the condition are kept; the others are removed.

A condition edited by hand in the flow's JSON is checked as the engine will write it: a part of an and / or, or what a not negates, that is not a comparison or a variable, and a variable reference without a name, are errors on the node before a Save.

The editor offers ALL of the following / ANY of the following, the operators = ≠ > ≥ < ≤ in, not in, chose, did not choose, and value pickers that show value labels. In the flow's inspector each of these is a list, with what each operator asks and each answer's label under its code; the codes of in and not in stay a choose codes checklist (see Parameters and variable pickers and Lists in the inspector). Closed, the condition reads as a sentence, region = Capital region (1). Its variables are those available at this node — the codebook's, and those made upstream, such as quality_score or a recode — so you can, for example, keep only quality_score = 0. A condition that nests groups opens as JSON. Conditions typed as raw text are refused ("raw string conditions cannot be evaluated on data").

Index / scale

prepare.index — combines several items into one index variable, the mean or the sum of the items. Check the items' internal consistency with Scale reliability first.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Index name	text	required	—	Name of the new index variable.
Items	variables (several)	required	—	The component variables.
Method	choice	mean	mean, sum	mean or sum of the items.
Label	text	—	—	Codebook label of the index.

MaxDiff scores

prepare.maxdiff_scores — "One counting-score variable per MaxDiff item, per respondent: best minus worst over the times the item was shown." With them, the preferences can go into a crosstab, a cluster or a regression.

In: data (SurveyData) → Out: data (SurveyData), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
MaxDiff question	text	required	—	The question's id or name. Its design is read from the questionnaire.
Variable prefix	text	—	—	Each item's score is <prefix><item code>. Defaults to the question and _score_, so item 3 of q_md becomes q_md_score_3.

• MaxDiff question is a list of the questionnaire's MaxDiff questions (— pick a MaxDiff question —), each by its name with its question text under it. A stored name the questionnaire does not have is kept and marked "(not in the questionnaire)" after it in the field, and the engine check names it at Check and at Save: "Parameter 'question' of md_scores: no MaxDiff question named 'q_mdx'; this questionnaire has: q_md." With no MaxDiff question in the questionnaire the field is a text box.
• One variable per item, q_md_score_1, q_md_score_2, …, labeled "MaxDiff score: item", interval, from −1 (picked worst every time it was shown) to 1 (picked best every time). A respondent who was never shown an item has a blank for it, not a 0: "never picked" and "never offered" are different answers. Later nodes offer the variables ("MaxDiff score: Price · made by node").
• The stat output gives the Question, Respondents scored, Not scored, Items, the Variables (q_md_score_1 … q_md_score_5), the Score ("best minus worst over times shown, −1 to 1; blank where never shown") and, when some answers could not be read against the design, Unreadable answers with the reasons.
• The scores are per respondent, so no weight enters them; on weighted data the statistics say "unweighted (the weight 'weight' is not applied)". A Group means of a score after Apply weight is weighted as usual.
• These are counting scores from a few tasks per person: coarse but real. The MaxDiff node gives the aggregate utilities; for individual utilities, export with Choice data for HB.

Missing values

prepare.missing — applies the codebook's declared missing codes ("Don't know", "Refused" with their codes): they become blanks, so they stop counting as real values in means and tests. Optionally drops the rows that are blank in chosen variables.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Action	choice	to_nan	to_nan, drop_rows	to_nan turns the codebook's declared missing codes into blanks; drop_rows does that and then drops rows blank in the variables listed below.
Drop rows missing in	variables (several)	empty	—	For drop_rows: rows with a blank in any of these variables are dropped.

The inspector shows Drop rows missing in only while Action is drop_rows (see Fields that depend on a choice).

Put this node before any analysis whose defaults count a missing code as an answer — Correlation with spearman, Compare groups without Dunn's test, Group means with Test auto, the Crosstab chi-square; each of them says when it did. A test you choose by hand (Correlation pearson or kendall, Compare groups with Dunn's test, Group means with another Test, Crosstab with fisher), and the t-test, Correlation matrix, Paired tests, Factor analysis, Descriptive statistics and Bands, leave the codes out on their own (see Analyze).

Response quality

prepare.quality — flags straightlining, contradictions, duplicate answer patterns and failed attention checks. By default it marks responses rather than dropping them, so you decide what to exclude and say so in your write-up.

In: data (SurveyData) → Out: data (SurveyData), table (Table)

Parameter	Type	Default	Allowed	Meaning
Battery to check	variables (several)	empty	—	A matrix or a battery of same-scale items. Straightlining and duplicate patterns are measured across these; a flat pattern (the same answer throughout) is straightlining, never a duplicate, since two straightliners of one column match whoever they are.
Duplicates also match on	variables (several)	empty	—	Other answers a duplicate must repeat as well as the battery, such as age, gender and a few questions of their own. Two honest respondents can answer a battery alike, and both are dropped as one person submitting twice; the more answers two responses must share, the surer the match. Empty, the battery alone decides.
Answers that must agree	JSON object	empty	—	left variable: right variable. A respondent answering them differently is flagged as contradictory.
Attention checks	JSON object	empty	—	The answer a reading respondent gives to each attention check, as variable: answer, e.g. {"q9": 3, "q12": [1, 2]}. Where several answers pass, list them: a nonsensical statement is passed by any answer on the correct side of its scale (the disagree side for a statement nobody agrees with). Any other answer is flagged; a skipped check is not. Fill from the questionnaire, below the box, fills it in from the questions marked as attention checks, lists included.
Straightlining tolerance	number	0.0	at least 0.0	Standard deviation across the battery at or below which a response counts as flat. 0 means literally identical answers.
Mode	choice	flag	flag, drop	Flag adds the columns and keeps everyone; drop also removes the flagged responses.
Flags column	text	quality_flags	—	New variable holding the names of the failed checks (empty when clean).
Score column	text	quality_score	—	New variable counting the failed checks (0 when clean).

The four checks:

Check (name in the flags)	Flagged when
straightlining	the standard deviation of a response across Battery to check is at or below Straightlining tolerance
inconsistency	a pair in Answers that must agree was answered differently
duplicate	the response's complete answer pattern across a battery of at least 5 items — not a flat one — is identical to another response's, and so are its answers to Duplicates also match on (two unanswered ones match) — every member of such a group is flagged
attention	an attention-check variable was answered with none of the answers that pass it (numbers compared as numbers, else as text); a skipped check is not a failure

• Flags column (quality_flags) holds the failed checks joined with ; , e.g. straightlining; duplicate, and is empty for a clean response; Score column (quality_score) counts them. Both enter the codebook (labels "Quality flags" and "Checks failed") and travel into exports and the generated script.
• Second output — table: one row per check with N and %, then Any check and Clean, as shares of everyone screened, counted before anything is dropped. Wire it into a Report section to put your exclusion numbers in the report. The counts are of responses whatever the weight; on weighted data the table says "Weight: unweighted (the weight 'weight' is not applied)".
• Later nodes offer quality_flags and quality_score in their variable lists, so a Filter rows can keep the clean responses (quality_score = 0) instead of Mode drop.
• For Attention checks, the inspector offers Fill from the questionnaire (N marked) when questions are marked as attention checks in the Builder; otherwise it says "Mark a question as an attention check in Builder to fill this in."

Rake weights

prepare.rake_weights — iterative proportional fitting: adjusts weights until several margins (region, gender, age group…) match their targets at once. It adds the weight column but does not apply it — follow it with Apply weight.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Targets (variable → code → share)	JSON object	required	—	One block per margin, e.g. {"region": {"1": 0.45, "2": 0.30, "3": 0.25}, "gender": {"1": 0.48, "2": 0.52}}.
Weight column	text	weight	—	Name of the new weight column.
Max iterations	whole number	50	at least 1	Maximum raking passes.
Tolerance	number	1e-06	—	Stop when no margin changes by more than this in a pass.
Cap	number	—	at least 1	Upper bound for weights of mean 1 (must be greater than 1).

Targets are normalized per margin (shares or counts). A category missing from a margin's targets is left unadjusted on that margin. Weights are scaled to a mean of 1. Multiple-answer variables are refused, as for Cell weights.

Recode

prepare.recode — maps the codes of a variable onto new codes, into a new variable with its own codebook entry: collapsing a 5-point scale to 3 points, merging small categories, harmonizing codes between waves.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	—	The variable to recode.
Old code → new code	JSON object	required	—	Old code → new code, e.g. {"1": 1, "2": 1, "3": 2}. Codes not listed are left blank in the new variable.
New variable	text	—	—	Defaults to <variable>_recoded.
Label	text	—	—	Codebook label of the new variable.
Scale	choice	—	nominal, ordinal, interval, ratio	Measurement scale of the new variable.

• Codes you do not list in the mapping become blank in the new variable, so list every code you want to keep ({"1": 1, "2": 2, "3": 3, "4": 3, "5": 3}).
• The new variable's value labels are its codes themselves (1, 2, …): there is no field for value labels, so tables of a recoded variable show codes. Say what the codes mean in the Label or in a caption.
• Scale defaults to the source variable's scale; Label defaults to the new variable's name.
• Later nodes offer the new variable in their variable lists, as "label · made by node".

Select columns

prepare.select — keeps only the listed variables.

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Variables	variables (several)	required	—	The variables to keep.

Everything else is dropped, including columns later nodes may need — the weight, respondent_id, timestamps. Select late, or include them.

Speeders & partials

prepare.speeders — quality screening on interview length and completeness. It adds two columns, duration_s and partial, and drops speeders. With Drop partials on, it also drops the interviews that miss any of the Required answers; with no required answers listed, it drops no partials. Later nodes offer both in their variable lists ("completion time · made by node", "partial response · made by node").

In: data (SurveyData) → Out: data (SurveyData)

Parameter	Type	Default	Allowed	Meaning
Minimum seconds	whole number	60	at least 0	Interviews completed faster than this are dropped; 0 drops none.
Required answers	variables (several)	empty	—	An interview is partial when any of these is blank. With none listed, no interview is.
Drop partials	checkbox	on	—	Also drop the interviews that Required answers marks as partial. Off, they stay, marked in partial.

The rules, precisely:

• A speeder is a response whose duration is known and shorter than Minimum seconds. Duration is the platform's interview length, or submitted_at − started_at when that is all the data has. A response whose duration is unknown is never treated as a speeder.
• A response is partial when any variable in Required answers is blank. With no required answers listed, nobody is partial, so Drop partials removes nothing. A required name that does not exist in the data marks every response partial — check the names.
• The node's partial replaces the table's own partial column, so an interview broken off after answering every required question stays in, marked false. To leave unfinished interviews out, check Only completed responses on the Responses node.

Code open answers

prepare.text_code — applies a codeframe to an open-text variable: each answer gets the themes a coder gave it by hand, else the ones the codeframe's word rules find, else it stays uncoded. The same way at every run, with no model and no network — and the rules code the answers collected after they were written too. You build the codeframe in the codeframe editor. See Coding Open Answers.

In: data (SurveyData) → Out: data (SurveyData), table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Codeframe	codeframe	required	—	The codeframe file (analysis/<name>.codeframe.json) — its themes, the answers coded by hand (kept as fingerprints, never as texts) and the rules that code the rest, including answers collected later.
Theme variable	text	—	—	Defaults to the name the codeframe carries. A codeframe that gives an answer several themes makes a multiple-choice variable (a list of codes per answer).
Also add sentiment	checkbox	off	—	A sentiment variable beside the theme, and each theme's negative / neutral / positive split in the table. Only when the codeframe was built with it — the stat says when it was not.

• Codeframe is a list of the project's codeframes (— choose a codeframe —), with a line under it, which a screen reader hears with the field: "Choose a codeframe document of this project." or "No codeframe document exists in this project." The button under it opens the codeframe editor: New codeframe… when none is chosen — it asks for the open-text variable and a name, sets the new file on the node at once and shows the answers this flow's Responses node reads — and Edit codeframe… for the chosen one. A new codeframe is a document only once saved in its editor; until then the field shows it with "(not saved)" after it, and the line reads "path is not saved yet: it is kept in this tab until you save it in its editor." (in another tab: "path is not saved in this project: it was started and never saved. Start it again, or choose another codeframe."). There is no button to code with the AI assistant: AI coding of open answers is switched off on this platform.
• Theme variable left empty takes the name the codeframe carries (set in the editor's Codeframe settings → Theme variable, <variable>_theme by default).
• Output data: adds the theme variable (label Theme: <variable>, value labels = your theme labels): the code, or with several themes an answer the list of codes (a multiple-choice variable). An answer coded as no theme and an uncoded one are empty there. With Also add sentiment, also a <theme variable>_sentiment variable coded −1 / 0 / 1 (Negative / Neutral / Positive).
• Output table: one row per theme and per net ("net (net)", its themes under it; a net counts a respondent once), ordered by their counts, with N and %; then No theme (answers a coder decided have none, when there are any), Coded, Coded by hand, Coded by rules and Uncoded. Every % is of the respondents who answered; with several themes an answer the themes add up to more than 100 %. With Also add sentiment and a codeframe built with it, each row adds Negative %, Neutral % and Positive % of its answers.
• Output stat — the table's statistics, which it also prints under it: Variable, Answered, Themes, Coverage ("87.5 % of the answers are coded"), Coded by hand, Coded by rules, Distinct uncoded answers (how many different answers are left to read), Percentages ("of the respondents who answered", and with several themes an answer "; a respondent can have several themes, so the themes add up to more than 100 %"), Nets when there are nets, and Codeframe (the model and when) for a codeframe the assistant built earlier. With sentiment: Sentiment ("negative 66.7 %, neutral 0.0 %, positive 33.3 % of 3 answers") and Net sentiment (positive minus negative, in points); asked of a codeframe built without it, Sentiment reads "not in this codeframe". Wire it into a Live tile to watch the coverage as new answers arrive.
• An older codeframe (version 1) — one the assistant built, or the example study's — gives the table it always did: theme rows as shares of the coded answers, Coded and Uncoded as shares of everyone who answered, Coverage "… % of the answers have a theme" and Percentages "a theme: of the coded answers; Coded and Uncoded: of all answers".
• The table counts answers, not weights; on weighted data it says "Weight: unweighted (the weight 'weight' is not applied)".
• Answers are matched by their text with case, spacing and Unicode form set aside, so a coder's decision holds wherever the same answer appears. The rules read the words: English negations, clauses, word forms (delay*), alternatives (slow|late), negated mentions (not_late) and words near each other (staff ~3 rude); no regular expressions.
• Later nodes offer the theme variable in their variable lists ("made by node"), whether its name is typed in Theme variable or left to a codeframe saved in the project. A multiple-choice theme variable is treated as a multiple-choice question: a donut, a Split by, a banner column or a Likert chart of it is refused. The sentiment variable is not offered.
• The flow's check reads the project's codeframes: a codeframe the run could not apply is an error of the flow ("Parameter 'codeframe' of node: path cannot be applied: …"), and so is one that codes a variable the questionnaire does not have ("… codes 'variable', which is not a variable of this questionnaire.").

---

Analyze

SurveyData in; a Table, a Stat, or both (Cluster (k-means) and Factor analysis also pass the data on). Frequencies and Crosstab wired straight to a Responses node also show instant counts in the inspector — see Instant counts. How each node treats an applied weight is summed up under Apply weight.

Which test. You can let a node choose, or name the test yourself:

Question	Node and choice
Does a mean differ between two groups?	t-test (Welch's by default, or Student's), or Group means with Test welch / student
… between three or more groups?	Group means with Test anova or welch_anova, and Post-hoc tukey / games_howell for every pair
Do ratings (ranks) differ between groups?	Compare groups (Mann-Whitney U, Kruskal-Wallis H, with Dunn's test for the pairs), or Group means with Test mannwhitney / kruskal
Do the same respondents answer two questions differently?	t-test with Design paired; Paired tests (Wilcoxon signed-rank, or McNemar for yes/no)
… three or more questions?	Paired tests (Friedman, with pairwise Wilcoxon tests)
… three or more yes/no questions (brands they know, channels they saw)?	Paired tests with Test cochran (Cochran's Q, with pairwise McNemar tests)
Is a mean different from a fixed value?	t-test with Design one_sample
Are two answers related?	Crosstab (chi-square, or Fisher's exact test for small counts); Correlation (Pearson, Spearman or Kendall)
How do several variables correlate?	Correlation matrix
What does an ordered answer (dissatisfied … satisfied) depend on?	Regression with Model ordinal
Which attributes drive an overall rating?	Key drivers
Which brands go with which attributes, or which regions with which answers?	Perceptual map
What would respondents pay?	Price sensitivity (Van Westendorp or Gabor-Granger)
Which items belong together?	Factor analysis, Principal components, Scale reliability

To draw a result — means with their intervals, a scree plot, a reach curve, odds ratios, a map, price curves — connect the node's table to a Result chart.

Three rules hold for all of these tests:

• Missing codes are not answers. A test you choose by hand — and the t-test, Correlation matrix, Paired tests, Factor analysis, Key drivers, Perceptual map, Price sensitivity and the ordinal Regression — leaves the codebook's missing codes (a "Refused" coded 9) out and says how many: "Missing codes left out = Trust: Acme: 44 (9 = Refused)" (in Paired tests, Factor analysis, Key drivers, a Perceptual map and Price sensitivity: "Missing codes = 44 answers with a missing code (9 = Refused) left out"; in the ordinal Regression: missing_codes = "Trust: Acme: 74 (9 = Refused)"). Descriptive statistics leaves them out too, counts them in its Missing column with the blanks and names the codes ("Missing codes = trust_acme: 9"). The defaults — Correlation spearman, Compare groups without Dunn's test, Group means with Test auto, the Crosstab chi-square — read the data as they always have, so a flow keeps its numbers, and say when they counted a missing code as an answer: "Missing codes counted as answers = Trust: Acme: 44 (9 = Refused); run Missing values first to leave them out" (missing_codes_counted among the statistics of Correlation and Compare groups). That is why Group means can give a Kruskal-Wallis H with Test auto and a different one with kruskal. Put Missing values first to have every node leave them out.
• What the data cannot carry is said in words. A two-group test asked of three groups, one respondent in a group, or no spread at all gives a line such as "Test = not run: Welch's t-test (unequal variances) compares two groups and Region has 3 — choose anova or welch_anova", not a number or a crash. A request that cannot work at all (a t-test of a grouping with three groups and none named) stops the run with the reason; the check warns about it before the run ("Gender has 3 answers (1 = Male, 2 = Female, 3 = Other); a t-test compares two — name them in Group A and Group B, unless the data this node reads holds only two of them."), and a Paired tests node given a number of variables its test cannot compare is an error of the check, in the run's own words ("McNemar compares exactly two variables; 3 were given. For three or more yes/no variables, use Cochran's Q.").
• The weight is used where there is a standard weighted form (Pearson's correlation) and otherwise the result says "unweighted (the weight 'weight' is not applied)" — see Apply weight.

Banner table

analyze.banner — the cross-break: several questions down the page, several breakdowns across it, each cell a column percentage with its count, and letters marking significant differences.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Questions (down)	variables (several)	required	nominal / ordinal variables	One block of rows per question.
Breakdowns (across)	variables (several)	required	nominal / ordinal variables	Each becomes a block of columns. Letters run across the whole table, but columns are only compared inside their own block.
Significance letters	checkbox	on	—	Mark significant differences with column letters.
Level	number	0.05	0.001–0.2	Significance level for the letters.
Multiple comparisons	choice	none	none, bonferroni	bonferroni corrects for the number of comparisons; none is the industry default.

A letter says this column is significantly higher than the column that letter names — and three rules keep that claim honest, all restated under the table:

• Columns are compared only inside their own block. The values of one breakdown are mutually exclusive, which is what the test assumes; columns from two different breakdowns overlap (a northerner is also under 35), so they are never compared.
• On weighted data the test uses the effective base (Kish), because weights make a sample behave like a smaller one. The banner uses the applied weight.
• A column with fewer than 30 respondents is not tested, and the table says which ones those were.

The same test puts letters on a Bar chart split into groups (Significance letters) and into every sheet of a Tab book (Excel), which writes each question of the study by the banner to a workbook. Those two leave the codebook's missing codes out and test each column on its respondents who answered; the Banner table counts a missing code as an answer and tests on everyone in the column — the same numbers when everyone answered and no missing code was given.

Cluster (k-means)

analyze.cluster — segments respondents on a set of items with k-means (the best of ten k-means++ starts, deterministic for a given seed). It adds a labeled nominal cluster variable to the data and returns the centroids. Wire its data output on to use the segments: later nodes offer the cluster variable in their variable lists (a Crosstab by segment, say), and Export file or Write table take it further.

The segmentation is drawn on the respondents as they are, never on the weight: on weighted data its statistics carry weight = "unweighted (the weight 'weight' is not applied)". To give the segments their weighted sizes, run a Frequencies of the cluster variable on the weighted data.

In: data (SurveyData) → Out: data (SurveyData), table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	ordinal / interval / ratio variables	Variables to segment on.
Clusters	whole number	3	2–12	Number of clusters.
Cluster variable	text	cluster	—	Name of the new cluster variable.
Seed	whole number	42	—	The random draws of ten k-means++ starts; the start with the smallest within-cluster sum of squares is kept, so the clusters do not depend on the order of the rows.
Standardize items	checkbox	on	—	Put the items on a common scale before clustering.
Number clusters by	variable	—	one of the Items	Optional. Empty, the clusters are numbered by size, largest first; set, by this item's mean, lowest first. Two segments of close sizes swap numbers when a few respondents come or go, and a name given to a number (with Derive) then lands on the other segment; numbered by the item that tells them apart, each keeps its number.

Compare groups

analyze.compare_groups — a nonparametric test of whether a variable differs between groups: "Mann-Whitney U for two groups, Kruskal-Wallis H for more — with Dunn's test on every pair of groups after Kruskal-Wallis when asked."

In: data (SurveyData) → Out: stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	ordinal / interval / ratio variables	The variable to compare.
Group	variable	required	nominal / ordinal variables	The grouping variable.
Test	choice	auto	auto, kruskal, mannwhitney	auto uses Mann-Whitney U for two groups and Kruskal-Wallis H for more.
Post-hoc	choice	none	none, dunn	Dunn's test compares every pair of groups after Kruskal-Wallis — one line per pair with z and the adjusted p. With it the codebook's missing codes are left out of the test too, and the result says how many; without it they are counted as answers, as they always were, and the result says so (missing_codes_counted) — so the Kruskal-Wallis result can change. Run Missing values first to have both leave them out.
Dunn p adjustment	choice	holm	holm, bonferroni	Shown only with Post-hoc dunn.

• Without Dunn's test the statistics are statistic and p_value, with groups (Kruskal-Wallis) or group_a and group_b (Mann-Whitney), and missing_codes_counted when a missing code was counted as an answer. With it they name the test ("Kruskal-Wallis H"), statistic, p_value, groups, n and posthoc ("Dunn's test (Holm)"), then one entry per pair of groups, by their value labels — Capital vs North = "z = 0.193, p = 1.0" (the adjusted p) — and missing_codes for the codes left out.
• Rule: "Dunn's test follows Kruskal-Wallis — set Test to kruskal or auto, or Post-hoc to none." (an error with Test mannwhitney).
• For the same test with the means table, the effect size and a table of the pairs, use Group means with Test kruskal and Post-hoc dunn.

Rank tests have no standard weighted form: on weighted data the test runs on the respondents as they are, and the statistics add weight = "unweighted (the weight 'weight' is not applied)".

Conjoint

analyze.conjoint — part-worths and attribute importance from a choice-based conjoint question: what people gave up to get what.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Conjoint question	text	required	—	The question's id or name. Its attributes and design are read from the questionnaire.

Every level is measured against the first level of its own attribute, which sits at zero. Importance is an attribute's range of part-worths over the sum of all ranges — of the levels you tested: price from £10 to £12 will look unimportant beside price from £10 to £100, and that is a fact about your design. Estimates are aggregate; for individual-level part-worths, export with Conjoint data for HB. The statistics give the Question, the Base ("812 respondents"), Tasks read, the Method ("conditional logit (aggregate)"), the Reference, Pseudo R² and a Note that importance is of the levels tested. After Apply weight the part-worths — and so the importances — are fitted on the weighted choices, the Base gives the weighted total beside the people ("812 respondents (798 weighted)") and Weight names the column. See MaxDiff and Conjoint.

Share of preference

analyze.conjoint_shares — a market simulator: what the part-worths predict a market of specific products would do.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Conjoint question	text	required	—	The conjoint question's id or name.
Products	JSON	required	—	Name → one level code per attribute, e.g. {"Ours": {"brand": 1, "price": 15}}. A half-specified product has no utility.
Include "none of these"	checkbox	off	—	Only when the question offered it. Leaving it out rescales everyone who would have walked away into buyers.

A half-specified product has no utility, so give every attribute a level. Leaving "none of these" out when the question offered it rescales the people who would have walked away into buyers.

The table has one row per product (its utility and share) and a footer; the stat output carries the same footer: Question, Base ("812 respondents"), Method ("logit rule on aggregate conditional-logit part-worths") and Note — "shares of the products listed, not market shares", or with Include "none of these" "shares of the products listed and of choosing none, not market shares". After Apply weight the shares come from part-worths fitted on the weighted choices, as in Conjoint; the Base adds the weighted total ("812 respondents (798 weighted)") and Weight names the column.

Correlation

analyze.correlation — "Pearson, Spearman or Kendall correlation between two variables, with its p-value and N — Pearson with a 95 % CI, and weighted when a weight is applied."

In: data (SurveyData) → Out: stat (Stat)

Parameter	Type	Default	Allowed	Meaning
X	variable	required	ordinal / interval / ratio variables	First variable.
Y	variable	required	ordinal / interval / ratio variables	Second variable.
Method	choice	spearman	pearson, spearman, kendall	Spearman and Kendall (tau-b) correlate ranks and suit rating scales; Pearson measures a straight-line relationship and gives a 95 % CI. Pearson is weighted when a weight is applied, the rank methods are not and say so. Pearson and Kendall leave the codebook's missing codes (a 9 = Refused) out and say how many. Spearman, the default, counts them as answers — as this node always has, so flows made before keep their numbers — and says so in missing_codes_counted; run Missing values first to leave them out.

The list names each code under it: pearson ("Pearson r"), spearman ("Spearman rho"), kendall ("Kendall tau-b"). What each gives:

Method	Statistics
spearman	rho, p_value, n; missing_codes_counted when a missing code was counted as an answer
pearson	method ("Pearson"), r, p_value, n, the confidence interval lower – upper and confidence (0.95); missing_codes for the codes left out
kendall	method ("Kendall tau-b"), tau, p_value, n; missing_codes

p is two-sided. Kendall's tau-b corrects for ties, which a rating scale is full of. A flow saved before Method existed has no value there and runs Spearman, exactly as before.

On weighted data Pearson's r is the weighted coefficient; its p and interval are computed on Kish's effective base (n_effective), so a weight never makes a correlation look more certain than the respondents behind it, and weight names the column. Equal weights give the unweighted result. Spearman and Kendall run on the respondents as they are and add weight = "unweighted (the weight 'weight' is not applied)".

Correlation matrix

analyze.correlation_matrix — "Pearson, Spearman or Kendall correlations between every pair of several variables — coefficients with significance marks, or one row per pair with p and N."

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variables	variables (several)	required	ordinal / interval / ratio variables	The variables to correlate, every pair of them.
Method	choice	spearman	pearson, spearman, kendall	Spearman and Kendall (tau-b) correlate ranks and suit rating scales; Pearson measures straight-line relationships and is weighted when a weight is applied.
Missing answers	choice	pairwise	pairwise, listwise	Pairwise uses, for each pair, everyone who answered both, so N can differ between pairs; listwise keeps only the respondents who answered every variable.
p adjustment	choice	none	none, holm, bonferroni, fdr_bh	Allows for the number of pairs tested — Holm and Bonferroni hold the chance of any false finding at 5 %, fdr_bh (Benjamini-Hochberg) the share of false findings among the significant ones. The marks then follow the adjusted p.
Layout	choice	matrix	matrix, pairs	Matrix prints the lower triangle with * p < .05, ** p < .01, *** p < .001; pairs gives one row per pair with the coefficient, p and N.

matrix — the lower triangle, variables by their labels, — on the diagonal:

| Variable             | Age    | Trust: Acme | Trust: Globex | Overall satisfaction |
| Age                  | —      |             |               |                      |
| Trust: Acme          | -0.031 | —           |               |                      |
| Trust: Globex        | -0.010 | 0.072       | —             |                      |
| Overall satisfaction | -0.042 | 0.045       | -0.067        | —                    |

pairs — one row per pair: Variable 1, Variable 2, the coefficient (r, rho or tau), p, the adjusted p when there is one (p (Holm)) and N. Read this one when N differs from pair to pair.

Under the table, and in the stat output: Method ("Spearman rank correlation", "Pearson correlation", "Kendall rank correlation (tau-b)"), Missing ("pairwise: each pair uses everyone who answered both", or "listwise: only respondents who answered every variable"), N (a range such as 112–194 when pairs differ), p adjustment ("none", "Holm, over 3 pairs", or "Holm, over the 2 pairs computed (of 3)" when a pair could not be computed — it reads n/a in the matrix), Marks (matrix layout; "(adjusted p)" when adjusted) and Missing codes left out — the matrix leaves the codebook's missing codes out with every method. On weighted data a Pearson matrix is weighted and adds Weight and Base ("weighted coefficients; p on Kish's effective base"); Spearman and Kendall say "Weight = unweighted (the weight 'weight' is not applied)".

To draw the matrix, connect its table to a Result chart (Kind heatmap): the lower triangle with the table's significance marks.

Perceptual map

analyze.correspondence — "Correspondence analysis of a crosstab, or of the attributes checked for each brand." It gives the dimensions and their share of the inertia, and where each row and column lies on the map, with contributions and quality. A table of counts says cell by cell which brands are seen as modern and which as good value; the map shows it at a glance: a brand lies toward the attributes it gets more of than the average brand does. Connect any of its tables to a Result chart to draw the map.

In: data (SurveyData) → Out: table (Table), rows (Table), columns (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Table	choice	crosstab	crosstab, attributes	crosstab counts the respondents in each pair of answers of Rows and Columns (a multiple-choice variable counts each answer chosen). attributes counts, for each answer of Rows — the brand, in data with a row per respondent and brand — the respondents who checked each of the Attributes.
Rows	variable	required	nominal / ordinal variables	The points of one kind on the map — brands, regions, segments. A respondent whose answer is blank or one of the codebook's missing codes is left out.
Columns	variable	—	nominal / ordinal variables	Crosstab: the points of the other kind — the answers crossed with Rows. Shown only with Table crosstab.
Attributes	variables (several)	—	—	Attributes: two or more 0/1 variables, one per attribute ("is modern", "good value"), such as the ones Explode multiple choice makes; a blank or a missing code counts as unchecked. Shown only with Table attributes.
Counts as yes (attributes)	answer code	—	—	The code, or a list of codes, that counts as checking an attribute. Empty works for 0/1 variables. Shown only with Table attributes.
Dimensions in the tables	whole number	2	1–10	How many dimensions the rows and columns tables give coordinates, contributions and cos² for; the map shows the first two, and the inertia table lists them all.

The Table list has crosstab ("Rows by Columns") and attributes ("Attributes checked for each answer of Rows"); the node's card reads region × satisfaction, or brand × modern, good_value, friendly for attributes.

• Two layouts. crosstab crosses two questions: Rows region, Columns brand_used. attributes needs data with one row per respondent and brand, the brand in a column of its own (Rows) and one 0/1 column per attribute. No node reshapes a brand grid into that layout: bring such a file in with a Data file node. Rows can also be a question about the respondent — a region, a segment — with the 0/1 columns of an exploded multiple-choice question as Attributes: the map then shows which regions check which options.
• Rules: "A crosstab map crosses Rows with Columns — choose the Columns variable.", "An attribute map counts the Attributes checked for each answer of Rows — choose two or more attribute variables." and "A perceptual map of attributes needs two or more attribute variables; 1 was given." (errors); "Counts as yes is read only with Table = attributes — set it, or clear Counts as yes." (a warning).

The outputs. The preview shows the three tables one under another, the rows and columns tables each under its output's name.

Output	What it holds
table	one row per dimension — Dimension, Singular value, Principal inertia, % of inertia, Cumulative % — with the statistics under it
rows	one row per answer of Rows, by its label — Mass, Quality, Inertia %, then per dimension Dim 1, Contribution 1 %, cos² 1, Dim 2, …; under it Coordinates "principal (symmetric map)", Contribution "% of the dimension's inertia", Quality "cos² of the 2 dimensions shown"
columns	the same for the answers of Columns, or for the Attributes by their labels
stat	Map ("Region × Overall satisfaction", or "Region × Attributes"), Rows, Columns, N, Counts as yes (attributes: "1 = Yes"), Total inertia, Dimensions, Dimension 1 %, Dimension 2 %, …, Map % (the share of the inertia the map shows); for a crosstab of two single-answer questions Chi-square, df and p; Note, Not in the map, Excluded, Excluded because ("no answer (or a missing code) to Rows or Columns") and Missing codes when there is something to say

• Reading the map. Rows and columns are both in principal coordinates — the symmetric map R's ca and FactoMineR draw by default. Each axis is a dimension, with its share of the table's inertia (its departure from independence). A point's Quality says how well the dimensions shown represent it; its Contribution how much it shapes a dimension.
• Signs are arbitrary in any correspondence analysis: here each dimension points toward the row that contributes most to it, so a map can be the mirror image of another package's.
• The chi-square test of independence is given for a crosstab of two single-answer questions. Where a respondent can be counted in several cells — a multiple-choice variable, an attribute map — there is no test.
• An answer nobody gave is named, not drawn: "Not in the map = South (nobody counted in them)". A table with two rows or two columns has one dimension: "Note = the table has one dimension (two rows or two columns), so the map is a line".
• Refusals, with the reason: a table smaller than two rows by two columns ("Region × Gender: A perceptual map needs at least two rows and two columns with counts; this table has 3 × 1."), rows that all have the same profile ("… so there is nothing to map: the table is independent."), attributes that are not 0/1 without Counts as yes ("The attributes hold 1, 2, 3, 4, 5, 9: name the code (or codes) that counts as checking an attribute in Counts as yes — yes outside a flow."), and a multiple-choice attribute (run Explode multiple choice first, or use the crosstab layout with it as Columns).
• On weighted data every cell is a sum of weights, so the map is weighted, and the statistics add Weight and Weighted N. The chi-square test counts respondents and says so: "Chi-square counts = respondents (unweighted), as a test of a table must".

Crosstab

analyze.crosstab — "Cross-tabulation with a chi-square test and Cramér's V, or Fisher's exact test for small tables."

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Rows	variable	required	nominal / ordinal variables	Row variable.
Columns	variable	required	nominal / ordinal variables	Column variable.
Percentages	choice	col	none, row, col, total	Which percentages to show: none, row, col (column %) or total.
Significance test	checkbox	on	—	Add the test chosen under Test. (It was labeled Chi-square test before Fisher's test was added; stored flows keep their setting.)
Test	choice	chi2	chi2, fisher	Chi-square with Cramér's V, on Kish's effective base when weighted. Fisher's exact test suits small counts; for 2 × 2 it gives the odds ratio and its 95 % CI, larger tables use the Fisher-Freeman-Halton test (exact, or from 20,000 random tables with a fixed seed when there are too many to sum). Fisher counts respondents, unweighted, and leaves the codebook's missing codes out of the table and the test; chi-square keeps them as rows and columns of their own, as it always has, and says so.

For a multiple-answer row variable the table counts respondents, not answers, and does not offer a chi-square test (its independence assumption does not hold).

Chi-square (chi2, the default) gives χ², df, p, Cramér's V and N. A declared missing code (a "Refused") stays a row or column of its own, as it always has, and the statistics then say "Missing codes counted as answers = Trust: Acme: 12 (9 = Refused); run Missing values first to leave them out".

Fisher's exact test (fisher) is the test to use when cells are small — the chi-square's approximation is poor when expected counts fall below 5. The codebook's missing codes are left out of the table and the test ("Missing codes left out = …").

• A 2 × 2 table gets Test "Fisher's exact test", the two-sided p, the Odds ratio (conditional maximum likelihood) with its exact OR 95% CI, and Odds ratio of, which says which way round it is read: "Under 50 (vs 50 or over) for Male over Female". Estimate reads "conditional maximum likelihood with its exact interval, as R's fisher.test defines them (R stops its root search sooner, so its printed values can differ slightly on sparse tables)" — the p-values agree with R's.
• A larger table gets Test "Fisher-Freeman-Halton exact test" and p, summed exactly over every table with the observed margins when there are at most 200,000 of them (p method "exact, over every table with these margins"), and otherwise estimated from 20,000 random tables drawn from a fixed seed, so a rerun gives the same p: p method "Monte Carlo, 20,000 random tables with these margins from a fixed seed (± 0.0021); too many tables to sum exactly", the ± being the Monte Carlo error.
• A table the test cannot use reads "Test = Fisher's exact test: not run — the table has fewer than two rows or columns with answers in them".
• Rule: "Fisher's exact test is not run while Significance test is off." (a warning).

After Apply weight, the cells and totals are sums of weights (to one decimal) and the percentages are taken from them the same way. The chi-square test is run on the counts scaled down to the effective (Kish) sample size, as the Banner table does: weights make a sample behave like a smaller one. With Significance test on, the statistics keep N (the respondents counted) and add Weighted N, Weight and Base: for chi-square also Effective N, with Base "effective (Kish) for the test; weighted counts shown"; for Fisher's test, which needs whole counts, Base "the test counts respondents (an exact test needs whole counts); weighted counts shown". With the test off they are Weighted N and Weight. For a multiple-answer row variable the percentages are of each column's weighted base, and an Unweighted base row follows the "Base (respondents answering)" row.

Data check

analyze.data_check — "The data against its codebook before any analysis: each problem, with how many rows have it and examples." It finds problems such as values outside the valid range, codes without a label, duplicate IDs and columns nobody declared.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variables	variables (several)	empty	—	Check only these. Empty checks every column and every codebook variable.

One row per problem, errors first, with the columns Severity (error or warning), Variable, Problem, Rows (how many rows have it), Examples (up to five offending values with their counts — 17 (1), 999 (1) — and "… (3 more)" beyond five) and Code:

Problem	Code
outside the valid range 18–75	OUT_OF_RANGE
codes the codebook has no label for	INVALID_LABEL_VALUE
the same ID on more than one row	DUPLICATE_ID
weight values that are not numbers	INVALID_WEIGHT
missing codes without a label	MISSING_VALUE_WITHOUT_LABEL
the weight column is not in the data	MISSING_WEIGHT_COLUMN
in the codebook but not in the data	MISSING_COLUMN
in the data but not in the codebook	EXTRA_COLUMN

• A declared missing code is not "outside the valid range": a 999 "Refused" is not flagged for an age of 18–75.
• Columns the codebook does not know, and codebook variables the data lacks (common after a Select columns), are gathered into one row each.
• The weight column and the response metadata beside the answers — respondent_id, duration_s, partial, started_at, the url_* parameters and the like — are expected there, so they are not problems; the statistics name them under Not in the codebook, as expected (for example "weight (the weight)").
• The statistics: Checked ("9 variables, 240 rows"), Errors, Warnings, and with nothing wrong Result "no problems found" (the table is then empty).
• It counts rows, so on weighted data it adds "Weight: unweighted (the weight 'weight' is not applied)".

Put it right after the source and wire its table into a Report section or a Live tile: it is the check a methods section's "data were screened for out-of-range values" rests on.

Describe

analyze.describe — a summary table of the data: one row per codebook variable with its label, scale, N, missing and unique values — or the codebook itself.

In: data (SurveyData) → Out: table (Table)

Parameter	Type	Default	Allowed	Meaning
Show	choice	variables	variables, codebook	variables: one row per variable with N, missing and unique values; codebook: the codebook table.

The rows include the variables earlier nodes made. The counts are of rows (records), which is what a completeness check is about; after Apply weight a weighted_n_valid column stands beside them — the sum of the weights of the rows that have a value, the weighted base a table of that variable would report.

Descriptive statistics

analyze.descriptives — "N, missing, mean, SD, median, minimum and maximum of several numeric variables — by group if you like, with quartiles, skewness and kurtosis on request."

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variables	variables (several)	required	ordinal / interval / ratio variables	The codebook's missing codes (a 99 "Don't know") are not answers and count as missing.
By group	variable	—	nominal / ordinal variables	Optional. One row per variable and group; a blank or a missing code of the group variable is no group. A multiple-choice question gives one group per option, and they overlap.
Quartiles, skewness and kurtosis	checkbox	off	—	Only with Layout long.
Layout	choice	long	long, means	With By group. long: a row per variable and group with N, mean, SD, median, minimum and maximum. means: a row per variable and a column per group with its mean, a compact profile of the groups; their sizes are under the table.

One row per variable — or per variable and group — with Variable, Label, the group, N, Missing, Mean, SD, Min, Median and Max; the checkbox adds Q1 and Q3 (linear interpolation, R's default) and Skewness and Kurtosis (the bias-corrected G1 and excess G2 that SPSS and Excel report).

• A missing code is not an answer. A declared missing code, and a value that is not a number, count in Missing; the statistics name them (Missing codes = trust_acme: 9, Not numbers). This is the difference from Describe, which counts what is in the cells.
• Undefined is blank. The SD of one answer, skewness below three answers and kurtosis below four are empty cells, never nan.
• By group: a respondent whose group is blank or a missing code is in no group (Not in a group). A multiple-choice By group gives one group per option, of everyone who chose it; a respondent who chose two is in both, and Groups says the groups overlap.
• On weighted data the mean, SD, median and quartiles are weighted — the same formulas as Group means, so the two never disagree — beside a Weighted N column, while N and Missing stay counts of respondents and Min and Max those of every answer. A respondent weighted 0 (or with no weight) is set aside from the weighted statistics, the SD's n included, so one weighted answer has no SD. The statistics add Weight, Weighted N, Effective N (Kish), the Design effect and the Note "mean, SD and median are weighted; N and Missing count respondents" (with Quartiles, skewness and kurtosis checked: "mean, SD, median and quartiles are weighted; N and Missing count respondents; skewness and kurtosis are unweighted") — with "; rows weighted 0 are left out of the weighted statistics" after "count respondents" when any row weighs 0.

To chart the means, connect its table to a Result chart: means (each mean with its 95 % interval) or means_sd (± 1 SD), a row per variable with its base; with By group, a color per group, the legend giving each group's base (Capital (n = 170–197) when it differs between variables).

Key drivers

analyze.drivers — "Which of several predictors matter most for an outcome." It gives each one's share of R², by Johnson's relative weights or the Shapley value (LMG), beside the correlations and standardized betas. A regression's coefficients answer another question — what changes when one rating moves and the others stay put — and when the ratings correlate, as ratings of one brand do, a coefficient can shrink or even flip sign because a neighbor took its share. Key drivers splits the model's R² between the drivers instead, so the shares add up to R² (100 %).

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Outcome	variable	required	—	What the drivers explain — a rating, a score, a likelihood to recommend. A respondent missing it or any driver is left out; the codebook's missing codes count as missing.
Drivers	variables (several)	required	—	Two or more numeric predictors — ratings of attributes, scores, 0/1 variables. A nominal variable with more than two answers has no amounts to weigh; make a 0/1 variable per answer with Explode multiple choice or Derive.
Importance	choice	relative_weights	relative_weights, shapley	relative_weights is Johnson's method, which splits R² through the orthogonal variables closest to the drivers (as R's rwa). shapley averages each driver's gain in R² over every order in which the drivers could enter the model (LMG, as R's relaimpo) — exact, for at most 15 drivers. Both add up to R², shown as 100 %; they agree closely.

The Importance list has relative_weights ("Johnson's relative weights") and shapley ("Shapley value (LMG)"); the card reads satisfaction ~ price, service, range, staff.

The table has one row per driver, the largest share first — Rank, Driver (by its label), r (its correlation with the outcome), Beta (the standardized coefficient), Beta p (its t-test in the regression), VIF, Relative weight (or Shapley value) and % of R²:

| Rank | Driver        | r      | Beta   | Beta p  | VIF  | Relative weight | % of R² |
| 1    | Trust: Acme   | -0.177 | -0.175 | 0.02088 | 1.01 | 0.0308          | 88.1    |
| 2    | Trust: Globex | -0.068 | -0.06  | 0.4218  | 1.0  | 0.0041          | 11.7    |
| 3    | Age           | -0.011 | 0.006  | 0.9381  | 1.01 | 0.0001          | 0.2     |

Under it, and in stat: Method ("Johnson's relative weights", or "Shapley value decomposition of R² (LMG)"), Outcome, Drivers, N, R², Adjusted R², the model's F, df ("3, 173") and p, Excluded, Excluded because ("a missing value in the outcome or any predictor (listwise)") and Missing codes ("138 answers with a missing code (9 = Refused) left out").

• Warnings (in Warning): a VIF of 10 or more ("strong collinearity — VIF …: their betas are unstable, and the importance splits the variance they share between them"), and a suppressor, a driver whose beta has the other sign than its correlation ("… the beta has the opposite sign of the correlation (a suppressor), so the share says little about it on its own"). Drivers that explain the outcome exactly (R² of 1) get no test: "Note = the drivers explain the outcome exactly, so there is no residual to test against".
• Rules (errors of the check, before the run): "Key drivers splits R² between two or more predictors; 1 was given. For one predictor, its correlation with the outcome is the whole story." and, for shapley, "The Shapley value decomposition averages over every order of the predictors: with 16 that is 65,536 subset regressions, and it is computed for at most 15 predictors. Use Johnson's relative weights, which come close to it, or fewer predictors."
• Refusals — the run stops on this node with the reason: a nominal driver with more than two answers ("Gender is nominal with 3 answers (1 = Male, 2 = Female, 3 = Other): its codes are not amounts to weigh. Make a 0/1 variable per answer (Explode multiple choice, or Derive) and use those."), a driver or outcome that is the same for everyone, drivers that are a combination of each other ("… are collinear: one is a combination of the others, so their shares cannot be told apart. Leave one of them out."), and fewer respondents than the regression needs.
• On weighted data the correlations, betas, R² and shares come from the weighted correlation matrix, and the tests are taken on Kish's effective N: the statistics add Weight, Effective N and Tests on ("Kish's effective N"), and df has decimals ("3, 167.33").

Connect its table to a Result chart for the bars: each driver's share of R², largest first, a driver with a negative beta in a second color (legend "positive beta", "negative beta"), titled "Key drivers of Overall satisfaction" with the method, R² and N on the line under it.

Factor analysis

analyze.factor — "Exploratory factor analysis of a set of items — loadings, communalities, variance explained, KMO and Bartlett's test, factor correlations, and factor scores if wanted." It is the check before a set of items is averaged into a scale: which items move together, how strongly each belongs to each factor, and whether the items share enough to be factored at all.

In: data (SurveyData) → Out: data (SurveyData), loadings (Table), variance (Table), correlations (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	ordinal / interval / ratio variables	Three or more, analyzed as standardized scores (their correlations). A respondent missing any item is left out; the codebook's missing codes count as missing.
Factors	whole number	—	at least 1	Empty chooses the number by the rule below.
Number of factors by	choice	kaiser	kaiser, parallel	When Factors is empty. kaiser keeps the eigenvalues above 1; parallel keeps the factors whose eigenvalue beats the 95th percentile of 100 random data sets of the same size, drawn from Seed.
Extraction	choice	minres	minres, principal, ml	minres (minimum residual) as psych and factor_analyzer; principal is iterated principal axis factoring; ml is maximum likelihood, as R's factanal, and adds a test of fit; it is started from 14 fixed points and keeps the best, and warns when they disagree (often a sign of too many factors).
Rotation	choice	varimax	varimax, promax, oblimin, none	varimax keeps the factors uncorrelated. promax and oblimin let them correlate and report how much (the correlations output); their loadings are the pattern matrix.
Sort items by factor	checkbox	off	—	—
Hide loadings below	number	0	0–1	Blank out the loadings smaller than this in the table (0 shows every loading), so the structure reads at a glance.
Add factor scores	checkbox	off	—	Regression-method scores, one variable per factor, missing for the respondents left out.
Score variable prefix	text	factor_	—	The scores are named 1, 2, … — with Add factor scores on. The prefix itself is not a variable. Shown only with Add factor scores checked.
Seed (parallel analysis)	whole number	42	—	Shown only with Number of factors by parallel.

The lists name each code under it: minres ("minimum residual"), principal ("principal axis"), ml ("maximum likelihood"); varimax ("orthogonal"), promax ("oblique"), oblimin ("oblique, quartimin"); kaiser ("eigenvalues above 1"), parallel ("parallel analysis").

The outputs. The preview shows the three tables one under another, each under its output's name.

Output	What it holds
loadings	one row per item — Variable, Label, Factor 1, Factor 2, …, Communality, Uniqueness and MSA (the item's measure of sampling adequacy)
variance	one row per possible factor — Eigenvalue, % of variance, Cumulative %, then for the factors kept Extracted SS, Extracted %, Extracted cumulative % and, after a rotation, Rotated SS; with parallel analysis also Random 95th percentile
correlations	the correlations between the factors — after promax or oblimin; an orthogonal solution notes "an orthogonal solution: the factors are uncorrelated by construction"
stat	Extraction, Rotation ("promax (power 4), Kaiser-normalized"), Factors, Factors chosen by ("fixed", "Kaiser criterion (eigenvalues above 1)", "parallel analysis (95th percentile of 100 random data sets, seed 42)"), Items, N, Variance explained %, KMO, Bartlett chi-square, Bartlett df, Bartlett p, RMSR; with ml the test of fit, Fit chi-square, Fit df and Fit p; Warning when there is something to warn about; Scores with Add factor scores; Excluded and Excluded because; Missing codes
data	the data passed on — with Add factor scores, plus factor_1, factor_2, …

• Each factor is signed so its loadings sum positive, and the factors are ordered by the variance they carry. The numbers reproduce R's psych::fa and the factor_analyzer package (and factanal for maximum likelihood).
• Refusals, with the reason: fewer than three items ("A factor analysis needs at least three items; 2 were given."), no more respondents than items, an item everyone answered the same, an item that copies or totals others, as many factors as items, and with ml more factors than maximum likelihood can fit ("Maximum likelihood cannot fit 3 factors to 5 items: … Use at most 2.").
• Warnings (in Warning): a KMO below 0.5, an item with (almost) no uniqueness left, a rule that suggests more factors than can be fitted (the most that fit is used), with minres or principal more factors than the items' correlations can identify, and with ml, starts that disagree — a reason to compare a solution with fewer factors.
• Factor scores. With Add factor scores checked the data output carries factor_1, factor_2, … (interval), blank for the respondents left out, and later nodes offer them ("factor 1 score · made by node") — a Group means of factor_1 by region, a Regression on them. With Factors empty the pickers list every score the analysis could make (one fewer than the items), the ones after the first as "factor 2 score (empty unless the rule keeps it) · made by node". The run makes the scores of the factors it kept, and of the others the ones a node downstream reads, empty and labeled "Factor 2 score (not made: the Kaiser criterion kept 1 factor)"; Scores names them ("factor_1 (regression method); factor_2 empty: the Kaiser criterion kept 1 factor"). A score nothing reads is not added, so the data, its exports and its tables carry only the scores made and the ones in use.
• It is unweighted: on weighted data each table says "Weight: unweighted (the weight 'weight' is not applied)". The scores of a later Group means are weighted as usual.

Frequencies

analyze.freq — the distribution of one variable with value labels, %, and cumulative %.

In: data (SurveyData) → Out: table (Table)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	—	The variable to tabulate.
Exclude missing	checkbox	on	—	Leave blank answers out of the table.
Sort	choice	value	value, freq, label	value (by code), freq (largest count first) or label (alphabetical).

Cumulative % is the running share of the counts (after Apply weight, of the weights), rounded once, so the last answer always reads 100.0, even when the % column, each row rounded on its own, adds up to 99.9 or 100.1.

For a multiple-answer question, each option's share is of the respondents who answered, so the % column sums above 100 %; the base is a row of its own and there is no cumulative column.

After Apply weight, N is the sum of the weights (shown to one decimal) and the percentages are taken from the sums as they are, not from the rounded N — so, over the same respondents, they agree with the Bar chart's and the Likert chart's to the last digit. An Unweighted N column beside N gives the respondents actually counted — the number a reader judges a percentage by — and the statistics add Weighted N and Weight. For a multiple-answer question the weighted counts are rounded to whole numbers, the base row carries the unweighted base under Unweighted N, and the statistics read, for example, "812 respondents (798 weighted)".

MaxDiff

analyze.maxdiff — what a best–worst question found: one row per item, best first, with the counting score, the conditional-logit utility and the share it implies.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
MaxDiff question	text	required	—	The question's id or name. Its design is read from the questionnaire, so nothing has to be re-entered here.
Estimate	choice	both	both, counts, utilities	How each item is scored. both: the counting score and the utilities, with the share each implies. counts: the counting score alone — best minus worst over the times shown, which anyone can recount. utilities: the utilities, a conditional logit on the choices the design actually showed, with the counting score still beside them. A Result chart's utilities and shares need both or utilities.

The design is read from the questionnaire, so the only thing to name is the question. Utilities come from an aggregate conditional logit; individual ones need hierarchical Bayes, which Choice data for HB exports for. Each respondent's own counting scores, as variables a crosstab or a cluster can use, come from MaxDiff scores (Prepare). The footer gives the Question, the Base ("812 respondents"), Tasks read, the Method, the Reference item and Pseudo R².

After Apply weight every column is weighted: Shown, Best and Worst are sums of weights (rounded to whole numbers), the Score follows them, and the Utility and Share % are fitted on the weighted choices. The footer adds Weight and gives the weighted total beside the people: "812 respondents (798 weighted)". Before this was fixed the Utility and Share % columns stayed unweighted; a weighted flow run again gives different numbers there.

Group means

analyze.means — "Mean, SD, median and N of a variable by group, with a significance test — chosen for you, or by hand with post-hoc comparisons of every pair of groups."

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	ordinal / interval / ratio variables	The variable whose mean is compared.
By	variable	required	nominal / ordinal variables	The grouping variable.
Significance test	checkbox	on	—	Add the significance test.
Test	choice	auto	auto, student, welch, anova, welch_anova, mannwhitney, kruskal	Auto chooses by scale and groups (t-test or ANOVA for interval data, Mann-Whitney or Kruskal-Wallis for ordinal). Student and Welch compare two groups (Welch does not assume equal variances), ANOVA and Welch's ANOVA two or more, Mann-Whitney and Kruskal-Wallis compare ranks. A test chosen here reports df and an effect size and leaves the codebook's missing codes out. Auto counts them as answers, as it always has, and says so (Missing codes counted as answers); run Missing values first to leave them out.
Post-hoc	choice	none	none, tukey, games_howell, dunn	Every pair of groups compared after the test — Tukey's HSD after ANOVA, Games-Howell after Welch's ANOVA, Dunn after Kruskal-Wallis. The pairs appear under the table.
Dunn p adjustment	choice	holm	holm, bonferroni	How Dunn's p-values allow for the number of pairs. Tukey and Games-Howell allow for it themselves. Shown only with Post-hoc dunn (and a Test other than auto).

The Test list names each code under it: student ("Student's t"), welch ("Welch's t"), anova ("one-way ANOVA"), welch_anova ("Welch's ANOVA"), mannwhitney ("Mann-Whitney U"), kruskal ("Kruskal-Wallis H"); Post-hoc has tukey ("Tukey HSD"), games_howell ("Games-Howell"), dunn ("Dunn's test").

auto (the default, and what every flow saved before Test existed runs) picks by the variable's scale and the number of groups: Student's t-test or a one-way ANOVA for interval and ratio data, Mann-Whitney U or Kruskal-Wallis H for ordinal data. Its footer is a single line such as "Kruskal-Wallis H = 2.417; p = 0.2986; N = 120; Variable = Satisfaction".

A test chosen by hand is named in the footer with its statistic, df, p and an effect size, and leaves the codebook's missing codes out of the table and the test ("Missing codes left out = …"):

Test	Footer
student, welch	Test ("Welch's t-test (unequal variances)"), t, df, p, Mean difference, Difference ("Male − Female"), 95% CI, Cohen's d
anova, welch_anova	Test ("One-way ANOVA", "Welch's ANOVA"), F, df ("2, 237"; Welch's has decimal df), p, η²
mannwhitney	Test, U, p, rank-biserial r
kruskal	Test ("Kruskal-Wallis H"), H, df, p, ε²

A two-group test asked of more groups is not run and says what to choose instead: "Test = not run: Welch's t-test (unequal variances) compares two groups and Region has 3 — choose anova or welch_anova".

Post-hoc comparisons. With Post-hoc set, the footer adds a summary ("Post-hoc = Tukey HSD: 0 of 3 pairs differ at p < 0.05") and a second table prints under the means table — in the preview and in the report — headed Post-hoc: Tukey HSD, one row per pair of groups by their value labels:

Post-hoc	Columns
tukey (Tukey-Kramer for unequal groups)	Pair, Difference, 95% CI low, 95% CI high, q, p
games_howell	the same, with df before p
dunn	Pair, Mean rank difference, z, p (unadjusted), p (Holm) (or p (Bonferroni))

Tukey and Games-Howell give the difference of the means with its simultaneous interval, the studentized range q and a p that already allows for the number of pairs; Dunn gives the difference of the mean ranks, z, and p before and after the adjustment. A p below 0.0001 keeps four significant digits (3.363e-07); a Tukey or Games-Howell p below 1e-07 reads < 1e-07, because the studentized range is not computed finely enough to give a smaller one — past that point it came out as the same tiny number for every strong pair. Under the pairs: Method, Groups, Difference ("mean of the first group minus the second", or for Dunn "mean rank of the first group minus the second") and, for Tukey and Games-Howell, p "adjusted for the number of pairs by the method itself" — with "; < 1e-07 where it is smaller than SciPy computes the studentized range to" when a pair reads < 1e-07.

Rules (errors, then warnings):

• "Tukey's HSD follows a one-way ANOVA — set Test to anova, or Post-hoc to none."
• "Games-Howell follows Welch's ANOVA — set Test to welch_anova, or Post-hoc to none."
• "Dunn's test follows Kruskal-Wallis — set Test to kruskal, or Post-hoc to none."
• "Test is not run while Significance test is off." (warning)
• "Post-hoc is not run while Significance test is off." (warning)

After Apply weight, each group's mean, SD and median are weighted (the SD is scaled so that equal weights give exactly the ordinary sample SD, counting only the answers that carry weight — a respondent weighted 0 is set aside, and a group with one weighted answer has no SD, a blank cell; the median is the value at which the cumulative weight reaches half). N stays the number of respondents, and the significance test and the post-hoc pairs stay unweighted; the statistics add Weight and the note "means, SD and medians are weighted; N and the test are not", and the post-hoc table says "Weight: unweighted (the weight 'weight' is not applied)".

Connect its table to a Result chart to draw each group's mean with its 95 % interval (means) or ± 1 SD (means_sd), each group with its base (North (n = 97)). After a post-hoc test each mean carries letters: "Means sharing a letter do not differ (Tukey HSD, p < .05)."

Net Promoter Score

analyze.nps — detractors (0–6), passives (7–8) and promoters (9–10) of a 0–10 item, and the score with its standard error and 95 % confidence interval. Uses the applied weight.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
0–10 item	variable	required	ordinal / interval / ratio variables	A 0–10 likelihood-to-recommend item.

After Apply weight the group shares and the score are weighted and the standard error uses Kish's effective base; the N column still counts respondents, and the statistics add Weight. A Result chart of its table draws the detractors, passives and promoters as one stacked bar, with the score and its interval (Kind stacked).

Paired tests

analyze.paired — "The same respondents answering two or more questions — Wilcoxon signed-rank or McNemar for two, Friedman or Cochran's Q for three or more, with pairwise comparisons." Use it where Compare groups would be wrong because the answers are not from different people: one scale asked about two brands, a rating before and after a message, three concepts each rated by everyone, three brands each respondent has heard of or not.

In: data (SurveyData) → Out: table (Table), pairs (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variables	variables (several)	required	—	Two for Wilcoxon or McNemar, three or more for Friedman or Cochran's Q — asked of the same respondents, on the same answer scale. Differences are the first minus the second (for Friedman's pairs, A − B), as in the paired t-test. A respondent missing any of them is left out of all of them; the codebook's missing codes count as missing.
Test	choice	auto	auto, wilcoxon, mcnemar, friedman, cochran	Auto runs Wilcoxon signed-rank for two variables and Friedman for more. McNemar is for two yes/no questions, Cochran's Q for three or more. Wilcoxon and Friedman rank the answers, so they refuse a nominal variable.
Counts as yes (McNemar, Cochran's Q)	answer code	—	—	The answer code, or a list of codes, that counts as yes; every other answer counts as no. Empty works for 0/1 variables such as the ones Explode multiple choice makes. Shown only with Test mcnemar or cochran.
Same answer twice (Wilcoxon)	choice	wilcox	wilcox, pratt	wilcox drops the respondents who gave both the same answer before ranking, as R and SPSS do; pratt ranks them with the others and leaves them out of the sums. Shown with every Test but mcnemar and cochran (Friedman's pairwise tests use it too).
p-value	choice	auto	auto, exact, approximate	Whether the p-value is exact or an approximation. auto is exact for small samples — Wilcoxon: up to 50 pairs with no ties or zeros, or up to 13 with them; McNemar (and the pairs of Cochran's Q): fewer than 25 respondents who answered the two differently — and the normal (Wilcoxon) or chi-square (McNemar) approximation otherwise. exact is exact, for Wilcoxon up to 1,000 pairs (past that the normal approximation, with a note); approximate always takes the approximation.
Pairwise comparisons (Friedman, Cochran's Q)	choice	holm	holm, bonferroni, none	A Wilcoxon signed-rank test (after Friedman) or a McNemar test (after Cochran's Q) for every pair of variables, its p-value adjusted for the number of pairs by Holm's step-down method or by Bonferroni. The pairs output holds them. Shown with Test friedman, cochran or auto (auto is Friedman for three or more).

The Test list names each code under it: wilcoxon ("Wilcoxon signed-rank"), mcnemar ("McNemar, yes/no"), friedman ("Friedman, three or more"), cochran ("Cochran's Q, three or more yes/no"). auto never picks Cochran's Q: choose it by hand.

Counts as yes (McNemar, Cochran's Q) is a checklist of the first variable's answers (4 High, 5 Full) — check 4 and 5 for a top-two box. McNemar or Cochran's Q on variables that are not 0/1 without it stops with the answers to choose from: "McNemar needs to know which answer counts as yes: the variables hold 1 = No trust, 2 = Low, 3 = Medium, 4 = High, 5 = Full. Name that code (or a list of codes) in Counts as yes …" ("Cochran's Q needs to know …" the same way). Rule (a warning): "Counts as yes is read only by McNemar and Cochran's Q — set Test to mcnemar or cochran, or clear it." A number of variables the test cannot compare is an error of the check, before the run, in the run's own words: "McNemar compares exactly two variables; 3 were given. For three or more yes/no variables, use Cochran's Q.", "Friedman's test compares three or more variables; 2 were given. For two, use Wilcoxon signed-rank (or McNemar for yes/no).", "Cochran's Q compares three or more yes/no variables; 2 were given. For two, use McNemar."

Who is compared. Each respondent is compared with themselves, so a respondent who did not answer every variable is left out of all of them (listwise), and the codebook's missing codes count as missing. The footer says how many: Excluded (91), Excluded because ("a missing value in either variable (listwise)") and Missing codes ("95 answers with a missing code (9 = Refused) left out").

Direction. Every difference is the first variable minus the second — as in the t-test with Design paired, so the two agree in sign. W+ sums the ranks of the respondents whose first answer is higher.

What each test gives (the table output, with its statistics under it and in stat):

Test	Table	Statistics
Wilcoxon signed-rank	per variable and for the difference: N, Mean, SD, Median	Difference ("Trust: Acme − Trust: Globex"), N, Positive differences, Negative differences, Zero differences, W+, W-, Z, p, p-value ("exact" or "normal approximation, tie-corrected"), r (Z/√n), Rank-biserial r, Zeros ("dropped before ranking (Wilcoxon)")
McNemar	the 2 × 2 table of yes and no on the two variables	Counts as yes ("4 = High, 5 = Full"), N, % yes of each, Difference (points), Yes only for each variable (the two discordant counts), Chi-square and df, or the exact binomial p, p, p-value ("chi-square with continuity correction", or "exact binomial (12 discordant pairs)"), Cohen's g, Odds ratio
Friedman	per variable: N, Mean, SD, Median, Mean rank	Variables, N, Chi-square, df, p, Kendall's W, Pairwise ("Wilcoxon signed-rank, Holm-adjusted p")
Cochran's Q	per variable: N, Yes, % yes	Test ("Cochran's Q"), Counts as yes ("1 = Yes"), Variables, N, Q, df, p, Pairwise ("McNemar, Holm-adjusted p")

The pairs output holds the pairwise comparisons, one row per pair:

• after Friedman — Variable A, Variable B, N, Zero differences, W+, W-, Z, p, p adjusted, r, Rank-biserial r — with "Difference = A − B" and "Adjustment = Holm (3 comparisons)" under it;
• after Cochran's Q — a McNemar test of every pair: Variable A, Variable B, N, % yes A, % yes B, Difference (points), Yes only A, Yes only B, Chi-square (or the exact p), p, p adjusted — with "Test = McNemar for each pair", "Difference = A − B", "Adjustment = Holm (3 comparisons)" and the p-value method under it.

The preview shows the pairs under the main table. With two variables the output is empty and says "no pairwise comparisons: two variables are one comparison; see the table's statistics".

Cochran's Q asks McNemar's question of three or more yes/no variables: is the share saying yes the same for all of them? Q = (k − 1)(k ΣCⱼ² − N²) / (k N − ΣRᵢ²) on k − 1 df, as R's DescTools::CochranQTest and statsmodels' cochrans_q compute it.

• Everyone giving the same answer twice is a result, not an error: p is left out and a Note says why.
• A nominal variable is refused by Wilcoxon and Friedman, which rank answers; use McNemar or Cochran's Q for yes/no questions.
• None of these tests has a weighted form: on weighted data the statistics say "Weight: unweighted (the weight 'weight' is not applied)".
• Connect the table to a Result chart for a picture of it: after Wilcoxon or Friedman the means of the measurements with their 95 % intervals (means, or means_sd), noting "The same 393 respondents answered each; the test compares their ranks — Wilcoxon signed-rank, p = 0.9402."; after McNemar or Cochran's Q the share saying yes to each, with Wilson's interval (shares).

Principal components

analyze.pca — loadings and explained variance of a set of items. Three outputs: loadings (Table), variance (Table: eigenvalues and explained variance) and stat; the node's preview shows the loadings, with the variance table under them (headed variance). For latent factors rather than components — communalities, rotations that let factors correlate, KMO and Bartlett's test, factor scores — use Factor analysis.

In: data (SurveyData) → Out: loadings (Table), variance (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	ordinal / interval / ratio variables	The items to analyze.
Components	whole number	—	at least 1	Empty keeps the components with an eigenvalue above 1.
Standardize items	checkbox	on	—	Analyze the correlation rather than the covariance matrix.

After Apply weight the components are those of the weighted covariance (or, standardized, correlation) matrix, so the loadings, eigenvalues and explained variance are weighted; equal weights give the unweighted result. The n statistic stays the rows analyzed, and weight names the column.

A Result chart draws the scree plot from the variance output (with the Kaiser line at an eigenvalue of 1) and a heatmap of the loadings from the loadings output — the same two charts as for Factor analysis, whose scree plot fills the factors kept and adds the parallel analysis's line when it ran.

Price sensitivity

analyze.price — "What respondents would pay, by one of two methods": Van Westendorp's price sensitivity meter, from four price questions — the optimal and indifference price points and the range of acceptable prices, with Newton-Miller-Smith trial and revenue if asked — or a Gabor-Granger demand and revenue curve, from purchase intent at set prices.

In: data (SurveyData) → Out: table (Table), curves (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Method	choice	van_westendorp	van_westendorp, gabor_granger	van_westendorp reads four prices each respondent named and finds where their cumulative curves cross. gabor_granger reads a yes or no to buying at each of a set of prices and finds the demand and the revenue-maximizing price among them.
Too cheap	variable	—	interval / ratio variables	Van Westendorp: "At what price would it be so cheap that you would doubt its quality?" A respondent missing any of the four prices, or whose prices are not in the order too cheap ≤ cheap ≤ expensive ≤ too expensive, is left out and counted.
Cheap (a bargain)	variable	—	interval / ratio variables	"At what price would it be a bargain — a great buy for the money?"
Expensive (getting expensive)	variable	—	interval / ratio variables	"At what price would it start to seem expensive, though still worth considering?"
Too expensive	variable	—	interval / ratio variables	"At what price would it be so expensive that you would not consider buying it?"
Likelihood at the cheap price (NMS)	variable	—	—	Optional, for the Newton-Miller-Smith extension: how likely the respondent would be to buy at the price they called cheap, on a scale that Likelihood as probability (NMS) turns into a probability.
Likelihood at the expensive price (NMS)	variable	—	—	How likely the respondent would be to buy at the price they called expensive.
Likelihood as probability (NMS)	JSON	—	—	Each likelihood code with its purchase probability, e.g. {"5": 0.7, "4": 0.5, "3": 0.3, "2": 0.1, "1": 0}. Left empty, these are used, for a five-point scale from 5 (definitely would buy) to 1 (definitely would not).
Would buy at each price	variables (several)	—	—	Gabor-Granger: one question per price — "Would you buy it at 9.99?" — asked of every respondent. A respondent missing any of them is left out; a sequential design that skips prices should fill in the implied answers first.
Prices	JSON	—	—	The price of each question, as a list in the same order, e.g. [4.99, 6.99, 8.99].
Counts as would buy	answer code	—	—	The answer code, or a list of codes, that means would buy — a top-two box such as [4, 5] on a likelihood scale. Empty works for 0/1 questions.

The fields follow Method. Its list has van_westendorp ("four price questions") and gabor_granger ("buy or not at set prices"):

Method	Fields shown	Marked required
van_westendorp	Too cheap, Cheap (a bargain), Expensive (getting expensive), Too expensive, the two likelihood questions and Likelihood as probability (NMS)	the four price questions; the other likelihood question once one is chosen
gabor_granger	Would buy at each price, Prices, Counts as would buy	Would buy at each price, Prices

The card reads "Van Westendorp: p_too_cheap, p_bargain, p_expensive, p_too_expensive" or "Gabor-Granger: buy_499, buy_699, buy_899". The empty Likelihood as probability (NMS) and Prices boxes show the examples of their help; Counts as would buy is a checklist of the first purchase question's answers when it has value labels.

Rules (errors, then a warning):

• "Van Westendorp needs the too-cheap price — choose it in Too cheap.", and the same for the other three ("… needs the cheap (bargain) price — choose it in Cheap.", "… the getting-expensive price — choose it in Expensive.", "… the too-expensive price — choose it in Too expensive.").
• One variable named for two of the four prices: "p_price answers more than one of the four price questions; each is its own variable."
• "Newton-Miller-Smith needs both likelihood questions — choose the one at the expensive price too, or clear both." (or "… the one at the cheap price too …").
• "Gabor-Granger needs a purchase-intent question per price — choose them in Would buy at each price." and "Gabor-Granger needs the price of each question — list them in Prices, in the same order."
• Prices against the questions: "Prices lists 3 prices for 4 purchase-intent questions; give one price per question, in the same order.", "Prices is a list of numbers, one per purchase-intent question, e.g. [5, 7.5, 10].", "Prices must all be numbers, one per purchase-intent question.", "Each price can be asked once; Prices lists one twice.", "Prices cannot be negative.", "A Gabor-Granger demand curve needs two or more prices."
• "Calibration is read only with the two likelihood questions — choose them, or clear Calibration." (a warning; the message's Calibration is the Likelihood as probability (NMS) field).

Van Westendorp. At every price anyone named, the curves give the share who would call it too cheap, cheap (and not cheap, the rest), expensive (and not expensive) and too expensive, joined by straight lines. The price points are where they cross; PMC to PME is the range of acceptable prices:

Point	Where
Point of marginal cheapness (PMC)	too cheap meets not cheap
Optimal price point (OPP)	too cheap meets too expensive
Indifference price point (IPP)	not cheap meets not expensive
Point of marginal expensiveness (PME)	not expensive meets too expensive

Where two lines run together for a stretch, the point is the middle of it; where they do not meet within the prices named, the point is left out and a note says so. The outputs:

Output	What it holds
table	one row per point — Point, Price, Share % (the curves' value there), Where ("too cheap = not cheap"); with the likelihood questions also Highest trial (NMS) ("the highest mean purchase probability") and Highest revenue (NMS) ("the highest price × trial")
curves	one row per price named — Price, Too cheap %, Cheap %, Not cheap %, Expensive %, Not expensive %, Too expensive %, and with NMS Trial % and Revenue per respondent; "Curves = % of the respondents at each price named"
stat	Method ("Van Westendorp price sensitivity meter"), N, Inconsistent, PMC, OPP, IPP, PME, Range of acceptable prices ("6.4 – 14.34"); with NMS Highest trial (NMS), Trial % at it, Highest revenue (NMS), Revenue per respondent at it and Calibration ("5 → 0.7, 4 → 0.5, 3 → 0.3, 2 → 0.1, 1 → 0"); Excluded, Excluded because, Missing codes

• A respondent whose four prices are not in order misunderstood a question: they are left out and counted in Inconsistent. With the likelihood questions, the base is the respondents who answered all six ("Excluded because = a missing value in any of the price or likelihood questions (listwise)"), so the price points and the trial curve share one base.
• Newton-Miller-Smith turns the likelihood answers into purchase probabilities — 5 → 0.7, 4 → 0.5, 3 → 0.3, 2 → 0.1, 1 → 0 when Likelihood as probability (NMS) is empty — and gives the trial curve and the prices with the highest trial and the highest revenue. A likelihood answer the calibration has no probability for stops the run: "The likelihood answers include 6, which the calibration does not turn into a probability (…); give Calibration a probability for each answer code."
• Nobody answering the four in order stops the run: "No respondent answered all four price questions in order (too cheap ≤ cheap ≤ expensive ≤ too expensive): …".

Gabor-Granger. The table (and curves, which holds the same rows) has one row per price, lowest first — Price, Question (by its label), Would buy %, Revenue per respondent (price × share), Revenue index (100 at the best price) and Elasticity (the arc elasticity from the price before). The statistics: Method "Gabor-Granger", Counts as yes, Prices, N, Revenue-maximizing price, Would buy % at it, Revenue per respondent at it, Not monotone (respondents who would buy at a higher price but not at a lower one), Elasticity and Excluded. Every respondent should answer every price: in a sequential design that stops asking after a no, fill in the implied answers first (Derive). Questions that are not 0/1 without Counts as would buy stop the run: "Gabor-Granger needs to know which answer means would buy: the questions hold 1, 2, 3, 4, 5. Name that code (or a list of codes, such as a top-two box) in Counts as yes — yes outside a flow."

On weighted data every curve and share is a share of the weights; N stays the number of respondents, and the statistics add Weight and Weighted N.

Connect the table or the curves to a Result chart (Kind curves): Van Westendorp's four curves with the points named at their prices and the acceptable range shaded — with NMS, the trial curve in a panel below, marking the highest trial and revenue — or Gabor-Granger's demand above its revenue, "highest revenue at 8" marked. Each measure has its own panel: no chart puts two scales on one axis.

Proportion CI

analyze.proportion_ci — the share of respondents who gave one answer, with its confidence interval.

In: data (SurveyData) → Out: stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	—	The variable.
Answer code	answer code	required	—	The answer whose share you want, picked from the Variable's value labels, such as 1 ("Yes"). For a variable without value labels it is typed as JSON: 1 or "yes".
Confidence	number	0.95	0–1	Confidence level between 0 and 1.
Weighted	checkbox	off	—	Use the applied weight.

The statistics are p (the share), lower, upper and n. The base is the respondents who answered the variable — someone who skipped it is not counted as "did not choose".

• Weighted checked: the share is weighted and n is Kish's effective base of those respondents; a missing weight counts as 0, and weights of 1 give the unweighted result. The statistics add weight with the column's name. Earlier, a weighted share also counted the respondents who did not answer, which made it too small; a flow run again reports the corrected share.
• Weighted unchecked on weighted data: the share is of the respondents as they are, and weight reads "unweighted (the weight 'weight' is not applied)".
• Its stat output is the one statistic a Result chart draws (Kind interval): the share on a 0–100 % track with its interval, titled by the variable and the answer ("Gender: Female") and noting the level and base ("95 % confidence interval 28.4 – 35.9 %, base 600 respondents").

Regression

analyze.regression — "Linear (OLS, weighted when a weight is applied), logistic or ordinal logistic regression with a coefficient table." With a weight applied, OLS becomes weighted least squares and the logistic and ordinal models are weighted too, and the statistics add weight. Nominal predictors are dummy-coded against their first category, using the codebook's labels.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Outcome	variable	required	—	Two values → logistic (auto); otherwise linear. Ordered answers (dissatisfied … satisfied) → choose ordinal; a nominal outcome (a region) has no order, and ordinal refuses it.
Predictors	variables (several)	required	—	Nominal predictors are dummy-coded against their first category.
Model	choice	auto	auto, ols, logit, ordinal	auto fits a logit when Outcome has two values and a linear model otherwise. ordinal is the proportional-odds (cumulative logit) model of three or more ordered answers — thresholds between the answers, one coefficient per predictor with its odds ratio and 95 % Wald interval, McFadden's pseudo-R²; a positive coefficient makes the higher answers more likely, as R's MASS::polr. It leaves the codebook's missing codes out and says how many. All three use the weight when one is applied.

The Model list names each code under it: ols ("linear"), logit ("logistic, two answers"), ordinal ("ordinal logit, ordered answers"). auto never picks the ordinal model, so a flow saved before it existed runs as it did.

The ordinal model is for an outcome of ordered answers — very dissatisfied to very satisfied — which is neither a number an ordinary regression may average nor a yes/no for a logit. It fits one threshold between each pair of neighboring answers and one coefficient per predictor, the same at every cut: logit P(y ≤ j) = θⱼ − xβ, as R's MASS::polr and ordinal::clm, Stata's ologit and statsmodels' OrderedModel define it. A positive coefficient makes the higher answers more likely; its odds ratio is the odds of answering above any cut rather than at or below it.

• The table lists the coefficients, then the thresholds: term, type (coefficient or threshold), estimate, std_error, statistic (z), p_value, odds_ratio, odds_ratio_lower and odds_ratio_upper (the 95 % Wald interval). A threshold is named by the two answers it cuts between — Very dissatisfied / Dissatisfied — and its odds-ratio cells are blank. A nominal predictor's terms read region = North.
• The statistics: model ("ordinal logit (proportional odds)"), outcome, categories, order ("Very dissatisfied < Dissatisfied < Neutral < Satisfied < Very satisfied"), n, log_likelihood, pseudo_r_squared (McFadden's), the likelihood-ratio test against the thresholds alone (lr_chi_square, lr_df, lr_p), aic, converged, coefficients ("a positive coefficient makes the higher answers more likely: logit P(y ≤ j) = threshold j − xβ, as R's MASS::polr"), interval ("95 % Wald interval of the odds ratio") and missing_codes ("Trust: Acme: 74 (9 = Refused)"); note when a labeled answer nobody in the model gave is not a category, warning when there is something to warn about.
• A nominal outcome is refused, before the run by the check (VARIABLE_SCALE; a warning when a node upstream makes the variable nominal) and by the run itself: "Region is nominal: its answers (Capital, North, South) have no order, and the ordinal model would take one from their codes. Use the logit for an outcome of two answers, or recode it onto an ordered scale (Recode with Scale = ordinal) first."
• Other refusals, with the reason: fewer than three answers ("… the ordinal model needs three or more ordered answers — with two, use the logit."), more than 20 different values ("… For a scale of numbers use the linear model (ols), or group the values into bands first."), text that is not a code, a predictor that does not vary, predictors that are a combination of each other. Warned (in warning): a fit that did not converge, and a predictor that separates the answers — its estimate runs off toward infinity and cannot be read.
• On weighted data each respondent's weight multiplies their part of the likelihood, as the logit's does; when the weights do not average about 1, weights says what they sum to, because the standard errors count that sum. The proportional-odds assumption itself is not tested.

Connect the table to a Result chart (Kind coefficients) for a forest plot without the intercept: each coefficient with its 95 % interval, or for a logit and the ordinal model each odds ratio on a log scale, the thresholds left out. Terms are named by their labels and a note says what each nominal predictor is compared with ("compared with Region = Capital"). Connect the stat too: a linear model's intervals then use its t distribution ("t with 589 df").

Scale reliability

analyze.reliability — Cronbach's alpha for a set of items, with item–total correlations and alpha-if-deleted per item.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	ordinal / interval / ratio variables	The items of the scale.

After Apply weight alpha, the item means, the item–total correlations and alpha-if-deleted are all computed with the weight, and the statistics add weight. A row is dropped for a missing item, never for a missing weight (that weighs 0), so the respondents counted (n) are the same either way.

t-test

analyze.ttest — compares means: of two groups (Welch's or Student's t-test), of two measurements of the same respondents (paired), or of one variable against a value. "It reports t, df, p, the mean difference and its CI, and Cohen's d."

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Design	choice	independent	independent, paired, one_sample	Independent compares two groups of respondents; paired compares two variables answered by the same respondents; one_sample compares a mean with a fixed value.
Variable	variable	required	ordinal / interval / ratio variables	—
Groups	variable	—	nominal / ordinal variables	Independent: the variable that splits the respondents into the groups compared — one answer each; for a multiple-choice question, run Explode multiple choice and use one option's 0/1 column.
Group A	answer code	—	—	Independent: the code of the first group — needed only when Groups has more than two values.
Group B	answer code	—	—	The code of the second group.
Variances	choice	welch	welch, student	Welch's test does not assume the two groups vary equally and is the safer default; Student's pools the variances.
Second measurement	variable	—	ordinal / interval / ratio variables	Paired: the variable compared with Variable, respondent by respondent.
Test value	number	0	—	One sample: the value the mean is tested against.
Confidence	number	0.95	0.5–0.999	—

The fields follow Design. Its list names each code under it: independent ("two groups"), paired ("two variables, same people"), one_sample ("a mean against a value"); the inspector shows only what that design reads:

Design	Fields shown	Marked required
independent	Variable, Groups, Group A, Group B, Variances, Confidence	Groups; Group B once Group A is set (and the other way round)
paired	Variable, Second measurement, Confidence	Second measurement
one_sample	Variable, Test value, Confidence	—

Group A and Group B list the answers of the Groups variable — 1 ("Male"), 2 ("Female"), … — including the bands of a Bands node. Leave both empty when Groups has exactly two values.

Rules (errors):

• "An independent-samples t-test compares two groups — choose the variable that splits them in Groups."
• "A paired t-test compares two measurements of the same respondents — choose the second in Second measurement."
• "Name both groups to compare in Group A and Group B, or neither when Groups has only two values."

The table has one row per group — or per measurement, plus a Difference row for a paired test — with N, Mean, SD and SE, the groups by their value labels:

| Gender | N  | Mean   | SD     | SE    |
| Male   | 89 | 54.045 | 23.356 | 2.476 |
| Female | 86 | 55.291 | 24.349 | 2.626 |

Test = Welch's t-test (unequal variances); t = -0.345; df = 172.0; p = 0.7303;
Mean difference = -1.246; Difference = Male − Female; 95% CI = -8.369 – 5.877;
Cohen's d = -0.052; Hedges' g = -0.052; N = 175; Variable = Age

Design	Footer
independent	Test ("Welch's t-test (unequal variances)" or "Student's t-test (equal variances)"), t, df (Welch's df has decimals), p, Mean difference, Difference (Group A − Group B), the CI (95% CI), Cohen's d, Hedges' g, N, Variable
paired	Test "Paired t-test", t, df, p, Mean difference, Difference ("Trust: Acme − Trust: Globex"), the CI, Cohen's d (d_z), N (the complete pairs), Incomplete pairs left out
one_sample	Test "One-sample t-test", Test value, t, df, p, Mean difference, Difference ("mean − 3"), the CI, Cohen's d, N

The difference is always the first minus the second — Group A − Group B, Variable − Second measurement — as in Paired tests, so a t-test and a Wilcoxon test of the same two variables agree in sign. The codebook's missing codes are left out ("Missing codes left out = …").

Refusals — the run stops on this node with the reason:

• a Groups variable with more than two groups and none named: "Gender has 3 groups (1 = Male, 2 = Female, 3 = Other); a t-test compares two — name them in Group A and Group B.";
• the same group twice: "Group A and Group B are both 1 = Male; a t-test compares two different groups — name another in one of them.";
• a missing code named as a group ("… is a missing code of Gender, not a group — name two groups that are answers.");
• a multiple-choice Groups: "… holds several answers per respondent, so its groups overlap and a t-test, which compares two separate groups, cannot use them. Run Explode multiple choice and compare by one option's 0/1 column (chose it or not)."

What the data cannot carry — a group of one, no spread at all — is said in the footer ("Test = not run: …") rather than as a number.

The t-test has no standard weighted form here: on weighted data it runs on the respondents as they are and says "Weight: unweighted (the weight 'weight' is not applied)". For two groups, Group means with Test welch or student runs the same test beside weighted means.

TURF

analyze.turf — how many different people a shortlist of options reaches together: "which three flavors should we stock?" rather than "which three are most popular" — or, with Search fixed, what a portfolio you already have reaches.

In: data (SurveyData) → Out: table (Table), stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Options	variables (several)	required	—	One 0/1 column per option. Run Explode multiple choice on a multiple-choice question to get them.
Largest portfolio	whole number	3	at least 1	—
Search	choice	best	best, greedy, fixed	Best tries every combination. Greedy extends the previous winner and is fast, but can miss the best portfolio — the table says which ran. Fixed searches nothing and reads the Portfolio below.
Always include	variables (several)	empty	—	Options that are in the portfolio whatever they add — shelf space already committed.
Portfolio	variables (several)	empty	—	For Search = fixed: exactly these options — each one's reach, what only it reaches, and the reach and frequency of all of them together. Empty reads every option.

The Search list names each code under it: best ("every combination"), greedy ("extend the winner"), fixed ("the Portfolio as it is"). The inspector shows Largest portfolio and Always include for best and greedy, and Portfolio for fixed; the node's card reads "up to 3, best", or for a fixed portfolio "fixed portfolio" and its options.

• Each respondent is counted once however many options they chose — that is what makes it unduplicated reach rather than a sum of percentages. The base is the respondents who answered; the applied weight is used, and the statistics then add Weight.
• best and greedy: one row per portfolio size with its options, reach, reach %, what it adds to the previous size and its frequency (the mean number of a portfolio's options a reached respondent chose). The table and the statistics (Search) name which search ran, because an exhaustive answer and a greedy one are not the same claim.
• fixed: one row per option of the Portfolio — its reach, its unique reach (the respondents no other option of the portfolio reaches: what dropping it would lose) and frequency — then a (portfolio) row, "All 3 together", with the reach and frequency of the whole portfolio. The statistics give Base, Search "none: a fixed portfolio", Reach ("31.8 %"), Frequency and the Note. The base is the same as a search over Options would use: keep all the question's options in Options.
• Rules (warnings): "Always include is not read with Search = fixed, which evaluates exactly the Portfolio — add those options to Portfolio." and "Portfolio is read only with Search = fixed; best and greedy search for a portfolio themselves."
• On weighted data the reach is a sum of weights and the frequency a weighted mean.
• Options must be 0/1 columns. An exhaustive search that would not finish says so and suggests the ways out.
• Where the columns come from: an Explode multiple choice node upstream (its indicator columns are offered in Options), a Multiple choice question with Data layout wide, or a set of yes/no questions coded 0/1.
• Chart it with a Result chart fed by the table: best and greedy draw the reach curve (Kind reach) — reach by portfolio size, each size named by the option it adds, by its label (+ <label>), with its reach and gain ("33.6 % (+10.5)"); a best portfolio that is not the one before plus an option reads "a new set" and is listed in full under the chart. fixed draws each option's reach beside what only it reaches (Kind items), with the whole portfolio's reach in the legend ("All 3 together: 39.0 %").

---

Visualize

Chart out (a Trend gives its points as a Table too). Six of the seven draw from the data (SurveyData in); the Result chart draws what an analysis computed (its table and statistics in). All seven accept a Title, a Figure width (in) and Figure height (in) (2–30 inches, 10 × 6 by default) and a Palette — Heatmap takes a Color map instead, and the Likert chart offers diverging palettes. Width and height resize the figure the engine draws, not the picture of it, so the axis labels keep their proportion. The newer charts — the Bar chart's newer forms, the Pearson and Kendall heatmaps, a heatmap of means in the theme's colors, the Likert chart, the Trend and the Result chart — also fit themselves to what they show: long labels wrap and shrink, crowded ticks are thinned, and a figure too small for its labels grows taller rather than print them over each other. Counts and bases separate thousands (n = 182,128).

Which chart. A quick look needs no flow (Data → Insights); these nodes are for reports and Live tiles.

You want to show	Node	Set
how the answers to one question are spread	Bar chart	Show percent; Sort value for the largest first
only the answers given most, of a long list	Bar chart	Top N, and Combine the rest as Other for one gray bar of the rest
each share with its margin of error	Bar chart	Show percent, Confidence intervals
the parts of a whole	Bar chart	Layout donut
how a number spreads — an age, an amount	Bar chart	Layout histogram, Bins
a question's answers within the groups of another	Bar chart	Split by; Layout stacked_100 for the shares of each group
… and which groups differ	Bar chart	Split by, Show percent, Significance letters
a mean in each group	Bar chart	By (and Confidence intervals)
a battery of statements on one scale	Likert chart	—
a share, a mean or a count over waves or months	Trend	Time, Period, Measure
a number within groups; two numbers against each other	Box plot; Scatter plot	—
how items correlate, or their means by group	Heatmap	Method; By
the result of an analysis — means with intervals, a scree plot, key drivers…	Result chart	Kind auto

A chart is drawn as its node runs: one that cannot be drawn (a Likert chart of items without a scale) fails its own node, with the reason, in a preview and in a run — not the Save report or the preview after it.

Palette theme. Every chart's Palette offers theme, with "the report's chart colors (Save report's Look)" under it in the list, and a Heatmap's Color map suggests theme: the chart takes its colors from the Look of the Save report it is saved in — the Chart colors you set there (see Chart colors), with its text color, grid lines and typeface. Left empty, those are the engine's default chart colors, eight colors any two of which readers with protanopia or deuteranopia can tell apart. In the report — its .md figures and its .html alike — the chart is drawn in that report's colors. In a preview, and on a Live tile after a run, it is drawn in the look of the flow's own Save report node — in the default chart colors when the flow has none, or when its Save report nodes name different looks. Its colors never come from the house style; only the p in a Result chart's note follows the house style's P values there (see P values). A named palette (muted, RdBu, YlOrRd, …) keeps its own colors whatever the report's Look says. With theme:

• a series takes the next color of Series, an answer of an ordered scale a step of Magnitude (light to dark), and a Likert chart, a correlation heatmap or the red, gray and blue of NPS and sentiment the two Diverging ends;
• a series past the palette's colors gets a darker, then a lighter, shade of them — every one still at least 2:1 on the white background;
• a Heatmap of means by group is drawn as the newer charts are: each cell the mean of the group's respondents who answered that item (as Group means gives it), the codebook's missing codes left out and counted under the chart (see Heatmap).

After Apply weight a chart either draws the weighted numbers or says under its title that it does not — so a picture never disagrees in silence with a weighted table beside it:

Chart	On weighted data
Bar chart	weighted: bars are sums of weights, weighted percentages, or weighted means with By — matching the Frequencies, Crosstab and Group means tables of the same data; a histogram's heights and a donut's slices are weighted too, and its intervals and letters use Kish's effective base
Heatmap with By	weighted means by group; the color bar reads "Weighted mean"
Heatmap without By, Method pearson	weighted coefficients; the color bar reads "Weighted Pearson r"
Likert chart	weighted shares; the axis reads "% of respondents (weighted)"
Trend	weighted percents, means and counts; the axis reads "% of respondents (weighted)", "Mean (weighted)" or "Respondents (weighted)"
Result chart	as the result it draws: its title's second line reads "weighted by 'weight'" or "unweighted (the weight 'weight' is not applied)"
Box plot, Scatter plot, Heatmap without By with Method spearman or kendall	unweighted — a box, a point per respondent and a rank correlation have no standard weighted form — with a second title line "unweighted (the weight 'weight' is not applied)", under your own Title too

Bar chart

visualize.bar — "Distribution of a variable as counts or percentages, or its mean by group; a histogram of a number, or a donut. Split by a second variable, grouped or stacked, it is the chart of a crosstab."

In: data (SurveyData) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	—	The variable to chart.
By	variable	—	nominal / ordinal variables	The mean of Variable in each group of this one. For the answers themselves in each group, use Split by.
Show	choice	count	count, percent	Count, or percent of the respondents who answered — for a multiple-choice question the share who named each option, so the bars add up to more than 100 % (the chart says so). With By the bars are means.
Split by	variable	—	nominal / ordinal variables	The answers within each group of this variable — the chart of a crosstab, with percentages of each group (column percentages), weighted when a weight is applied.
Layout	choice	grouped	grouped, stacked, stacked_100, histogram, donut	With Split by: the groups' bars side by side, stacked, or stacked to 100 % of each group (percentages, whatever Show says); a multiple-choice question's options overlap and are drawn side by side only. histogram: an interval or ratio variable in Bins, as counts or percent — with Split by, a panel per group. donut: one variable's answers as the parts of a whole, their percentages on the slices and the base in the middle (not with Split by, By or a multiple-choice question).
Sort	choice	code	code, value	Code keeps the codebook's order; value draws the largest bar first (with Split by, the answer most given overall — or, when the answers are a scale, the groups with the largest share of its top answer first, the scale kept in order). Colors follow the answer, not its place — the same in a Top N chart and a donut of the same question. Any setting but the defaults (Show count, no Split by, Sort code, no Top N, intervals, histogram or donut) draws the newer chart, with its base and notes under the plot.
Horizontal	checkbox	off	—	Bars across, labels beside them. With Show, Split by or Sort set, vertical bars whose labels cannot be read under them (many long answers) are drawn across too.
Show values	checkbox	on	—	Print the value on each bar — in a donut, each slice's percentage.
Top N	whole number	—	1–100	Only the N answers given most — with Split by, given most overall; for a multiple-choice question, the options named most; for a number of many values, its N values given most (up to 30). Empty draws every answer. In a donut the rest are combined as Other.
Combine the rest as Other	checkbox	off	—	With Top N, one more bar for the answers after the top N — for a multiple-choice question, the respondents who named any of them — drawn gray and last. Off, the rest are left out; the note under the chart says how many.
Confidence intervals	checkbox	off	—	An error bar on each percentage (Wilson's interval, on Kish's effective base when weighted — each group's own base with Split by) and on each mean by group (the interval Group means draws when it leaves the missing codes out — Student's t, weighted the linearization interval; the Bar chart always leaves them out). Drawn on bars side by side (Layout grouped) only; counts have none.
Confidence	number	0.95	0.5–0.999	The intervals' confidence level.
Significance letters	checkbox	off	—	With Split by, Show percent and Layout grouped. The groups of Split by are lettered A, B, … (under their names) in the Banner table's order, and a bar carries the letters of the groups whose share of that answer is significantly lower — the Banner table's two-sided z-test of column proportions, on the group's respondents who answered (Kish's effective base when weighted); a group below 30 is not tested. The Tab book shows the same comparisons, under the letters its banner gives these columns — the same letters only when Split by is the banner's first variable.
Level	number	0.05	0.001–0.2	The letters' significance level.
Multiple comparisons	choice	none	none, bonferroni	bonferroni divides Level by the number of pairs of groups tested.
Bins	text	auto	—	auto: Freedman and Diaconis's width (whole-number answers take a whole width); a number: that many bins of equal width; two numbers or more: the bins' edges, increasing, separated by commas (0, 18, 25, 35, 50, 65). At most 100 bins.
Other below (%)	number	3	0–50	Slices smaller than this percentage are combined as Other (two or more of them); 0 keeps every slice.
Title	text	—	—	Chart title.
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	—
Palette	choice	muted	muted, deep, pastel, dark, colorblind, Set2, tab10, theme	Theme takes the colors of the report the chart is saved in — the chart colors of its Save report's Look (by default eight colors any two of which readers with protanopia or deuteranopia can tell apart), with its text color, grid and font. The others are seaborn's palettes.

The lists name each code under it: count ("respondents"), percent ("% of those who answered"); grouped ("side by side"), stacked ("stacked, as Show says"), stacked_100 ("stacked to 100 %"), histogram ("histogram of a number, in Bins"), donut ("donut: the parts of a whole"); code ("the codebook's order"), value ("largest first"); bonferroni ("Bonferroni"); theme ("the report's chart colors (Save report's Look)"). Bins suggests auto, 10 and 0, 18, 25, 35, 50, 65. The card reads satisfaction by region with Split by set, and the variable alone for a donut.

The fields each form reads. The inspector shows only those (see Reading this page): By with Show count and a bar layout (grouped, stacked, stacked_100); Show and Split by for every layout but donut; Sort, Show values and Top N for every layout but histogram; Horizontal and Combine the rest as Other for bars only (not a histogram or a donut); Confidence intervals with grouped (and with a stacked layout at Show count and Sort code, where a rule warns that it is not drawn), Confidence once it is checked; Significance letters with Show percent and grouped, its Level and Multiple comparisons once it is checked; Bins for a histogram and Other below (%) for a donut.

Seven charts in one node:

You want	Set	The chart
the distribution of one question	Variable (and Show)	a bar per answer: respondents, or % of those who answered
the answers given most, of a long list	Top N (and Combine the rest as Other)	the N answers given most; the rest left out, or one gray Other bar, last
the answers within groups — the chart of a crosstab	Split by, Layout	per group of Split by, its distribution: counts, or % of the group (column percentages); stacked_100 stacks each group to 100 %
which groups differ on an answer	Split by, Show percent, Significance letters	the grouped bars, each carrying the letters of the groups whose share is significantly lower
a mean by group	By	the mean of Variable in each group of By, with its interval when Confidence intervals is checked
how a number spreads	Layout histogram, Bins	the number in bins, as counts or % — a panel per group with Split by
the parts of a whole	Layout donut	a ring of the answers' shares, the base in the middle

The classic chart and the newer one. At the defaults — Show count, no Split by, Sort code, no Top N, Confidence intervals off and a bar layout — the node draws the chart it always has, picture for picture, and a stored flow renders the same code: each new option is written into the script only when it is set. Any other setting (Sort value alone included) draws the newer chart, whose look differs:

• It leaves the codebook's missing codes out of the bars and says so under the plot ("Left out as missing: Trust: Acme: 47 (9 = Refused).").
• It writes the base under the plot — "Base: 385 respondents who answered." ("… who answered both. Each group's n is under its name." with Split by, each group's n under its name on the axis) — and, with percentages of groups, "Percentages are of each group of Region."; the axis reads "% of respondents", "% within Region" or "Count".
• A multiple-choice question counts respondents, not answers: "Several answers allowed: each bar is the share of respondents who named the option, so the bars add up to more than 100 %." It is drawn by the newer chart whatever the settings, and its options are never stacked.
• One series is one color; the answers of an ordered scale are one hue, light to dark. A color belongs to its answer, not its place, so a sorted chart, a Top N chart and a donut of the same question color the same answer alike; Other is gray.
• Sorted with Split by a scale (ordinal and up), the answers keep the scale's order and the groups go largest first, by their share of the top answer: "Groups in order of their share of Very satisfied; the answers keep the scale's order."
• Long labels wrap; every label beside horizontal bars gets a row of its own, and vertical bars whose labels cannot be read under them, even turned, are drawn across. A legend too tall or too wide to sit beside the plot goes under it, a percent axis puts its ticks on whole percents, and a value axis drops ticks rather than print their numbers into each other.
• It draws a bar (or a slice) per value given, so it refuses a number with more than 30 different values given: "Age is a number with 84 different values given, and this chart draws a bar for each: layout='histogram' draws its distribution (or band it first with Bands)." ("… a slice for each …" in a donut). With Top N it draws the N values given most, so N is what counts: up to 30 are drawn, in steps of one hue light to dark among themselves (among all 84, neighbors would read as one color); past 30, "Age is a number with 84 different values given, and top=40 draws a bar for each of the 40 given most: give top=30 or fewer, or layout='histogram' draws its distribution (or band it first with Bands)." A question nobody answered stops the node with "No respondent answered Trust: Acme."

Top N and Other. The bars keep Sort's order — the codebook's, or the largest first — and the note under the chart says what was left out: "The 3 answers given most of 5 are drawn; the other 2 are left out." With Combine the rest as Other: "…; Other combines the other 2." For a multiple-choice question: "The 2 options named most of 4 are drawn; Other is the respondents who named any of the other 2." With Split by the answers are those given most overall ("The 2 answers given most overall of 5 are drawn; …"). A donut always combines the rest: "…; Other combines the other 3 — a donut's slices make a whole." A Top N of as many answers as there are draws them all.

Confidence intervals. Error bars on the bars side by side, at Confidence, and a note naming them: "Error bars: 95 % confidence intervals (Wilson score)." — "(Wilson score, on Kish's effective base)" on weighted data — or, for means by group, "Error bars: 95 % confidence intervals of the mean (Student's t)." ("(weighted: the linearization interval)"). A share is the one Proportion CI gives for the same respondents (Proportion CI's interval is the normal approximation, so its bounds can differ a little, most near 0 % and 100 %); a mean's interval is the one Group means gives when it leaves the missing codes out (a Test chosen by hand) — its auto test counts them as answers, so its chart can differ from this one.

Significance letters. The groups of Split by are lettered under their names on the axis (Capital (A), North (B), …), and a letter over a bar names a group whose share of that answer is significantly lower. The note reads "A letter over a bar names a group (its letter is under its name) whose share of that answer is significantly lower (two-sided z-test of column proportions, p < 0.05) — the Banner table's and the Tab book's letters.", adding "Bonferroni-corrected" and "on Kish's effective base" where they apply; "No group's share of any answer is significantly higher than another's." when no bar carries a letter; and "Not tested, fewer than 30 respondents who answered: Other (C)." for a small group. The comparisons are the Banner table's and the Tab book's for the same cells; a Tab book letters every column of its banner in one run (A, B, C for gender, then D, E, F for region), so its letters match the chart's only when Split by is the banner's first variable.

Histogram. An interval or ratio variable in bins, as counts (Show count) or % of those who answered (percent), with the bins' rule under the plot:

Bins	The bins	Note
auto	Freedman and Diaconis's width; whole-number answers take a whole width	"Bins: 11 of width 8 (Freedman–Diaconis), each holding 8 whole numbers."
a number, 10	that many of equal width (1–100)	"Bins: 10 of equal width."
edges, 16, 25, 35, 50, 65, 100	from one edge to the next, each holding its left edge, the last its right edge too; commas, semicolons or spaces between them; at most 100 bins	"Bins: the edges given; each bin holds its left edge, the last its right edge too." — and, when the bins differ in width, "The bins differ in width: a bar's height is its percentage, not its density."

With Split by, each group is a panel of its own on the same bins, its n over it ("Male (n = 404)"), and the base note says "Each group's n is over its panel." After Apply weight the heights are sums of weights ("the width is of the answers as they are, the heights weighted"). Bins it cannot read is an error on the field as you edit: "bins: The bins' edges must increase from one to the next, and 5 is followed by 3.", "bins: A number of bins is a whole number from 1 to 100; got '0'.", "bins: At most 100 bins: 102 edges give 101." or "bins: Bins is auto (Freedman and Diaconis's width), a number of bins, or the bins' edges in increasing order, separated by commas (0, 18, 35, 65); got 'many'."

Donut. The answers of one question as the slices of a ring, their percentages on the slices (beside a thin one), the base in the middle ("981 respondents"). Slices under Other below (%) are combined as one gray Other when there are two or more of them: "Other combines 3 answers under 20 % each: Low, Medium and High." When every answer is under it, the node stops rather than draw one gray ring: "Each of the 40 answers to Brand drawn is under 3 % of the respondents who answered, so Other would fill the whole ring: draw them as bars (layout='grouped'), or lower min_slice." The codebook's missing codes are left out, as in every newer form.

Rules (an error stops the flow; a warning names a setting the chart ignores):

• Errors: "By draws the mean of Variable in each group and Split by its answers in each group — clear one of them."; "Top N keeps the answers given most, and with By the bars are means of groups — clear one of them."; "Split by must be another variable than Variable."
• Warnings: "By (the mean in each group) is not drawn when Show is percent; to show the answers in each group, use Split by." (on bars); "Stacked layouts apply only when Split by is set."; "Combine the rest as Other applies with Top N — set Top N."; "Top N keeps the answers given most; a histogram draws bins of a number, so it is not applied."; "By (the mean in each group) is not drawn in a histogram; for a histogram of each group, use Split by."; "By (the mean in each group) is not drawn in a donut, which shows the shares of Variable's answers."; "Split by is not drawn in a donut, which shows one variable's answers as the parts of a whole; Layout stacked_100 shows the answers within each group."; "Confidence intervals are drawn on bars side by side (Layout grouped) only."; "Confidence intervals are drawn for percentages and for means by group; counts have none — set Show to percent."; "Significance letters compare the groups of Split by — set Split by."; "Significance letters are drawn on bars side by side (Layout grouped), not on stacks."; "Significance letters compare percentages, as the Banner table's do — set Show to percent."
• The rules on the intervals and the letters, and on By with Show percent, speak of bars (grouped, stacked, stacked_100): a histogram or a donut draws neither intervals nor letters, and when they are left checked from the bars it is not warned of them — they are named under the fields as not used with these choices. Of a donut with By and Split by left over, the two warnings are that it draws neither.

The check also refuses, before the run, what the questionnaire settles: a Split by that allows several answers ("Split by needs one answer per respondent, and Brands heard of (unaided) allows several: draw it as the Variable, or split by one of its options after Explode multiple choice."), a stacked layout of one ("Brands heard of (unaided) allows several answers, so its options overlap and cannot be stacked: draw them side by side (Layout = grouped)."), a histogram of a question with answers ("A histogram draws the distribution of a number, and Region is nominal: draw its answers as bars (Layout = grouped).", "Brands heard of (unaided) allows several answers; a histogram draws one number per respondent.") and a donut of a multiple-choice question ("Brands heard of (unaided) allows several answers, so its shares add up to more than 100 % and are not the parts of a whole: draw them as bars (Layout = grouped)."). A number whose valid range holds more than 30 whole numbers, drawn in a newer form, is a warning: "Age is a number of up to 84 values, and bars draw each value given: Layout histogram draws its distribution." (a donut: "… and a donut draws a slice for each value given …"; a Top N over 30: "… and Top N draws a bar for each of the 40 given most: set Top N to 30 or fewer, or Layout histogram draws its distribution.") — the data decides whether the newer chart refuses it. A range of 0 to 29.5 holds 30 whole numbers, and is not warned of; nor is the classic chart, which draws every value as it always did.

After Apply weight, a distribution's bars are sums of weights — the axis reads "Weighted count", and the values on the bars have one decimal when they are fractional (the newer chart writes a count from 100 up whole, 18,848) — or weighted percentages ("% of respondents (weighted)", "% within Region (weighted)", weighted as the Frequencies and Crosstab tables are), and with By the bars are weighted means, on an axis "Weighted mean label". A histogram's and a donut's shares are weighted too, and the intervals and letters use Kish's effective base. The newer chart's base adds the weighted total ("Base: 385 respondents who answered (weighted: 380.9).") and the note "Weighted by 'weight'; n counts respondents." Unweighted, the axes read "Count" and "Mean label".

Box plot

visualize.boxplot — the distribution of a variable across groups.

In: data (SurveyData) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
Variable	variable	required	ordinal / interval / ratio variables	The variable whose distribution is shown.
By	variable	required	nominal / ordinal variables	The grouping variable (required).
Show points	checkbox	off	—	Overlay the individual points.
Title	text	—	—	Chart title.
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	—
Palette	choice	muted	muted, deep, pastel, dark, colorblind, Set2, tab10, theme	Theme takes the colors of the report the chart is saved in — the chart colors of its Save report's Look (by default eight colors any two of which readers with protanopia or deuteranopia can tell apart), with its text color, grid and font. The others are seaborn's palettes.

Quartiles and whiskers are of the respondents as they are; on weighted data the title's second line says "unweighted (the weight 'weight' is not applied)".

Heatmap

visualize.heatmap — "Correlation matrix of several items (Spearman, Pearson or Kendall), or their means by group."

In: data (SurveyData) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	—	The items to show.
By	variable	—	nominal / ordinal variables	Each item's mean in each group of this variable (weighted when a weight is applied). With a named Color map it is drawn as it always was, over the respondents who answered every item, a missing code such as 99 = Not applicable counting as an answer (run Missing values first to leave it out). With Color map theme each cell is the mean of the group's respondents who answered the item, as Group means gives it, the codebook's missing codes left out and counted, long items numbered and each group's base under its name. Without By, the heatmap is the correlation matrix of the items.
Method	choice	spearman	pearson, spearman, kendall	The correlation drawn without By, over the respondents who answered every item. Spearman, the default, reads the answers as it always has (a missing code such as 9 = Refused counts as an answer; run Missing values first to leave it out). Pearson and Kendall leave the codebook's missing codes out and say so, as Correlation matrix does; Pearson is weighted when a weight is applied, Spearman and Kendall are not — with a weight applied, Spearman's title says so, and Kendall says it is not.
Title	text	—	—	Chart title.
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	—
Color map	text	YlOrRd	a matplotlib colormap, or theme	A matplotlib colormap, used when means are shown by a group; a correlation matrix keeps its own diverging scale. Type theme for the colors of the report the chart is saved in — its Save report Look's sequential color for the means, its diverging pair for a correlation matrix, with its text color, grid and font.

The Method list names each code under it: pearson ("Pearson r"), spearman ("Spearman rho, answers as they are"), kendall ("Kendall tau-b"). spearman draws the chart the node always has; a flow saved before Method existed keeps it. Color map suggests theme, YlOrRd, Blues, viridis and RdBu_r as you type.

Means by group in the theme's colors. With By and Color map theme, the heatmap is drawn as the newer charts are, and its numbers are those of Group means for the same respondents: each cell is the mean of the group's respondents who answered that item, not only of those who answered every item. The notes say so — "Base: 1,159 respondents who answered Region and an item; each group's n is under its name. Each cell is the mean of the group's respondents who answered the item, as Group means gives it." — and count the missing codes left out ("Left out as missing: Trust: Acme: 194 (9 = Refused); …"). Long item labels are numbered (1. Trust: Acme), each group's n is under its name, and the means are written at a size their cells hold. With a named color map (YlOrRd, the default), the means are drawn as they always were — over the respondents who answered every item, a missing code counting as an answer — so run Missing values first there.

Pearson and Kendall draw the numbers of a Correlation matrix table with Missing answers listwise, titled "Pearson Correlation Matrix" or "Kendall tau-b Correlation Matrix":

• the codebook's missing codes are left out and counted under the plot, with N: "N = 251 respondents who answered every item (listwise)." and "Left out as missing: Trust: Acme: 108 (9 = Refused); Trust: Globex: 109 (9 = Refused).";
• a pair that cannot be computed — an item everyone answered the same, its own diagonal too — is a blank cell, and the note says why;
• long labels are numbered — the rows read 1. label, the columns 1, 2, … — and the coefficients are written at a size their cells hold (at most 10 pt); below 6 pt they are left to the table, and the note says so.

Rule (a warning): "Method applies to the correlation matrix drawn without By; with By the heatmap shows means."

After Apply weight, the means by group are weighted and the color bar is labeled "Weighted mean". A Pearson matrix is weighted too — color bar "Weighted Pearson r", and the note "Weighted by 'weight': the coefficients are weighted; N counts respondents." A Spearman or Kendall matrix is never weighted: its title's second line says "unweighted (the weight 'weight' is not applied)". For the coefficients with their p-values and N, use Correlation matrix — and a Result chart of its table for the same heatmap with the table's significance marks.

Likert chart

visualize.likert — "A battery of items on one ordered scale as diverging stacked bars centered on the neutral answer, with each item's top-2 and bottom-2 shares." Each item is a bar: the answers below the middle of the scale stack to the left of a center line, those above it to the right, so a battery's lean reads at a glance.

In: data (SurveyData) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
Items	variables (several)	required	ordinal / interval variables	Items with one answer each on the same scale — the same value labels in the codebook (a battery such as agree–disagree statements). Codes run low to high, left to right; recode a scale that runs the other way first. The codebook's missing codes (a 9 "Don't know") are left out and counted under the chart.
Neutral answer	choice	split	split, side	Split draws the middle answer of an odd scale half on either side of the center; side draws it apart, in a panel at the right. An even scale has no neutral answer, and its center falls between the two middle answers.
Sort items	choice	top2	top2, listed	top2 puts the item with the largest share in the top two answers first (the top answer alone on a scale of two or three); listed keeps the order of Items.
Show values	checkbox	on	—	Each answer's share in its segment where it fits. The top-2 and bottom-2 shares are always written at the ends.
Title	text	—	—	Empty takes the words the items' labels start with ("Trust" of "Trust: Acme" and "Trust: Globex"), the rest of each label naming its bar.
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	A battery whose labels need more room makes the figure taller.
Palette	choice	RdBu	RdBu, BrBG, PuOr, RdYlBu, PiYG, coolwarm, theme	A diverging palette, the low answers in its first color; the neutral answer is gray. Theme takes the colors of the report the chart is saved in — the diverging pair of its Save report's Look (low end first), with its text color, grid and font.

The lists name each code under it: split ("half on either side"), side ("in a panel at the right"), top2 ("largest top-2 share first"), listed ("in the order of Items").

                                  Trust
                  Bottom-2                                    Top-2
Acme (n = 485)       42%    [ 22% | 20% |  18%  | 21% | 19% ]    40%
Globex (n = 484)     41%    [ 20% | 21% |  21%  | 20% | 18% ]    38%
                           60%   40%   20%   0%   20%   40%   60%
                                     % of respondents
               ■ No trust  ■ Low  ■ Medium  ■ High  ■ Full

The center line (0 %) runs through the middle of the neutral answer, here Medium, half of which lies on either side.

What counts as the scale. Each item's labeled answers in the codebook, its missing codes left out; without value labels, a valid range of whole numbers (2 to 11 of them); failing that, a Likert scale question's points, the two ends named by its end labels. Every item must have the same answers. The check says so before the run, and a chart that cannot be drawn fails its own node with the same words:

• "The items of a Likert chart must share one scale, and these do not: Trust: Acme has 1 = No trust, 2 = Low, 3 = Medium, 4 = High, 5 = Full; Overall satisfaction has 1 = Very dissatisfied, 2 = Dissatisfied, 3 = Neutral, 4 = Satisfied, 5 = Very satisfied. Draw them in separate charts, or recode them onto one scale first."
• "Brands heard of (unaided) allows several answers; a Likert chart draws items with one answer each on a scale."
• "A Likert chart draws the answers of a scale, and none of the items has value labels (or a valid range of whole numbers) in the codebook, or a Likert scale question, to say what the scale is."

Under the chart, notes say what it shows:

• "Base: the respondents who answered each item on the scale (n beside it). Items in order of their top-2 share." — each bar is labeled with its base (Acme (n = 485));
• "Top-2: 4 = High, 5 = Full; bottom-2: 1 = No trust, 2 = Low." — the shares at the ends of each bar;
• "The neutral answer (3 = Medium) is split around the center." (or "… is drawn apart, at the right.");
• "Left out as missing: Trust: Acme: 108 (9 = Refused); Trust: Globex: 109 (9 = Refused)." — and "Not on the scale, left out: …" for values off the scale, "No answer on the scale, not drawn: …" for an item nobody answered.

A single item draws one bar, titled by its label. The item labels are fitted to the figure — smaller and wider on a narrow one — before it grows taller. After Apply weight the shares are sums of weights, the axis reads "% of respondents (weighted)" and a note says "Weighted by 'weight'; n counts respondents."

Result chart

visualize.result_chart — "The chart that suits an analysis's result, drawn from the numbers it computed: group means with their confidence intervals, a scree plot, TURF's reach curve, key drivers, a perceptual map, price sensitivity curves." Because it draws the analysis's own numbers rather than the data again, the picture never disagrees with the table beside it.

In: result (Table or Stat, several) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
Kind	choice	auto	auto, means, means_sd, interval, stacked, reach, items, utilities, scores, shares, importance, partworths, scree, loadings, profile, coefficients, heatmap, sentiment, map, curves	Auto draws the chart the result suits; the others are listed below.
Title	text	—	—	Empty takes the result's own title ("Age by Region", "Key drivers of Overall satisfaction").
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it. A chart with more rows than its height holds legibly grows taller.
Figure height (in)	number	6	2–30	—
Palette	choice	muted	muted, deep, pastel, dark, colorblind, Set2, tab10, theme	The colors of the bars, points and series. Net Promoter Score and sentiment keep their own red–gray–blue, the heatmaps their diverging scale, and Key drivers, the Perceptual map and Price sensitivity their own color-blind-safe pairs. Theme takes every color from the report the chart is saved in — the chart colors of its Save report's Look (by default eight colors any two of which readers with protanopia or deuteranopia can tell apart), its series, its diverging pair for the red–gray–blue and the heatmaps, its text color, grid and font.

Connecting it. Wire the analysis's table into result — for Factor analysis and Principal components the variance output (scree) or loadings, for Proportion CI its stat, for a Perceptual map any of its three tables, for Price sensitivity its table or curves. The input takes several wires: connect the analysis's stat too and the chart learns its base and weight from it (a regression's intervals then use its t distribution). The table alone says how the weight was used.

The Kind field says under it what the connected result suits and what auto draws — "Group means (means.table) suits means — means with 95 % intervals; means_sd — means ± 1 SD. Auto draws means." — and in its list every other kind has "(not for this result)" after its name, under its code. With nothing connected it reads "Connect an analysis's table (or Proportion CI's stat) to see the charts it suits." A screen reader hears that line with the field. The list names each kind under its code: auto ("the chart the result suits"), means ("means with 95 % intervals"), means_sd ("means ± 1 SD"), interval ("a share and its interval"), stacked ("NPS: detractors, passives, promoters"), reach ("TURF reach curve"), items ("TURF: each option's reach"), utilities ("MaxDiff utilities"), scores ("MaxDiff scores"), shares ("shares"), importance ("importance: Conjoint, Key drivers"), partworths ("Conjoint part-worths"), scree ("scree plot"), loadings ("loadings heatmap"), profile ("cluster profiles"), coefficients ("coefficients or odds ratios"), heatmap ("correlation heatmap"), sentiment ("sentiment split"), map ("perceptual map"), curves ("price curves"). Its list has more than ten entries, so it opens with a filter box ("Filter chart kinds…"): type "maxdiff" to keep utilities and scores.

What it draws — the first kind of each row is what auto draws:

Analysis (output)	Kinds	The chart
Group means, Descriptive statistics, t-test (table)	means, means_sd	each mean with its 95 % confidence interval, or ± 1 SD, a row per group or measurement with its base (North (n = 97)); after a post-hoc test, letters: "Means sharing a letter do not differ (Tukey HSD, p < .05)."; a t-test notes its test ("Welch's t-test (unequal variances): difference Male − Female = 2.5 (95% CI -2.101 – 7.100), p = 0.2861")
Paired tests with Wilcoxon, Friedman or auto (table)	means, means_sd	the means of the measurements, noting that the test compares ranks
Paired tests with McNemar or Cochran's Q (table)	shares	the share saying yes to each, with Wilson's interval
Proportion CI (stat)	interval	the share and its interval on a 0–100 % track, titled "Gender: Female"
Net Promoter Score (table)	stacked	detractors, passives and promoters as one bar, with the score and its interval
TURF (table)	reach; items with Search fixed	reach by portfolio size, each size named by the option it adds; or each option's reach and what only it reaches
MaxDiff (table)	utilities, scores, shares; scores with Estimate counts	the utilities with 95 % intervals against the reference item at 0, the counting scores, or the shares
Conjoint (table)	importance, partworths	each attribute's importance, or the part-worths of its levels in the design's order
Share of preference (table)	shares	each product's share
Principal components, Factor analysis (variance)	scree	the eigenvalues with the Kaiser line at 1 — the factors kept filled, and the parallel analysis's line when it ran
Principal components, Factor analysis (loadings)	loadings	a heatmap of the loadings
Cluster (k-means) (table)	profile	each cluster's means down the items, the legend giving its size ("Cluster 1 (n = 141, 44.9 %)")
Regression (table)	coefficients	a forest of the coefficients with their 95 % intervals, without the intercept; odds ratios on a log scale for a logit and the ordinal logit (its thresholds left out)
Correlation matrix (table)	heatmap	the lower triangle with the table's significance marks
Code open answers (table)	shares; sentiment too with sentiment checked	each theme's share (of the respondents who answered; of the coded answers for an older, version 1 codeframe) — not the nets, nor the Coded and Uncoded rows; its answers' sentiment
Key drivers (table)	importance	each driver's share of R², largest first, a negative beta in a second color
Perceptual map (table, rows or columns)	map	the symmetric map of the first two dimensions, each axis with its share of the inertia
Price sensitivity (table or curves)	curves	Van Westendorp's curves and price points (with the NMS trial curve), or Gabor-Granger's demand above its revenue

A Frequencies table is not a result it draws: use a Bar chart (Show percent) for a distribution.

Checks before the run, as you edit and at Check:

• An output it cannot draw is an error, RESULT_NOT_DRAWABLE: "A Result chart cannot draw the table output of Frequencies (f); it draws the results of Group means, Descriptive statistics, t-test, Paired tests, Proportion CI, Net Promoter Score, TURF, MaxDiff, Conjoint, Share of preference, Principal components, Factor analysis, Cluster (k-means), Regression, Correlation matrix, Code open answers, Key drivers, Perceptual map, Price sensitivity." A statistic alone that only says the weight and base: "The stat output of Regression (r) only tells a chart its weight and base; connect the output it draws, table, too."
• A Kind the result does not suit is an error, RESULT_KIND, naming what it does draw — and the output of the same analysis that draws the kind you chose: "Kind 'scree' does not suit the loadings output of Principal components (p), which draws 'loadings'; its variance output draws 'scree'."
• Results of two analyses are a warning, RESULT_SOURCES: "The results connected come from m, t; a Result chart draws one of them — the table output of Group means (m)."

What only the data can tell fails the node, with the reason, when it runs.

Titles, weight and labels. The title's second line says how the result used the weight — "weighted by 'weight'", or "unweighted (the weight 'weight' is not applied)" for a t-test, a paired test, a factor analysis or a cluster — under your own Title too. Axis labels say what is drawn ("Mean with its 95 % confidence interval", "Weighted mean with …", "Odds ratio with its 95 % confidence interval (Wald), log scale"). Categories are rows, first at the top; their labels wrap and stay whole, the font shrinks as rows multiply, and a chart that still cannot hold them grows taller rather than cut them — only a label past four lines is cut, and never so that two read alike. The title, axis titles and notes wrap to the plot; values never leave it. Many series each get a color of their own. A perceptual map whose labels would overlap is drawn taller, and one too crowded even then numbers its points and lists the numbered names under the legend.

Scatter plot

visualize.scatter — two variables against each other, optionally colored by a third, with a trend line.

In: data (SurveyData) → Out: chart (Chart)

Parameter	Type	Default	Allowed	Meaning
X	variable	required	interval / ratio / ordinal variables	Horizontal axis.
Y	variable	required	interval / ratio / ordinal variables	Vertical axis.
Color by	variable	—	nominal / ordinal variables	Optional variable that colors the points.
Trend line	checkbox	on	—	Draw a fitted trend line.
Title	text	—	—	Chart title.
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	—
Palette	choice	muted	muted, deep, pastel, dark, colorblind, Set2, tab10, theme	Theme takes the colors of the report the chart is saved in — the chart colors of its Save report's Look (by default eight colors any two of which readers with protanopia or deuteranopia can tell apart), with its text color, grid and font. The others are seaborn's palettes.

Every respondent is one point and the trend line is fitted unweighted; on weighted data the title's second line says "unweighted (the weight 'weight' is not applied)".

Trend

visualize.trend — "A measure over waves or dates, as a chart and as a table: the percent choosing an answer, a mean, or the count of respondents. One line per group, with its confidence band and the base of every point." The chart of a tracking study: satisfaction month by month, awareness wave by wave, completes per week.

In: data (SurveyData) → Out: chart (Chart), table (Table)

Parameter	Type	Default	Allowed	Meaning
Time	variable	required	any variable, or a response timestamp	A wave code (one point per wave, ordered by code, its labels on the axis) or a date — a date column, or text in ISO 8601 such as the responses' created_at (2026-05-25 09:00:00+00:00). Times with a zone are read in UTC.
Period	choice	month	day, week, month, quarter, year	How dates are grouped into points. A week is an ISO week, Monday to Sunday, labeled by its ISO year and number (2026-W22). Every period between the first and the last is on the axis, an empty one as a gap. Not used for a wave code.
Measure	choice	percent	percent, mean, count	percent: the share of those who answered that gave one of the Answer codes. mean: the mean of a numeric (or ordinal) variable. count: the number of respondents — the sum of weights when weighted.
Measure variable	variable	—	—	The question the percent or the mean is of. Its missing codes are left out of every base and counted.
Answer codes	answer code (several)	—	answers of Measure variable	The answer counted, or a list of them counted together (a top-2 box), e.g. [4, 5]. For a multiple-choice question, choosing any of them.
Split by	variable	—	nominal / ordinal variables	One line per group. Respondents with a missing code, or no answer, are in no line.
Confidence band	checkbox	on	—	The 95% interval of each point — a percent's Wilson score interval, as the Bar chart's intervals, or the mean's t interval, on Kish's effective base when weighted. A point with fewer respondents than Minimum base is drawn without it; the table gives its interval.
Minimum base	whole number	30	1 or more	A percent or a mean of fewer respondents than this is drawn hollow and noted in the table. Not used for the count, which is its own base.
Title	text	—	—	Empty takes what is tracked ("Overall satisfaction: % Satisfied or Very satisfied", "Mean Trust: Acme", "Respondents"), with "by Split by".
Figure width (in)	number	10	2–30	The figure itself, in inches — the axis labels scale with it.
Figure height (in)	number	6	2–30	—
Palette	choice	muted	muted, deep, pastel, dark, colorblind, Set2, tab10, theme	Theme takes the colors of the report the chart is saved in — the chart colors of its Save report's Look (by default eight colors any two of which readers with protanopia or deuteranopia can tell apart), with its text color, grid and font. The others are seaborn's palettes. Whatever the palette, past four lines each line's points also take a shape of their own.

The lists name each code under it: percent ("% choosing the Answer codes"), mean ("mean of the Measure variable"), count ("respondents (weighted: sum of weights)"), and Period's week ("ISO week, Monday to Sunday"). Measure variable is shown with percent and mean, Answer codes with percent, Confidence band and Minimum base with percent and mean — a count has no band and is its own base. The card reads satisfaction = 4, 5 over created_at, mean trust_acme over wave or respondents over created_at.

Time. The picker lists the codebook's variables — Waves and dates first (labeled codes, ordinal variables, date columns, a Date question's answers), then Other variables — and, under Beside the answers, the timestamps the survey's responses carry: created_at ("Response date (created_at)" under it; when the response came in — the usual choice), updated_at ("Last change (updated_at)") and started_at ("Start time (started_at)"). Each group is a heading in the list. The axis names them by those labels ("Response date (month)"). It leaves out a multiple-choice question, a ranking and an open answer (not a Date question's), as Split by does; with Measure mean, Measure variable leaves out nominal and multiple-choice questions. A variable already stored stays listed, with the reason after it in the field (aware "(several answers: not one wave or date)"). A variable with value labels, or of numbers, is read as waves: a point per code, in the order of the codes, each named by its label ("Spring 2026", "Summer 2026"); a wave the codebook declares between the first and the last, with no data yet, is a gap. A column of dates, or text most of which reads as ISO 8601 dates, is grouped by Period: day (2026-01-05), week (the ISO week, 2026-W02), month (Jan 2026), quarter (2026 Q1) or year; a value that is not a date is left out and counted under the chart. Every period from the first to the last is on the axis, so a week without respondents is a gap in a percent or a mean, and a 0 in a count.

Answer codes is a checklist of the Measure variable's answers (its missing codes are not offered): check one answer, or several to track them together — 4 and 5 for a top-2 box. Without value labels it is a JSON box ([4, 5]).

Under the chart, notes say what it shows: the base ("Base: 821 respondents who answered; 114 to 150 per point."), the gaps ("Gaps: no respondents in 6 of 18 points."), the band ("Band: 95% confidence interval.", "Bands: 95% confidence intervals."), the hollow points ("Hollow points: fewer than 30 respondents, drawn without a band (the table gives their intervals)."), the weight, the missing codes left out ("Left out as missing: Trust: Acme: 194 (9 = Refused).") and the rows no line or point could take ("Left out: 25 without Trust: Acme."). The band is drawn for up to four lines; with more, "No bands: the 95% intervals of 5 lines would hide one another; the table gives each point's.", and each line's points take a shape of their own, which the legend shows too. Every line has a color of its own, however many there are.

The table output gives the same points, for a report section or a Live tile: Period, the Split by group, Percent (or Mean, or Count), Lower 95% and Upper 95% (not for a count), Base (respondents), on weighted data Weighted base and Effective base, and Note ("base below 30", "no respondents", "their weights sum to 0"). Under it: Measure ("% choosing 4 = Satisfied, 5 = Very satisfied — Overall satisfaction"), Base, Time ("created_at, by week — ISO weeks, Monday to Sunday, labeled by ISO year and week number"), Interval ("95% Wilson score interval, as the Bar chart draws a share's", "95% t interval of the mean"), Weight, Low base ("6 of 18 points have fewer than 30 respondents (drawn hollow)") and Missing codes left out. The node's preview shows the table under the picture. The Methods draft describes the node in words: "the percentage of respondents who answered Overall satisfaction choosing 4 (Satisfied) or 5 (Very satisfied) was tracked over the date each response came in, by month, one line per group of Region, with its 95% confidence band (Wilson score intervals, on Kish's effective base when weighted), points of fewer than 30 respondents drawn hollow and without a band, missing codes left out". Over a wave variable — labeled codes, or numbers without labels, a point per code — it says "was tracked wave by wave (Wave)".

A percent's point is the share Proportion CI gives for the same respondents, and its band Wilson's interval — the one the Bar chart draws for a share. A mean's band is Student's t interval. After Apply weight the percents, means and counts are weighted, the intervals use Kish's effective base, the axis adds "(weighted)" ("% of respondents (weighted)", "Mean (weighted)", "Respondents (weighted)"), the base note adds the weighted total ("… who answered (weighted: 822.8); …") and the note "Weighted by 'weight'; the bases count respondents." follows.

Rules. Errors, before the run: "Name the Measure variable whose answers are counted.", "Name the Answer codes whose percent is tracked — one code, or a list such as [4, 5]." (percent) and "Name the Measure variable whose mean is tracked." (mean). What the questionnaire settles is refused before the run too:

• "Region is nominal: its codes are names, not amounts, so their mean says nothing. Track the percent choosing an answer instead (Measure = percent)."
• "Brands heard of (unaided) holds multiple-choice answers (lists of codes), which have no mean. Track the percent choosing an answer instead (Measure = percent)."
• "Brands heard of (unaided) holds multiple-choice answers (lists of codes), and Time is one wave or one date per respondent: choose the wave's variable or a date."
• "Split by needs one answer per respondent, and Brands heard of (unaided) allows several: split by one of its options after Explode multiple choice, or choose another variable."
• "9 (Refused) is a missing code of Trust: Acme, not an answer: missing codes are left out of the base. Name an answer."

A timestamp Studio's responses do not carry is an unknown variable (submitted_at: "time: "submitted_at" is neither in the codebook nor made by this flow or a table it reads."). One the node reads from Simulated data, which has none, is a warning: "time: "created_at" is a timestamp of the survey's responses, and this node reads Simulated data: the run stops unless that data has a column created_at. Read the Responses source, or choose a column the data has." Below a Data file whose columns have been read, a timestamp is known only if the file has that column (see The file's columns in the flow); otherwise it is an error like any name the file lacks ("time: "created_at" is not a column of assets/panel_wave2.csv, nor made by this flow."), and the Save also gives the warning above, naming Data file. A file whose read columns include the timestamp gets no warning, on the canvas or at Save; one not read yet gets the warning at Save.

What only the data can tell stops the node when it runs, with the reason: a code that is no answer ("Trust: Acme has no answer 7; its answers are 1 = No trust, 2 = Low, 3 = Medium, 4 = High, 5 = Full."), a Time with no dates in it ("created_at holds no dates to draw a trend over."), and more than 500 points — a date range too long for its Period ("created_at spans 905 days: too many points for one chart. Choose a longer Period.") or a Time of more than 500 different values ("… has 731 different values: not wave codes. Time is a wave code or a date; for dates, the column must hold dates (ISO 8601 text such as 2026-05-25 is read as one).").

---

Output

Where results leave the flow.

Choice data for HB

output.choice_data — writes a MaxDiff's answers in the long format that R's hierarchical-Bayes packages read.

In: data (SurveyData) → Out: none (writes files, a table or a tile)

Parameter	Type	Default	Allowed	Meaning
MaxDiff question	text	required	—	The question's id or name. Its design is read from the questionnaire.
File name	file name	required	—	The name before a fixed .csv, in outputs/; a node added from the palette gets <flow>_maxdiff_choices. The dictionary and R script are written beside it.

Three files, so that estimating individual-level utilities on your own machine needs no rewriting: the choices (<name>.csv), a dictionary saying what every column means (<name>.dictionary.json) and a script that runs the model and writes the per-respondent estimates back (<name>.hb.R). The field's ⓘ adds, in a paragraph of its own: "It has no weight column, since the HB packages take none, so an applied weight is not in it. Weight the individual utilities when you aggregate them." A run of the flow keeps all three under Files; Run all keeps them only when the flow lists them among its outputs (as the example study's segments flow does), and the line under the field says so: "Run all doesn't keep these files: run this flow on its own to get them."

Why export rather than estimate here: hierarchical Bayes takes minutes of MCMC, and a flow run has one CPU and a few minutes for everything. A cut-down chain would be a worse answer under the same name. The tables MaxDiff and Conjoint produce are aggregate estimates, which is a different — and honestly labeled — thing.

Conjoint data for HB

output.conjoint_data — the same for a choice-based conjoint: the choices in long format, a column dictionary and an R script.

In: data (SurveyData) → Out: none (writes files, a table or a tile)

Parameter	Type	Default	Allowed	Meaning
Conjoint question	text	required	—	The question's id or name. Its attributes and design are read from the questionnaire.
File name	file name	required	—	The name before a fixed .csv, in outputs/; a node added from the palette gets <flow>_conjoint_choices. The dictionary and R script are written beside it.

As for MaxDiff, the file has no weight column: weight the individual part-worths when you aggregate them. Run all keeps the files only when the flow lists them among its outputs; a run of the flow keeps them always.

Export file

output.export_file — "Write the data as a file (format by extension: .parquet, .csv, .xlsx, .sav, .dta) with its dictionary, as an R bundle (.R), or the dictionary alone (.json)."

In: data (SurveyData) → Out: none (writes files, a table or a tile)

Parameter	Type	Default	Allowed	Meaning
File name	file name	required	—	The file's name, in outputs/; a node added from the palette gets <flow>_data. Format under it picks the kind of file — its ending.

Format offers CSV (.csv) (a new node's), Excel (.xlsx), SPSS (.sav), Stata (.dta), Parquet (.parquet), R bundle (.R) and Codebook only (.json). There is no .xls: an .xls name would hold an .xlsx workbook, which Excel warns about, so a typed clean.xls is answered "Choose Excel (.xlsx): an .xls name would hold an .xlsx workbook, and Excel warns about that." with Use clean as Excel (.xlsx). What each format writes, for File name clean:

Format	Files
CSV, Excel, Parquet	the data (clean.csv, clean.xlsx, clean.parquet) and clean.dictionary.json
SPSS, Stata	the data with its labels inside (variable labels, value labels, declared missing values) and clean.dictionary.json
R bundle (.R)	clean.csv, clean.dictionary.json and clean.R
Codebook only (.json)	the data dictionary (codebook) alone, no data — an error when the data has no codebook

The R bundle. clean.R reads clean.csv (as UTF-8) and its dictionary (with the jsonlite package) into a data frame named survey_data: the codebook's missing codes become NA, labeled codes become factors, and each column's label attribute is the variable's label from the codebook (not the question's text). A code the codebook has no label for keeps a level of its own rather than turning into NA; a multiple-choice column (codes joined by ;) stays text, because a factor holds one value per respondent; a text answer that reads NA stays an answer. The script finds its two files beside itself, whether you run it with Rscript clean.R or source("path/to/clean.R") from R.

An .xlsx keeps text as text: an open answer such as =HYPERLINK(…) is written as the string it is, never as a formula Excel would run.

A run of the flow keeps the files: they appear in Files and as download chips on the run's card. Run all keeps them only when the flow lists them among its outputs (as the example study's cleaning flow lists its R bundle); for a flow built on the canvas the line under the field says "Run all doesn't keep these files: run this flow on its own to get them." Written after the cleaning and weighting steps, the export carries the variables the flow made (recodes, bands, factor scores, a weight column) with their labels.

Live tile

output.live_tile — publishes whatever is connected to it — a number, a table, a chart or a statistic — as a tile on the Live screen. See Live Monitoring.

In: input (Any) → Out: none (writes files, a table or a tile)

Parameter	Type	Default	Allowed	Meaning
Kind	choice	number	number, table, chart, stat, text	How the tile shows the value: number, table, chart, stat or text.
Label	text	required	—	The tile's caption on the Live screen, shown in capitals; a statistic's symbol — a Greek letter on its own, with any index written onto it (the χ² of "χ² test", ηp²) — keeps its case.
Show	choice	value	value, rows	rows: the number of respondents in the data connected.

• Tiles are published by every run of the flow (a manual run, a scheduled run of this flow, a live recompute). Run all and Run to here do not publish tiles, and do not clear them either: the Live screen shows the tiles of the flow's latest completed run. After a rename it goes on showing them until the flow runs under its new name.
• Show = rows is the usual way to show "respondents after cleaning": connect the SurveyData output of the last cleaning step.
• A Trend makes a tracking tile: its chart with Kind chart, or its table — the points with their intervals and bases — with Kind table. A chart whose Palette is theme is drawn on the tile in the look of the flow's Save report, as its report draws it.
• A chart tile carries the chart's picture and its interactive form: the Live tab and the public page draw it with a tooltip on every mark and a legend that hides and shows a series, whatever the flow's Save report says about Interactive charts in HTML; the public page shows a Scatter plot, or a Box plot with outliers or Show points, as its picture only. See Interactive chart tiles.
• A tile's size on the Live screen is not set from the canvas; tiles appear at the standard size; on a phone a chart tile takes the whole row at its own height.

Report section

output.report_section — collects tables, charts and statistics into one section of a report, with a heading, introductory text, a caption per item and a closing note. Without a heading it is a plain block of Markdown between the sections around it. Usually edited in the flow's Report view — see Reports.

In: items (Table or Chart or Stat, several, optional) → Out: report (Report)

Parameter	Type	Default	Allowed	Meaning
Heading	text	—	—	Section title, on one line. Leave empty for a plain block of text between sections.
Text	Markdown	—	—	Introductory text (Markdown).
Captions	one caption per input	—	—	A caption for each connected table, chart or statistic.
Size and placement	size per input	—	—	Per item — a width (60%, 320px), an alignment, or a page break before it. The Markdown is unaffected.
Note	Markdown	—	—	A closing note (Markdown) — methods, source, base.

• The order you connect items in is the order they appear.
• A statistic (a Stat output) is printed as one line — Caption: key = value; … — where its caption is the label before the values. It has no size or placement: under Size and placement its row reads "one line" ("A statistic is one line of the report: size and placement apply to tables and charts only").
• Size and placement reach the HTML only; the Markdown is unaffected.

Save report

output.save_report — combines sections, in the order you connect them, into one report and saves it: Markdown (the content, with each chart as a PNG beside it named by the report — report_fig_1.png, report_fig_2.png for outputs/report.md), by default an HTML twin (the look, stylesheet and images inside the file), and on request the report's tables in an Excel workbook. Two reports in one folder therefore never share a figure. A report produced by a run on the platform ends with a provenance footer while Settings → Reports → End every report with the provenance footer is on (the default).

In: sections (Report, several) → Out: report (Report)

Parameter	Type	Default	Allowed	Meaning
Title	text	required	—	The report's title.
File name	file name	outputs/report.md	—	The report's name before a fixed .md, in outputs/. A node the Report view creates, or one added from the palette, gets the flow's name (outputs/<flow>.md).
Also save HTML	checkbox	on	—	Also write the styled .html twin next to the Markdown.
Interactive charts in HTML	checkbox	off	—	The HTML draws its charts in the reader's browser — a tooltip on every bar, point and cell with its value and base, a legend whose entries hide and show their series, zoom where it helps — and still opens offline, so it can be sent to a client as it is. For that it carries the chart libraries (Vega, Vega-Lite and Vega-Embed, about 0.8 MB, written in once) and each chart's numbers — what it draws, and a scatter plot's points and a box plot's outliers, the respondents' values it plots. Each chart's picture stays in the HTML for printing and for readers without scripts; the Markdown and the Excel workbook are unchanged. Applies only with Also save HTML.
Table of contents	checkbox	off	—	Add a table of contents.
Also save tables to Excel	checkbox	off	—	Every table of the report in one workbook beside it, named as the report with .xlsx — a sheet per table, named by its caption or its section's heading, with its statistics under it, and a Contents sheet first. Banner tables keep their significance letters and Group means its post-hoc pairs on a sheet of their own; charts are left out.
Look	report look	—	—	Typefaces, measure, table style and page size of the rendered report, and the chart colors and font of every chart in it whose Palette is theme (in the Markdown's figures too). Its p_values says how the report writes a p-value — exact, as computed (the default); 0.01, one below 0.01 as < 0.01; or 0.001, one below 0.001 as < 0.001 — in its tables' cells and statistics lines, the Markdown as the HTML, the Excel workbook (a number format, so the cell keeps its number) and the charts' notes (< .01); the results and their exports keep the exact p. Otherwise the Markdown's text is unaffected.

• The Excel workbook (<name>.xlsx beside <name>.md) is the report's tables as a spreadsheet: numbers stay numbers, and text stays text — an answer, label or caption that begins with = is never turned into a formula. Its first sheet, Contents, carries the report's title and one row per table (Sheet, Section, Table), each a link to its sheet; a table without a caption is named by what it is ("Group means: Age", "Perceptual map: Region × Overall satisfaction — rows (Region)"). Sheet names are cut to Excel's 31 characters, lose [ ] : * ? / \ and are made unique ("Satisfaction", "Satisfaction (2)"). Charts, text and statistics lines (a Stat in a section) are not in it. The report composer's Also Excel is the same box; see Tables in Excel.

• Interactive charts in HTML (the Report view's box of the same name) makes the .html draw its charts in the reader's browser: a tooltip with each value and its base, a legend that hides and shows a series, zoom on a Scatter plot, a Trend of more than 24 periods and a Perceptual map, and a menu that saves a chart as PNG or SVG. Beside the Markdown, each figure's Vega-Lite spec is written next to its picture (report_fig_3.png, report_fig_3.vl.json); the .md and the Excel workbook keep the pictures. Unchecked (the default), the node writes exactly what it wrote before the option existed. With Also save HTML unchecked it does nothing, and Check and the node warn: "Interactive charts in HTML draws the charts of the HTML — turn on Also save HTML; the Markdown and the Excel workbook keep their pictures." What the file then carries — the chart libraries and each chart's numbers, a scatter plot's and a box plot's points included — is under Interactive charts.

• The Look parameter is edited in the Look tab of the Report view (Reports). A node created by the Report view or added from the palette starts with the project's house style as its look, when the project has one. A node with no look renders with the engine's defaults — on the platform, in previews and in a research bundle alike; the house style is not applied to it behind the scenes.

• The Look also holds the report's Chart colors (the theme's chart_palette, chart_sequential, chart_diverging, chart_text_color, chart_grid_color and chart_font): every chart in the report whose Palette is theme (a Heatmap's Color map theme) is drawn in them, in the .md's figures as in the .html. A chart with a named palette keeps its own. What the engine refuses is named on the node as you type, in its words: "theme: chart_palette: 'purple' is not a hex color such as '#2a78d6'."; "theme: chart_palette: give between 2 and 12 colors, in the order the series take them; got 1."; "theme: chart_palette: '#2a78d6' is given twice; two series would look alike."; "theme: chart_palette: '#ffe8b2' on the charts' white background has a contrast of 1.2:1; a bar or a line in it needs at least 1.3:1 to be seen."; "theme: chart_diverging: the two ends are the same color, so the scale would not diverge."; "theme: chart_text_color: '#cccccc' on the charts' white background has a contrast of 1.6:1; text needs at least 4.5:1."; "theme: chart_font: a list of font names separated by commas, without ; { } < >." See Chart colors.

• The Look also holds the report's P values (the theme's p_values: exact, the default, 0.01 or 0.001), the one choice of the look besides the chart colors that reaches the Markdown: a p below the threshold is written as the bound — < 0.01 in a table's p column, p < 0.01 in a statistics line, a number format in the Excel workbook, and a chart's note. The results, the tables the node receives and every export keep the exact p, and a Proportion CI's p, the proportion, is never changed. It travels with the node, so the downloaded script writes theme={"p_values": "0.01"}. Any other value is named on the node: "theme: p_values: '0.05' is not one of exact, 0.01, 0.001." See P values.

• The line under File name says where a run leaves the report ("After a run: Files → outputs/flow/name.md") and what is written beside it (<name>.html with Also save HTML, <name>.xlsx with Also save tables to Excel, and <name>_fig_1.png, "a picture per chart").

• Choose this node as the flow's Report path (Flow settings, with no node selected) if you want this report kept by Run all and in its combined report; the Report view does both for you when it creates the node. A node added from the palette sets no Report path, and its field says "Run all keeps a report only when it is the flow's Report path (in the Flow panel, with no node selected): run this flow on its own to get this one." How the Report path follows the node, and the warning for one that names no node's file: The combined report.

• There is no PDF output. A typed .pdf is answered under the field — "Reports are saved as Markdown and HTML — for a PDF, open the HTML and print it." — and so are .html ("The report is saved as Markdown (.md); Also save HTML puts an .html copy beside it."), .docx ("Reports are saved as Markdown and HTML — Word opens the .html copy.") and .xlsx. A path ending in .pdf saved earlier fails at the run; the field shows it under "Other location, kept as it was written", with a fix.

Tab book (Excel)

output.tabbook — "Every chosen question crossed by a banner of segments, in an Excel workbook with a contents page and notes. Each question has a sheet of its own, with its bases, counts, percentages and the Banner table's significance letters." The tab book an agency hands a client after fieldwork: the whole study by the segments that matter, in one file.

In: data (SurveyData) → Out: stat (Stat)

Parameter	Type	Default	Allowed	Meaning
Questions	variables (several)	—	—	Empty is every nominal, ordinal and multiple-choice variable of the codebook, except the banner variables, the weight, the response metadata, open answers (an OpenText question, or words without answer labels — they would go into the workbook word for word) and rankings (every option would read 100 %); the Notes sheet says why each is left out. Named here, any question is tabulated as asked, and an interval or ratio one shows its mean.
Banner	variables (several)	required	nominal / ordinal variables	The columns are Total, then every code of each banner variable. Letters run across the whole banner, but columns are only compared inside their own variable.
Percentages	choice	column	column, row, none	column: of the column's respondents who answered. row: of the answer's respondents across each banner variable. none: counts only.
Counts	checkbox	on	—	The count of each cell (the sum of weights on weighted data).
Significance letters	checkbox	on	—	The Banner table's test — a two-sided z-test of column proportions within each banner variable, on the base of those who answered (Kish's effective base when weighted); a column below 30 is not tested. The letters sit in their own cells beside the column percentages, and are shown only with them.
Level	number	0.05	0.001–0.2	The letters' significance level.
Multiple comparisons	choice	none	none, bonferroni	bonferroni divides Level by the number of pairs of columns tested.
Means	checkbox	on	—	The mean and standard deviation of an interval or ratio question per column (not tested).
File name	file name	outputs/tabbook.xlsx	—	The workbook's name before a fixed .xlsx, in outputs/. A node added from the palette gets <flow>_tabbook (outputs/<flow>_tabbook.xlsx), numbered when the flow already has one.

The Percentages list names each code under it: column ("of each column's respondents"), row ("of each answer's respondents"), none ("counts only"). Level and Multiple comparisons are shown only while Significance letters is checked. The card reads the banner and the file: gender, region → outputs/tabbook.xlsx.

The workbook.

Sheet	What it holds
Contents (first)	"Tab book", a line such as "5 questions × Total, Gender, Region", then one row per question — #, Question (its label), Variable, Base, Sheet — each linked to its sheet; a link to the notes; and Not tabulated, each question left out with Why
one per question, named after its variable	the question's label; a line with its variable, scale and base ("trust_acme · ordinal · Base: respondents who answered", "· multiple answers" for a multiple-choice question); ← Contents; the banner — Total, then each banner variable over its columns, each column lettered (Male (A), Female (B), Other (C), Capital (D), …); a Base row; per answer its count and, under it, its percentage with the letters in the cell beside it; for an interval or ratio question Mean and Standard deviation; footnotes ("Percentages are of the column's respondents who answered.", "Several answers were allowed, so the percentages add to more than 100%.", the letters' test, "Left out, as missing codes: Trust: Acme: 194 (9 = Refused).", "Means are not tested.")
Notes	Created (UTC), Respondents, Weight, Banner ("Total, Gender (A–C), Region (D–F)"), Percentages, Counts, Base, Test, Alpha, Bonferroni, Minimum base for a test, Letters, Means, Missing codes left out, Not tabulated

On each question's sheet the rows down to the bases and the column of answers are frozen, so they stay in view as you scroll. Numbers are numbers (percentages formatted as percentages), and text stays text: a label or an answer that begins with = is never turned into a formula. A sheet name is the variable's name cut to Excel's 31 characters, made unique.

What the numbers are. The counts, column percentages and letters are the Banner table's, with two differences, both said on the Notes sheet: the codebook's missing codes are left out — of the question (not in its base) and of a banner variable (a respondent with a missing code for region is in Total, not in a region column) — and a column's base, for its percentages and its test, is its respondents who answered the question (the Banner table tests on everyone in the column; the two agree when everyone answered). A letter beside a cell names a column of the same banner variable whose percentage this one's is significantly higher than. A multiple-choice question is one table of every option, its base those who chose at least one.

On weighted data each sheet has Base (unweighted) (respondents) and Base (weighted); counts and bases are sums of weights shown to one decimal, the percentages are of those sums, and the test uses Kish's effective base ("Minimum base for a test: 30 respondents (Kish's effective base)"). A footnote says "Weighted by 'weight'; the unweighted base is the number of respondents."

What is left out, and why, on the Contents and Notes sheets and in the statistic:

• with Questions empty, an open answer — "an open answer: code it first (Code open answers)" — and a ranking — "a ranking: every respondent orders every option, so each would be 100 % — derive its first choice (Derive) and tabulate that";
• always, a question nobody answered ("nobody answered it"), and one without answer labels holding more than 30 different answers ("… different answers and no answer labels — an open answer? Code it first (Code open answers) or band it (Bands)") — an interval or ratio question then shows only its mean and standard deviation, and is left out with Means off ("… turn on Means").

Interval and ratio variables are not taken by default; name them in Questions for their means.

The statistic says what was written — Workbook, Sheets written, Questions skipped, Skipped (each with its reason), Banner, Percentages, Test ("two-sided z-test of column proportions at 0.05, within each banner variable") and, on weighted data, Weight. A preview runs the node and shows it, with "Not kept: a preview never keeps the files nodes write. A run writes outputs/tabbook.xlsx." (the node's file). A run keeps the workbook as outputs/<flow>/<name>.xlsx, as the line under File name says: on the run's card, in Files, and on the Reports screen as Tab book (see The Reports screen). Run all keeps it too. The Methods draft says what it holds: "every nominal, ordinal and multiple-choice question (open answers and rankings left out) was cross-tabulated by a banner of Gender and Region into a tab book in Excel (outputs/tabbook.xlsx, a sheet per question with its bases, column percentages and counts), with significance letters beside each column percentage …".

Rules. Errors: a path that is not a workbook ("path: A tab book is an Excel workbook: its path must end in .xlsx (got 'outputs/tabs.csv')."), a banner variable with several answers ("Brands heard of (unaided) holds multiple-choice answers, and a banner column is a group of respondents that no one else is in. Explode it first (prepare.explode) and use its columns, or choose another banner variable.") and "With Percentages none and Counts off the tables would be empty — turn one of them on." Warnings: a path outside outputs/ ("path: 'tabs.xlsx' is not under outputs/, where a run keeps what it writes."), "Significance letters compare column percentages, so the sheets show them only with Percentages column; with row percentages or counts only they are left out.", and a ranking or an open answer named in Questions ("Free comment is an open answer: its answers would go into the workbook word for word — code it first (Code open answers)."; "… is a ranking: every respondent orders every option, so each option would read 100 % in every column — derive its first choice (Derive) and tabulate that."). When it runs, a banner variable of more than 30 different values stops the node: "Postcode has 212 different values: too many columns for a banner. Band it first (prepare.bands) or choose a variable of a few groups."

The File name field keeps outputs/ and .xlsx fixed, so the not-a-workbook error and the outside-outputs/ warning come only from a path saved earlier, which the field shows as it was written, under "Other location, kept as it was written", with a one-click fix. A typed tabs.csv is answered under the field: "A tab book is an Excel workbook: it is saved as .xlsx."

Write table

output.write_table · platform — saves the data as a project table: a source for other flows (Project table) and a table on the Data screen.

In: data (SurveyData) → Out: none (writes files, a table or a tile)

Parameter	Type	Default	Allowed	Meaning
Table name	text	required	—	Table name: lowercase letters, digits and _, starting with a letter or _, at most 63 characters.
If it exists	choice	replace	fail, replace, append	replace (drop and recreate), append (add rows) or fail (stop the run with "table already exists").

• The table keeps the variables of the columns it holds — labels, scales and value labels, including those of variables the flow created — so a flow that reads it gets them back (see Project table). With append, the variables are merged with those the table already carries. Written from a Data file, it brings the file's columns to the flows that read it.
• The columns keep their names as they are — a Qualtrics export's StartDate and Duration (in seconds), a header in Cyrillic, a name with % in it. A name longer than 63 bytes (a Cyrillic letter takes two) cannot be a table column, and neither can an empty one or two columns of one name: the node stops before anything is written, in a preview as in a run — "column name … is longer than 63 bytes, which a table column cannot be: rename it (Compute or Recode into a shorter name, then Select columns) before Write table", or "columns named twice: …".
• A preview (Run to here, Preview all) never writes the table. The node's Preview pane says what a run would do: "Not written: a preview never writes project tables. A run writes 812 rows to table 'clean_responses' (if it exists: replace)."
• In a research bundle, when the script is given a data file, this node is skipped: there is no project database outside Studio.

---

What is deliberately absent

• A Python or SQL node. Every node is a documented engine call, which is what keeps the generated script honest and the sandbox safe. Need custom code? Download the .py and continue in your own environment.
• Branches and loops. A flow is a directed acyclic graph: it reads top to bottom, like the script it becomes. The canvas refuses a connection that would close a cycle.
• A model anywhere inside a run. Open answers are coded by Code open answers, which applies a codeframe's decisions by hand and its word rules — so a flow always produces the same numbers and never calls out to anything.
• A PDF writer. Reports are Markdown and HTML (with their tables in Excel on request); print or convert the HTML.

See also

• Analysis Flows
• Cleaning and Weighting Data
• Coding Open Answers
• Reports
• MaxDiff and Conjoint

---

← Analysis Flows · Studio contents · Cleaning and Weighting Data →