-
Notifications
You must be signed in to change notification settings - Fork 9
What's Changed?
In a word: everything
Nearly every line of code is new.
New, however, doesn't mean throwing out the old – your previous code should still work.
So, what's changed?
The new {EGAnet} is fast. Nearly all functions were optimized, resulting in 2-15x faster code than previous versions. Analyses that used to take minutes now take seconds.
The biggest difference? Fast computation of tetrachoric and polychoric correlations
These correlations are at the core of most {EGAnet} functions. Unfortunately, there aren't many implementations in R that quickly compute these correlations. So, we've implemented our own tetrachoric and polychoric correlations in C.
Here's a comparison between {EGAnet}'s auto.correlate, its former implementation of {qgraph}'s cor_auto, {psych}'s polychoric, and recently published and fast implementation in R: {Turbofuns}'s PolychoricRM (using 1 core). This demonstration used the bfi dataset in {psych} with 100 repetitions:
Unit: milliseconds
expr min lq mean median uq max neval cld
EGAnet 75.1616 75.4386 77.57433 75.74405 78.19075 201.0211 500 a
Turbofuns 79.5288 79.7113 80.45382 79.87415 80.30210 97.3525 500 a
qgraph 437.7368 449.0530 490.07021 456.72130 560.37485 743.6972 500 b
psych 2555.8532 2586.6863 2701.75927 2603.60790 2727.74910 4682.7300 500 c
auto.correlate's median time was on par with PolychoricRM. auto.correlate includes several checks to compute Pearson's correlations for continuous data and polyserial correlations for continuous with ordinal data similar to cor_auto; however, auto.correlate is 6x faster. Compared to polychoric, auto.correlate is 33x faster.
To put the speed difference in context: auto.correlate completes one full bootEGA and gets 100 iterations into a second bootEGA by the time cor_auto completes its first 100 iterations in its first bootEGA.
Functions are more modular and build on each other to allow for combinations that were previously unthinkable.
Want to use {igraph}'s cluster_spinglass in EGA? You can do that
Want to see itemStability of your hierEGA? You can do that
Want to perform a bootEGA with EGA.fit using BGGM and consensus clustering Louvain (community.consensus)? That's a lot... but you can do that too
Such flexibility warrants a disclaimer: Just because you can, doesn't mean you should! Keep best practices in mind.
Functions that had S3 print, summary, and plot methods have been updated. Functions that didn't have print, summary, and plot methods now do.
print and summary methods are designed to give you the information you want without hassle. Here's EGA's summary output:
Model: GLASSO (EBIC with gamma = 0.5)
Correlations: auto
Lambda: 0.0764652282008741 (n = 100, ratio = 0.1)
Number of nodes: 25
Number of edges: 117
Edge density: 0.390
Non-zero edge weights:
M SD Min Max
0.046 0.119 -0.269 0.548
----
Algorithm: Walktrap
Number of communities: 5
A1 A2 A3 A4 A5 C1 C2 C3 C4 C5 E1 E2 E3 E4 E5 N1 N2 N3 N4 N5 O1 O2 O3 O4 O5
1 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 4 5 5 5 5 5
----
Unidimensional Method: Louvain
Unidimensional: No
----
TEFI: -27.335
Model and algorithm information, network descriptives, community memberships, and fit (Total Entropy Fit Index) are provided with every *EGA function. This information is everything (and perhaps more!) than you need to write your reproducible Methods section.
What about plot? There's so much flexibility there that it needs its own Wiki
Like polychoric correlations, we've implemented high quality open-source pseudorandom number generation (PRNG) in C instead of R. This change avoids potential conflicts with the seed you set in your script and the seed you set in an {EGAnet} PRNG function.
In short, your results can be replicated, time and time again.
For the gritty details, head over to Reproducibility and PRNG