Weighted tables?

[tylcole]

Would it be difficult to incorporating weighting into this package? This would be incredibly helpful for both administrative datasets as well as propensity score matching reporting. Thanks!

[ewenharrison]

There are lots of different glm() options that could be added. Poisson/weights/propensity score as you rightly say. Rather than make things more complicated, ff_merge() allows any table to be created. Here are some examples of going from a simple logistic regression model, to a weighted model.

Standard approach:

explanatory = c("age.factor", "age", "sex.factor", "nodes", "obstruct.factor", "perfor.factor")
dependent = 'mort_5yr'
colon_s %>%
	finalfit(dependent, explanatory)

Build each column separately approach:

colon_s %>% 
	summary_factorlist(dependent, explanatory, fit_id=TRUE) %>% 
	ff_merge(
		glmuni(colon_s, dependent, explanatory) %>% 
			fit2df(estimate_suffix = " (univariable)")
	) %>% 
	ff_merge(
		glmmulti(colon_s, dependent, explanatory) %>%
			fit2df(estimate_suffix = " (multivariable)")
	) %>% 
	dplyr::select(-fit_id, -index) %>% 
	dependent_label(colon_s, dependent)

Incorporate more complex model:

colon_s %>% 
	summary_factorlist(dependent, explanatory, fit_id=TRUE) %>% 
	ff_merge(
		glmuni(colon_s, dependent, explanatory) %>% 
			fit2df(estimate_suffix = " (univariable)")
	) %>% 
	ff_merge(
		glm(ff_formula(dependent, explanatory), data=colon_s, 
				family = "binomial", weights = NULL) %>% 
			fit2df(estimate_suffix = " (multivariable)")
	) %>% 
	dplyr::select(-fit_id, -index) %>% 
	dependent_label(colon_s, dependent)

Let me know what you think.

[tylcole]

That's great for multivariate models, thanks!

What do you think about weighted univariate stats? Would that require using different functions within the glmuni() function?

[ewenharrison]

lmuni(), lmmulti(), glmuni() and glmmulti() now supported weights.

explanatory = c("age.factor", "age", "sex.factor", "nodes", "obstruct.factor", "perfor.factor")
dependent = 'mort_5yr'
wgts = runif(dim(colon_s)[1], 0, 1)

library(finalfit)
colon_s %>% 
	summary_factorlist(dependent, explanatory, fit_id=TRUE) %>% 
	ff_merge(
		glmuni(colon_s, dependent, explanatory, weights = wgts) %>% 
			fit2df(estimate_suffix = " (univariable)")
	) %>% 
	ff_merge(
		glm(ff_formula(dependent, explanatory), data=colon_s, 
				family = "binomial", weights = wgts) %>% 
			fit2df(estimate_suffix = " (multivariable)")
	) %>% 
	dplyr::select(-fit_id, -index) %>% 
	dependent_label(colon_s, dependent)

[tylcole]

That's incredible! Also supporting strata and clusters (as in function svydesign) would be amazing.

…

On Mon, Mar 25, 2019 at 07:40 ewenharrison ***@***.***> wrote: Closed #13 <#13>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#13 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AL7HSB7jMXLIy0Rt446LtbkjlnitSaQuks5vaN_fgaJpZM4aI4lV> .

[ewenharrison]

Thanks.
Could you provide an example of what you suggest.

[tylcole]

Certainly, an example is in here: https://www.hcup-us.ahrq.gov/reports/methods/2017-01.pdf In the US, this is a commonly used set of databases. From page 10 of the pdf: install.packages("survey"); library(survey) # Specify the sampling design with sampling weights DISCWT, # hospital clusters HOSP_NRD, and stratification NRD_STRATUM strnrddsgn <- svydesign(ids = ~HOSP_NRD, weights = ~DISCWT, strata = ~NRD_STRATUM, data = core)

…

On Wed, Mar 27, 2019 at 4:26 AM ewenharrison ***@***.***> wrote: Closed #13 <#13>. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#13 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AL7HSLg2SO3q1wCT9jnzAy27yTxqpPKiks5va1VJgaJpZM4aI4lV> .

[ewenharrison]

Hi,

I don't know too much about this area. I've added these functions and would appreciate if you could let me know if this is what you were thinking of. Also, are the examples correct.

Many thanks for your help with this.

Ewen

library(dplyr)
library(survey)

# Examples from survey::svyglm() help page

data(api)

# Label data frame
apistrat = apistrat %>% 
  mutate(
    api00 = ff_label(api00, "API in 2000 (api00)"),
    ell = ff_label(ell, "English language learners (percent)(ell)"),
    meals = ff_label(meals, "Meals eligible (percent)(meals)"),
    mobility = ff_label(mobility, "First year at the school (percent)(mobility)"),
    sch.wide = ff_label(sch.wide, "School-wide target met (sch.wide)")
    )
		
# Linear example
dependent = "api00"
explanatory = c("ell", "meals", "mobility")

# Stratified design
dstrat = svydesign(id=~1,strata=~stype, weights=~pw, data=apistrat, fpc=~fpc)

# Univariable fit
fit_uni = dstrat %>%
    svyglmuni(dependent, explanatory) %>%
    fit2df(estimate_suffix = " (univariable)")

# Multivariable fit
fit_multi = dstrat %>%
    svyglmmulti(dependent, explanatory) %>%
    fit2df(estimate_suffix = " (multivariable)")
	
# Pipe together
apistrat %>%
    summary_factorlist(dependent, explanatory, fit_id = TRUE) %>%
    ff_merge(fit_uni) %>% 
    ff_merge(fit_multi) %>% 
    select(-fit_id, -index) %>%
    dependent_label(apistrat, dependent)

#               Dependent: API in 2000 (api00)             Mean (sd)    Coefficient (univariable)  Coefficient (multivariable)
#     English language learners (percent)(ell)  [0,84] 652.8 (121.0) -3.73 (-4.35--3.11, p<0.001)  -0.48 (-1.25-0.29, p=0.222)
#              Meals eligible (percent)(meals) [0,100] 652.8 (121.0) -3.38 (-3.71--3.05, p<0.001) -3.14 (-3.70--2.59, p<0.001)
# First year at the school (percent)(mobility)  [1,99] 652.8 (121.0)  -1.43 (-3.30-0.44, p=0.137)   0.23 (-0.54-1.00, p=0.567)

# Binomial example
## Note model family needs specified and exponentiation if desired
dependent = "sch.wide"
explanatory = c("ell", "meals", "mobility")

# Univariable fit
fit_uni = dstrat %>%
    svyglmuni(dependent, explanatory, family = "quasibinomial") %>%
    fit2df(exp = TRUE, estimate_name = "OR", estimate_suffix = " (univariable)")

# Multivariable fit
fit_multi = dstrat %>%
    svyglmmulti(dependent, explanatory, family = "quasibinomial") %>%
    fit2df(exp = TRUE, estimate_name = "OR", estimate_suffix = " (multivariable)")

# Pipe together
apistrat %>%
    summary_factorlist(dependent, explanatory, fit_id = TRUE) %>%
    ff_merge(fit_uni) %>% 
    ff_merge(fit_multi) %>% 
    select(-fit_id, -index) %>%
    dependent_label(apistrat, dependent)

# Dependent: School-wide target met (sch.wide)                    No         Yes          OR (univariable)        OR (multivariable)
#     English language learners (percent)(ell) Mean (SD) 22.5 (19.3) 20.5 (20.0) 1.00 (0.98-1.01, p=0.715) 1.00 (0.97-1.02, p=0.851)
#              Meals eligible (percent)(meals) Mean (SD) 46.0 (29.1) 44.7 (29.0) 1.00 (0.99-1.01, p=0.968) 1.00 (0.98-1.01, p=0.732)
# First year at the school (percent)(mobility) Mean (SD)  13.9 (8.6) 17.2 (13.0) 1.06 (1.00-1.12, p=0.049) 1.06 (1.00-1.13, p=0.058)

[tylcole]

That's fantastic, I should have time in the next week to go through this and confirm. I will also test on my own data. I'll respond as soon as I do. Thank you for this great update!

[ewenharrison]

Thanks. One issue is that the summary statistics from summary_factorlist() are not weighted. They cannot be weighted using this function. So to be useful it might be necessary to incorporate svymean() or similar.

What would be useful is to know how you would actually use this. What your final table would actually look like. What variations would be common in practice etc. Best wishes.

[tylcole]

Ewen I just took a look and that looks correct, though like you mention a key feature that would be really nice is the weighted means and SD in the table using svymean(). The final table would be just like you are demonstrating, except weighted mean/SD.

This is a basic example of what those tables would look like in a paper: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5624560/pdf/cureus-0009-00000001536.pdf