In timdisher/neoDecision: Create slides for neonatal decision modeling talk

Generate Pseudo Data

In the final section of this presentation we will walk through a proposal that pretty much all trial data can and should be turned into portable pseudo data that is shared freely without restrictions. We has originally planned to show two versions of this: one using {synthpop} and the other using {mvProbit}. After reviewing the synthpop documentation more closely however it looks like you can actually just save the model objects which means you can interact with them using local data. This is much faster/more general than the multivariate probit approach so we won't move that work forward any further. Note that this repository also includes the code necessary to do a bayesian version of the multivariate probit model which is primarily of interest if you'd like to do something like specify treatment effects across binomial outcomes as shared/exchangeable.

Generate data from synthpop

We will use vanilla default settings for synthpop since the data here is simulated but in real applications there is a lot more fine tuning available. We can get a nice simple default output compare synthetic to original data using the compare function.

data("ex_dat")

dat <- as.data.frame(ex_dat) %>% dplyr::select(trt, mort, sev_ivh, sepsis, cld, nec)

set.seed(124)
synth_dat <- xfun::cache_rds({
  syn(dat, minnumlevels = 5, m = 10, models = FALSE, method = "parametric")
}, dir = cache,  file = "synth_dat.rds")  

compare(synth_dat, data = dat)

We can also take a quick look at correlations for a pooled version of the synthetic datasets vs the observed data.

synth_dat$syn %>%
  do.call(rbind, .) %>% 
  mutate_all(as.numeric) %>% select(-trt) %>% cor()

dat[,-1] %>% cor()

For the rest of the presentation we will just use on example dataset. For the comparison at the end we will create a composite with sev_ivh + cld

pres_dat_syn <- synth_dat$syn[[10]] %>% 
  mutate_all(~ as.numeric(as.character(.))) %>% 
  mutate(cld_ivh = ifelse(sev_ivh == 1 & cld  == 1, 1, 0),
         cld_no_ivh = ifelse(cld == 1 & sev_ivh == 0, 1, 0),
         ivh_no_cld = ifelse(cld == 0 & sev_ivh == 1, 1, 0))

pres_dat <- dat %>%
  mutate(cld_ivh = ifelse(sev_ivh == 1 & cld  == 1, 1, 0),
         cld_no_ivh = ifelse(cld == 1 & sev_ivh == 0, 1, 0),
         ivh_no_cld = ifelse(cld == 0 & sev_ivh == 1, 1, 0))

colMeans(pres_dat)
colMeans(pres_dat_syn)