assumption of order

[pengguanya]

@gowerc
It seems pool_bootstrap_percentile relies on the assumption of the order of the visits which were intrinsically determined by where they appeared in the drawed or imputated dataframes. How did it guarantee that data were initially ordered as expected for pool_bootstrap_percentile, the same for pool_bootstrap_normal

rbmi/R/pool.R

Lines 431 to 433 in a45df71

	pool_bootstrap_percentile <- function(est, conf.level, alternative) {
	est_orig <- est[1]
	est <- est[-1]

[nociale]

Hi @pengguanya

I think these functions you mention rely on the fact that the first element of est is the estimate from the original dataset (and not from a leave-one-out or a bootstrapped dataset). This is internally managed based on the user's choice of method (already at the call to draws()), and cannot be changed by the user. So, as long as there is not any internal error, the first estimate should always be referred to the estimate from the original dataset for these particular methods. Also pool_internal.jackknife() rely on the same assumption. I don't think it is necessary to check that the order of est is equivalent throughout the call to all the sequential functions from draws() to pool() (and I don't know how this could be done with the current implementation).

This is the line of code that first insert the estimate from the original dataset at the beginning of samples in draws() (which then is used to create imputations objects and est...):

if (first_sample_orig) {
        samples <- append(list(initial_sample), samples)
}

est[1] is the treatment effect estimate derived from the analysis of the original dataset which was imputed based on a MMRM fit that produced initial_sample in draws().