Model development: using D1, 2 or 3, versus significant p-values in pooled results

[stuartstewart]

Hi,

I have a dataset containing ~160,000 patients with kidney disease and about 50 variables (demographics, clinical details) and have run linear and logistic regression models to assess the impact of different predictors on outcomes.

I have read the following chapters by SVB: https://stefvanbuuren.name/fimd/sec-stepwise.html and https://stefvanbuuren.name/fimd/sec-multiparameter.html.

Now assuming I have understood those two chapters, I wanted to know when selecting the final model - is it methodologically sound to use a stepwise process and remove predictors that have non-significant estimates from the pooled results, instead of testing each predictor prior to pooling (using e.g. D1, D2 or D3)?

Many thanks in advance.

Stuart

[gerkovink]

Have a look at this link. It details exactly what you want.

[stuartstewart]

Hi,

This looks like it could work - only problem being that I had to reshape my data post-imputation and put the data in a long format following these general instructions:

long workflow using a dplyr list-column

library(dplyr)
est7 <- nhanes %>%
mice(seed = 123, print = FALSE) %>%
mice::complete("long") %>%
group_by(.imp) %>%
do(model = lm(formula = chl ~ age + bmi + hyp, data = .)) %>%
as.list() %>%
.[[-1]]

The solution you recommended, where fit <- with(imp, expr), requires the model to be run on 'imp' which is class mids. However my data is now in a long list format.

Is there a modification of your solution for my data structure?

Thanks

Stuart

[gerkovink]

See mice::as.mids().

[stuartstewart]

Brilliant - this is what I was looking for. Thanks a lot