Is it possible to imput values using mice package, reshape and perform GEE in R?

Question

I have a longitudinal database that has more than 50% of the missing data of the MAR type.

This amount of missing values was a surprise to me because I did not foresee this in the study design, and for that reason I cannot delete those that have missing values because my sample N will be very small. however, these imputation methods are new to me, and I am facing many difficulties.

I saw some tutorials on multiple imputation through the MICE package. So the following points were not clear:

How to specify categorical variables (The categorical variables have no missing data). In my case I have two categorical variables, one with two levels (before and after) and the other with 4 levels (grp1,grp2, grp3, grp4). and eight continuous variables with missing values (D0 to D7).

Just like in the example below:

When imputing data I would like to specify that an imputation should occur through categorical variables (time and group).

I used the code below but I don't know if considered the categorical variables

 library(mice)
 imputed_Data <- mice (my.data, m = 5, maxit = 50, method = 'pmm', seed = 500)
 

After making the imputations, I would like to change from the wide format to the long format, where there would be only the columns: ID, Name, Time, group and a two columns one with the repeated mensure (D0 to D6) and other with Values. I know how to do this with DAtaFrames, but the object generated by mice is of the mids type. As in the example below:

As my data are repeated (dependent) measures, I want to perform the analysis of generalized equation estimation.

fit<- with(geeglm(value ~ Time+ group,
              data= IM.imputed_Data ,family=gaussian,id=ID,
              corstr="ar1"))


  summary(pool(fit))

Answer 1

There are two ways to do this in the mice package. First, you could use complete() to turn the imputed mids object into a dataframe containing the impute values, reshape the dataframe into long format, turn it into a new mids object with as.mids(), then fit the models with the new mids object. Second, you could place the code to reshape the data inside the with() block. I show an example of how to do that below

library(geepack)
library(mice)
#> 
#> Attaching package: 'mice'
#> The following object is masked from 'package:stats':
#> 
#>     filter
#> The following objects are masked from 'package:base':
#> 
#>     cbind, rbind
library(tidyverse)
# This data comes in long format and has no missing values, so I
# first transform it to wide and make some values missing
koch2 <- koch
koch2$id <- factor(koch2$id)
koch2$trt <- factor(koch2$trt)
koch2 <- pivot_wider(koch2,
                     names_from = day,
                     values_from = y,
                     names_prefix = "day")
set.seed(1)
koch2[rbinom(nrow(koch2),1,0.1)==1, "day3"] <- NA
koch2[rbinom(nrow(koch2),1,0.1)==1, "day7"] <- NA
koch2[rbinom(nrow(koch2),1,0.1)==1, "day10"] <- NA
koch2[rbinom(nrow(koch2),1,0.1)==1, "day14"] <- NA
# Create imputations, using predictorMatrix to specify the
# variables used in imputations
ini <- mice(koch2,
            maxit = 0)
pred <- ini$predictorMatrix
pred[,"id"] <- 0
imp_dat <- mice(koch2,
                printFlag = FALSE,
                predictorMatrix = pred)
# fit the models to the imputed data
mods <- with(imp_dat,
             {
               dat <- data.frame(trt = trt,
                                 id = id,
                                 day3 = day3,
                                 day7 = day7,
                                 day10 = day10,
                                 day14 = day14)
               dat_long <- pivot_longer(dat,
                                        cols = contains("day"),
                                        names_to = "time",
                                        values_to = "y")
               geeglm(y ~ time + trt,
                      id = id,
                      data = dat_long)
             })
summary(pool(mods))
#>          term   estimate  std.error  statistic       df      p.value
#> 1 (Intercept)  2.0861111 0.09938558 20.9900775 259.9958 0.000000e+00
#> 2   timeday14 -0.3055556 0.09866652 -3.0968514 264.0978 2.166670e-03
#> 3    timeday3  0.4500000 0.07020265  6.4100143 118.7481 3.066805e-09
#> 4    timeday7  0.0500000 0.08668541  0.5767983 196.6644 5.647353e-01
#> 5        trt1 -0.3444444 0.10217856 -3.3710051 250.4968 8.672982e-04

If there are many variables in the imputed data, you can select all of them using dat <- data.frame(mget(ls())). If you only want some of the variables, the ls() could be replaced with a vector of character names of variables, which could be generated programmatically.

To use the complete() method I initially mentioned you need to set include = TRUE and also recalculate the .id column created by mice. Below is an example of how to do that.

all_dats <- complete(imp_dat, action = "long", include = TRUE)
working_dats <- list()
for(i in 0:max(all_dats$.imp)) {
  working_dats[[i+1]] <- 
    all_dats %>%
    subset(.imp == i) %>%
    pivot_longer(cols = contains("day"),
                 names_to = "time",
                 values_to = "y") %>%
    mutate(.id = 1:nrow(.))
}
imputed_long <- as.mids(do.call(rbind, working_dats))
mods_complete <- with(imputed_long,
                      geeglm(y ~ time + trt,
                             id = id))
summary(pool(mods_complete))
#>          term   estimate  std.error  statistic       df      p.value
#> 1 (Intercept)  2.0861111 0.09938558 20.9900775 259.9958 0.000000e+00
#> 2   timeday14 -0.3055556 0.09866652 -3.0968514 264.0978 2.166670e-03
#> 3    timeday3  0.4500000 0.07020265  6.4100143 118.7481 3.066805e-09
#> 4    timeday7  0.0500000 0.08668541  0.5767983 196.6644 5.647353e-01
#> 5        trt1 -0.3444444 0.10217856 -3.3710051 250.4968 8.672982e-04

Answer 2

You could use tidyr::gather() to transform your data into long format and then use the package jomo for multiple imputation accounting for the nested structure of the data.

Answer 3

Have you tried the following:

fit<- with(imputed_Data, geeglm(value ~ Time+ group,
              family=gaussian,id=ID,
              corstr="ar1"))


summary(pool(fit))