CODEWITHSUNDEEP
rregressionmissing-dataanalysisr-mice
Lea
Lea

Reputation: 85

Probe Interaction with multiple imputed Data


I would like to probe an interaction in a moderation analysis between a categorical predictor and a continuous variable in a multiple imputed data set. I would like to look at the effect of different levels of the continuous moderator (-1SD, Mean, +1SD).

I know that this is possible in SPSS with the PROCESS macro, but unfortunately, it cannot handle multiple imputed data. The interaction package can't handle it either. Does anyone have any ideas on how to solve this problem?

I have imputed the data with the mice package, with m = 50.

My model looks like this:

lm(outcome ~ group * continuous_moderator)

Here is a reproducible example, using the windspeed dataset integrated in mice:

data <- mice::windspeed
data$group <- sample(c(0:1), replace = T, size = 433)

imp_data <- mice(data,
                 m=50,
                 maxit=20,
                 meth='cart',
                 seed=12345)

model1 <- with(imp_data,
               lm(RochePt ~ group * Clones))

Thanks in advance for all your help and tips.

Upvotes: 0

Views: 371

Answers (1)

Vincent
Vincent

Reputation: 17805

You can do this using [the marginaleffects package.] (Disclaimer: I am the author.) See this vignette for details:

https://vincentarelbundock.github.io/marginaleffects/articles/multiple_imputation.html

The quick illustration below uses the by argument which is available in the dev version (should be out on CRAN in the next couple weeks). Install it:

library(remotes)
install_github("vincentarelbundock/marginaleffects")

Then,

library(mice)
library(marginaleffects)

data <- mice::windspeed

# {marginaleffects} reserves the name "group", so I "group_id" instead
data$group_id <- sample(c(0:1), replace = T, size = 433)

fit <- function(x) {
    mod <- lm(RochePt ~ group_id * Clones, data = x)
    out <- comparisons(mod, variables = "Clones", by = "group_id")
    # see vignette linked above for why we need the next line.
    out$term <- out$group_id
    return(out)
}

imp_data <- mice(data,
    m = 5,
    maxit = 20,
    meth = "cart",
    seed = 12345,
    printFlag = FALSE)
imp_data <- complete(imp_data, "all")

mod <- lapply(imp_data, fit)

mice::pool(mod)
#> Class: mipo    m = 5 
#>   term m  estimate        ubar b           t dfcom       df riv lambda
#> 1    1 5 0.9394599 0.002593162 0 0.002593162   433 430.9702   0      0
#> 2    0 5 1.0443336 0.003134239 0 0.003134239   433 430.9702   0      0
#>           fmi
#> 1 0.004608611
#> 2 0.004608611

Edit: Here’s a minimal example using the newdata argument and the datagrid() function to assess contrasts conditional on a continuous moderator. You can then modify the fit function I gave you before to integrate this in your multiple imputation workflow.

Please read the documentation to learn more. It is very detailed and the functions are super flexible: https://vincentarelbundock.github.io/marginaleffects/

library(marginaleffects)

mod <- lm(mpg ~ hp * factor(cyl), data = mtcars)

comparisons(
    mod,
    variables = "cyl",
    newdata = datagrid(hp = c(100, 110, 120)))
#>   rowid     type        term contrast_cyl comparison std.error
#> 1     1 response interaction        6 - 4  -4.792913  2.016159
#> 2     2 response interaction        6 - 4  -3.741286  2.025602
#> 3     3 response interaction        6 - 4  -2.689660  2.253712
#> 4     1 response interaction        8 - 4  -8.049775  2.314779
#> 5     2 response interaction        8 - 4  -7.064458  2.393402
#> 6     3 response interaction        8 - 4  -6.079140  2.563475
#>   statistic      p.value   conf.low  conf.high predicted_lo
#> 1 -2.377250 0.0174422739  -8.744511 -0.8413143     24.70544
#> 2 -1.847000 0.0647471270  -7.711393  0.2288198     23.57768
#> 3 -1.193436 0.2326988568  -7.106855  1.7275346     22.44992
#> 4 -3.477557 0.0005060048 -12.586658 -3.5128926     24.70544
#> 5 -2.951639 0.0031609267 -11.755439 -2.3734759     23.57768
#> 6 -2.371445 0.0177186636 -11.103458 -1.0548222     22.44992
#>   predicted_hi cyl  hp
#> 1     19.91252   8 100
#> 2     19.83639   8 110
#> 3     19.76026   8 120
#> 4     16.65566   8 100
#> 5     16.51322   8 110
#> 6     16.37078   8 120

Upvotes: 1

Related Questions