Dear colleagues
I am curious as to whether it is wise to impute missing values on background variables such as gender, age, GPA etc. Normally, I would only impute missing values on the latent factors' items before running CFA/SEM.
If, however, I did this, I wonder how that would affect my sample size description since multiple imputations do not just replace a missing value with another, but rather with a pooled average over n datasets.
Thank you in advance for your advice. Any insight is greatly appreciated.
Best
Marcel
Marcel Grieger I don't see why you shouldn't impute those background variables if you can reasonably assume that the missing scores are "missing at random" (MAR) or "missing completely at random" (MCAR). Otherwise, you may loose data (potentially resulting in a loss of statistical power) and/or rely on stricter assumptions than MAR (e.g., listwise deletion leads to unbiased results only under the assumption of MCAR, which is a stronger assumption than MAR).
In the sample size description, I would recommend being transparent about (1) what the total (original) sample size was that you started out with and (2) how many missing values occurred on each variable (and make clear that the latter were imputed to obtain complete data). This information could, for example, be reported in a summary table that also contains the descriptive statistics (e.g., M and SD) for all relevant variables.
Thank you Christian Geiser for never failing to come to the rescue. I thought so too and in fact have already done so. The Mplus output clearly shows that none of the background variables have any missing values any more. That is promising.
I wonder since the missings were replaced with a pooled average, does the output provide me with an absolute number of how many people are now classed as "male" and "female" in the case of the background variable "gender" so I can contrast pre- and post-imputation sample sizes? Also, I am completely with you on being transparent about the whole process.
Thanks again!
Marcel
Yes, you should be able to obtain pooled averages (means) for all variables in the output when requesting OUTPUT: SAMPSTAT with multiple imputation. Provided your gender variable is coded with 0 and 1, the mean would indicate the proportion of either males or females (depending on the coding) after imputation.
Swell! My problem is solved. It categorised "gender" (0 = female; 1 = male) as categorical. This is what I found in the output:
UNIVARIATE PROPORTIONS FOR CATEGORICAL VARIABLES
NOTE: These are average results over 50 data sets.
SEX_1
Category 1 0.587
Category 2 0.413
My n = 447. In absolute numbers this adds up to 262 = female and 185 = male.
Before the imputation, the numbers were as follows: 259 male, 180 = female and 8 = missing.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size