I am running analyses on a dataset with missing values. Running Multiple Imputation, MI, in SPSS is easy enough and quite handy My dilemma comes with imputing data on one variable that shows a conditional effect upon my main independent variable of interest (substantive model post imputation). It seems to me that the value in running MI, filling in missing data while producing accurate standard errors in regression models, is lost here. Is it correct to say that MI is appropriate for independent variables, dependent variable, and covariates only? But not for variables showing an interaction effect (conditional effect)?
One variable (binary) shows an interaction with my independent variable. In the original data there is a strong effect in one subgroup with no effect in the other. This variable is missing 18% of its values, so imputing these results in considerable misclassification. For inclusion as a simple covariate, this misclassification would be addressed and the standard errors for the regression coefficient would be accurate.
When pursing analyses to illustrate this interaction term, my effects seen in the original complete data now are washed out. In the subgroup that had showed a strong, clear effect, it now is not strong and no longer statistically significant. I think that is due to the misclassification, not due to having a fuller more accurate dataset.
Your thoughts? Any references? Need clarification?
No. If imputed values are good for one variable in a regression they are good for all. We don't get to pick and choose.
Btw it looks like you might be willing to do almost anything to get a p-value less than 0.05 If so, You are coming close to dishonesty. BE CAREFUL
Interaction terms are 'created'. Imputed terms are 'created'.
I would advise against compounding the creativity in your dataset and look for alternatives to testing moderation perhaps (for ex - full-model comparisons with equality constraints on/off the focal effect to-be moderated.).
Hope this helps,
Matt
As I understand it the recommendation is to include the interaction term in the imputation model - at least for linear regression models. X1*X2 is a predictor that can be imputed just like X1 and X2. Apparently imputing X1 and X2 and creating the product term after imputation can bias things.
e.g., see https://thestatsgeek.com/2014/05/10/multiple-imputation-with-interactions-and-non-linear-terms/
The quality of imputation also depends on having variables that predict missing ness - so its important to include auxiliary variables that might predict missingness.
@Thom. Don't disagree with your reply but how else would you do such a piece of research as was described in the question?. Am I missing something here?
If the missing-data pattern is monotone, one can use stata command 'mi impute monotone' to impute missing values of 2 variables, say, X1 and X2 simultaneously
Not sure I understand your Q David. I agreed with your answer that imputation applies to interactions as well as other predictors. Just noting that there are different ways to impute that might impact the outcome.
- Badges
- Science topic