0
$\begingroup$

I am using inverse probability weighting to calculate stabilized weights to try to account for the loss-to-follow-up in my cohort study of children from birth to 5 years. I used multiple imputation to impute missing data for the small percentage of observations that were missing (usually <5% for each variable) to be able to calculate the stabilized weights for "DROPOUT" in the entire population.

My outcome of interest is a neurological test children not lost to follow up took at 5 years. I will be performing logistic regression models adjusted with the stabilized weights calculated from IPW. Am I supposed to include only the children with data available for the test at 5 years in my SW weighted logistic regression? Or should I have used multiple imputation to impute the missing test results and include the entire population + IPW weights?

Thank you in advance for your response

$\endgroup$
1
$\begingroup$

Short answer: The final model (the logistic regression models) should only include those not lost-to-follow-up.

Longer answer: Setting aside the multiple imputation (MI) for other variables, let's focus on loss-to-follow-up (LTFU) exclusively. Informative LTFU (a common cause of the outcome and LTFU) can result in bias from the parameter of interest. There are two approaches we can consider to account for informative LTFU: re-weight all the values not-LTFU, or 'fill-in' those LTFU.

Inverse probability of censoring weights (IPCW) are to adjust for informative LTFU. IPCW do this by re-weighting the data, so that the variables related to LTFU are no longer related in the weighted population. Therefore, you can conduct a complete case analysis of the data with those weights. Under this approach, we don't need to 'fill-in' any of the missing data.

Multiple imputation is one approach to fill-in the values for those LTFU. Instead of re-weighting the data, we 'make up' the missing outcome data for those LTFU. We propose some model that is predictive of the outcomes, estimate the parameters for that model based on the observed data, and then fill-in the missing outcome values for those LTFU. This is another approach and relies on a different modeling strategy (i.e. we are filling in the missing values rather than re-weighting).

Under the assumption that the model for IPCW is correctly specified, there is no need to use any of the imputed outcome data for those LTFU. You can combine these two approaches (augmented-IPW) but they are done in a special way.

$\endgroup$
1
  • $\begingroup$ Thank you so much for your response! Very helpful! $\endgroup$
    – Courtney
    Aug 14 '20 at 11:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.