I wonder whether it is reliable to conduct an imputation of full scales/questionnaires in our study.
We conducted a longitudinal data collection of 3 waves. There's a time of 3-4 years between each wave (wave 2 was collected 3 years after wave 1, wave 3 was collected 3-4 years later after wave 2). However, some participants collected in wave 3 left some of the questionnaires unanswered, which means that those answers are missing.
Something important to add is that all the participants of the three waves were emerging adults, which means the participants are undergoing numerous and significant personal, social, work-related, and other life changes over the years (and wave 2 was collected during Covid lockdown). Longitudinal analyses between waves 1 and 2 have shown that the relative stability of most scales and subscales ranges from 0.3 to 0.6, which is not especially high. Additionally, some participants were not present in wave 2, meaning that the data we have for the imputation of the responses of these participants would come from only one collection. Is it methodologically reliable to impute responses to entire questionnaires based on the responses provided by those same participants in previous waves to those same questionnaires? (and given the characteristics of the developmental stage in which they are). And, if it is possible, which would be the more robust methods to do so?
It is almost certainly better to use multiple imputation (or full information maximum likelihood [FIML] estimation with missing data, which may be more straightforward than multiple imputation in your case if you are dealing with continuous variables) as compared to throwing data away (e.g., through listwise or pairwise deletion). Multiple imputation requires the less restrictive assumption of missing at random (MAR) as compared to missing completely at random (MCAR; which is required by listwise & pairwise deletion).
You may have relevant auxiliary variables at Wave 1 in your data set (i.e., variables that are correlated with dropout and/or your DVs) that you can include in your imputations and/or FIML estimation to make MAR more likely to be met. See
Enders, C. K. (2022). Applied missing data analysis. 2nd edition. New York: Guilford Press.
Graham, J. W. (2003). Adding missing-data relevant variables to FIML-based structural equation models. Structural Equation Modeling, 10, 80-100.
No, it is generally not appropriate to take the simple average of variances in time-to-event meta-analysis (e.g., survival analysis or meta-analysis of hazard ratios).
Variances represent the degree of uncertainty or variability in an estimate. In meta-analysis, instead of averaging the variances directly, it is more appropriate to use a weighted approach. Specifically, the inverse-variance weighting method is commonly employed. This method gives more weight to studies with smaller variances (i.e., those that provide more precise estimates) and less weight to studies with larger variances.
Here’s why you shouldn’t take a simple average of the variances:
The correct way is to compute a weighted average of the effect sizes (such as hazard ratios) using the inverse of their variances as weights. This approach ensures that studies with more precise estimates (lower variance) contribute more to the final pooled estimate. Then, the overall variance (and confidence interval) can be calculated based on these weights.
Longitudinal imputation can be reliable if: Missing Data Mechanism: The data is missing completely at random (MCAR) or missing at random (MAR). Method Choice: Advanced methods like multiple imputation or model-based approaches (e.g., Bayesian or mixed models) are used rather than simpler methods like LOCF. Data Trends and Time Intervals: Consistent time intervals and trends improve reliability. If these factors align with your study design, longitudinal imputation should yield robust results. Otherwise, be cautious, as results may be less stable.
- Badges
- Science topic
- Similar topics
- Psychology