Hello,
I am currently working on a project involving secondary data analysis from a survey that used a complex sampling design. Weights are provided to correct for stratification, clustering, and attrition. As this is a longitudinal study, I have to apply the appropriate weight for the last wave of data collection that I am using. However, several thousand participants are missing data on the attrition-adjusted weighting variable. I am using Mplus, and from what I understand, Mplus requires cases with missing data on weighting variables to be deleted.
Is there an alternative approach to address the missing data on the weighting variable and ensure a representative sample without excluded all these participants from analyses? Or, is there no other option? Also, will excluding these individuals affect the accuracy of the weights, or will the final sample still be representative, despite the reduced sample size?
Any suggestions would be greatly appreciated.
Take a look at the last post by Tihomir Asparouhov in the following Mplus discussion thread:
http://www.statmodel.com/discussion/messages/22/3495.html?1541263309
Perhaps multiple imputation prior to constructing the weight variable is an option for you. It would certainly be undesirable to delete several thousand cases.
Thank you for the quick response Christian. I will look into the options he suggested in his post.
Jourdan Wilson It might be difficult to deal with missing data on weighting variables in a sophisticated survey design. Multiple imputation (MI) is one method for estimating missing values in Mplus when there is missing data on weighting variables. MI is a statistical strategy that uses observable data to produce numerous plausible values for missing data. After that, the imputed values may be used to estimate the parameters of interest while accounting for the uncertainty provided by missing data.
Another option is to employ Mplus's "full information maximum likelihood" (FIML) estimate method, which can accommodate missing data on the weighting variable without deleting instances. FIML estimation estimates the parameters of interest using all available information.
If you choose to utilize the MI or FIML methodology, make sure to evaluate the assumptions of these approaches and report the results with prudence.
Excluding individuals, with incomplete data on weighing factors might impair weight accuracy and result in a non-representative sample. However, if the missing data is not random, the final sample may still be representative despite the smaller sample size.
It is also necessary to consider the weighting variable's proportion of missing data. The impact on the final sample may be modest if the missing data is a tiny fraction of the whole sample. However, if the missing data constitutes a considerable proportion of the entire sample, the influence on the final sample may be greater.
Finally, missing data on weighting factors can be addressed using multiple imputations or the Mplus "full information maximum likelihood" (FIML) estimate approach to assure a representative sample without eliminating all participants from the analysis. It is critical to thoroughly analyze the method's assumptions and to publish the results with caution.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis