As said above. My main question is how to do pairwise deletion in ANCOVA. However, I have included some extra information for context if required.
I am running two-way mixed model ANCOVAs, and I want to do pairwise deletion. I have a sample of parents and children, where one case = a parent-child dyad. e.g. case 50 = one parent and child set. I have two time points - baseline and 12 months. I have deleted dyads where both parent and child have not completed the 12 months assessments. However, I do have remaining cases I have retained because the parent did the assessment, but the child didn't, and I want to use that parent's data. Therefore, i want to do pairwise deletion on the ANCOVA rather than listwise, in order to retain these parents data which did complete the assessment, but their child didn't.
Deanndra Mekail This is much more clear if you write the model in regression form rather than in ANOVA. form See the section on ANCOVA in: https://b-ok.cc/book/3591248/4df2ac . Best, David Booth
Hello Deanndra,
From your query, I don't fully understand how you're handling the data (or even what the covariate might be, if it's not baseline scores). However, I can offer my two cents' worth on the missing data issue.
You didn't indicate your total sample size or the rate of missing values for parent-child pairs. If data points appear to be MCAR (missing completely at random: no discernable association of presence/missing status with scores on any other variable in the data set, and no data collection reason to unequally result in present/missing data points) and the rate of missing values is low (some say 10%, some 5% or less), then the functional impact of listwise deletion will be minimal as regards accuracy of parameter estimates. Loss of statistical power can be a concern, however, when N is modest to begin with.
Pairwise deletion as a method for coping with missing data has several possible vulnerabilites: among these is the distinct possibility that relationships among variables will no longer be coherent, and so-called Heywood cases (which inhibit your ability to evaluate the model) can arise.
So, you may find that the need to apply pairwise deletion isn't all that strong. Even if the statistical power issue is what is driving your need to use all available data points, you might be better served by: (a) checking data for MCAR; (b) based on that, selecting and applying a different method for imputation; and (c) reporting the results with and without the imputation method you select, and letting the reader decide which result is the more credible.
Good luck with your work!
David Morse thank you for your detailed response! I have a large sample size, but did not want to lose data if it's possible to keep it. I have 288 parents, and 277 children with complete data at both timepoints. The problem is that listwise deletion reduces the sample to 265, by deleting any cases where a parent or child has missing data due to not doing the 12-month assessment.
I have already done mean imputation for some cases that had small amounts of missing data, so the only 'missing data' remaining is for cases where the parent did the assessment but their child did not. There's no problem with that data, so it should be retained for analysis...
I suppose you're saying that this isn't a big problem if I already have a large enough sample size?
- Badges
- Science topic
- Similar topics
- Medicine
- Epi
- Age Groups
- Children