Repeated measures with non-random missing data (aka death)

We have submitted a large paper in a high impact journal within basic/translational cardiology, and we received positive feedback with the opportunity for a revision. The statistical reviewer had some minor comments, most of which we will handle easily. However, one tricky question remains and I hope to get some ideas in this forum:

We have a large cohort of mice (about 150) that underwent transaortic constriction, and were subsequently followed for 20 weeks. The mice were randomized into six different treatment groups. We then performed serial measurements of cardiac contractility using echocardiography and MRi at baseline, 4w, 8w, 12w, 16w, and 20w. In the original submission, we did area under the curve for all groups compared using one-way ANOVA with Sidak’s post hoc test.

The reviewer suggested that we should use two-way repeated measures ANOVA instead. However, some mice died during the study, so we have missing data. Only including the surviving mice would introduce a huge bias. We though about using mixed model approach, but this requires random missing data, and when a mouse dies, all data points are missing after the time of death, so the data is obviously not missing at random.

Any ideas on how to resolve this?


TS Contributor
AFAIK the question for MAR is whether missingness is associated with cardiac contractility.
Whether those with missing data at time point tx would have shown cardiac contractility different
from those without missing data.

With knd regards



Active Member
yes, death is weird as a source of missing data. To me it seems that the application of missing data methods, be they mixed models or otherwise, would result in an adjusted estimate that attempts to impute what the deceased animals would have been if alive. But that is kind of questionable science, in my view.

I would say mixed model is ok.