How to run repeated-measures ANOVA when data points per subject vary?

Question

I am attempting to run statistics in the following scenario:

I have a list of hospital visits by approx. 60 patients to the emergency room. Each patient gets an EKG done of their heart, and I have four measurements from each EKG. Then, each person gets their potassium level taken. That makes five variables: V2, V3, V4, QRSD, and K (short for potassium).

I would like to test the hypothesis that potassium level affects V2, V3, V4, and/or QSRD. That is, I'd like to run a test that, for instance, could detect that if potassium goes up, so does the QSRD etc.

I am not sure if ANOVA is correct here, or if regression is... Anyway, one of the big issues is that repeated-measures tests calls for a fixed number of measurements per participant... But in this case, each patient has a variable number of visits. Patient A might have 3 visits in the data set, and thus 3 sets of data. Meanwhile, patient B might have 7 visits and thus 7 sets of data.

Any guidance here?

Answer 1

I was looking into something similar recently and stumbled upon these papers:

• About "situations where the observed sample consists of a combination of correlated and uncorrelated data due to missing responses (partially correlated data).": Samawi, H.M., Yu, L., & Vogel, R. (2015). On some nonparametric tests for partially observed correlated data: Proposing new tests. Journal of Statistical Theory and Applications, 14(2), 131-155, available at http://www.atlantis-press.com/php/download_paper.php?id=23226
• Editorial that mentions repeated measures data with data points missing completely at random (MCAR) and missing at random (MAR): Samawi, H.M., & Vogel, R. (2015). On inference of partially correlated data. Biometrics & Biostatistics International Journal, 1 (2), available at http://medcraveonline.com/BBIJ/BBIJ-02-00019.pdf
• "Partially correlated data are the result of a matched pair which is missing one of the two correlated values or when a subject in a repeated measures design is missing at least one of their repeated observations, but not all observations.": Hani M. Samawi & Robert Vogel (2014) Notes on two sample tests for partially correlated (paired) data, Journal of Applied Statistics, 41:1, 109-117, DOI: 10.1080/02664763.2013.830285

They didn't fit my case, so I didn't study them thoroughly enough to explain them properly. But hopefully these papers will bring you closer to your answer!