How drop out cases are handled in mixed effects models

[lpThao]

I am conducting a longitudinal analysis using mixed effect models to estimate the evolution of some biomarker over time. As my cohort consists people who are older than 70 years old, drop-out due to death is an issue I should take into account, I think. I wonder how mixed effects model handling drop-out. I was looking into joint modelling for longitudinal outcome and survival outcome, but my impression is that the coefficient estimates of the mixed model part in the joint model does not change much. What are your thoughts on this?

[jroon]

Hi @lpThao,

@lpThao

I am conducting a longitudinal analysis using mixed effect models to estimate the evolution of some biomarker over time. As my cohort consists people who are older than 70 years old, drop-out due to death is an issue I should take into account, I think. I wonder how mixed effects model handling drop-out. I was looking into joint modelling for longitudinal outcome and survival outcome, but my impression is that the coefficient estimates of the mixed model part in the joint model does not change much. What are your thoughts on this?

It is hard to give you definite opinions as you didn’t give us many details to go on (of course we understand if you cannot share all the details). For instance, what kind of JM using which software? Also, how long are you folllowing them up for and what % of the cohort died in that time? You say that you are studying the evolution of the biomarker over time - do I take it then that you are not interested in death as an outcome / the biomarker as a predictor of death?

[f2harrell]

My approach may not be popular with others but I suggest

• fit the usual use-all-available-data longitudinal model (random effects + serial correlation pattern) that ignores death
• fit a similar model but using a semiparametric ordinal model (e.g. proportional odds with random effects and optimally also with a serial correlation pattern)
• repeat the second model counting death within a time interval as the worst biomarker value
• compare 2 and 3

[lpThao]

Thank you @jroon. The average follow-up time is about 5 years. Measurement was taken annually. Drop-out is about 1%. You are correct that I am mainly interested in the longitudinal outcome, not the survival outcome. I used JMbayes package in R to fit the jointmodel.

[jroon]

@lpThao

Thank you @jroon. The average follow-up time is about 5 years. Measurement was taken annually. Drop-out is about 1%. You are correct that I am mainly interested in the longitudinal outcome, not the survival outcome. I used JMbayes package in R to fit the jointmodel.

Dropout of 1% is very low. At that level of dropout I wouldn’t expect a joint model to alter the longitudinal parameter estimates to any meaningful extent!

@f2harrell

My approach may not be popular with others but I suggest

• fit the usual use-all-available-data longitudinal model (random effects + serial correlation pattern) that ignores death
• fit a similar model but using a semiparametric ordinal model (e.g. proportional odds with random effects and optimally also with a serial correlation pattern)
• repeat the second model counting death within a time interval as the worst biomarker value
• compare 2 and 3

I’m having trouble to conceptualise this - are there any worked examples in your written material ?

[f2harrell]

No other than a very brief example in BBR where a mock-up of an analysis of serum creatinine is given, where dialysis overrides SCr with a worse valuye of SCr.