I have data on family care for elderly people. Data stem from 6 EU counries. People were asked at baseline and followed-up one year later. Now I'd like to find predictors that explain why people stopped caring for their elder relatives (or continued caring), or in short: to find barriers and facilitators that keep care-at-home settings stable (in 6 country comparison).

My idea was to use mixed effects models with country as random intercept. But how would I include the time comparison (i.e. all people were caring at baseline, but a certain percentage stopped caring one year later).

Would this be another random intercept? Or a random slope? Or would I include this predictor as interaction?

I'm planning to analyse my data with R and the lme4 package.


IMPORTANT EDIT: I think I've misread your question. As described, you're not interested at all in the effect of time, but rather in what other factors predictors predict if people will stop being carers (assuming that everyone at time $A$ is a carer at the time). Your independent variable in this case is just whether or not a given person is still a carer at time $B$. Time shouldn't be in your model.

Original answer

As described, time would be a fixed effect (i.e. a predictor), although if you believe that countries might differ in this regard (the effect of time varies from country to country) you might also include a random slope, which allows for this variation.

In lme4, the first model (random intercepts per country) would be, roughly

glmer(is_caring ~ time + (1|country), data=your.data, family=binomial)

while the second (allowing by-country variation in the effect of time) would be

glmer(is_caring ~ time + (time|country), data=your.data, family=binomial)

Obviously, you'll want to include your other predictors alongside time here, and consider

You might want to take a look at some of Andrew Gelman's publications (i.e. this, or this excellent book) on the topic for a good primer.

  • $\begingroup$ Thanks for your answer. I realized that my question lacks of some information, which make my thinking about time comparison (pre-post) clearer. Some of my predictors also may change over time, eg dependency of elderly, perceived burden of care... How can I model such changes over time? (I have this picture in my mind of different "slopes" for each case that indicates individual changes over time, eg elder relative became more dependent, thus one stopped caring one year later). Simply by calculating the differences of predictors at baseline and follow-up? $\endgroup$ – Daniel Nov 10 '14 at 16:26
  • $\begingroup$ I don't think your data really models anything over time - you're predicting a single outcome, specifically if a person has stopped being a carer at time $B$. As a general suggestion, you might consider calculating, for each subject, the change in each of your predictors (i.e. burden_of_care_t2 - burden_of_care_t1) and seeing how this predicts your outcome. $\endgroup$ – Eoin Nov 10 '14 at 16:27
  • $\begingroup$ "Slopes" in this context refer to the relationship between your predictors and what you're predicting, not changes in your predictors over time. A nice rule of thumb to remember is that random effects are supposed to be normally distributed samples from a larger population, like the six countries you've sampled from the possible set of all countries/all EU countries. $\endgroup$ – Eoin Nov 10 '14 at 16:33
  • $\begingroup$ Thanks a lot! Sorry for confusing edits, but I added comments with my iPhone - German spell correction and auto-suggestion, when writing in English - which was not very fast and effective. So changes over time will be included (like you suggested) as difference of predictor_t2 - predictor_t1 and I have a random-intercept model then. $\endgroup$ – Daniel Nov 10 '14 at 18:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.