Multilevel Longitudinal Models and Correlated Errors

Dear Talk Stats

I have done some analysis modelling national GDP, national Gini coefficients, and mortality rates for low and middle income countries over time from 1990 to 2016, to assess how income inequality and GDP are associated with mortality in different age groups. This is the model I have used (i have simplified it a bit for illustration)

xtmixed mortalityrate_log year gdp_log gini || country_name: year, cov(un) variance mle

One of the criticisms I have had is that this model incorporates uncorrelated errors, yet nearly all of the variables used in this study exhibit very strong temporal autocorrelation. i.e. GDP and mortality in Brazil in 2003 will be correlated to GDP and mortality in Brazil in 2004, and should not be treated as independent cases.

My understanding is that this model includes a random effect for country_name, which will then take in to account this within country autocorrelation over time. Am I correct in this?

Best Wishes



Less is more. Stay pure. Stay poor.
Is the critique which variance/covariance structure you used? You can try different ones then see how much it changes fit measures, but your decision should also be based on your content knowledge of the topic.

Controlling for random effects will account for additional variability, but I am unsure in time series if it is going to pick up the right approximate structure if it is not provide.
Last edited: