Two quick additions. Another coding option I have seen:      Random int / subject = countries    Random int / subject = couples(countries)   Also, regardless of which coding option I use, there is one independent variable that is so highly associated with the binary outcome that the covariance parameter estimate for all random effects is zero. The intercept and fixed effect estimates do vary depending on the code I use. If the covariance parameters estimates are zero, does that suggest that nesting is not needed or does it mean there may be a broader problem with the model?  ... View more