Fitting mixed-effects models when data are left truncated