BLUP is failing following xtmixed

[Taya Collyer]

Hi Folks,

I have a longitudinal dataset, which includes ~1800 total observations across 30 individuals. I am using xtmixed to understand whether the group means calculated across three time epochs are different (0- 11 years, 11-17y, 18y+). When I ran the below, to just compare times under 18y to above 18y I had no issues running predict to store fitted values (timedummy=1 for measures where the individual is over 18) and was able to produce lovely plots including both the fixed and random components.

xtmixed measure i.timedummy || record_id: i.timedummy, covarance(unstructured)

but when I switch to the below factor with three levels (stage = 0 in childhood, 1 for teenage years, 2 in adulthood) I can use predict to store the fixed effects but not the random effects. I get the perplexing error "BLUP calculation failed; estimation data have changed".

I have tried varying the covariance structure but this did not help.

I am sure I am overlooking something obvious... Any advice received with thanks.

Taya

[Joseph Coveney]

I don't know what the "estimation data have changed" is all about, but does each of your record IDs have adequate observations at all three stages (0–11, 12–17 and 18+ years)?

Perhaps each patient has enough observations on either side of 18-year divide, but observations are too sparse in the childhood years to fit a random slope for that stage.

[Taya Collyer]

Thanks for this good question. I've discovered that there is one individual who only has observations in 2/3 time periods, but when I re-ran excluding them I got the same error. (I used preserve and dropped the ID to ensure I hadn't messed up an if statement...) all other ids have a minimum of 3 measurements per level of "stages" (and most have 10-20 available)