NaN confidence intervals from fitlme()

Hi all,

I have made a mixed effects model in the following form:

lm=fitlme(data,'y ~ x1 + x2*x3 + (1|x4)')

and the output in the command window looks sensible except for one thing: for the 'Random effects covariance parameters (95% CIs):', both the Lower and Upper confidence interval values are NaN.

Does anyone know what this means? I am pretty sure CIs shouldn't be NaN.

Interestingly, when I add a random slope to the random effects term in the model like this,

lm=fitlme(data,'y ~ x1 + x2*x3 + (1+x5|x4)')

the NaNs disappear and I get numbers for the random effects confidence intervals.

Any insight would be appreciated. Thanks!

-----

Edit: if it helps, the data is from an experiment with different groups of people repeating a task and getting measured on their performance. Each person belongs to only one of the groups. The random effects term x4 is 'person', x1 is 'group', every person was tested under the same range of different experimental parameters defined by x5.

[Christophe KInnard]

I am having the exact same problem and changing the starting option does not solve the problem. The estimate on the intercept random term is small (1e-12) but confidence intervals are NaN. Any explanation? Fitlme converged after 7 iterations.

Answer 1

[Gautam Pendse]

Hi KJ,

Have a look at the 'StartMethod' name/value pair:

https://www.mathworks.com/help/stats/fitlme.html#namevaluepairarguments

Try using 'StartMethod','random' in the call to fitlme - it may be that the default initialization converges to a poor local minimum.

Hope this helps,

Gautam