Issue when building the null model

[yuliu426]

Hello,

When using GMMAT on my dataset, I got a warning message in building null model step:

(Warning messages: 1: Large variance estimate observed in the iterations, model not converged... 2: Average Information REML not converged, refitting model using Brent method...)

And then I got NAs for all the GMMAT results (P-value, test scores).

In the beginning, I have multiple variables(age, sex, PCs of genotype data) in the model, to do the diagnosis I tried removing variables and using different combinations of variables to build the null model. It seems that the model won't converge as long as the second PC is in the model. Otherwise, it will be fine.

So I tried to build the model only using AD status(outcome) and the second principal component of my genotype data(variable). The warning message is still there and the null model does not converge.

Could you help me with this issue? I can share my sample data if needed.

Thank you very much! Yu

[hanchenphd]

Hi Yu,

Thank you for your interest in GMMAT. Can you turn on verbose = TRUE and let me know what you see?

Best, Han

[yuliu426]

Hi Han,

Thank you for the quick response! By turning on the verbose = TURE, I got 500 iterations providing different fixed-effect and coefficients, like the following:

Iteration 1 : Variance component estimates: [1] 99999.5 Fixed-effect coefficients: [1] 107.5223 -12623.7554

By the way, there are some statistics of my sample data: cor(AD,PC2) = 0.156535

summary(data$AD) Min. 1st Qu. Median Mean 3rd Qu. Max. 1.000 1.000 2.000 1.605 2.000 2.000

summary(data$PC2) Min. 1st Qu. Median Mean 3rd Qu. Max. -0.033159 -0.005110 -0.002341 0.009807 0.001964 0.691251 var(PC2) = 0.001588948

Thank you! Best Yu

[hanchenphd]

Hi Yu,

Unfortunately it was a convergence issue in your dataset that we could not really help with as the software developer. The likelihood function was monotonous with respect to the variance component parameter, which converged to 1e5 (the upper bound of the search space). This is usually seen in small to moderate samples.

If you cannot change your sample size, you would need to revise your analysis plan to add/delete some covariates to get a logistic mixed model that converges on your data.

Best, Han

[yuliu426]

Hi Han,

Thanks for your clarification.

I did have multiple covariates in the model in the first place(PC2 is definitely not the only covariate) but the model won't converge. If the PC2 was taken out then the model can converge on the data. I then tried different combinations of covariates and it turned out that as long as the PC2 is in the model won't converge.

Anyways thanks for your help! I may try a different sample size and see what will happen. Best Yu