Fitting NULL model for binary outcomes about staarpipeline-tutorial HOT 5 CLOSED

Hi @GACGAMA,

Thanks for reaching out. When fitting the null model for binary outcomes, please use family=binomial(link="logit") instead of family=gaussian(link="identity"). Also, it is generally recommended to code your phenotype variable as 0/1 numerical values for control/case status instead of factor values.

Best,
Xihao

from staarpipeline-tutorial.

Converting the outcome to 0/1 and using binomial(link="logit") did work, thanks Xihao!
For some covariables, I get the following error with the STAAR fitnull:

Error in h(simpleError(msg, call)) :
  error in evaluating the argument 'x' in selecting a method for function 'chol2inv': LAPACK routine 'dpotrf': leading principal minor of order 4 is not positive

Do you have any suggestions for this?

from staarpipeline-tutorial.

Could I ask what you mean by covariables? Are they "outcomes" or "covariates"?

from staarpipeline-tutorial.

Sorry, I meant covariates. I think the error means that my covariates don't have enough variance for the model to converge or something like that. My dataset is pretty small (60 WGS samples).

from staarpipeline-tutorial.

Hi @GACGAMA,

Thanks for clarifying. Yes, you are correct. Generally, the null model fitting step is more prone to convergence issues for very small datasets, since the number of observations (e.g. n = 60) may not be sufficiently larger than the number of parameters in the model.

Best,
Xihao

from staarpipeline-tutorial.