Se in logistic 'firth' model

Feature Requests
1

output from my ‘firth’ model run with command

log_reg = hl.logistic_regression_rows(
    test='firth',
    y=mt.pheno.trait1,
    x=mt.GT.n_alt_alleles(), 
    covariates=covariates 
)

does not have standard errors (se) but gives me

locus
alleles
beta
chi_sq_stat
p_value
n_iterations
converged
exploded

when is do log_reg.show()

is there a way to get se as part of the output. is there a particular command I need to issue

2

Hey @hirbokazi ,

As you’ve noted, Hail does not currently have this functionality but we could add it. Just to be clear, are you looking for standard errors on the coefficients (i.e. on beta) or standard errors on the predicted values (i.e. beta * genotype)? In the latter case, do you also want the predicted values?

3

yes - would love to get se on beta

4

Thanks a ton for a quick response. For purpose of reported se on the coefficients (beta) is good - I am not sure some other users might be interested in the latter

5

Hi @hirbokazi,

None of us are statisticians, so please correct me if I’m wrong, but my understanding is that in the Firth model, the betas are not normally distributed, and can even be very asymmetric, so standard errors don’t really make sense. There does appear to be a method to compute approximate confidence intervals, but it isn’t trivial and we probably won’t have the bandwidth to implement it in the near future.