Hi,

I was wondering what is the best way to estimate the confidence of a likelihood estimate. I am doing a high throughput analysis, so I am looking for an automatic criterion. So far what I am doing is to pick in random the starting value of the parameters to estimate (e.g. P_0 = Random(0.0, 1.0);), I optimize and repeat it N times. I keep the best likelihood score of N repeats and print a warning flag if the variance of likelihood score of N samples is too high.
Does it sound reasonable? ???

Thanks,

Olivier

Dear Olivier,

I am not entirely sure what you are after.

Are you trying to verify that the optimization procedure converges? In this case your approach is sensible, albeit computationally expensive.

If you want to estimate the variance of parameter estimates, then either asymptotic normal or likelihood profile methods (implemented with CovarianceMatrix in HyPhy) are more appropriate.

Cheers,
Sergei