I asked the following question in the Julia-Opt mailing list, but was advised to ask for some help here.
I am trying to estimate the confidence intervals of parameters that have been estimated using the JuMP optimisation framework and the Ipopt solver.
For now, I've been using parametric and non-parametric bootstrapping, in the first case resampling my dataset with a given PDF for the y values or in the second case just resampling my dataset with replacement.
However, for one of my problems, it takes too much time: I have approximately 50 different large datasets to fit with the same procedure, such that bootstrapping starts to consume too much time, and hence, becomes non-realistic to perform.
I would like to have another "quicker" but still robust way to get the confidence interval for my parameters, and I was wondering if anybody has an idea for implementing such function...
From the LsqFit.jl package (https://github.com/JuliaOpt/LsqFit.jl/blob/master/src/curve_fit.jl), I was thinking about writing something like:
r = y - y_calc # the residuals
dof = length(r) - length(parameters) # the degrees of freedom
mse = sumabs2(r)/dof # mean square error is: standard sum square error / degrees of freedom
Q,R = qr(J) # compute the covariance matrix from the QR decomposition
Rinv = inv(R)
covar = Rinv*Rinv'*mse
std_error = sqrt(diag(covar)) # standard errors
If I continue following the example of the LsqFit package:
std_error *=quantile(dist, alpha)
Does such code seem good? Could I think about other alternatives, such as Log-likelihood or F-test? Any advise to implement them? I never implemented those tests yet...
Thanks in advance!
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to [hidden email].
For more options, visit https://groups.google.com/d/optout.