BUG: full_output in optimizers Likelihood model

[josef-pkt]

full_output=False

In this case, the default fit function (newton) returns one array (newparams). However, LikelihoodModel.fit() expects 2. fit() then tries to unpack newparams into two variables, and thus gives the ValueError "too many values to unpack" (unless for some reason newparams is of length 2). (Ian Langmore on mailinglist)

if params is of length 2, using newton silently cuts off one parameter.

looking at the code, I also found that xopt is not defined in any of the full_output=False cases:

    return xopt, retvals
UnboundLocalError: local variable 'xopt' referenced before assignment

[josef-pkt]

needs regression tests to avoid refactoring bugs.

[jseabold]

Is there a small example that demonstrates this that can be used for testing?

[jseabold]

It looks to me like a lot of these are going to be broken if full_output = False. I wonder if it's a refactoring victim or if it ever worked... Need to think a bit about what we want and look at the new optimize interface in scipy for guidance.

[jseabold]

I started a branch to fix this here https://github.com/jseabold/statsmodels/tree/fix-opt-fulloutput

I think it's okay now but needs some tests. Might just be able to steal some from scipy.optimize and just have some smoke tests.