Rewrite MLE code to Bayes code in Stata using bayesmh

[Simon Heß]

I have a simple MLE estimation in Stata 14, which I want to move towards a Bayesian estimation using `bayesmh`.

program define mylikelihood
args lnf Xb1 Xb2
qui replace `lnf' = ln(1-normal(`Xb1'))+ln(1-normal(`Xb2')) if $ML_y1==0
qui replace `lnf' = ln(normal(`Xb1'))+ln(1-normal(`Xb1'))+ln(1-normal(`Xb2')) if $ML_y1==1 & $ML_y2==0
qui replace `lnf' = ln(1-(1-normal(`Xb1')*normal(`Xb1'))*(1-normal(`Xb2'))) if $ML_y1==1 & $ML_y2==1
end

MLE is quite easy then:

ml model lf mylikelihood (link=treatment x1 x2 x3 x4 x5) (bilink=treatment x5 x6 x7), ///
vce(cluster clustvar)

Now I want to move towards simple Bayesian analysis. For starters I wanted to try out the simplest possible version with flat priors:

By attempt so far:

bayesmh (link=treatment x1 x2 x3 x4 x5) (bilink=treatment x5 x6 x7) ,///
llevaluator(mylikelihood) ///
prior({eco_link: treatment x1 x2 x3 x4 x5} {bilink: treatment x5 x6 x7},flat)

Which gives me the error:

==0 invalid name
mylikelihood could not be evaluated

I suspect I need to specify the parameters passed on the mylikelihod somehow, but I'm not sure how.

[Simon Heß]

I think I solved it. Turns out that I overlooked that bayesmh takes the likelihood in a different form (not by obs. but overall). So I rewrote:

program define mylikelihoodevaluator
    args lnf Xb1 Xb2
    tempvar lnfj
    qui gen `lnfj' = .
    qui replace `lnfj' = ln(1-normal(`Xb1'))+ln(1-normal(`Xb2'))                        if $MH_y1==0         
    qui replace `lnfj' = ln(normal(`Xb1'))+ln(1-normal(`Xb1'))+ln(1-normal(`Xb2'))        if $MH_y1==1 & $MH_y2==0
    qui replace `lnfj' = ln(1-(1-normal(`Xb1')*normal(`Xb1'))*(1-normal(`Xb2')))            if $MH_y1==1 & $MH_y2==1
    summarize `lnfj' if $MH_touse, meanonly
    if r(N) < $MH_n {
        scalar `lnf'=.
        exit
    }
    scalar `lnf' = r(sum)
end

Now I get a lot of:
Variable omissions (I didn't get these when doing plain ML):

note: parameter eco_link:7.i_enum is omitted
note: parameter eco_link:2.j_enum is omitted
note: parameter eco_link:3.j_enum is omitted
note: parameter eco_link:4.j_enum is omitted
note: parameter eco_link:5.j_enum is omitted
note: parameter eco_link:6.j_enum is omitted
note: parameter eco_link:7.j_enum is omitted
note: parameter eco_link:2.i_ethnicity is omitted
note: parameter eco_link:3.i_ethnicity is omitted
note: parameter eco_link:4.i_ethnicity is omitted
note: parameter eco_link:5.i_ethnicity is omitted
note: parameter eco_link:6.i_ethnicity is omitted

And after some quite outlandish parameter estimates (posterior summary statistics) I get the following remark:

Note: There is a high autocorrelation after 500 lags.
Note: Adaptation tolerance is not met.

Any clues as to what might be going on here?

[Nikolay Balov (StataCorp)]

Simon,

The model parameters {eco_link:} are dropped because they are not used anywhere else; specifically, they are not used in your likelihood specification. Maybe you want to specify a prior for the {link:} parameters?

The latter notes you receive after running bayesmh simply mean that your model has not converged. Of course, in such cases the posterior estimates are useless. It is either a problem with the likelihood or with your prior. Using flat prior is generally not recommended.

Nikolay

[Simon Heß]

Thank you for you reply Nikolay. Unfortunatel the difference between {eco_link:} and {link:} only arose when I was simplifying my code for posting it here. In my code they are the same.

The parameters which are dropped are those for the factor variables i.foo, I ask about this in a separate question however: here

OT: When/why/where was it decided that past posts can not be edited? Can I vote against this somewhere?