Bayesmh droppes factor variables/dummies

[Simon Heß]

This is a follow-up question to this one, but since it's a different question, and to increase searchability of the forum I'm posting it separately.

I'm setting up an estimation of a two-equation model using bayesmh. Some of the model's parameters are ~fixed effects, i.e. call bayes mh as:

bayesmh (link=treatment x1 i.group) (bilink=treatment x2 i.group) ,///
  llevaluator(mylikelihoodevaluator) ///
  prior({eco_link: treatment x1 i.group} {bilink: treatment x2 i.group},flat)

But bayesmh drops these and writes:

note: parameter link:1.group is omitted
note: parameter link:2.group is omitted
note: parameter bilink:1.group is omitted
note: parameter bilink:2.group is omitted

I can work around this by using dummies, but I suspect this is telling me that something about doing this is bad.


My likelihood is:

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

[Simon Heß]

I noticed some typos in the code above that only happened when simplifying my code to be posted here. Hence, here comes an updated version of my question:


Bayesmh drops factor variables/dummies:

I'm setting up an estimation of a two-equation model using bayesmh. Some of the model's parameters are ~fixed effects, meaing I call bayesmh as

bayesmh (link=treatment x1 i.group) (bilink=treatment x2 i.group) ,///
  llevaluator(mylikelihoodevaluator) ///
  prior({link: treatment x1 i.group} {bilink: treatment x2 i.group},normal(0,1))

But bayesmh drops these and writes:

note: parameter link:1.group is omitted
note: parameter link:2.group is omitted
note: parameter bilink:1.group is omitted
note: parameter bilink:2.group is omitted

I can work around this by using dummies, but I suspect this is telling me that something about doing this is bad.

My likelihood is:

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

[Nikolay Balov (StataCorp)]

Simon,

I believe you misspecified the likelihood equations of your model. I see the equal sign, =, in your two equations

(link=treatment x1 i.group) (bilink=treatment x2 i.group)

However, an equal sign is only used when you specify multiple outcomes with common regressors or in non-linear specifications. You should not use it in your case.

A right specification of your model would be

. bayesmh (link treatment x1 i.group) (bilink treatment x2 i.group), ///
    llevaluator(mylikelihoodevaluator)                               ///
    prior({link: treatment x1 i.group _cons},   normal(0, 1))        ///
    prior({bilink: treatment x2 i.group _cons}, normal(0, 1))        ///
    block({link:}) dots

Please let me know if you still have problems.

Nikolay