st: Problems re biprobit partial observability

["Deller, David R D" <drddel@essex.ac.uk>]

Dear statalist,

I am currently using a short and highly unbalanced panel data set to estimate a model of individuals committing property crime. Having run basic logit and probit regressions successfully I'm now considering more sophisticated ways to analyse the data. One is to try to control for the notion that some individuals may never commit crime under any circumstances. Hence I split my variables into two: one set of variables fixed through time to explain whether an individual is a criminal/non-criminal type and another set of time varying variables to explain whether criminal types commit a crime in a given time period. Given that all we ever observe is whether a crime gets committed in a given time period this structure would seem to suit a biprobit model with partial observability in the spirit of Poirer (1980).

The problem I am having relates to convergence in the estimations (the full code is included below). I know that in the biprobit model with partial observability identification can be a problem. However if this is going on it would appear severe in my case

If the independent variables used to explain the criminal/non-criminal type and offending in a particular time period are completely different stata has no problem providing estimates. However as soon as one independent variable is used to estimate both concepts stata has problems. If I use either the default Newton-Raphson or bhhh maximisation techniques there never appears to be convergence and for virtually all iterations (not concave) is reported. I have tried running the regressions with the option difficult. If I set the maximisation technique to dfp or bfgs after a few hundred iterations estimated results are produced. The problem with dfp and bfgs is that whilst estimates of the standard errors etc are obtained for the fixed independent variables they are missing from the table of estimation results for the time varying variables. Instead only the co-efficients are reported.

Any suggestions about possible ways to overcome this problem would be much appreciated.

The code which results in the missing standard errors is:
gen anyoff12m11=anyoff12m1
#delimit ;
biprobit (anyoff12m11 i.male ib2.breaklawfea ib26.L1pfa i.Risk9fe i.L1religionfea i.L1ethnicfe
i.L1brokenhome i.L1immigrant1 i.parprisa i.L1homelessa
i.L1expelleda i.everarstb i.eversentb i.everprisb i.helpmenhfe1
i.freeschlmealfe1) (anyoff12m1 i.male  i.L1hhmanage ib2.L1l1statgh ib5.L1hhincomeg i.L1tenureb
i.L1edqualifb i.L1single i.L1drugtake ib4.L1alcoholb i.L1ownchild i.L1parenta i.L1h1frenc L1respage
ib3.L1ahholdb i.L1morpha i.L1wave i.L1vpersoncrm i.L1vhhldcrm i.L1literacyprob i.L1h1genh2
i.L1g1safea1 i.L1g1returna1 ib4.L1rundown1 i.L1cominvolve2
i.L1druglie i.L1crimelie), partial technique(dfp) vce(cluster caseref);
#delimit cr

Regards

David Deller
University of Essex

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/