[clarify] Error caused by clustered SEs in combination with fixed effects

Hello, I have been receiving an error when using -estsimp- for a model with fixed effects and clustered SEs. I am hoping that one of you might be able to shed some light on why this error is occurring. *Data* I have annual data on the FDI choices of 50+ firms for a 10-year period. Each firm chooses one or more countries in which to do FDI each year. The total number of countries in each firm's annual choice set is approximately 60. Thus, for each firm-year, there are roughly 60 records, with the dependent variable equal to one or zero for each record. *Model* I am running the model as a logit with firm-year intercepts (similar to a conditional logit), and after dropping some firm-years, there are roughly 440 "fixed effects." My matsize is large enough and I can run the model without problems using the -logit- command. I can also use -estsimp logit- to estimate the exact same model without any problem. (I run -logit- first to figure out which intercepts Stata is dropping, and then exclude those intercepts when I run -estsimp logit-. The number of obs, log-likelihood etc. are identical in both cases.) *Error* An error occurs when I run the model with standard errors clustered by firm. Although I can estimate the model using -logit-, I receive the error "matrix not positive definite r(506);" when I use -estsimp logit-. I traced the error, and it occurs in estsimp._chol (please see below). The rrror also occurs when I include even a single fixed effect in combination with clustering, and when I run the model using "estsimp regress-. Removing all of the fixed effects *or* the "cluster" option eliminates the error. There is no intuitive reason I can think of why the simulation shouldn't work when the error covariance matrix is estimated using clustering and fixed effects are include, so I'm a bit of a loss. I'd much appreciate help that you can give me on with this. Happy holidays, and thanks in advance. Regards, Bennet Zelner begin estsimp._chol --- - version 6.0 - args V k - tempname A - capt matrix `A' = cholesky(`V') = capt matrix __0000PF = cholesky(__000002) - if _rc == 506 { - tempname eye transf varianc vcd tmp - mat `eye' = I(`k') = mat __0000PG = I(454) - mat `transf' = J(1,`k',0) = mat __0000PH = J(1,454,0) - local i 1 - while `i' <= `k' { = while 1 <= 454 { - scalar `varianc' = `V'[`i',`i'] = scalar __0000PI = __000002[1,1] - if `varianc' ~= 0 { = if __0000PI ~= 0 { - mat `tmp' = `eye'[`i',1...] = mat __0000PK = __0000PG[1,1...] - mat `transf' = `transf' \ `tmp' = mat __0000PH = __0000PH \ __0000PK - } [OUTPUT SUPPRESSED--LOOPS THROUGH ALL RHS VARIABLES] - local i = `i' + 1 = local i = 1 + 1 - } - mat `transf' = `transf'[2...,1...] = mat __0000PH = __0000PH[2...,1...] - mat `vcd' = `transf'*`V'*`transf'' = mat __0000PJ = __0000PH*__000002*__0000PH' - mat `A' = cholesky(`vcd') = mat __0000PF = cholesky(__0000PJ) matrix not positive definite mat `A' = `transf''*`A'*`transf' } ------------------------------------------------------------------------------------ end estsimp._chol --- -------------------------------------------------------------------------------------- end estsimp._simp --- local newvars `r(newvars)' _simsig2, s(`sims') g(`genname') `antisim' m(`modname') n(`newvars') local newvars `r(newvars)' matrix `sig' = r(Sigma) scalar `dfsig' = r(dfSigma) } ---------------------------------------------------------------------------------------------- end estsimp --- r(506); -- Prof. Bennet A. Zelner Duke University Fuqua School of Business Box 90120 Durham, NC 27708-0120 bzelner(a)duke.edu Tel +1 919 660-1093 Fax +1 919 681-6244 -- Clarify mailing list served by Harvard-MIT Data Center [Un]Subscribe/View Archive: http://lists.gking.harvard.edu/?info=clarify