Dear Statalist,
I want to estimate the effect of income on the binary decision to work or not to work, much like the example in the Stata ivprobit help file. In contrast to the example in the help file, however, I have multiple instruments: the continuous IV variable itself - the IV is husband’s income standardized across countries - and additionally the IV interacted with 10 country dummies.
* The data are pooled with ca. 350,000 observations.
When I only include the IV without country interactions, ivprobit (MLE) in the Stata version 14.1 followed by the margins command works well, although it takes 10 minutes to converge. When I add the multiple IVs, however, I get the following error message:
“matrix not symmetric”
When I run the same model with 2SLS without error, all of my instruments have a near-zero p-value, so I am not sure what is causing this error. I am not keen on using the ivprobit twostep option instead because I cannot cluster by country or use sample weights and the marginal command, as I understand, only gives me the marginals from the equation with the latent variable (equal to the coefficients, which I cannot interpret) rather than the derivatives of the actual probability of work that I am interested in.
Two questions result:
Many thanks in advance for any guidance you have!
Cortnie
I want to estimate the effect of income on the binary decision to work or not to work, much like the example in the Stata ivprobit help file. In contrast to the example in the help file, however, I have multiple instruments: the continuous IV variable itself - the IV is husband’s income standardized across countries - and additionally the IV interacted with 10 country dummies.
Code:
ivprobit E i.group $x i.c i.y (y2 y2_interaction2-y2_interaction10 = iv iv2-iv10) [pw=weight], vce(cluster c)
* where E is the probability of employment
*y2 is the endogenous explanatory variable and y2_interaction2-y2_interaction10 are the endogenous ///
explanatory variable interacted with each country dummy.
*iv is the excluded instrument and iv2-iv10 are the instrument interacted with each country dummy
*i.group is a factor variable with 18 groups ranging in size from 30-2,100 observations each
*i.c and i.y denote 10 country and 4 year FEs
*$x is a matrix of included exogenous variables
When I only include the IV without country interactions, ivprobit (MLE) in the Stata version 14.1 followed by the margins command works well, although it takes 10 minutes to converge. When I add the multiple IVs, however, I get the following error message:
“matrix not symmetric”
When I run the same model with 2SLS without error, all of my instruments have a near-zero p-value, so I am not sure what is causing this error. I am not keen on using the ivprobit twostep option instead because I cannot cluster by country or use sample weights and the marginal command, as I understand, only gives me the marginals from the equation with the latent variable (equal to the coefficients, which I cannot interpret) rather than the derivatives of the actual probability of work that I am interested in.
Two questions result:
- Should I be worried about incidental parameters given that I have 10 country dummies, in addition to 18 group dummies and use 2SLS rather than ivprobit?
- If nothing is theoretically wrong with this model and the problem is merely computational, would it be possible to use David Roodman’s -cmp- command as follows:
Code:
cmp setup
svyset c [pw=weight]
cmp (E = y2 y2_interaction2-y2_interaction10 $x i.group i.c i.y) ///
(y2 = iv iv2-iv10 $x i.group i.c i.y) ///
(y2_interaction2 = iv iv2-iv10 $x i.group i.c i.y) ///
(y2_interaction3 = iv iv2-iv10 $x i.group i.c i.y) ///
*… adding one equation for each interaction of the IV*country up to the last one:
*
*
(y2_interaction10 = iv iv2-iv10 $x i.group i.c i.y), ///
ind($cpm_probit $cpm_cont $cpm_cont … $cpm_cont ) svy
* And then for the marginal effects:
margins, predict(pr eq(#1) dydx(*)
Cortnie
Comment