Post estimation with endogenous switching regressions

[Martin Paul]

Dear STATALIST,

I am using an endogenous switching regression model to understand how selection affects a program's outcome. I use the movestay command by Lokshin and Sajaia (2004). It works fine though I get convergence issue with very large regression coefficients.

I now want to do some post estimation calculations in order to compute ATT and ATU. I use their exact code in the STATA journal which is "predicted fam, yc1 _1". However, STATA tells me the option

yc1 _1 not allowed.

What could be the issue?

[FernandoRios]

Hi Martin
I suspect that has to do with the version you are using. I do not know which version is available from "ssc", but if you do "search movestay" you will find the st0071_2 version of the command, which may be the one you need to replicate and bypass the problem you report.
Fernando

[Martin Paul]

Thanks
Fernando,

mspredict works. However, when I tried computing the ttest, STATA tells me 'no observation'

[Martin Paul]

dataex income_future age_hh educ_hh marital_status income1 asset1 dependency_ratio credit_access labour_days extension_contact cropland TLU radio_use mobilephone_use

mspredict fam, yc1_1
mspredict yam, yc2_1
ttest fam==yam

The last option comes with "no observation"

[FernandoRios]

Hi Martin
My guess is that the way -mspredict- works, it only creates the variable for its corresponding group.
In other words, when - fam!=. ->yam==. and when yam!=. -> fam!=.

I would say that obtaining the predictions should be relatively easy, but requires some work with the output from movestay. You will need to follow their paper closely, to see how things are actually created.
The other option, as I suggested previously, is to use the two step approach. In that case, you could simply estimate two separate heckman regressions to obtain the predicted outcomes.
Best
Fernando

[Martin Paul]

Yes, I could as well do the two-stage but I am also interested in understanding the outcomes of both the treated and the non-treated outcomes. The two-stage only estimates the treatment and the outcome of those treated.

You're very right. A close look into the predicted data shows many missing values are predicted, This is definitely the cause of the 'no observations'. Would it be appropriate to replace the missing values?

[Iyasu Demeke]

Dear STATALIST,

I am working a paper on “adoption of climate smart agricultural practice on HH farm income” conducted at household level.

I choose 4 climate smart technologies improved seeds, crop rotation, organic fertilizer and push-pull pest management. My selection model is binary logit. However, the outcome dependent variables for each combination of technologies is ordinal outcome variables. So, how I can use this ordinal data in the endogenous switching regression model to estimate the outcome of each 15 combination (16^th one it reference) of technologies.

I hope you can help me!!!