Can we use weights or do svy for didregress?

[Lars Pete]

Dear all,

I have svyset. I am trying to run:

didregress (wjnilf  i.racecat i.sex age c.age#c.age i.educat fullpart incwage) (treat14), group(statefips) time(year) aggregate(dlang)

wjnilf is binary. I am also doing a probit and glm logit. However, I also want to run a didregress. However, unlike probit and logit, I can't svy didregress. How can I use weights in did? Thanks.

[Maarten Buis]

 
 didregress (wjnilf  i.racecat i.sex age c.age#c.age i.educat fullpart incwage) (treat14) [pweight=weightvar], group(statefips) time(year) aggregate(dlang)

[Lars Pete]

Thanks a lot Maarten Buis . So, since we can't svy, we can't use the replicate weights.
A follow-up question: Is this model alright if wjnilf is a binary variable?
Here treat14 is treated#post14. What would be the difference between this model and:

reg wjnilf treat14 i.racecat i.sex age c.age#c.age i.educat fullpart incwage i.statefips i.year [pw=weightvar]

OR

probit wjnilf treat14 i.racecat i.sex age c.age#c.age i.educat fullpart incwage i.statefips i.year [pw=weightvar]

I get different results. Which would be better?

[Jeff Wooldridge]

Lars: What do you mean by "different results." Do you mean after you compute the average partial effect in the probit it's fairly different from the coefficient in linear regression?

[Lars Pete]

Dear Jeff Jeff Wooldridge : Thank you for your comment.

#A. Actually, I was using i.statefips#i.year where they differed. But without the interaction of FEs, they remain same. So, I conclude all 3 models get same results while using state and year FE separately? The treatment is state level policy in 2014. Will deem these (health i.racecat i.sex age c.age#c.age i.educat fullpart ln_incwage ln_incnonwage disabwrk) as `controls'. Treat14 = treated_state * post14.

reg wjnilf treat14 `controls'  i.statefips i.year [pw=asecwt] //(coeff =  .0102327   P value = 0.404 )
probit wjnilf treat14  `controls' i.statefips i.year [pw=asecwt]
margins, dydx(treat14) // (dy/dx =   .0106622  P value 0.371 )
didregress (wjnilf `controls')(treat14) [pw=asecwt], group(statefips) time(year) // (coeff .01 Pvalue .318)

While using interactions however, ( I am confused whether I should use # or not or I should use # which just does istate#i.year or should do ## which does i.state i.year i.state#i.year), they differ (I am not able to use interaction in didrregress):

reg wjnilf treat14 `controls' i.statefips#i.year [pw=asecwt] // (coeff = -.003 P value = 0.962)
probit wjnilf treat14 `controls' i.statefips#i.year [pw=asecwt]
margins, dydx(treat14) // (coeff = -.01 Pvalue = .898)

#B. Its a pooled cross sectional data i.e. cps march. I cannot xtset. However, If i do this:

isid cpsidp year   //variables cpsidp and year do not uniquely identify the observations
bysort cpsidp year: assert _N == 1   //9 contradictions in 791,864 by-groups assertion is false
duplicates report cpsidp year   //--------------------------------------
*   Copies | Observations       Surplus
*----------+---------------------------
*       1 |       791855             0
*    39984 |        39984         39983
*    42299 |        42299         42298
*    42614 |        42614         42613
*    42814 |        42814         42813
*    43905 |        43905         43904
*    43963 |        43963         43962
*    44077 |        44077         44076
*    45435 |        45435         45434
*    45441 |        45441         45440
*--------------------------------------

duplicates tag cpsidp year, gen(isdup)
edit if isdup
drop isdup

xtset cpsidp year   //. xtset cpsidp year repeated time values within panel r(451);


duplicates drop cpsidp year, force
duplicates tag cpsidp year, gen(isdup)
drop if isdup == 1
xtset cpsidp year   // Panel variable: cpsidp (unbalanced), Time variable: year, 2011 to 2019, Delta: 1 unit

I could get around it. Then the results are:

*xtreg wjnilf treat14 `controls', fe // (coeff =  -.1826311 P value =0.023 )
*eststo fix1
*xtreg wjnilf treat14 `controls', re // (coeff = .0229734, P value =  0.001)
*eststo ran1
hausman fix1 ran1 //Test of H0: Difference in coefficients not systematic. chi2(16) = (b-B)'[(V_b-V_B)^(-1)](b-B)=  21.49. Prob > chi2 = 0.1605


xtprobit wjnilf treat14 `controls', fe // fixed-effects model not available r(198);
xtprobit wjnilf treat14 `controls', re // (coeff = .075845, P value = 0.215 )  
margins, dydx(treat14)   // (coeff = .1173941 , P value = 0.002)  
xtdidregress (wjnilf `controls')(treat14), group(statefips) time(year) // (coeff -.1988181  Pvalue .002)

Note: Here probit re and reg re differ.
Note: I can use Probit with RE which is not possible without xtset in #A .....BUT none of these models in #B would allow me to use weights. The errors are:"pweight not allowed in random-effects caser(101);" or for xtreg: "weight must be constant within cpsidp r(199);"



#C. Another approach which was taken in a QJE labor paper for DID using the same data was used: collapsing at the state level (they did take population weighed averages I suppose) and then doing DID where wjnilf will be average weighted proportion by state and year.
The same paper then used DDD to see ATET for a subsample. There they did (without collapsing or taking state level weighted averages) something like:

 reg wjnilf treat14*1(subsample indicator) `controls' i.cpsidp#i.state i.cpsidp#i.year i.state#i.year [pw=asecwt]

I believe this can't be done in probit because of the incidental parameter problem?


It is hard to decide which one should I go with? #A or #B or #C? In #A, probit and LPM give same results without i.state#i.year but different results with i.state#i.year. In #B Probit and LPM give different results. But the major problem in #B is I can't use weights. This is just one dep var among many, so I have to stick with a minimalist approach. I will also need to use DDD for subsample analysis. I'd really appreciate any feedback. Thank you.

[Lars Pete]

Jeff Wooldridge what do you think?