problem with didregress using a continuous treatment

[Daniel Perez Parra]

Dear statusers,

I'm currently estimating the effect of a large public program that opened an outstanding lower high school number of classes some countries in Africa. My outcome variable (Y = mdschoolattendance, binary 0-1) is lower secondary school attendance and my treatment variable (T = newclasses, ranges from 0-750) is equal to the number of schools opened in the village of my observations during the 8 years of the program, therefore, the treatment is continuous. The treatment, as I specified, is defined at village level (id_geo, 1-2202) but also depending on the cohort (treatment_cohort, 0-1) in the sense that some were too old to be affected by the policy and others enough young.

With this setup I've tried to use didregress, with cross-sectional data and clustering by village, however I get the same message: "The treatment variable newclasses was omitted because of collinearity".

didregress (mdschoolattaind) (newclasses, continuous), group(id_geo treatment_cohort) vce(cluster id_geo)

Someone knows why this happens? I'm sure it is something obvious but I can't see it....

Thanks in advance!

Daniel.

[Andrew Musau]

Show us the results of

xtset id_geo
xtsum newclasses
xtset treatment_cohort
xtsum newclasses

[Daniel Perez Parra]

Show us the results of

xtset id_geo
xtsum newclasses
xtset treatment_cohort
xtsum newclasses

Thanks for your reply Andrew, here you have it:

 xtset id_geo

Panel variable: id_geo (unbalanced)

. xtset id_geo

Panel variable: id_geo (unbalanced)

. xtsum newclasses

Variable         |      Mean   Std. dev.       Min        Max |    Observations
-----------------+--------------------------------------------+----------------
newcla~m overall |  92.93385   137.9326          0        707 |     N =   52619
         between |             146.3207          0        707 |     n =    2204
         within  |                    0   92.93385   92.93385 | T-bar = 23.8743

. xtset treatment_cohort

Panel variable: treatment_cohort (unbalanced)

. xtsum newclasses

Variable         |      Mean   Std. dev.       Min        Max |    Observations
-----------------+--------------------------------------------+----------------
newcla~m overall |  92.93385   137.9326          0        707 |     N =   52619
         between |             4.624628   90.56356   97.10378 |     n =       2
         within  |             137.8967   -4.16993   707.3703 | T-bar = 26309.5

[Andrew Musau]

So there you have it:

. xtsum newclasses

Variable         |      Mean   Std. dev.       Min        Max |    Observations
-----------------+--------------------------------------------+----------------
newcla~m overall |  92.93385   137.9326          0        707 |     N =   52619
         between |             146.3207          0        707 |     n =    2204
         within  |                    0   92.93385   92.93385 | T-bar = 23.8743

Your treatment variable does not vary within id_geo. So it is collinear with this variable. In other words, you cannot identify the effect of "newclasses" in the presence of id_geo fixed effects.

[Daniel Perez Parra]

Andrew Musau thank you so much. As I said, it was something silly, have a good day!

Daniel.