Incorrect Stata Error Message for xtdidregress

[Jared Greathouse]

When I estimate

import delim "https://raw.githubusercontent.com/SucreRouge/synth_control/master/basque.csv", clear

egen id = group(regionname)

drop if inlist(id,18)

g weight = 0

replace weight = 0.5327967 if id ==9 // catalan

replace weight = 0.0225190 if id ==13 // madrid

replace weight =  0.1151287 if id == 16 // asturias 

replace weight = 0.2892709 if id ==17 //la rioja

replace weight = . if weight ==0


drop v1

xtset id year, y

g treat = cond(id==4 & year >=1975,1,0)

cls

xtdidregress (gdpcap) (treat) [aw=weight], group(id) time(year)

Stata tells me

invalid treatment variable
Either the treatment does not vary or it is never zero. The treatment variable is either zero for
controlled observations or 1 for treated observations. To use a continuous treatment, specify (treat,
continuous) for your treatment specification.

even though this is preposterous, it's 0 for all states before 1975, 1 for the Basque Country after 1975, and 0 for all other states after 1975. However, when I estimate

import delim "https://raw.githubusercontent.com/SucreRouge/synth_control/master/basque.csv", clear

egen id = group(regionname)

drop if inlist(id,18)

g weight = 0

replace weight = 0.5327967 if id ==9 // catalan

replace weight = 0.0225190 if id ==13 // madrid

replace weight =  0.1151287 if id == 16 // asturias 

replace weight = 0.2892709 if id ==17 //la rioja

replace weight = . if weight ==0


drop v1

xtset id year, y

g treat = cond(id==4 & year >=1975,1,0)

cls

xtdidregress (gdpcap) (treat), group(id) time(year)

the error goes away and the DD estimates as we'd want it to. Why? The weights have nothing to do with the treatment variable, so it's unclear to me why use of analytic weights would affect it in one instance, but not the other.

[FernandoRios]

Hi Jared
It seems to me that in your first example, county "id==4" (the treated country) is getting a weight of zero.
Now, when a variable has weight of zero, stata assumes its out of the sample. So appears there isn't a treated unit. Thus explaining the error you report.
HTH

[Jared Greathouse]

Yes, but, I replace all weights as missing if they originally had the 0 value before running the analysis. Presumably this includes The Basque Country, no? FernandoRios

[William Lisowski]

. g treat = cond(id==4 & year >=1975,1,0)

. 
. tab id treat if weight!=.

group(regi |   treat
   onname) |         0 |     Total
-----------+-----------+----------
         9 |        43 |        43 
        13 |        43 |        43 
        16 |        43 |        43 
        17 |        43 |        43 
-----------+-----------+----------
     Total |       172 |       172

[Jared Greathouse]

I should listen to people like this more often, who advise that the treatment unit should have a weight of one. Problem solved, thank you all.