Hello,
I am trying to use the command acreg (see https://www.stata.com/meeting/switzerland20/slides/Switzerland20_Colella.pdf) in Stata 17 to calculate Conley spatial standard errors. I am getting a warning about a matrix not being full rank and I am not sure if this is something I need to address or can ignore.
Here is a sample of my data using dataex:
I am using cohort variation in a single cross-section to run a DiD analysis with some high dimensional fixed-effects. My regression code is as follows:
I am regressing an indicator for literacy onto some controls, year of birth fixed effects, and am absorbing locality fixed effects (agg_obj_id_encode) and municiplity-by-cohort fixed effects (mun_yob). For reference, municipalities are the next highest administrative level above localities in my context. I also include the locality-level treatment variable indicator (rural_sch), an indicator for being part of the younger cohort exposed to treatment (young), and an interaction between the two. I am ultimately interested in the coefficient and standard errors on rural_schXyoung. (Dummies for year of birth (yob_fe*) are not included in the dataex sample.)
Leaving out most of the ommitted and estimated coefficients on the year of birth fixed effects, this is the output:
I am concerned about the warning, but the coefficient on the interaction is quantitatively similar to estimates using other commands. My feeling is that the warning is more about the year of birth fixed-effects, which, because of omission and collinearity, are probably not reliable, though these estimates are not of interest for me. If this is the case, does it make sense to ignore the warning? Or is this indicating a problem with the estimation of the standard errors for my variable of interest?
I am trying to use the command acreg (see https://www.stata.com/meeting/switzerland20/slides/Switzerland20_Colella.pdf) in Stata 17 to calculate Conley spatial standard errors. I am getting a warning about a matrix not being full rank and I am not sure if this is something I need to address or can ignore.
Here is a sample of my data using dataex:
Code:
* Example generated by -dataex-. For more info, type help dataex clear input float(literate male disabled rural_sch young rural_schXyoung lat lon yob) long agg_obj_id_encode float mun_yob 0 1 0 1 0 0 22.1291 -104.657 1900 10 70233 0 1 0 1 0 0 22.1291 -104.657 1900 10 70233 0 1 0 1 0 0 22.1291 -104.657 1900 10 70233 0 0 0 1 0 0 22.1291 -104.657 1905 10 79913 0 0 0 1 0 0 22.1291 -104.657 1905 10 79913 0 1 0 1 0 0 22.1291 -104.657 1908 10 85814 0 0 0 1 1 1 22.1291 -104.657 1922 10 113496 0 1 0 1 1 1 22.1291 -104.657 1924 10 117520 0 0 0 0 0 0 20.9253 -100.738 1890 279 51524 0 1 0 0 0 0 20.9253 -100.738 1894 279 58828 0 0 0 0 0 0 20.9253 -100.738 1902 279 73608 1 1 0 0 0 0 20.9253 -100.738 1905 279 79447 0 0 0 0 0 0 20.9253 -100.738 1905 279 79447 0 1 0 0 0 0 20.9253 -100.738 1905 279 79447 0 1 0 0 0 0 20.9253 -100.738 1905 279 79447 0 0 0 0 0 0 20.9253 -100.738 1906 279 81448 0 1 0 0 0 0 20.9253 -100.738 1907 279 83397 0 1 0 0 0 0 20.9253 -100.738 1907 279 83397 0 1 0 0 0 0 20.9253 -100.738 1908 279 85348 0 1 0 0 1 0 20.9253 -100.738 1921 279 111029 end label values agg_obj_id_encode agg_obj_id_encode label def agg_obj_id_encode 10 "18083400013", modify label def agg_obj_id_encode 279 "cve110030027", modify
Code:
acreg literate male disabled yob_fe* rural_sch young rural_schXyoung, spatial latitude(lat) longitude(lon) dist(5) id(agg_obj_id) time(yob) pfe1(agg_obj_id_encode) pfe2(mun_yob) dropsingletons
Leaving out most of the ommitted and estimated coefficients on the year of birth fixed effects, this is the output:
Code:
SPATIAL CORRECTION
DistCutoff: 5
LagCutoff: 0
No HAC Correction
Absorbed FE: agg_obj_id_encode and mun_yob
554 singleton observations dropped
Included instruments: male disabled yob_fe1 yob_fe2 yob_fe3...yob_fe114 yob_fe115 rural_sch young rural_schXyoung
Number of obs = 22120
Total (centered) SS = 1600.728742 Centered R2 = 0.0299
Total (uncentered) SS = 1600.728742 Uncentered R2 = 0.0299
Residual SS = 1552.847495
---------------------------------------------------------------------------------
literate | Coefficient Std. err. z P>|z| [95% conf. interval]
----------------+----------------------------------------------------------------
male | .0959288 .0049651 19.32 0.000 .0861973 .1056603
disabled | -.0824189 .0245307 -3.36 0.001 -.1304982 -.0343396
yob_fe1 | 0 (omitted)
yob_fe2 | 0 (omitted)
yob_fe3 | 0 (omitted)
...
yob_fe105 | 4860.244 2.19e+08 0.00 1.000 -4.29e+08 4.30e+08
yob_fe106 | -4936.966 1.74e+08 -0.00 1.000 -3.40e+08 3.40e+08
yob_fe107 | -6484.07 1.88e+08 -0.00 1.000 -3.69e+08 3.69e+08
yob_fe108 | -18760.33 1.75e+08 -0.00 1.000 -3.44e+08 3.44e+08
yob_fe109 | -46066.02 2.27e+08 -0.00 1.000 -4.45e+08 4.45e+08
yob_fe110 | 0 (omitted)
yob_fe111 | 0 (omitted)
yob_fe112 | 0 (omitted)
yob_fe113 | 0 (omitted)
yob_fe114 | 0 (omitted)
yob_fe115 | 0 (omitted)
rural_sch | 0 (omitted)
young | 25670.82 1.49e+08 0.00 1.000 -2.93e+08 2.93e+08
rural_schXyoung | .075119 .0154737 4.85 0.000 .0447912 .1054468
_cons | -5.80e-09 .0003258 -0.00 1.000 -.0006386 .0006386
---------------------------------------------------------------------------------
nb: total SS, model and R2s are after partialling out.
To get the corrected ones use the option correctr2
Warning: X'X matrix not of full rank. Some variables might be omitted.
Beta Coefficients and Standard Errors should be interpreted with caution.
