Hello,
I am estimating regressions that include different fixed effects and variables whose coefficients I am not interested in. I want to control them in the estimations. To do so, I am using the command reghdfe. My understanding of this command is that it accounts for the variables declared in the "absorb" option, but it does not estimate their coefficients. Then, I would expect to obtain the same coefficients of the variables of interest when using the command reg, but controlling for all the variables (those inside and outside "absorb") and when using the command reghdfe.
However, it is not the case; when estimating an equation using the command reg, I obtain a coefficient for each one of the variables of interest. When using the command reghdfe, it omits the coefficients of some of the variables of interest. If I use a big dataset, the estimated coefficients of non-omitted variables are the same as those obtained using reg. If the sample is small (such as the one below), the coefficients are quite different, and Stata omits most of the variables of interest.
The following are examples of the estimations:
trend_* --> These are statefip-birthyr (state-year o birth) specific time trends. Therefore, they should be equivalent to what Stata would generate with statefip#c.birthyr
The following is an example of the dataset I used to generate the estimations above:
The following are the results of such estimations:
As you can see, when using reghdfe, Stata omits all of the variables of interest, except for one of them, while when using reg, I obtain a coefficient for each one of these variables.
So, what I would like to understand is:
1) What are the differences between reg and reghdfe that are generating different coefficients?
1) Why does reghdfe omit variables, whereas reg estimated coefficients for each variable?
2) How can I prevent Stata from omitting the variables of interest and instead make it omit the coefficients of the variables I declared in "absorb"?
I would appreciate any help!
I am estimating regressions that include different fixed effects and variables whose coefficients I am not interested in. I want to control them in the estimations. To do so, I am using the command reghdfe. My understanding of this command is that it accounts for the variables declared in the "absorb" option, but it does not estimate their coefficients. Then, I would expect to obtain the same coefficients of the variables of interest when using the command reg, but controlling for all the variables (those inside and outside "absorb") and when using the command reghdfe.
However, it is not the case; when estimating an equation using the command reg, I obtain a coefficient for each one of the variables of interest. When using the command reghdfe, it omits the coefficients of some of the variables of interest. If I use a big dataset, the estimated coefficients of non-omitted variables are the same as those obtained using reg. If the sample is small (such as the one below), the coefficients are quite different, and Stata omits most of the variables of interest.
The following are examples of the estimations:
Code:
reg manager nonblack admit_exp20 admit20_nb admit_noexpbord20 admit20free_nb i.cpuma0010 i.birthyr trend_*, vce(cluster stateyear) reghdfe manager nonblack admit_exp20 admit20_nb admit_noexpbord20 admit20free_nb, absorb(cpuma0010 birthyr statefip#c.birthyr) vce(cluster stateyear)
The following is an example of the dataset I used to generate the estimations above:
Code:
* Example generated by -dataex-. To install: ssc install dataex clear input float(statefip birthyr manager nonblack admit20_nb admit_noexpbord20 admit20free_nb admit_exp20) int cpuma0010 float(stateyear trend_1 trend_2 trend_3 trend_4 trend_5 trend_6) 12 1968 0 0 0 0 0 0 255 445 0 1968 0 0 0 0 13 1966 0 0 0 1 0 1 262 491 0 0 1966 0 0 0 15 1965 0 1 0 . . 0 284 538 0 0 0 1965 0 0 15 1965 0 1 0 . . 0 288 538 0 0 0 1965 0 0 12 1968 0 1 0 0 0 0 253 445 0 1968 0 0 0 0 12 1964 0 1 0 1 1 0 204 441 0 1964 0 0 0 0 15 1967 0 1 0 . . 0 284 540 0 0 0 1967 0 0 12 1968 0 1 0 0 0 0 257 445 0 1968 0 0 0 0 15 1968 0 1 0 . . 0 285 541 0 0 0 1968 0 0 15 1967 0 1 0 . . 0 282 540 0 0 0 1967 0 0 12 1967 0 1 0 0 0 0 239 444 0 1967 0 0 0 0 17 1968 0 1 1 1 1 1 310 637 0 0 0 0 0 1968 17 1965 0 0 0 1 0 1 335 634 0 0 0 0 0 1965 15 1966 0 1 0 . . 0 283 539 0 0 0 1966 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 13 1967 0 1 1 1 1 1 273 492 0 0 1967 0 0 0 17 1965 0 0 0 1 0 1 337 634 0 0 0 0 0 1965 12 1967 0 1 0 0 0 0 239 444 0 1967 0 0 0 0 11 1964 0 1 0 1 1 0 201 393 1964 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1965 0 0 0 1 0 0 201 394 1965 0 0 0 0 0 11 1964 1 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1968 0 1 0 1 1 0 200 397 1968 0 0 0 0 0 11 1966 0 1 0 1 1 0 200 395 1966 0 0 0 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1966 1 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1967 0 0 0 1 0 0 202 396 1967 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1965 0 0 0 1 0 0 202 394 1965 0 0 0 0 0 11 1966 0 1 0 1 1 0 202 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1967 0 1 0 1 1 0 201 396 1967 0 0 0 0 0 11 1966 0 0 0 1 0 0 202 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1965 0 0 0 1 0 0 202 394 1965 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1967 0 1 0 1 1 0 201 396 1967 0 0 0 0 0 11 1968 0 0 0 1 0 0 202 397 1968 0 0 0 0 0 11 1967 1 0 0 1 0 0 201 396 1967 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1968 1 1 0 1 1 0 200 397 1968 0 0 0 0 0 11 1965 0 1 0 1 1 0 200 394 1965 0 0 0 0 0 11 1966 0 0 0 1 0 0 201 395 1966 0 0 0 0 0 11 1968 1 0 0 1 0 0 201 397 1968 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1965 0 0 0 1 0 0 202 394 1965 0 0 0 0 0 11 1965 0 1 0 1 1 0 201 394 1965 0 0 0 0 0 11 1966 0 1 0 1 1 0 200 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1966 0 0 0 1 0 0 201 395 1966 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1968 0 0 0 1 0 0 201 397 1968 0 0 0 0 0 11 1965 0 0 0 1 0 0 201 394 1965 0 0 0 0 0 11 1965 0 0 0 1 0 0 201 394 1965 0 0 0 0 0 11 1965 0 0 0 1 0 0 202 394 1965 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1968 0 0 0 1 0 0 201 397 1968 0 0 0 0 0 11 1967 0 0 0 1 0 0 202 396 1967 0 0 0 0 0 11 1966 0 0 0 1 0 0 202 395 1966 0 0 0 0 0 11 1968 1 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1966 0 0 0 1 0 0 202 395 1966 0 0 0 0 0 11 1965 0 1 0 1 1 0 201 394 1965 0 0 0 0 0 11 1965 0 0 0 1 0 0 201 394 1965 0 0 0 0 0 11 1966 0 0 0 1 0 0 202 395 1966 0 0 0 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1968 1 1 0 1 1 0 200 397 1968 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1967 0 1 0 1 1 0 201 396 1967 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1965 0 0 0 1 0 0 201 394 1965 0 0 0 0 0 11 1964 0 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1965 0 0 0 1 0 0 202 394 1965 0 0 0 0 0 11 1966 1 1 0 1 1 0 200 395 1966 0 0 0 0 0 11 1966 0 1 0 1 1 0 200 395 1966 0 0 0 0 0 11 1968 1 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1968 0 0 0 1 0 0 202 397 1968 0 0 0 0 0 11 1968 0 0 0 1 0 0 201 397 1968 0 0 0 0 0 11 1967 0 0 0 1 0 0 202 396 1967 0 0 0 0 0 11 1967 1 1 0 1 1 0 201 396 1967 0 0 0 0 0 11 1966 0 0 0 1 0 0 202 395 1966 0 0 0 0 0 11 1967 0 0 0 1 0 0 202 396 1967 0 0 0 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1968 0 0 0 1 0 0 202 397 1968 0 0 0 0 0 11 1967 0 1 0 1 1 0 201 396 1967 0 0 0 0 0 11 1967 0 0 0 1 0 0 201 396 1967 0 0 0 0 0 11 1968 1 1 0 1 1 0 200 397 1968 0 0 0 0 0 11 1966 0 1 0 1 1 0 201 395 1966 0 0 0 0 0 11 1965 0 1 0 1 1 0 200 394 1965 0 0 0 0 0 11 1967 0 0 0 1 0 0 202 396 1967 0 0 0 0 0 11 1964 0 0 0 1 0 0 202 393 1964 0 0 0 0 0 11 1964 0 1 0 1 1 0 201 393 1964 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1966 0 0 0 1 0 0 201 395 1966 0 0 0 0 0 11 1964 1 0 0 1 0 0 201 393 1964 0 0 0 0 0 11 1968 0 1 0 1 1 0 201 397 1968 0 0 0 0 0 11 1966 0 0 0 1 0 0 201 395 1966 0 0 0 0 0 11 1967 0 1 0 1 1 0 201 396 1967 0 0 0 0 0 end label values statefip statefip_lbl label def statefip_lbl 11 "District of Columbia", modify label def statefip_lbl 12 "Florida", modify label def statefip_lbl 13 "Georgia", modify label def statefip_lbl 15 "Hawaii", modify label def statefip_lbl 17 "Illinois", modify
The following are the results of such estimations:
Code:
. reg manager nonblack admit_exp20 admit20_nb admit_noexpbord20 admit20free_nb i.cpuma0010 i.birthyr tren > d_*, vce(cluster stateyear) note: 255.cpuma0010 omitted because of collinearity note: 257.cpuma0010 omitted because of collinearity note: 310.cpuma0010 omitted because of collinearity note: 337.cpuma0010 omitted because of collinearity note: trend_1 omitted because of collinearity note: trend_2 omitted because of collinearity note: trend_3 omitted because of collinearity note: trend_4 omitted because of collinearity note: trend_5 omitted because of collinearity Linear regression Number of obs = 94 F(2, 11) = . Prob > F = . R-squared = 0.1838 Root MSE = .3353 (Std. Err. adjusted for 12 clusters in stateyear) Robust manager Coef. Std. Err. t P>t [95% Conf. Interval] nonblack -1.02e-14 1.19e-14 -0.85 0.411 -3.63e-14 1.60e-14 admit_exp20 -.2653784 .1433278 -1.85 0.091 -.5808408 .050084 admit20_nb -.2726897 .0501327 -5.44 0.000 -.383031 -.1623483 admit_noexpbord20 .5479404 .1346726 4.07 0.002 .251528 .8443529 admit20free_nb -.0098723 .0563795 -0.18 0.864 -.1339628 .1142181 cpuma0010 201 -.2675532 .1452092 -1.84 0.092 -.5871564 .05205 202 -.3605872 .1864338 -1.93 0.079 -.7709252 .0497509 204 -.4246173 .1518185 -2.80 0.017 -.7587674 -.0904671 239 .0862023 .0371662 2.32 0.041 .0044 .1680046 253 7.72e-16 1.02e-15 0.75 0.466 -1.48e-15 3.02e-15 255 0 (omitted) 257 0 (omitted) 262 -.0786729 .0147373 -5.34 0.000 -.1111095 -.0462362 273 .0862023 .0371662 2.32 0.041 .0044 .1680046 310 0 (omitted) 335 -5.81e-15 7.51e-15 -0.77 0.456 -2.23e-14 1.07e-14 337 0 (omitted) birthyr 1965 -.1691112 .0222369 -7.60 0.000 -.2180542 -.1201682 1966 -.0904383 .0334711 -2.70 0.021 -.1641078 -.0167689 1967 .0272485 .0172123 1.58 0.142 -.0106356 .0651326 1968 .1134508 .0399003 2.84 0.016 .0256308 .2012708 trend_1 0 (omitted) trend_2 0 (omitted) trend_3 0 (omitted) trend_4 0 (omitted) trend_5 0 (omitted) _cons -.1134508 .0399003 -2.84 0.016 -.2012708 -.0256308 . reghdfe manager nonblack admit_exp20 admit20_nb admit_noexpbord20 admit20free_nb, absorb(cpuma0010 birt > hyr statefip#c.birthyr) vce(cluster stateyear) (dropped 9 singleton observations) note: admit_noexpbord20 is probably collinear with the fixed effects (all partialled-out values are close > to zero; tol = 1.0e-09) (MWFE estimator converged in 5 iterations) note: admit_exp20 omitted because of collinearity note: admit20_nb omitted because of collinearity note: admit_noexpbord20 omitted because of collinearity note: admit20free_nb omitted because of collinearity HDFE Linear regression Number of obs = 85 Absorbing 3 HDFE groups F( 1, 5) = 0.03 Statistics robust to heteroskedasticity Prob > F = 0.8692 R-squared = 0.1709 Adj R-squared = 0.0589 Within R-sq. = 0.0001 Number of clusters (stateyear) = 6 Root MSE = 0.3398 (Std. Err. adjusted for 6 clusters in stateyear) Robust manager Coef. Std. Err. t P>t [95% Conf. Interval] nonblack -.0098723 .0569517 -0.17 0.869 -.1562713 .1365266 admit_exp20 0 (omitted) admit20_nb 0 (omitted) admit_noexpbord20 0 (omitted) admit20free_nb 0 (omitted) _cons .14559 .0254608 5.72 0.002 .080141 .2110389 Absorbed degrees of freedom: Absorbed FE Categories - Redundant = Num. Coefs - cpuma0010 4 0 4 birthyr 5 1 4 statefip#c.birthyr 2 0 2 ? ? = number of redundant parameters may be higher
So, what I would like to understand is:
1) What are the differences between reg and reghdfe that are generating different coefficients?
1) Why does reghdfe omit variables, whereas reg estimated coefficients for each variable?
2) How can I prevent Stata from omitting the variables of interest and instead make it omit the coefficients of the variables I declared in "absorb"?
I would appreciate any help!
Comment