Code:
xtset id time xtreg salary i.treatment#i.time i.time, fe cluster(id) // NOTE #, NOT ##
-
Login or Register
- Log in with
- You are not logged in. You can browse but not post. Login or Register by clicking 'Login or Register' at the top-right of this page. For more information on Statalist, see the FAQ.
-
should give you a model that will support estimation of -margins-. If it does not, post back showing the exact regression and margins command(s) used, and the exact, complete results you got from Stata.
-
Hi again,
Thanks for your answer. When I do not apply noestimcheck, the margins are "not estimable". But when I use noestimcheck, I get margins, including from the first timeperiod, but not the last (given that the regression omits the last timeperiod). However, I reckon that the noestimcheck is ill-advised?
Here you have my dataset:
And here you have my code and results:Code:* Example generated by -dataex-. For more info, type help dataex clear input float(id time treatment salary) 1 98 1 20000 1 99 1 20000 1 100 1 15000 1 101 1 15000 1 102 1 15000 1 103 1 15000 2 98 1 20000 2 99 1 20000 2 100 1 17000 2 101 1 17000 2 102 1 17000 2 103 1 17000 3 98 1 21000 3 99 1 21000 3 100 1 24000 3 101 1 24000 3 102 1 24000 3 103 1 24000 4 98 1 19000 4 99 1 19000 4 100 1 15000 4 101 1 15000 4 102 1 15000 4 103 1 15000 5 98 1 20000 5 99 1 20000 5 100 1 21000 5 101 1 22000 5 102 1 22000 5 103 1 22000 6 98 0 24000 6 99 0 24000 6 100 0 24000 6 101 0 24000 6 102 0 24000 6 103 0 24000 7 98 0 20000 7 99 0 20000 7 100 0 20000 7 101 0 20000 7 102 0 20000 7 103 0 20000 8 98 0 20000 8 99 0 20000 8 100 0 20000 8 101 0 20000 8 102 0 20000 8 103 0 20000 9 98 0 21000 9 99 0 21000 9 100 0 21000 9 101 0 21000 9 102 0 21000 9 103 0 21000 10 98 0 21000 10 99 0 21000 10 100 0 20000 10 101 0 20000 10 102 0 20000 10 103 0 20000 12 98 1 20000 12 99 1 20000 12 100 1 20000 12 101 1 20000 12 102 1 19000 12 103 1 19000 13 98 0 23000 13 99 0 23000 13 100 0 23000 13 101 0 23000 13 102 0 23000 13 103 0 22000 14 98 0 20000 14 99 0 22000 14 100 0 20000 14 101 0 20000 14 102 0 20000 14 103 0 20000 end
Code:. xtset id time Panel variable: id (strongly balanced) Time variable: time, 98 to 103 Delta: 1 unit . . xtreg salary i.treatment#i.time i.time, fe cluster(id) note: 1.treatment#103.time omitted because of collinearity. Fixed-effects (within) regression Number of obs = 78 Group variable: id Number of groups = 13 R-squared: Obs per group: Within = 0.1642 min = 6 Between = 0.2194 avg = 6.0 Overall = 0.0026 max = 6 F(6,12) = . corr(u_i, Xb) = -0.1448 Prob > F = . (Std. err. adjusted for 13 clusters in id) -------------------------------------------------------------------------------- | Robust salary | Coefficient std. err. t P>|t| [95% conf. interval] ---------------+---------------------------------------------------------------- treatment#time | 1 98 | 1047.619 1371.415 0.76 0.460 -1940.438 4035.676 1 99 | 761.9048 1392.428 0.55 0.594 -2271.936 3795.745 1 100 | -142.8571 301.5741 -0.47 0.644 -799.9306 514.2163 1 101 | 23.80952 224.942 0.11 0.917 -466.2969 513.9159 1 102 | -142.8571 147.5766 -0.97 0.352 -464.399 178.6847 1 103 | 0 (omitted) | time | 99 | 285.7143 295.1533 0.97 0.352 -357.3694 928.798 100 | -142.8571 147.5766 -0.97 0.352 -464.399 178.6847 101 | -142.8571 147.5766 -0.97 0.352 -464.399 178.6847 102 | -142.8571 147.5766 -0.97 0.352 -464.399 178.6847 103 | -285.7143 190.5206 -1.50 0.160 -700.823 129.3945 | _cons | 20208.79 259.4511 77.89 0.000 19643.5 20774.09 ---------------+---------------------------------------------------------------- sigma_u | 2364.618 sigma_e | 1171.6652 rho | .80287816 (fraction of variance due to u_i) -------------------------------------------------------------------------------- . . margins treatment, at(time =(98 99 100 101 102)) Adjusted predictions Number of obs = 78 Model VCE: Robust Expression: Linear prediction, predict() 1._at: time = 98 2._at: time = 99 3._at: time = 100 4._at: time = 101 5._at: time = 102 ------------------------------------------------------------------------------- | Delta-method | Margin std. err. z P>|z| [95% conf. interval] --------------+---------------------------------------------------------------- _at#treatment | 1 0 | . (not estimable) 1 1 | . (not estimable) 2 0 | . (not estimable) 2 1 | . (not estimable) 3 0 | . (not estimable) 3 1 | . (not estimable) 4 0 | . (not estimable) 4 1 | . (not estimable) 5 0 | . (not estimable) 5 1 | . (not estimable) -------------------------------------------------------------------------------Last edited by Gustav Egede Hansen; 25 Apr 2023, 01:01.Comment
-
Yes, it is inadvisable to ignore the -noestimcheck- here.
What you need to do is create a new variable, let's call it treated, that takes on the value 1 in observations that are both in the treatment group and at a time when the treatment has been administered, and 0 elsewhere. Then you do:
Code:xtreg salary i.treated i.time, fe margins treated, at(time =(98 99 100 101 102))
Comment
-
Hi again,
Thanks for your reply. Everything is no running; however, something strange appears to be happening. First, here is a dataset with the new variable.
As you said, I am now able to calculate margins after running the regression:Code:* Example generated by -dataex-. For more info, type help dataex clear input float(id time treatment salary treated) 1 98 1 20000 0 1 99 1 20000 0 1 100 1 15000 0 1 101 1 15000 1 1 102 1 15000 1 1 103 1 15000 1 2 98 1 20000 0 2 99 1 20000 0 2 100 1 17000 0 2 101 1 17000 1 2 102 1 17000 1 2 103 1 17000 1 3 98 1 21000 0 3 99 1 21000 0 3 100 1 24000 0 3 101 1 24000 1 3 102 1 24000 1 3 103 1 24000 1 4 98 1 19000 0 4 99 1 19000 0 4 100 1 15000 0 4 101 1 15000 1 4 102 1 15000 1 4 103 1 15000 1 5 98 1 20000 0 5 99 1 20000 0 5 100 1 21000 0 5 101 1 22000 1 5 102 1 22000 1 5 103 1 22000 1 6 98 0 24000 0 6 99 0 24000 0 6 100 0 24000 0 6 101 0 24000 0 6 102 0 24000 0 6 103 0 24000 0 7 98 0 20000 0 7 99 0 20000 0 7 100 0 20000 0 7 101 0 20000 0 7 102 0 20000 0 7 103 0 20000 0 8 98 0 20000 0 8 99 0 20000 0 8 100 0 20000 0 8 101 0 20000 0 8 102 0 20000 0 8 103 0 20000 0 9 98 0 21000 0 9 99 0 21000 0 9 100 0 21000 0 9 101 0 21000 0 9 102 0 21000 0 9 103 0 21000 0 10 98 0 21000 0 10 99 0 21000 0 10 100 0 20000 0 10 101 0 20000 0 10 102 0 20000 0 10 103 0 20000 0 12 98 1 20000 0 12 99 1 20000 0 12 100 1 20000 0 12 101 1 20000 1 12 102 1 19000 1 12 103 1 19000 1 13 98 0 23000 0 13 99 0 23000 0 13 100 0 23000 0 13 101 0 23000 0 13 102 0 23000 0 13 103 0 22000 0 14 98 0 20000 0 14 99 0 22000 0 14 100 0 20000 0 14 101 0 20000 0 14 102 0 20000 0 14 103 0 20000 0 end
Hereafter, I ran a test where I changed the salary for the treatmentgroup after the intervention (so that the treatment effect should be bigger), but very strangely, the salary also changes in the pre-period.Code:xtset id time xtreg salary i.treated i.time, fe margins treated, at(time =(98 99 100 101 102))
For instance, in timeperiode 98 the average new_salary should be 20.000 for the treatmentgroup (sum salary_new if treatment == 1 & time == 98) but according to margins it appears to be 38819.29.Code:gen salary_new = salary replace salary_new = salary_new*2 if treated == 1 xtset id time xtreg salary_new i.treated i.time, fe margins treated, at(time =(98 99 100 101 102))
Do you have any idea what is going on?
Best
GustavComment
-
I cannot replicate your problem in the example data:
Code:. xtreg salary i.treated i.time, fe Fixed-effects (within) regression Number of obs = 78 Group variable: id Number of groups = 13 R-squared: Obs per group: Within = 0.1360 min = 6 Between = 0.2194 avg = 6.0 Overall = 0.0893 max = 6 F(6, 59) = 1.55 corr(u_i, Xb) = 0.1448 Prob > F = 0.1789 ------------------------------------------------------------------------------ salary | Coefficient Std. err. t P>|t| [95% conf. interval] -------------+---------------------------------------------------------------- 1.treated | -595.2381 522.4779 -1.14 0.259 -1640.714 450.2379 | time | 99 | 153.8462 451.1385 0.34 0.734 -748.8799 1056.572 100 | -692.3077 451.1385 -1.53 0.130 -1595.034 210.4183 101 | -340.6593 511.543 -0.67 0.508 -1364.254 682.9358 102 | -417.5824 511.543 -0.82 0.418 -1441.178 606.0127 103 | -494.5055 511.543 -0.97 0.338 -1518.101 529.0896 | _cons | 20692.31 319.0031 64.87 0.000 20053.98 21330.63 -------------+---------------------------------------------------------------- sigma_u | 2230.8089 sigma_e | 1150.182 rho | .78999382 (fraction of variance due to u_i) ------------------------------------------------------------------------------ F test that all u_i=0: F(12, 59) = 20.64 Prob > F = 0.0000 . margins treated, at(time =(98 99 100 101 102)) Adjusted predictions Number of obs = 78 Model VCE: Conventional Expression: Linear prediction, predict() 1._at: time = 98 2._at: time = 99 3._at: time = 100 4._at: time = 101 5._at: time = 102 ------------------------------------------------------------------------------ | Delta-method | Margin std. err. z P>|z| [95% conf. interval] -------------+---------------------------------------------------------------- _at#treated | 1 0 | 20692.31 319.0031 64.87 0.000 20067.07 21317.54 1 1 | 20097.07 612.1652 32.83 0.000 18897.25 21296.89 2 0 | 20846.15 319.0031 65.35 0.000 20220.92 21471.39 2 1 | 20250.92 612.1652 33.08 0.000 19051.09 21450.74 3 0 | 20000 319.0031 62.70 0.000 19374.77 20625.23 3 1 | 19404.76 612.1652 31.70 0.000 18204.94 20604.58 4 0 | 20351.65 399.8915 50.89 0.000 19567.88 21135.42 4 1 | 19756.41 425.3374 46.45 0.000 18922.76 20590.06 5 0 | 20274.73 399.8915 50.70 0.000 19490.95 21058.5 5 1 | 19679.49 425.3374 46.27 0.000 18845.84 20513.13 ------------------------------------------------------------------------------ . xtreg salary_new i.treated i.time, fe Fixed-effects (within) regression Number of obs = 78 Group variable: id Number of groups = 13 R-squared: Obs per group: Within = 0.9034 min = 6 Between = 0.5968 avg = 6.0 Overall = 0.7879 max = 6 F(6, 59) = 92.01 corr(u_i, Xb) = -0.1697 Prob > F = 0.0000 ------------------------------------------------------------------------------ salary_new | Coefficient Std. err. t P>|t| [95% conf. interval] -------------+---------------------------------------------------------------- 1.treated | 18126.98 1038.707 17.45 0.000 16048.54 20205.43 | time | 99 | 153.8462 896.8814 0.17 0.864 -1640.809 1948.502 100 | -692.3077 896.8814 -0.77 0.443 -2486.963 1102.348 101 | -289.3773 1016.968 -0.28 0.777 -2324.325 1745.571 102 | -443.2234 1016.968 -0.44 0.665 -2478.172 1591.725 103 | -520.1465 1016.968 -0.51 0.611 -2555.095 1514.802 | _cons | 20692.31 634.1909 32.63 0.000 19423.29 21961.32 -------------+---------------------------------------------------------------- sigma_u | 3241.4483 sigma_e | 2286.6079 rho | .66772236 (fraction of variance due to u_i) ------------------------------------------------------------------------------ F test that all u_i=0: F(12, 59) = 11.57 Prob > F = 0.0000 . margins treated, at(time =(98 99 100 101 102)) Adjusted predictions Number of obs = 78 Model VCE: Conventional Expression: Linear prediction, predict() 1._at: time = 98 2._at: time = 99 3._at: time = 100 4._at: time = 101 5._at: time = 102 ------------------------------------------------------------------------------ | Delta-method | Margin std. err. z P>|z| [95% conf. interval] -------------+---------------------------------------------------------------- _at#treated | 1 0 | 20692.31 634.1909 32.63 0.000 19449.32 21935.3 1 1 | 38819.29 1217.009 31.90 0.000 36434 41204.59 2 0 | 20846.15 634.1909 32.87 0.000 19603.16 22089.15 2 1 | 38973.14 1217.009 32.02 0.000 36587.84 41358.43 3 0 | 20000 634.1909 31.54 0.000 18757.01 21242.99 3 1 | 38126.98 1217.009 31.33 0.000 35741.69 40512.28 4 0 | 20402.93 795.0004 25.66 0.000 18844.76 21961.1 4 1 | 38529.91 845.5879 45.57 0.000 36872.59 40187.24 5 0 | 20249.08 795.0004 25.47 0.000 18690.91 21807.26 5 1 | 38376.07 845.5879 45.38 0.000 36718.75 40033.39 ------------------------------------------------------------------------------Comment
-
Hi again,
I just copypasted your results from the example dataset. I have highlighted some of the results that are currious. For the treatmentgroup, new_salary was supposed to be higher only in the posttreatmentperiod (replace salary_new = salary_new*2 if treated == 1). Yet, the margins are also suddenly high in the pretreatmentperiod.
Here is the plain average new_salary for the treatmentgroup over time:Code:. xtreg salary_new i.treated i.time, fe Fixed-effects (within) regression Number of obs = 78 Group variable: id Number of groups = 13 R-squared: Obs per group: Within = 0.9034 min = 6 Between = 0.5968 avg = 6.0 Overall = 0.7879 max = 6 F(6, 59) = 92.01 corr(u_i, Xb) = -0.1697 Prob > F = 0.0000 ------------------------------------------------------------------------------ salary_new | Coefficient Std. err. t P>|t| [95% conf. interval] -------------+---------------------------------------------------------------- 1.treated | 18126.98 1038.707 17.45 0.000 16048.54 20205.43 | time | 99 | 153.8462 896.8814 0.17 0.864 -1640.809 1948.502 100 | -692.3077 896.8814 -0.77 0.443 -2486.963 1102.348 101 | -289.3773 1016.968 -0.28 0.777 -2324.325 1745.571 102 | -443.2234 1016.968 -0.44 0.665 -2478.172 1591.725 103 | -520.1465 1016.968 -0.51 0.611 -2555.095 1514.802 | _cons | 20692.31 634.1909 32.63 0.000 19423.29 21961.32 -------------+---------------------------------------------------------------- sigma_u | 3241.4483 sigma_e | 2286.6079 rho | .66772236 (fraction of variance due to u_i) ------------------------------------------------------------------------------ F test that all u_i=0: F(12, 59) = 11.57 Prob > F = 0.0000 . margins treated, at(time =(98 99 100 101 102)) Adjusted predictions Number of obs = 78 Model VCE: Conventional Expression: Linear prediction, predict() 1._at: time = 98 2._at: time = 99 3._at: time = 100 4._at: time = 101 5._at: time = 102 ------------------------------------------------------------------------------ | Delta-method | Margin std. err. z P>|z| [95% conf. interval] -------------+---------------------------------------------------------------- _at#treated | 1 0 | 20692.31 634.1909 32.63 0.000 19449.32 21935.3 1 1 | 38819.29 1217.009 31.90 0.000 36434 41204.59 2 0 | 20846.15 634.1909 32.87 0.000 19603.16 22089.15 2 1 | 38973.14 1217.009 32.02 0.000 36587.84 41358.43 3 0 | 20000 634.1909 31.54 0.000 18757.01 21242.99 3 1 | 38126.98 1217.009 31.33 0.000 35741.69 40512.28 4 0 | 20402.93 795.0004 25.66 0.000 18844.76 21961.1 4 1 | 38529.91 845.5879 45.57 0.000 36872.59 40187.24 5 0 | 20249.08 795.0004 25.47 0.000 18690.91 21807.26 5 1 | 38376.07 845.5879 45.38 0.000 36718.75 40033.39 ------------------------------------------------------------------------------
To me, it is quite strange that the margins are also high in the pretreatment period. Do you have any idea why or am I completely misunderstanding something?Code:. bysort time: sum salary_new if treatment == 1 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 98 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 20000 632.4555 19000 21000 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 99 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 20000 632.4555 19000 21000 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 100 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 18666.67 3614.784 15000 24000 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 101 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 37666.67 7527.727 30000 48000 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 102 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 37333.33 7447.595 30000 48000 ------------------------------------------------------------------------------------------------------------------------------------------------ -> time = 103 Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- salary_new | 6 37333.33 7447.595 30000 48000
Best
GustavComment
-
I see. Well, the expectation that only the margins for treated = 1 will change is not correct. At its simplest, this has nothing to do with fixed-effects, or the time variables. You can reproduce the same phenomenon with simple linear regressions against a single dichotomous predictor. And it's nothing really to do with -margins-. -margins- is just compiling the predicted values. Run:
Take look at those -rvfplot- graphs. When you change the dependent variable, the residual structure changes in both groups - it has to in order to be optimal for the entire sample under least squares. When the residual structure changes, the predicted values also necessarily change. And, as I noted, -margins- is just averaging predicted values, so those must change as well.Code:regress salary i.treated predict xbxb1, xb rvfplot, name(rvf1, replace) regress salary_new i.treated predict xbxb2, xb tabstat xbxb*, by(treated) rvfplot, name(rvf2, replace)
Comment
-
Yes that makes perfect sense (that the margins-command is just the predicted values from the regression output). So, to backup a bit, my goal is to have graph I can use to assess the trends for the treatment and control group before the intervention. And I need this with the fe-option. From our back-and-forth, I gather that the proposed option #18 does not allow for this because the coefficients in, say #21, are not enough to visualise the trends for both groups but just provide me with did-coefficient. Is there a way arrive at the following visualization (marginsplot), yet with the fe-option specified:
Code:reg salary_new i.treatment##i.time margins treatment, at(time =(98 99 100 101 102 103)) marginsplot
Last edited by Gustav Egede Hansen; 28 Apr 2023, 01:00.Comment
-
I think for this purpose you need to -collapse- the data to get the mean outcome in each group at each time point. Then plot those means.Comment
Comment