Code:
// TWFE regression
fvset base $baseYear year
reg `o' ib2010.year#c.Z ib2010.year i.$geo [aw=L${baseYear}], cluster($geo)
-
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.
-
I saw people say that is because I did not include the main effect, but I dont want to do that. How can I force Stata to do what I am asking it to? -
So can you explain what is it what you want to do? -
Stata doesn't know that you really want to omit all levels of a factor, and it's a sensible default not to do so unless explicitly instructed to do so. This is because it's most often an error.
Do you have a valid statistical reason to exclude the main effect of -z- ? If not, you model may not be interpretable.
Compare the output of these two completely absurd, but equivalent, models to show how Stata handles the default case, and convince yourself they're the same models.
If you really want to omit the level of an interaction, you need the 'o' prefix which stands for "omit". Example below, and -help fvvarlist- for details.Code:. reg price ib3.rep78##c.trunk Source | SS df MS Number of obs = 69 -------------+---------------------------------- F(9, 59) = 1.21 Model | 89992665.2 9 9999185.03 Prob > F = 0.3053 Residual | 486804294 59 8250920.23 R-squared = 0.1560 -------------+---------------------------------- Adj R-squared = 0.0273 Total | 576796959 68 8482308.22 Root MSE = 2872.4 ------------------------------------------------------------------------------- price | Coefficient Std. err. t P>|t| [95% conf. interval] --------------+---------------------------------------------------------------- rep78 | 1 | 5304.142 11917.1 0.45 0.658 -18541.92 29150.21 2 | 1526.613 4074.293 0.37 0.709 -6626.029 9679.255 4 | 3625.523 3009.916 1.20 0.233 -2397.305 9648.351 5 | -2802.15 4634.127 -0.60 0.548 -12075.02 6470.716 | trunk | 332.4264 148.5569 2.24 0.029 35.16469 629.688 | rep78#c.trunk | 1 | -578.7597 1362.207 -0.42 0.672 -3304.529 2147.01 2 | -121.3617 263.6709 -0.46 0.647 -648.9658 406.2425 4 | -251.5533 198.8251 -1.27 0.211 -649.4014 146.2947 5 | 310.197 372.4613 0.83 0.408 -435.0963 1055.49 | _cons | 1354.191 2327.813 0.58 0.563 -3303.752 6012.134 ------------------------------------------------------------------------------- . reg price ib3.rep78#c.trunk i.rep78 Source | SS df MS Number of obs = 69 -------------+---------------------------------- F(9, 59) = 1.21 Model | 89992665.2 9 9999185.03 Prob > F = 0.3053 Residual | 486804294 59 8250920.23 R-squared = 0.1560 -------------+---------------------------------- Adj R-squared = 0.0273 Total | 576796959 68 8482308.22 Root MSE = 2872.4 ------------------------------------------------------------------------------- price | Coefficient Std. err. t P>|t| [95% conf. interval] --------------+---------------------------------------------------------------- rep78#c.trunk | 1 | -246.3333 1354.082 -0.18 0.856 -2955.845 2463.178 2 | 211.0647 217.8375 0.97 0.337 -224.8271 646.9565 3 | 332.4264 148.5569 2.24 0.029 35.16469 629.688 4 | 80.87302 132.1449 0.61 0.543 -183.5482 345.2943 5 | 642.6234 341.5527 1.88 0.065 -40.82201 1326.069 | rep78 | 1 | 5304.142 11917.1 0.45 0.658 -18541.92 29150.21 2 | 1526.613 4074.293 0.37 0.709 -6626.029 9679.255 4 | 3625.523 3009.916 1.20 0.233 -2397.305 9648.351 5 | -2802.15 4634.127 -0.60 0.548 -12075.02 6470.716 | _cons | 1354.191 2327.813 0.58 0.563 -3303.752 6012.134 -------------------------------------------------------------------------------
Code:* now level 3 is omitted from the interaction. reg price i.rep78#c.trunk ib3.rep78 o3.rep78#oc.trunk
Comment
-
The way to program this is to use the "o.Z".
is
The reason I do not include the main effect is because it is captured by the unit fixed effects. This is commonly done in DiD regressions. You will have unit fixed effect, time fixed effect and then an interaction between treatment dummy and time fixed effect. I dont want to include the base level because it is collinear with the unit fixed effects.Code:// TWFE regression fvset base $baseYear year reg `o' ib2010.year#c.Z ib2010.year i.$geo o.Z [aw=L${baseYear}], cluster($geo)
Here is an example of a paper that runs such a regression: https://economics.mit.edu/sites/defa...lity_draft.pdfComment
-
Thanks for your answer on the o. category!
I don't share this position. Stata lets me run the regression which is presumably inappropriate, and it doesnt even give me any kind of warning or error message! Instead it just adds another issue by setting the base category to a different value than what the user wanted. Very weird behavior in my opinion. But good that there is at least a manual way to prevent it.Originally posted by Leonardo Guizzetti View PostComment
-
I tend to agree with Henry Strawforrd on this one. If the user specifies a base group, Stata should respect it. Whether the main effect of the interacted variable is included or not should not matter.
Note that just running a regression without an interaction is essentially a special case of a regression where the factor variable is interacted with the intercept. But even without intercept (the "main effect" in this case), Stata respects the choice of the base group, even though no base group is needed:
In my opinion, this should be brought to the attention of Stata's tech support.Code:reg price ib3.rep78, nocons
Comment
Comment