I will look at the manuals to see the usage of margins and try an obtain a more easily interpret-able result.Your input is extremely helpful. Thanks a lot Clyde!
-
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 am trying these two difference in difference design specifications using the information gained from above posts. I have a few questions pertaining to the following two set ups.
1. Why do I get different coefficients, and significance levels depending on the method I use? In the first method my three way interaction coefficient is insignificant while in the second form it is EXTREMELY significant
2. If I just had that one table, how would I interpret the three way interaction that is significantly positive in the second regression specification? I care about the 1,1,1 case.
When I add the coefficients all the way from the constant through the one way, two way and three way interactions, I do-0.157+0.014-0.030-0.028+...+ 0.0779 = -0.0531.
I am interpreting this as a firm that is a suspect firm and a peer and after is negatively related to the outcome variable (all are indicator variables). How then would I explain the positive value for the t-statistic?
Code:reg aexp_int_w size_dev_w m2b_dev_w ni_dev_w i.suspect_ni##i.peer##i.after i.years, vce(cluster sic)
PHP Code:Linear regression Number of obs = 3,941
F(29, 57) = .
Prob > F = .
R-squared = 0.2260
Root MSE = .13286
(Std. Err. adjusted for 58 clusters in sic)
---------------------------------------------------------------------------------------
| Robust
aexp_int_w | Coef. Std. Err. t P>|t| [95% Conf. Interval]
----------------------+----------------------------------------------------------------
size_dev_w | .0022026 .0016045 1.37 0.175 -.0010103 .0054154
m2b_dev_w | -.0001755 .000703 -0.25 0.804 -.0015832 .0012323
ni_dev_w | .0053463 .0252405 0.21 0.833 -.045197 .0558896
1.suspect_ni | .0263702 .0444291 0.59 0.555 -.0625974 .1153379
1.peer | -.0304873 .0393731 -0.77 0.442 -.1093306 .048356
|
suspect_ni#peer |
1 1 | .0682861 .0815506 0.84 0.406 -.0950161 .2315883
|
1.after | .0147882 .0129738 1.14 0.259 -.0111914 .0407678
|
suspect_ni#after |
1 1 | -.0601654 .0266696 -2.26 0.028 -.1135702 -.0067605
|
peer#after |
1 1 | -.0132005 .0389203 -0.34 0.736 -.0911371 .064736
|
suspect_ni#peer#after |
1 1 1 | .0723412 .0747779 0.97 0.337 -.0773988 .2220813
|
years |
1994 | .1397923 .0093843 14.90 0.000 .1210005 .1585841
1995 | .2482036 .0228263 10.87 0.000 .2024948 .2939124
1996 | .3055329 .0719437 4.25 0.000 .1614682 .4495977
1997 | .3844897 .0486397 7.90 0.000 .2870903 .4818891
1998 | .2503619 .016235 15.42 0.000 .2178519 .282872
1999 | .3539866 .0416915 8.49 0.000 .2705009 .4374723
2000 | .4317315 .0517075 8.35 0.000 .328189 .535274
2001 | .4081011 .0326505 12.50 0.000 .3427196 .4734825
2002 | .2753195 .0163614 16.83 0.000 .2425563 .3080827
2003 | .2666366 .0207415 12.86 0.000 .2251024 .3081708
2004 | .2023388 .012827 15.77 0.000 .1766531 .2280245
2005 | .2318076 .0234369 9.89 0.000 .1848761 .2787391
2006 | .3055193 .0174185 17.54 0.000 .2706392 .3403993
2007 | .2571196 .0300733 8.55 0.000 .1968989 .3173403
2008 | .2398641 .0259341 9.25 0.000 .1879319 .2917963
2009 | .2077649 .0160552 12.94 0.000 .1756149 .239915
2010 | .2221205 .0173986 12.77 0.000 .1872805 .2569605
2011 | .2164137 .0194042 11.15 0.000 .1775575 .2552699
2012 | .1751185 .0106036 16.52 0.000 .1538851 .1963518
2013 | .157368 .0238759 6.59 0.000 .1095574 .2051786
2014 | .1489964 .0239542 6.22 0.000 .101029 .1969639
2015 | .0713419 .0414502 1.72 0.091 -.0116606 .1543444
2016 | .1525649 .0416141 3.67 0.001 .069234 .2358958
|
_cons | -.1570958 .0049674 -31.63 0.000 -.1670427 -.1471488
---------------------------------------------------------------------------------------
Code:reg aexp_int_w size_dev_w m2b_dev_w ni_dev_w suspect_ni#peer#after i.years, vce(cluster sic)
PHP Code:Linear regression Number of obs = 3,941
F(30, 57) = .
Prob > F = .
R-squared = 0.2260
Root MSE = .13286
(Std. Err. adjusted for 58 clusters in sic)
---------------------------------------------------------------------------------------
| Robust
aexp_int_w | Coef. Std. Err. t P>|t| [95% Conf. Interval]
----------------------+----------------------------------------------------------------
size_dev_w | .0022026 .0016045 1.37 0.175 -.0010103 .0054154
m2b_dev_w | -.0001755 .000703 -0.25 0.804 -.0015832 .0012323
ni_dev_w | .0053463 .0252405 0.21 0.833 -.045197 .0558896
|
suspect_ni#peer#after |
0 0 1 | .0147882 .0129738 1.14 0.259 -.0111914 .0407678
0 1 0 | -.0304873 .0393731 -0.77 0.442 -.1093306 .048356
0 1 1 | -.0288996 .0197436 -1.46 0.149 -.0684356 .0106363
1 0 0 | .0263702 .0444291 0.59 0.555 -.0625974 .1153379
1 0 1 | -.0190069 .0329833 -0.58 0.567 -.0850549 .0470411
1 1 0 | .0641691 .0213909 3.00 0.004 .0213346 .1070036
1 1 1 | .0779326 .0131463 5.93 0.000 .0516077 .1042576
|
years |
1994 | .1397923 .0093843 14.90 0.000 .1210005 .1585841
1995 | .2482036 .0228263 10.87 0.000 .2024948 .2939124
1996 | .3055329 .0719437 4.25 0.000 .1614682 .4495977
1997 | .3844897 .0486397 7.90 0.000 .2870903 .4818891
1998 | .2503619 .016235 15.42 0.000 .2178519 .282872
1999 | .3539866 .0416915 8.49 0.000 .2705009 .4374723
2000 | .4317315 .0517075 8.35 0.000 .328189 .535274
2001 | .4081011 .0326505 12.50 0.000 .3427196 .4734825
2002 | .2753195 .0163614 16.83 0.000 .2425563 .3080827
2003 | .2666366 .0207415 12.86 0.000 .2251024 .3081708
2004 | .2023388 .012827 15.77 0.000 .1766531 .2280245
2005 | .2318076 .0234369 9.89 0.000 .1848761 .2787391
2006 | .3055193 .0174185 17.54 0.000 .2706392 .3403993
2007 | .2571196 .0300733 8.55 0.000 .1968989 .3173403
2008 | .2398641 .0259341 9.25 0.000 .1879319 .2917963
2009 | .2077649 .0160552 12.94 0.000 .1756149 .239915
2010 | .2221205 .0173986 12.77 0.000 .1872805 .2569605
2011 | .2164137 .0194042 11.15 0.000 .1775575 .2552699
2012 | .1751185 .0106036 16.52 0.000 .1538851 .1963518
2013 | .157368 .0238759 6.59 0.000 .1095574 .2051786
2014 | .1489964 .0239542 6.22 0.000 .101029 .1969639
2015 | .0713419 .0414502 1.72 0.091 -.0116606 .1543444
2016 | .1525649 .0416141 3.67 0.001 .069234 .2358958
|
_cons | -.1570958 .0049674 -31.63 0.000 -.1670427 -.1471488
---------------------------------------------------------------------------------------
Comment
-
You are completely misinterpreting the two models. The exercise you show adding up the coefficients is appropriate for the first model, but not the second: you applied it to the second and the result is an utterly meaningless number.
In the first model, you estimate separate effects for suspect, peer, and after conditional on the others being zero, as well as interaction terms that show how those effects are then modified when the others are not zero. Adding up the coefficients of all of the those terms would give you .064169, which, not coincidentally is precisely what the second model shows for the 1 1 1 level of suspect#peeer#after. That's because in the second model, there is no separate estimation given of the effects of suspect, peer, or after (neither conditional nor otherwise). Rather the second model gives you a series of 7 estimates of the combined effect of all possible combinations of suspect, peer, and after compared to the baseline case of all three being zero. The results of the two models are completely consistent with each other. But the models are entirely different ways of looking at the phenomenon under study.
Since the 1 1 1 term in the second model is a direct estimate of the effect of suspect = 1 when peer = 1 and after = 1 (compared to the baseline of all three being 0), whereas in the first model the three way interaction term is a difference in difference in differences of effects (and is difficult to describe in words), the two are completely different things and there is no reason to expect them to be the same, or even to have any close relationship to each other or the same sign, or anything else in common. They aren't even apples and oranges, they are apples and wristwatches.
If you run -margins suspect#peer#after- after each of those models, you will see that they give you exactly the same results: the two models are entirely equivalent in their predictions, they differ in how they represent the underlying variables.
I am interpreting this as a firm that is a suspect firm and a peer and after is negatively related to the outcome variable (all are indicator variables). How then would I explain the positive value for the t-statistic?
I don''t understand what you mean by this. It doesn't make any sense to me. I imagine it is a further extension in some way of the confusion that led you to the earlier part of your post, but I can't really figure out what you are trying to say in this part.Comment
-
Clyde,
That really makes things so much clearer in my head. I couldn't fathom why the same 1 1 1 situation gave different results, but now I understand that they are not the same. The DIDID coefficient is 1 1 1 in the first regression design but the 1 1 1 in the second regression is comparing relative to the 0 0 0 situation. I really appreciate your help. I was hoping and keeping my fingers crossed you would reply, since I get so much clarity from all your answers on other posts.
Thanks once again!Comment
-
Hi I was wondering if someone could help me. STATA keeps omitting my variable because of collinearity. I am using difference-in-differences.
My dependent variable is binary so I am using a probit model- it is 1 if the (political) party was not re-elected and 0 if it was.
My group variable is 0 if flooded (treatment group) and 1 if not flooded (control group). My time variable is 0 if 2001 and 1 if 2005 and i have a treatment effect which is the interaction between these two terms.
I would like to add another variable and interact it with the treatment effect which is also a dummy variable, flood size- 1 if large flood, 0 if small flood. But, everytime I add it, STATA omits the variables due to collinearity. Any advice how to fix this? This variable is important for my model.Comment
-
Hello, I'm in the process of designing a model, have already completed introductory and advanced courses positively - but this was a while ago.Originally posted by Clyde Schechter View Post
I have the following general question about the interpretation, the quoted topic.
* There is explained how to interpret (total) sales, how to interpret in this context sales/values per time unit e.g. units produced per month?
Another general question in connection with regressions and difference in difference analysis is for me if and how an index e.g. the CPI can/should be used as an influence variable to identify/take into account price influences. Is it useful to use the time series of the index as a basic variant with a base year, then you always have the change to the base year - can not imagine that you can derive an interpretation here. -> Rather, I imagine that one should draw the change to the previous time unit. (Also against the background that in a panel regression, the "first difference" model, exactly use these changes, to say it as a "low experienced").
And what happens or is the correct procedure when one "prepares" a time series e.g.: COICOP data by detrendig for the regression "prepares". Is the order decisive whether one generates first a LOG of the series and afterwards a "first difference", or vice versa, an important distinction? The interpretation of such data will then probably only be due to the influence of a reform but not directly to the data of the variable to be examined?
For timeseries, it is necessary to create staionarity before running regressions. Is it also possible to estimate transoformed time series with the fixed-effect model after converting them to first difference?
I have already researched a lot of theory and tutorials on the topics, but in some cases have not come across any information/solution/approach that is applicable or understandable for me. If there are any summary, introductory or extended theory papers on the subject where could I be referred to. I am also happy about such recommendations.
Sorry in advance if the question is a bit confused. It's just a general question without a direct model, I'm aware that for an accurate information, the model is always needed.
Hope there are some hints to the Issues.
Thanks in advance!Last edited by Tobias Ortner; 06 Sep 2023, 08:22.Comment
Comment