Fixed-Effects Individual Slopes (but only Stata 11) or simply adding interactions with time?

[Guest]

Dear Statalist,

I am using a three-wave panel data set. Right now I am conducting a Fixed-Effects Analysis, regressing wage on the participation in an educational program.

My dataset looks like this:

id	wage	participation (1=yes)	wave
1	2400	1	1
1	2500	0	2
2	1800	0	1
2	1900	1	2
2	1900	0	3

Because selection into treatment is not at random (people are selected on different criteria that I do not have data on) I would like to not only control individual intercepts but furthermore individual slopes. Because it is likely that those selected into treatment have a steeper slope than those not selected (probably because they are more motivated anyway, but I have no data on that). The Random-Trend Model by Wooldridge (2011, chapt 11.2) apparently is the same kind of model.

I have to use Stata 11. Xtfeis, an ado by Volker Ludwig, that de-trends the data is only available beginning with Stata 12. I cannot do anything about that unfortunately.

Is it possible to "de-trend" or to calculate individual slopes in a different way? I once heard that it might be enough to simply include interactions with time.

So instead of

xtset id wave

xtreg wage treatment covariates, fe vce (cluster id)

--> coefficient about 10.4 (0.09)***

I tried

xtreg wage treatment##c.wage covariates, fe vce (cluster id)

--> coefficient about 0.12 (0.03)***



But I am not sure if that is actually correct.

Are there any ideas or suggestions?


Thank you very much!

Kind regards
Guest


PS: Due to data restrictions I cannot post any data examples or any results, so I can only make examples up. Because of that, the coefficients above don't make any sense. I just wanted to show that my coefficient gets smaller.

[Guest]

I still haven't found any solution to the problem. Could somebody provide any help? I would be very grateful for that.I have found a book describing what I think is just what I want to do, however I am not sure how to "translate" that into syntax. It says:

"... the term g_it = g_i*t, an interaction of a country dummy and the time variable, introduces the possibility of an individual trend (...) g_it is a characteristic that has an additive effect over time. Thus, looking at the model specification, (...) at time t [the country] may deviate from the regional trend (..). This specification is analogous to the random trend model (..)."

So I wouldn't add an interaction of wave and treatment, but of time and id? The person than conducts a POLS regression.

But then I have also red that one first has to take first differences between adjacent observations, then apply the time-demeaned transformation, followed by OLS.

I am lost what to exactly do.


BTW, reading above I noticed a mistake, it should have said:

I tried xtreg wage treatment##c.wave covariates, fe vce (cluster id)

(The interaction is of course with wave, not with wage)

[Carlo Lazzaro]

Guest:
if your goal is to get randon intercepts as well as random slopes, I would take a look at -mixed-.

[Guest]

Thank you, Carlo! However, in xtmixed my units would be people (id) - otherwise I wouldn't get fixed-effects - and I would get random slopes for every single unit. But I want random slopes for my treatment groups, not my units. And, most importantly, I want to know how the coefficient change if I "de-trend", so subtract the time trend for the treatment group. I don't think xtmixed could do that. At least that's what I think after reading up on mixed.

[Guest]

Due to Josef Brüderl it would be like this:

First, first-differences then taking second-difference (I think this subtracts first the individual-specific intercepts, than the individual-specific trends), then running a POLS over it. This is in line with everything I am reading, but the question still is: how exactly does that look in my syntax?

So I would start like this:

xtset id wave
reg D.(wage treatment covariates), noconstant vce (cluster id)

reg D2.(wage treatment covariates), noconstant vce (cluster id)

reg wage treatment covariates, re vce (cluster id) Theta


I know the syntax by itself, but how do I combine it all?

[daniel klein]

I have to use Stata 11. Xtfeis, an ado by Volker Ludwig, that de-trends the data is only available beginning with Stata 12. I cannot do anything about that unfortunately.

Change the line

version 12

in xtfeis.ado to

version 11

if you want to use this command. I can run the examples in the help file on Stata 11.

Best
Daniel

[Guest]

Thank you so much, that absolutely worked! I can't believe it, that's just about perfect!

But I honestly wonder: Why is Stata 12 required? Do I need to be aware of any problems since it apparently wasn't written for Stata 11?

[daniel klein]

The version 12 statement probably just means that Volker Ludwig has developed (and tested) his command with Stata 12. Maybe he did not have access to to Stata 11 to verify his command would work, maybe he did not have the motivation to do so. You might want to contact him via the mail address given in the help file to get clarification on the issue.

For more on the topic read this FAQ.

Best
Daniel

[Guest]

I looked for such a FAQ, thank you very much! That is really helpful.