How can I test assumptions in fixed effects regressions

[Ninate Sum]

Hi all,
I have data on online hotel ratings, over a timespan of 5 years and including approx. 5600 hotels.
I am trying to run a fixed regression to see what influences hotel ratings (e.g. gender, age, etc.). I.e. I am group mean centering my variables at the hotel level.
Now I am trying to test regression assumptions, but the commands I use for normal regressions don't work for fixed effects regressions. Can anyone help with this?
Thanks!

[Laura Zutphen]

Did you already take a look at

xi: xtreg y x1 x2, fe

?

[Ninate Sum]

Hi Laura,
yes, thats the code I'm using for the regression. The output looks alright, but I don't know how to assess the underlying assumptions (homoscedasticity, no milticollinearity, etc.).
To me it looks as if quite a few users seem to have that problem, but no one found a soultion so far...
Can I maybe just group mean center my variables and then run a normal regression and test assumptions?

[Carlo Lazzaro]

Ninate:
welcome to the list.
Please, as per FAQ #12, post what you typed and what Stata gave you back.
Besides; please (as per FAQ again), note that statements along the lines of "did not/does not work" are seldom useful, as they do not convey any clue about the potential cause of the problem (if any problem exist). Thanks.

[Phil Bromiley]

You can use i.hotel in reg to run the fixed effect regression with hotel as the panel. It will take a little time and a large matsize, but then all the diagnostics from reg are available.

Phil

[Ninate Sum]

Perfect! That worked. Thanks so much!!
Now the next hurdle: The two approaches (xtreg vs. dummies) produce the same coefficients, but a different F-statistic and different R-squares. Which ones should I report?




---------------------------------------------------------------------------------------------------------------------------

[Carlo Lazzaro]

Ninate:
You should consider the output of -xtreg, fe-.
You should have clustered on hotel.id the standard errors of the pooled OLS, as your observations are not independent.

[Ninate Sum]

Hi Carlo,

Like this? To me it looks as if nothing changed...

[Carlo Lazzaro]

Ninate:
Pooled OLS ( -reg-), not -xtreg-.

[Ninate Sum]

Hi Carlo,
so now I ran the whole thing with pooled OLS. But now its not a fixed effect regression anymore, right?
Or should it be
reg x y i.HotelID, cluster() ?

[Carlo Lazzaro]

Ninate:
it should be:

reg depvar indepars i.HotelID, cluster(HotelID)

[Ninate Sum]

Just one more question regrading the output above (for xtreg x y, fe): What do rho and the "F-test that all u_i=0" tell me? Do the values indicate any problems? I think I remember that the F-test should not be significant. Is that correct?

[Carlo Lazzaro]

Ninate:
if the "F-test that all u_i=0" appearing as a footnote of the -xtreg,fe- outcome table is significant, that specification fits your data better than a pooled OLS.

[Ninate Sum]

Perfect! That is good news
But is the rho of 0.57 an issue? Or rather - what does that tell me?

[Carlo Lazzaro]

Ninate:
no, it is not.
-rho- (intraclass correlation) gives you the fraction of variance explained by individual effects (across panels variance).
Besides Stata .pdf manual related entries, see for further details: https://www.princeton.edu/~otorres/Panel101.pdf.