Time Dummies with Driscoll-Kraay

[Francesco Sama]

Dear Statalists,
I have estimated a fixed effect panel ( N=20, T=6) using xtscc. I'm experiencing some problems in interpreting the time-dummies coefficient: it seems strange to me that all thetime-dummies coefficients are positive and signficant. I'm considering Italian regional growth rates, as dependent variable, for the following decades: 1951-61, 1961-71, 1971-81, 1981-91, 1991-01. 2001-11. I leave out the first time dummies (1951-61) and I estimate a FE model using Driscoll-Kraay command, with 5 time dummies, since there is cross-sectional correlation. Here is my command end output:

xtscc GrowthRate LnGDPpc SocialCapital year2 year3 year4 year5 year6, fe lag(1)

Regression with Driscoll-Kraay standard errors Number of obs = 120
Method: Fixed-effects regression Number of groups = 20
Group variable (i): id F( 7, 19) = 14.46
maximum lag: 1 Prob > F = 0.0000
within R-squared = 0.9529

-------------------------------------------------------------------------------
| Drisc/Kraay
GrowthRate | Coef. Std. Err. t P>|t| [95% Conf. Interval]
--------------+----------------------------------------------------------------
LnGDPpc | -.0604739 .0055715 -10.85 0.000 -.0721352 -.0488125
SocialCapital | -.0039892 .0023667 -1.69 0.108 -.0089428 .0009645
year2 | .0270597 .0031496 8.59 0.000 .0204676 .0336519
year3 | .0411524 .0058991 6.98 0.000 .0288054 .0534994
year4 | .0524954 .0077714 6.75 0.000 .0362297 .0687612
year5 | .0585797 .0091433 6.41 0.000 .0394426 .0777168
year6 | .0501161 .0100311 5.00 0.000 .0291208 .0711115
_cons | .1421977 .0077877 18.26 0.000 .1258979 .1584976
-------------------------------------------------------------------------------



Since most of the Italian regions experienced higher growth rates in the first decade (1951-61), I don't understand why the time-dummies are all positive. What am I missing?

Best regards

Francesco Samà and Antonio Di Ruggiero

[Carlo Lazzaro]

Francesco (and Antonio):
- you used an user-written (i.e., non Stata built-in) programme: you should tell the list where you got it from (this remark is only seemingly pedantic: FAQ explains why the source of the user-written programme is relevant to report);
- year dummies: I'm not sure whether -xtscc- (I'm not familiar with that) supports -i.- notation for categorical variables (please, see -help fvvarlist- for further details on this wonderful command). If that were the case, -fvvarlist- outperforms by far creating categorical variables and intearctions by hand;
-last but not least: you're dealing with a large N, small T panel dataset. Hence, why not using -xtreg, fe- with robustified or clustered standard errors (in -xtreg-, unlike -regress-, they do the same job) to account for autocorrelation and/or heteroskedasticity?