> On Thu, 27 Dec 2012, Riccardo (Jack) Lucchetti replied: > > [...] >> When you write restrictions for the "restrict" command, no coefficients > must >> appear on the right-hand side of the equality. This is stated in the >> documentation. >> >>> Test statistic: F(3, 1016) = 3.27298, with p-value = 0.0205637 >>> >>> This test would appear to comprehensively reject the RE model in favour > of >>> retaining the FE model. Or does it? Is this an acceptable test in place > of >>> the Wald test for jointly equal parameters that I can't run in -gretl-? > Allin Cottrell wrote: > I'd just add to Jack's reply: this _is_ a Wald test for > jointly equal parameters (the F-form of the test). If you > multiply the test statistic by 3 and refer it to the > chi-square(3) distribution you get the same p-value as given > above. Excellent stuff - I thought I was on the right track! Nice to be able to deploy some correct logic for once. Thanks for pulling me up on not reading the documentation, but in my defence: (a) I've only just installed -gretl-; (b) it's Christmas, so I've not had much of a chance to peruse what is a very large users' guide; and (c) I did look at the program help on restrictions (although clearly they're not as comprehensive as that laid out in the guide). If I may, I had another quick query about panel-corrected standard errors. I'm delighted you've included them in -gretl-, but I've had no luck in fitting any fixed-effect models with PCSEs (N=47, T=36, NT=795, with ten X variables; all bar the first and last of the time dummies were included; there was no lagged dependent variable). Every time I do, the model does run, but with the message "Could not compute Beck-Katz errors". I'm sure the reasons for not doing so are valid, but no reason was given at all.

Also, which standard errors are reported instead?

Allin Cottrell <cottrell(a)wfu.edu&gt; replied: Beck-Katz does: > > Var(b) = A^{-1} W A^{-1} > > where A = \sum_{i=1}^n X'_i X_i, > W = \sum_{i=1}^n \sum_{j=1}^n \sigma_{ij} X'_i X_j > and \sigma_{ij} is estimated as (1/T) \sum_{t=1}^T e_i e_j, > with the e's being OLS (or fixed effects) residuals. > > The trouble is there's no guarantee that W (which is supposed > to be a variance measure) is positive definite in unbalanced > panels; if it's not, we "fail". > >> Also, which standard errors are reported instead? > > The "classical" ones. If the Arellano method is used you'll > see "Robust (HAC) standard errors". Fair enough, Allin, although I don't think I'll ever be fitting FE models to balanced panel data any time soon! Thanks once again.

On 27 December 2012 22:37, Allin Cottrell <cottrell(a)wfu.edu&gt; wrote: Not all unbalanced panels provoke a problem with PCSEs. From brief > experimentation it seems that strong collinearity among the > regressors is a contributory problem. Which is what I have experienced: the very last time dummy in my model is dropped due to collinearity, which provokes two questions: (1) how do I prevent that (T-2) dummies from being dropped? I don't necessarily want to drop any X variables from the model;

(2) as there is no LDV, how do I ensure that I select robust SEs that do not correct for autocorrelation?

On 28 December 2012 02:16, Allin Cottrell <cottrell(a)wfu.edu&gt; wrote: > Select a set of regressors that are as near orthogonal as possible > ;-). Seriously, gretl is better able to handle highly collinear > regressors than most econometric software; if a regressor is > dropped, it's probably not possible to estimate its coefficient > other than via multiple precision arithmetic (which gretl offers, > but for OLS only). Okay, that's sensible advice, but with this data, these are the only variables I have: they were chosen for their relevance, as opposed to their orthogonality! > > (2) as there is no LDV, how do I ensure that I select robust SEs that do > > not correct for autocorrelation? > > That seems to me a non-sequitur. Why should the non-inclusion of an > LDV immunize a model against an autocorrelated error? Rather the > reverse, I would think: if you include enough lags of the dependent > variable, then maybe you have "whitened" the error term to the point > where correction for autocorrelation is redundant. There is no LDV in the model because - correct me if I'm wrong - that would render the parameter estimates biased in a fixed-effects LSDV model (Judson and Owen, 1999). The solution to that would be to move to IV models, but (a) they're not for kids; and (b) I've not yet hammered out a requisite battery set of IVs to even think about making such a model work! > Allin Cottrell > _______________________________________________ > Gretl-users mailing list > Gretl-users(a)lists.wfu.edu > http://lists.wfu.edu/mailman/listinfo/gretl-users > -- Clive Nicholas (clivenicholas.posterous.com) [Please DO NOT mail me personally here, but at <clivenicholas(a)hotmail.com&gt;. Please respond to contributions I make in a list thread here. Thanks!] Judson RA and Owen AL (1999) "Estimating dynamic panel data models: a guide for macroeconomists", Economics Letters 65(1), 9-15

On Fri, 28 Dec 2012, Clive Nicholas wrote: > May I come back to (either of) you on my Wald/F-test query? When you say > the F-test score should be multiplied to give the Wald test score, > presumably the multiplier is the number of restrictions in the test? Yes. > Even if this is correct, I don't _quite_ get the same > P-values: > > (After running the GLS RE model on my pooled panel data:) > Test statistic: F(5, 749) = 4.65984, with p-value = 0.000339063 > > Chi-square(5): area to the right of 23.2992 = 0.000295879 > (to the left: 0.999704) > > They're close, but not exact! Well, no. The chi-square p-value, being asymptotic, is equivalent to the p-value of the corresponding F-test with infinite degrees of freedom in the denominator -- and 749 is not quite infinity. Allin Cottrell _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

On Fri, 28 Dec 2012, Clive Nicholas wrote: > On 28 December 2012 02:16, Allin Cottrell <cottrell(a)wfu.edu&gt; wrote: >>> (2) as there is no LDV, how do I ensure that I select robust SEs that do >>> not correct for autocorrelation? >> >> That seems to me a non-sequitur. Why should the non-inclusion of an >> LDV immunize a model against an autocorrelated error? Rather the >> reverse, I would think: if you include enough lags of the dependent >> variable, then maybe you have "whitened" the error term to the point >> where correction for autocorrelation is redundant. > > There is no LDV in the model because - correct me if I'm wrong - that would > render the parameter estimates biased in a fixed-effects LSDV model (Judson > and Owen, 1999). True. But I didn't mean that you have to include an LDV in your panel model. I was just questioning what I took to be your assumption: that since you don't have an LDV, you don't have to worry about autocorrelation of the error term when selecting a robust variance estimator. That doesn't follow. Allin Cottrell _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users