fixed and random effects model.

[Kenulina Schneider]

Dear all,

I am working with panel data( 9 countries and 12 years). I want to analyse the impact of stock market development on economic growth. For estimation I run OLS pooled, FE and RE models using Stata 12.0. I have 3 main explanatory variables and 3 control variables. By using robust and cluster(id) option when running regression I receive very insignificant results and moreover my coefficient are not consistent with most of the literature. I am trying to keep the same steps that is conducted in the literature my work based on.
In particular I am running the following codes:

regress dep_var indep_vars, robust cluster(id)
xtreg dep_var indep_vars, fe robust cluster(id)
xtreg dep_var indep_vars, re robust cluster(id)

And my data is suffering from heteroskedasticity and serial correlation. Therefore I use the robust cluster(id) option. But Hausman test does not allow those options, so run both FE and Re models without those options. And Both Hausman and LM tests show that RE model is appropriate. But the results are very ambigious. Also can anyone please explain the difference between xtreg dep_var indep_vars, fe robust cluster(id) and xtreg dep_var indep_vars i.year, fe robust cluster(id)? In both cases p-values are not significant. Now Im stuck here, not sure what to do. I also have doubts about my model, don't know whether to take log variables or not. The model is based on the scientific literature:
lnGDP_t=a₀+a₁lnX_t+a₂lnZ_t+a₃lnK_t+u_t

where a₀ estimated coefficients and u_t is an error, t is time. Is this model is Ok? I mean can I run OLS, FE and RE using log variables?
I don't have a clue in Stata and Econometrics, only started to learn in order to work on my thesis.

I appreciate any help. Thanks a lot!

Kenulina

[Carlo Lazzaro]

TKenulina:
1) robust clustered standar errors (SEs) account for serial correlation of residuals between observations, heteroskedasticity and dispersion of the coefficient estimates across clusters (not observations), since under the panel structure you have the same id measured multiple times across the time-series indentifier, That said, it is not surprising that most of your predictors fail to reach statistical significance (however, I do not consider this a problem, because, as reported by Altman & Bland "absence of evidence is not evidence of absence" (http://www.bmj.com/content/311/7003/...iant=full-text). Moreover, you do not report how many observations you have when contrasted against the number of predictors (the rule of thumb requires 20 observations per predictor¹): this may be another reason why statistical significance seems out of reach (but Altman & Bland's cautionary tale still holds).

2) Since your default SEs seem to widely differ from the clustered ones, you may want to further investigate this issue via a robust hausman test (http://www.stata.com/statalist/archi.../msg01069.html) which specification (i.e., -fe- or -re-) fits better your data under -xtreg-.

3) There's no diference between -cluster(id)- and -robust cluster(id) SEs, as you may have probably already noticed inspecting your results.

¹Katz MH. Multivariable Analysis. Second Edtion. NY: Cambridge University Press, 2006: 81.

[Kenulina Schneider]

Thank you very much!

I think the problem is due to multicolleniaruty among key explanatory variables. After running stepwise regression p-value seems significant. With regard to observations against the number of predictors, Stata exhibits the following:

Number of obs=107
number of groups=9
obs per group:min=11
avg=11.9
max=12
(I dont know if you mean that)


"3) There's no diference between -cluster(id)- and -robust cluster(id) SEs, as you may have probably already noticed inspecting your results"
Here i wanted to ask the difference between xtreg yvar xvar i.year, fe and xtreg yvar xvar, fe (without i.year). I mean if i shoud include time dummy variables. However the testparm i.year option showed that there is no need for time fixed effects. So, is that mean that I only need to run xtreg yvar xvar, fe without i.year?

xtoverid option however shows error (601) . I could not manage to solve the problem. I have to check in another computer maybe.

Best regards,

Kenulina

[Carlo Lazzaro]

Kenulina:
-as you can read in previous threads of this list, there's widespread and methodologically sound idea that -stwepwise- is hardly a good idea, in that you often end up finding statistical significance in a sort of window-dressing statistical model, potentially miles away from your original dataset;
-your problem may well be multicollinearity, but, given the number of your observations, I suspect that you are also asking too much out of your predictors in terms of expected statistical significance;
- about the dummy issue, if you have -xtset- your panel as -xtset idpanel year-, -i.year- might be dropped due to multicollinearity;
- as far as -xtoverid...- issue is concerned (and as a general rule), please report exactly what you typed and what Stata gave you back (as per FAQ).

[Kenulina Schneider]

I type the following commands:

Xtreg yvar xvar, fe robust
estimates store fixed
xtreg yvar xvar, re robust
estimates store random
xtoverid

but Stata shows: Error- must have ivreg2/ivreg29/ivreg28 version 2.1.15 or greater installed
I don't know how to fix this problem.

Thank you!
Best regards,

Kenulina

[Mark Schaffer]

Umm... you fix it by installing ivreg2 or one of the other named programs. You'll also need to install ranktest.

ssc install ivreg2
ssc install ranktest

[Kenulina Schneider]

Dear Schaffer,

Thank you very much!

Now it works!

Best, Kenulina

[Kenulina Schneider]

Dear all,

I have questions regarding the models to be estimated in my thesis. I decided to ask question here (did not create new topic), don't know if somebody will react.
As I said earlier, I am analysing the relationship between Financial development and economic growth. Can anybody explain the difference between the following models (examples from different literature): lnGDPt=a0+a1lnMCRt+a2lnNSTt+a3lnVSTRt+a4lnGIt+a6lnNTt+u (1)

Growth =α_t' +γ_t' +φ₁ (Growth_it-1 )+φ₂ (GDP _it-1)+φ₃ (INVF_it-1 )+φ₄(FDI_it-1 )+φ₅ (SE_it-1 )+ε' (2)

here I mean that in some literature the data are expressed in their logarithms to control for convergence(Model 1), while in Model 2 the author says that "we add to the set of regressors the lagged dependent variable in order to capture the dynamic effect in the stock market development and to get rid of possible autocorrelation". In latter model the instrumental variable 2SLS approach is applied ( since the presence of lagged dependent variable implies correlation between the error term and the lagged dependent variable, rendering the OLS estimator biased and inconsistent).

So, my questions are:
1). Can I run the model like the first one? Does ln mean that i need to generate log for all variables and then estimate model with using OLS pooled, FE or RE effects model.
2). In some articles they say that variables must be stationary in time series data, but since I am working with panel data can I ignore this assumption? (I don't use VAR panel model)

I would appreciate any help!

[Carlo Lazzaro]

Kenulina:
as far as your question #1) is concerned, yes, you can perform a log-log regression model.
However, logging both hand-side of the regression brings about some issues concerning the interpretation of your results, in that, quoting Stock JH and Watson MW. Introduction to Econometrics. Second Edition. Boston, MA: Pearson International Edition, 2007: 273: [QUOTE]A 1% change in X is associated with a β₁ % change in Y, so β₁ is the elasticity of Y with respect to β₁ [/QUOTE].
Although Authors referred to a simple log-log regresssion model in their textbook example, the substantive stuff holds for multiple log-log regression model, too.
I am quite unfamiliar with time series so I cannot reply to your question #2).

[Kenulina Schneider]

Carlo Lazzaro,

Thank you very much for such helpful suggestions!

Best regards,
Kenulina