Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# RE: st: RE: RE: comparing different means using ttest

 From Nick Cox To "'statalist@hsphsun2.harvard.edu'" Subject RE: st: RE: RE: comparing different means using ttest Date Fri, 17 Dec 2010 13:33:09 +0000

```I'd leave economists to discuss that one. My larger point remains that applying tests that ignore time series structure to data that are time series is a dubious and dangerous thing to do.

Nick
n.j.cox@durham.ac.uk

DE SOUZA Eric

" The regression still assumes independent error terms."

True. But GDP does often behave as a random walk (with structural breaks, may be). Hence the errror terms are very likely to be uncorrelated.

One could also robustify against serial correlation in the error terms.

Nick Cox

The regression still assumes independent error terms. There is more scope for doing something about that in a regression framework then within -ttest-, but in terms of what Eric suggested it is still a matter of six on one side and half-a-dozen on the other.

DE SOUZA Eric

It does, because it simply avoids the starting point of David Lempert which in my opinion is a false start: regressing GDP levels on a time trend will get you nowhere. If David is interested testing the equality of GDP growth rates across two time periods, you pool the data, calculate the GDP growth rate and regress this variable on two dummy (binary) variables for each time period. In order to avoid perfect collinearit you drop one of the two dummies and test whether the coefficient on the other is equal to zero.

Steven Samuels

But. Eric, I don't think that pooling will solve the dependence issues that Nick mentioned.

On Dec 16, 2010, at 1:26 PM, DE SOUZA Eric wrote:

Why not just pool your data and regress %GDP-growth on a dummy
(binary) variable (and a constant, of course) which takes the value of one for one of the two sub-samples and zero for the other; and test whether the coefficient on the dummy is significantly different from zero (or examine its confidence interval) ?
You can robustify for heteroscedasticity.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```