Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Felix Wädlich <fwaedlich@gmail.com> |
To | statalist@hsphsun2.harvard.edu |
Subject | Re: st: Test for autocorrelation - "sample may not include multiple panels" |
Date | Mon, 7 Mar 2011 18:05:46 +0100 |
Hi Markus, thanks for ur answer, I did not consider that. Unfortunately -dwstat2- and -durbina2- give me the same answer "sample may not include multiple panels". Regarding the unit root test: I also read, that non-stationarity could be a problem, and my regression does include GDP and GDP per capita, which are, as I understand, textbook variables for non-stationarity. The problem here is that most unit root test require a balanced panel. The one that seems to work for me, is the Dickey-Fuller-Test. But honestly, I am not sure how to interpret my findings, since the textbooks dont really cover this issue. When I use -xtunitroot fisher loggdppc, dfuller lags(1)-, I get this: Fisher-type unit-root test for loggdppc Based on augmented Dickey-Fuller tests --------------------------------------- Ho: All panels contain unit roots Number of panels = 135 Ha: At least one panel is stationary Avg. number of periods = 21.85 AR parameter: Panel-specific Asymptotics: T -> Infinity Panel means: Included Time trend: Not included Drift term: Not included ADF regressions: 1 lag ------------------------------------------------------------------------------ Statistic p-value ------------------------------------------------------------------------------ Inverse chi-squared(270) P 101.9727 1.0000 Inverse normal Z 12.1821 1.0000 Inverse logit t(674) L* 12.6890 1.0000 Modified inv. chi-squared Pm -7.2307 1.0000 ------------------------------------------------------------------------------ P statistic requires number of panels to be finite. Other statistics are suitable for finite or infinite number of panels. As I understand the literature a p-value>1 means nonstationarity. Am i right? Best regards, Felix 2011/3/7 Markus Eberhardt <markus.eberhardt@economics.ox.ac.uk>: > Hi Felix > > The tests routines you mention are for single time series only. For > the panel you can use the equivalents created by Kit Baum (panelauto: > http://ideas.repec.org/c/boc/bocode/s435102.html). > Given your data I should perhaps worry more about stationarity than > serial correlation. A detailed canon of panel unit root tests is > available here: > http://sites.google.com/site/medevecon/code#TOC-Panel-Time-Series-Tools > where you'll also find a number of other related issues (cross-section > correlation; coinegration; estimation). > If you insist on staying in the micro-panel estimator world, despite > your data being macro (which is the common attitude in the applied > literature), you should have a look at an explicitly dynamic model. > Bond (2002) in the Portuguese Journal discusses this is great detail > (naturally, for the micro panel case). > > > Markus Eberhardt > ESRC Post-doctoral Research Fellow, Centre for the Study of African > Economies, Department of Economics, University of Oxford > Stipendiary Lecturer, St Catherine's College, Oxford > > web: http://sites.google.com/site/medevecon/home > email: markus.eberhardt@economics.ox.ac.uk > twitter: http://twitter.com/sjoh2052 > mail: Centre for the Study of African Economies, Department of > Economics, Manor Rd, Oxford OX1 3UQ, England > > > > > On 7 March 2011 16:11, Felix Wädlich <fwaedlich@gmail.com> wrote: >> Hi Statalist, >> >> I have an unbalanced panel and need to test for autocorrelation (1978 >> to 2004, 100 to 140 countries). Due to my research design, I am very >> sure that I need to consider autocorrelation. Therefore I am including >> a lagged dependent variable (which also makes sense for theoretical >> reasons) as well as dummies for period (and unit effects). Basically I >> will first estimate an -xtpcse, corr(ar1)- and then -xtreg i.year, >> fe-. >> Judging from the literature the standard test for autocorrelation is >> -dwstat- (or -bgodfrey-). Unfortunately, Stata tells me "sample may >> not include multiple panels", and therefore cannot test for serial >> correlation.( No matter whether i use define my data set as -xtset- or >> -tsset-) >> I also tried -xtserial-, which works and indicates that my regression >> suffers from autocorrelation, but only without my fixed effects >> specification. So how can I tell, that autocorrelation is sufficiently >> adressed after my fe(timewise)-specification? >> Since I need such a test for my regression diagnostics, what can i do, >> what other options are there? Are there maybe graphical options as >> well? >> Thanks. >> >> Best regards, >> >> Felix >> * >> * 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/ >> > > * > * 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/ > * * 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/