Unit root testing - LLC test vs. IPS test

[Mateo Zokalj]

Dear all,

I am using Stata 14.1 to analyze balanced panel dataset of 11 countries over period 2002-2014 (143 observations). As a first step, I want to check for stationarity of the series, in this example stationarity of social protection expenditures-to-GDP ratio. However, problem is that LLC and IPS tests give different results, but in the way that more restrictive LLC indicates stationarity, while IPS non-stationarity.

 xtunitroot llc ln_socpexp, trend demean

p-value = 0,000

 xtunitroot llc ln_socpexp, trend demean lags(1)

p-value = 0,000

 xtunitroot ips ln_socpexp, trend demean

p-value = 0,34

 xtunitroot ips ln_socpexp, trend demean lags(1)

p-value = 0,69

How should I interpret those results?

Thanks in advance

Mateo

[Emad Shehata]

check the difference of null and alternative hypotheses

. xtunitroot ips y, trend demean

Im-Pesaran-Shin unit-root test for y
------------------------------------
Ho: All panels contain unit roots
Ha: Some panels are stationary

. xtunitroot llc y, trend demean

Levin-Lin-Chu unit-root test for y
----------------------------------
Ho: Panels contain unit roots
Ha: Panels are stationary

[Mateo Zokalj]

How is it possible that LLC test indicates that all panels are stationary, but IPS indicates that all panels have unit roots?

[Emad Shehata]

Check again some panels not all panels

[Emad Shehata]

For more details about these restrictive in both two tests you can read:

http://homepage.univie.ac.at/robert...._pres_nell.pdf

[Mateo Zokalj]

Dear Emad,

thank you for the link which explains different unit root tests. However, I still do not understand how the results are so different.

[Emad Shehata]

you can also check Stata user manual

http://www.stata.com/manuals14/xtxtunitroot.pdf

But the stata user guied needs some corrections in page 7

xtunitroot — Pane ldata unitroot tests 7
The Im–Pesaran–Shin (xtunitroot ips), Fisher-type (xtunitroot fisher), and Hadri LM (xtunitroot hadri) tests allow unbalanced panels, while the remaining tests require balanced panels

this is not correct because

Im-Pesaran-Shin test cannot have gaps in data, and
Hadri LM test requires strongly balanced data
So, I hope from Stata Corp. to correct this mistake


all best wishes

[Nchofua Protus]

please in the case where we reject the null hypothesis in the "trend demean" option what will be the implication? can we proceed to regress the variable like that or one must take the first differencing option?