Project CEMAPRE internal
|Title||Application of self-normalization approaches for statistical testing in panel data|
|Participants||Paulo M. M. Rodrigues, Nuno Sobreira (Principal Investigator)|
|Summary||Panel data is a fairly prevalent type of data structure used in applied econometrics and it is|
becoming more important nowadays with the increased availability of data. An important assumption
for panel data analysis and modeling is to ensure its structural stability throughout time.
Hence, in this project we start by proposing a new statistic to test for the presence of a mean
shift in a set of panel data observations.
We consider a type of panel CUSUM test where a self-normalization approach is used to correct for
serial correlation. Hence, instead of using one of the standard robust estimators for the long-run
variances, our suggestion is to divide the statistic by a quantity that cancels the long-run
variance in each panel so that the corresponding asymptotic distribution becomes free of these
nuisance parameters. This alternative route has already been proposed and developed for time series
by a number of authors and it has already been proven to be a relevant alternative to the
traditional HAC estimator in the context of time series analysis. In particular, self-normalization
avoids the always difficult choice of the kernel function and corresponding bandwidth parameter in
empirical applications. Moreover, the use of HAC estimators in mean shift test statistics is
associated with either severe size distortions or nonmonotonic power functions.
However applications of a self-normalization approach to panel data data are nonexistent or
relatively scarce. Hence we study in depth this alternative route to correct for serial correlation
in panel data analysis in the context of testing for the presence of shifts in the mean of at least
some of the panel units. The next step of this project is to analyse the relevance of this
self-normalization approach to other statistical testing problems in the context of panel data