However, in spite of their being less general, residual-based tests turned out to be the most popular in
empirical studies. Indeed, as explained by Westerlund and Basher (2008), parametric tests, such as the tests
by Larsson et al. (2001) and Groen and Kleibergen (2003), that can accommodate quite general
dependencies require to empirically determine the appropriate order p for the autoregression, which is
typically unknown (and only in this case these tests will not to depend on nuisance parameters). However, if
p is underestimated, there will still be a problem of nuisance parameter, whereas if p is overestimated, the
small-sample properties of the test will deteriorate, and this might have a significant impact on
cointegration test performance in small samples. On the contrary, the two residual-based panel data tests
for the null of no cointegration of Westerlund (2005), which can be viewed as ‘fully nonparametric’ (since
no correction for the temporal dependencies of the data is needed), do not require the empirical
determination of lag or bandwidth parameter, which ‘of course reduces the uncertainty and ambiguity of
the tests outcome’.