I am estimating some panel ARDL models, and wanted to ask where the value add comes from with cointegration analysis. From what I've read, one can estimate long-run effects using an ARDL in levels as long as variables are I(0) or I(1) (Pesaran and Shin papers from 1990s say as much). I'm aware that spurious regressions are normally an issue when running time series regressions, so I assume appropriate specification of lag structures gets around this issue (?)
If we also find variables that are I(1) and cointegrated, we can include an error correction term and estimate an ARDL in first differences with the ECM.
Is this correct? If so, what are the advantages of the second step over the first?
Thanks very much!
