I am conducting research on panel data where N=16 and T=24. After conducting panel unit root tests I would find that some of my variables are integrated of order one, however, they share no cointegrating relationship. As there is no cointegrating relationship between the variables that are integrated of order one I continued to proceed with the relevant tests (f-test, Breusch Pagan, Hausman) between pooled, fixed effects and random effects. However, I am confused in regard to if the relevant tests should be conducted using the regressions with variables that are in first differences, or should I conduct the tests using the variables in levels and then once I know which model is correct then regress the model using the variables in first differences.