Problem with dynamic panel data regression

[Alistair Alleyne]

I am trying to estimate dynamic panel data (DPD) models but when I do using the XTABOND command the lagged dependent variable in both instances is being dropped because of collinearity. I am assuming that I can no longer use DPD analysis and would be forced to use Random Effects or Fixed Effects Modelling followed by Hausman Testing to determine which model to report. Am I right?

[Richard Williams]

Why do you have perfect collinearity? If it is because the DV is constant across time you can't do much of a DPD analysis. I would figure out what the collinearity problem was due to before doing anything else.

[Alistair Alleyne]

Hi Richard, thanks for your reply. I generated a correlation matrix in STATA and the result concerning the relationship between the dependent variable and the lagged dependent variable was 0.67 in one case and 0.98 in the other case.. Previous readings have led me to understand that one cannot use two variables in a model once the relationship result exceeds 0.90 or 0.95. So I am thinking that it is now impossible for me to use DPD analysis in the second case but perplexed why the lagged dependent variable is being dropped from the model in the first case when the relationship result is 0.67..

[Phil Bromiley]

To expand on Richard's comment, you'll increase your chances of a useful answer by following to FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

The problem of colinearity is not the association of the iv with the dv, it is the association of one iv with the other iv's. Regress your lagged dv on the other variables in the model and you are likely to find a very high r-square. I'm not sure if this is fully the case with xtabond type estimators but it is the case in general.

It is also tricky to have a lagged dv with that explains almost all the dv because the parameters on the other variables are estimated holding the effect of the lagged dv constant. There just isn't much left to explain if your lagged dv correlates so highly with the dv. There is a literature on this (I think in political science) which notes that you can get very strange results if your iv's change slowly - the lagged dv can essentially use almost all the variance that the other variables might explain giving strange parameter estimates on the other variables.