Initial values not feasible

[Muhammad Azam]

I tried to run ARDL/PMG using 250 observations panel data but on typing the command below and its many variations i still got the error:

COMMAND
. xtpmg d.log_prod_cotton d.log_price_cotton d.log_price_rice d.log_price_wheat ,lr(l.log_prod_cotton log_price_cotton log_pric e_rice log_price_wheat) ec(ECT) replace pmg

ERROR
initial values not feasible
r(1400);

Please guide me.

[Nick Cox]

It means what it says, and good guidance is really hard without knowing your data -- and in my case without knowing this command at all.

Here are some generic possibilities. All are possibilities; there can be no certainty about any from a remote viewpoint. Finding out means being active and experimenting.

* You need to reach in and suggest initial values yourself. Sometimes the default starting points are absurd in terms of your data.

* You are trying to fit a model that is too complicated. Start with a (much) simpler model and build up. (You seem to be throwing everything into the model, including interaction terms, which is often not a good idea.)

* There are strong collinearities in your model that inhibit fitting. Have you looked at a scatter plot matrix, correlation matrix, PCA results? (Odds are that you're an economist and 99% of your training is on models and 1% (in your half-forgotten first course) was on data analysis. But watch out: some economists regard looking at the data as a bad idea.)

* There are other awkward details in your data that are inhibiting fit. In this case, everything seems to be based on taking logarithms first, but outliers and skewness can pop out of a log transformation too.

* Perhaps the model isn't suitable for the data or vice versa. (The first meeting of model and data is like any first date. It can go very badly, or very disappointingly, if the couple aren't suited at all.)

* The sample size may be too small. How many parameters are you estimating? Rules of thumb usually suggest having many, many points per parameter.

[Suborna Barua]

Hi Nick, thanks for you so detailed suggestions. For me, I have 102 countries with 52 years data on 8 variables. I followed your suggestions and nothing seems working. Except that I could not follow your First Suggestion. Could you please advise me on how I can set the initial value? Thank you indeed.

[Nick Cox]

Sorry, no. I can't add to the advice in #2 at all.

The email is flattering: the implication is that I have super powers and know what makes sense for panel data we can't see and an unspecified model and an unspecified command. I really don't.

More importantly, I don't see that anyone can help without much more detail.