Panel Data: Omitted Variables due to Collinearity

[Izzat Halim]

Hello everyone, referring to my current study mentioned here: https://www.statalist.org/forums/for...c-price-method

Currently, I have a total of 5 housing variable, 14 distance variable (12 are distance to parks), and 12 size of parks variables. The data is for 6 years and there are 645 observations. My initial plan was to only look at the distance of the closest park and its respective size (not to include all 12 parks). However, I noticed that the two specific variables are not significant. Explains why I decided to include all parks variables in the analysis.

The plan was to run three separate models (Model 1 consist of housing variables; Model 2 consist of housing and distance variables; Model 3 consist of housing and size of parks variables).

I did not experienced any problems with the first two models whereby I ran the Pooled OLS, xtreg, re and xtreg, fe. However, coming to the third model the minute I ran regress for housing and size variables, the results informed me all of the size variables were removed due to collinearity. Likewise for re. I know it does so for fe as the size variables are time invariant in which some of them are the same throughout the years (correct me if my definition is wrong).

I have also done the necessary test to identify which model between the three are more advisable to be used.

I don't plan on using xtreg, be as the OLS and random effect model is much efficient (?).

Would appreciate if anyone could provide comments or advice on the matter.

[Carlo Lazzaro]

Izzat:
I'm afraid I cannot reply positively with no further quantitative details about your data (example/excerpt shared via -dataex-).

That said, I fail to get why you ran pooled OLS if you actually had evidence of panel-wise effect.

[Izzat Halim]

Hi Carlo,
This is the result I have obtained when I first ran the 3rd model. I am unsure on how to share via dataex.

As to why I ran pooled OLS, I think its because I followed the manuals provided blindly on basics of running panel data analysis on stata.

[Izzat Halim]

This would be the result from RE

[Carlo Lazzaro]

Izzat:
there's something wrong in your data and/or in your model specification.
If the F-test appearing at the foot of the -fe- (no -re-) outcome table is statistically significant, get rid of of pooled OLS (that you actually did not perform, as you did not clustered the standard errors on your -panelid-) and test via -hausman- if -fe- or -re- are good for your data.