Coefficients become 100 times larger when adding the county fixed effects.

[Alex Carr]

Hello all.

I'm trying to regress the housing value at parcel level i on some average housing and socioeconomic characteristics at the county level j that the house i is located in.Because the property value of house i is not observed every year so I have a non-panel data. Now when I add some county fixed effects to the original OLS regression, the coefficients for those housing and socioeconomic variables at the county level become 100 times larger in magnitude and also become insignificant. Is this normal and how do I interpret this results.

Thanks all.

Alex

[Clyde Schechter]

I think you will get a more helpful answer if you show example data (use -dataex-) and also show the exact commands and Stata outputs from both versions of the regression. Speaking in generalities, what you describe is possible, although not often seen. Nothing more can be said without more information from you.

[Cem Payaslioglu]

Hello,

If your data has two periods or can be made as such, then I have seen a similar case right under the title "Policy Analysis with Pooled Cross Sections" from Wooldridge's "Introductory Econometrics" textbook ( the example on page 414 (2nd edition), see a portion attached below). The example is about the price effects of the proximity of the real estate to the planned incinerator project on real estate values . The author shows that a significant negative coefficient may not be an indicator of the depressing price of the incinerator as using only a cross section of the first period do not correspond to starting time of the project. He then goes on to show that using a two period model incorporating the aftermath of the project within the framework of treatment/ control setting (before and after periods) and using an interaction dummy, a proper assessment can be made . I guess your data is pooled cross section type, instead of using country effects, you might want to adapt this example to your case.

Wooldridge, do not explain how this can be extended to over two periods though ( Perhaps it is mentioned in his advanced book on Panel Data).

[Alex Carr]

I think you will get a more helpful answer if you show example data (use -dataex-) and also show the exact commands and Stata outputs from both versions of the regression. Speaking in generalities, what you describe is possible, although not often seen. Nothing more can be said without more information from you.

What I thought is that coefficients for county level characteristics, such as race, age and income are captured by the county fixed effects, which made them statistically insignificant. Do you concur ?

[Carlo Lazzaro]

Alex:
why not sharing wjat you typed and what Stata gave you back as per Clyde's helpful advice?

[Clyde Schechter]

In #1 you said that the coefficients for the county-level attributes became 100 X larger. Ordinarily you would expect them to become smaller when you add the county-level fixed-effects. But there is no guarantee of that. Also the addition of the county-level fixed effects turns the analysis into an analysis of within-county effects, which can be very different from the between-county effects that play a larger role when you don't have those fixed effects.

My point is that almost anything can happen when you add fixed-effects to a model. If you show the specifics it might be possible to figure out whether what you're seeing is within this very broad range of possible differences, or whether there is a mistake somewhere.