Fixed and random effects in cross sectional data

[Adeola Kolade]

Dear all,

1. I have a cross sectional data which has information on siblings. Will i be able to do a sibling fixed effect and how do i do this on Stata please.

can i do this in the cross sectional data:

clogit br003_ br002_ br001_ br005d2, group (siblings)

Or can i do this using the cross sectional data
xtset siblings
xtreg br003_ br002_ br001_ br005d2, fe

2. With a cross sectional data is it possible to run a random effects model and will it just be
reg br003_ br002_ br001_ br005d2, re

I will really appreciate your reply.

[Carlo Lazzaro]

Adeola:
if you have cross-sectional data with a continuous regressand, you should go -regress- instead of -xtreg-.
You can also plug in a categorical predictor for siblings (-i.siblings-).
Eventually, your code in 3) would simply not work and Stata will return an error message about the unfesibility of the -re- option with -regress-.

[Adeola Kolade]

Many thanks for your reply. If i get what you mean, you are saying i should run the first model with the cross sectional data like this:
reg br003_br002_ br001_ br005d2 i.siblings

and if we assume br003_ is binary can we also say:
probit br003_br002_ br001_ br005d2 i.siblings

we do not need "xtset siblings"

also random effects cannot work with a cross sectional data.

[Carlo Lazzaro]

Adeola:
- your first take about -regress- is correct (some concerns about using default standard errors may creep up, though). Obviously, no need to -xtset- your dataset;
- -clogit- implies something different from -probit- (please, see related entries in Stata .pdf manual for more details);
- -re- and -fe- are simply not appropriate with cross-sectional data.

[Rich Goldstein]

Carlo Lazzaro I'm not entirely sure what is meant by the last comment in #4; however, in both medicine and education (e.g., patients nested within hospitals or students nested within classes), it can and does make a lot of sense to use RE for cross-sectional data; this may be due to different meanings of RE in different disciplines

[Carlo Lazzaro]

Rich is obviously correct.
In replying, I was not thinking of Rich'examples, that actually call for -re- via mixed models due to the nested design.

[Adeola Kolade]

Many thanks for your replies, from @ Carlo Lazzaro second point if outcome if binary i suppose we can choose between using probit or logit and for fixed effects we can use xtprobit or clogit.

I am estimating the effect of parental smoking during childrens childhood on childrens adult outcomes. I have a retrospective data for parents smoking taken from wave 1 and i have outcomes of children during adult years from wave 2. Will i be able to estimate a random effect as data is practically cross-sectional.

[Carlo Lazzaro]

Adeola:
I think that you should go cross-sectional here.
If your regressand is binary (say, probability that children become regular smokers when adult), you should code something like:

logit adult_smoker i.parent_smoker <controls>

[Adeola Kolade]

Many thanks Carlo lazzaro