Test for sufficient sample size in linear probability model

[Nik Mayer]

Hi,

I run the following linear probability model:

reg nodad k1sex i.meduc i.state i.mdecade i.mrace mage mage2 mage3 [pweight=wf]

whereby nodad is a dummy indicating if the family is living with the father, k1sex is a dummy for the sex of the first child and the other variables are controls.

The variable of interest, k1sex, is insignficiant with a p-value of around 0.3, sample size is 19,588.

Is there a test to check whether the coefficient is insignificant because there is no statistical correlation or because the sample is just too small to infer significance?

And if so, how can I calculate the necessary number of observations?

Thanks in advance,
NM

[Carlo Lazzaro]

Nik:
I'm not aware of such a test.
However, your sample size seems large enough to capture stastistical significance. It way well be that the effect does not exist.
Just out of curiosity I would type -estat vce, corr- after -regress-, to take a look at coefficients correlation.

[Nik Mayer]

Thanks for your response Carlo!

[daniel klein]

You could take a look at

help power

However, I would first look at the estimated coefficient and its standard error. If the coefficient appears scientifically/substantially relevant and has a huge standard error, then power might be a concern; also I seriously doubt that will be the case with a sample size of ~ 20,000.

Also, mage2 and mage3 seem to be transformations of mage; if this is so, see

help fvvarlist

to learn how to include such transformations in your model in a way that post-estimation commands can understand them.

Best
Daniel