Index plot for residual diagtostics

[Anton Ivanov]

Hello dear forum members,

I came across this way to diagnose the variance of the residuals (please see attached):

- reg y x1 x2, vce(hc3) - robust SE for heteroskedastic data
- predict r1, res
- generate index=sum(1)
- twoway line r1 index

What is this "index" measure? I do see there is heteroskedasticity issue, but how shall I interpret the trend?

Thank you in advance,
Anton

[Rich Goldstein]

I can't see your attachment - please see the FAQ

the "index" is just the order of the data; this can be particularly useful when you think that problems might be related to the order of the data

[Clyde Schechter]

Your attachment is not visible, at least not on my set-up. Attachments tend to work poorly on this forum in any case.

The variable index is just the sequential observation number: it is 1 in the first observation, 2 in the second, etc. It could have just as easily been coded as -gen index = _n-. So the interpretation of the trend depends on the sort order of your data. If your data are sorted on some variable, then the trend will show how the residuals trend up or down with that particular variable.

If your data are not sorted on any particular variable, then I do not think the trend means anything at all.

You might also want to look at the built-in commands -rvfplot- and -rvpplot-, which you might find more helpful than this approach.

[Anton Ivanov]

I apologize for this technical issue. Please check now.

[Anton Ivanov]

Your attachment is not visible, at least not on my set-up. Attachments tend to work poorly on this forum in any case.

The variable index is just the sequential observation number: it is 1 in the first observation, 2 in the second, etc. It could have just as easily been coded as -gen index = _n-. So the interpretation of the trend depends on the sort order of your data. If your data are sorted on some variable, then the trend will show how the residuals trend up or down with that particular variable.

If your data are not sorted on any particular variable, then I do not think the trend means anything at all.

You might also want to look at the built-in commands -rvfplot- and -rvpplot-, which you might find more helpful than this approach.

Ok, I see. Thank you, Mr. Schechter. Once I sorted it on a particular variable, it made perfect sense.

Would (A) -boxcox- transformation, and/or (B) -vwls- with sigmahat=sqrt(resid^2) as estimate of conditional SD -- be proper ways to then check the robustness of the OLS estimation under vce(hc3)?