Regression assumption

[Brian Conway]

I'm running regression models (3 dummies and 2 controls (numerical)). To check for linearity, I ran graph matrix (attached). I'm wondering if this indicates issue or not.
Also, I read that the <robust> command can take care of non-normality in the residuals distribution. I was wondering if this is correct? Thank you.

[Carlo Lazzaro]

Brian:
1) sharing what you coded and what Stata gave you back can help (as per FAQ);
2) the -robust- option takes care of the heteroskedasticity of the residual distribution in -regress-, for which normality is a (weak) requirement.

[Brian Conway]

Thank you very much Carlo. I came across this earlier post (https://www.statalist.org/forums/for...ple-regression), which seems to be relevant to my issue. I have a small sample (obs=about 60 in some cases, 40 in others). This post mentions that the diagnostic tests assume a large enough sample size. I already corrected for AC, hetero and MC. Does this earlier post advice apply in my case? Thank you for your help.

[Nick Cox]

The question of linearity only arises for non-indicator variables.

[Brian Conway]

Thank you very much Nick. I ran acprplot for a non-indicator variable in two sets of models. x1 is numerical variable, x2 is the log of the same variable. The code/output is below/attached:

acprplot x1, lowess
acprplot x2, lowess

The pattern for the first plot didn't look linear. Using the log of the variable in the second plot, it didn't seem to make much difference.

The pattern for the third plot looked linear to me (as with the graph with the log of the same variable in graph 4).