1
$\begingroup$

I busy with a regression model that seems to have heteroscedasticity. The model has 6 independent variables and one dependent variable. I did the regression and noticed heteroscedasticity. I then eliminated the observations with very high residuals +/- 5% of the total observations which is 25000. I did the regression again without these high residual observations and found that this new regression does not have any signs of heteroscedasticity. I would like to know: is a process like this acceptable or not?

$\endgroup$
2
  • $\begingroup$ "I busy ...": I haven't edited what I don't understand: perhaps dialect or slang beyond my region or age group. "+/-" presumably means about. Did you delete observations with just large positive residuals, large positive and large negative, etc.? $\endgroup$
    – Nick Cox
    Commented Feb 22, 2015 at 14:56
  • $\begingroup$ Everything is acceptable, it just depends to whom;). When you say regression, you mean plain vanilla OLS? $\endgroup$
    – user603
    Commented Feb 22, 2015 at 17:55

1 Answer 1

2
$\begingroup$

You ask for a yes-no answer, but answers might vary greatly. I won't be the only member here unwilling to approve (or condemn) what I can't see.

From this description I'd say that the answer might vary from

  • that was brutal and utterly ad hoc, but it is just possible that you got a fair model with poor methods

through

  • we can't tell, because just mentioning heteroscedasticity does give any context to judge what is a good model; we need to see the data and know what model you tried (there is zero formal content in your post, although if I had to guess it's linear regression on the variables as given with no extra complications)

to

  • no; it's unacceptably poor practice to drop observations because they happen to be inconvenient compared with simplistic assumptions that would be nice if true.

Personally, I go with all of these, but my bottom line is as just given.

On this information, I would suggest some priorities:

  1. Heteroscedasticity is arguably a secondary problem to do with error structure; you may need much more emphasis on getting the functional form and estimation method right.

  2. Additionally, using transformations or a non-identity link function is a much better way to proceed than dropping some fraction of the data.

I'd add that as your analysis was demonstrably two-phase, the resulting standard errors and P-values from the second leg don't mean much and should certainly not be reported as if the dataset arrived as you left it.

$\endgroup$
6
  • $\begingroup$ As a counter example to the dangers of unwarranted transforming autobox.com/pdfs/vegas_ibf_09a.pdf slides 14-24 . Transformations should be based upon the distribution of the residuals from a model not on the distribution of the original series because parametric tests require a central ch-square distribution. Another possible translation for busy is "am currently working with" $\endgroup$
    – IrishStat
    Commented Feb 22, 2015 at 15:35
  • $\begingroup$ No one approves of unwarranted transformation, but I'd say your dictum (Diktat?) on transforming because of how residuals behave is contestable. If the pattern is (just one example) $y = a \exp(bx)$ it can make sense to work with $\ln y$, almost regardless of the exact pattern of residuals. $\endgroup$
    – Nick Cox
    Commented Feb 22, 2015 at 16:16
  • $\begingroup$ The rubber hits the road when you try and use standard parametric tests of significance and sufficiency when the residuals are non-normal. $\endgroup$
    – IrishStat
    Commented Feb 22, 2015 at 18:54
  • $\begingroup$ I don't how you test sufficiency; it's a mathematical property, not the subject of hypotheses. I guess I feel more comfortable than @IrishStat in working with distributions other than normal; we have had most of the useful ones for decades. Minor sparring apart, the process in this Q is a count process, so it's not especially clear that forcing such counts into a normal straitjacket would be a good idea. $\endgroup$
    – Nick Cox
    Commented Feb 22, 2015 at 19:23
  • $\begingroup$ The idea of model sufficiency is that if there is identifiable information in the residuals then there is an argument for adding/moving structure to the model whereas necessity is the concept of do I need the structure I have. $\endgroup$
    – IrishStat
    Commented Feb 22, 2015 at 19:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.