0

This may be a simple question, but my data (once transformed) passes normality, however does not pass homoscedasticity. Can I look at the residuals to see if they pass this? or am I over looking this?

AyAyRon166
  • 11
  • 2
  • what analytical technique e. g. regression , you want to implement ? what are your residuals representing? –  Dec 20 '17 at 05:33
  • Which tests do you run to analyze these two things and what is the structure of your data? – Jeremias K Dec 20 '17 at 10:18

1 Answers1

1

Since you mention residuals I assume you are talking about some sort of regression. In this case, you should be using residuals for all diagnostics.

Specifically, linear regression assumes the residuals are i.i.d. Normally distributed, not the dependent variable.

If the residuals severely lack constant variance, then one possible solution is a weighted regression.

P.S. How do you know your data do not pass homoscedasticity but you have not looked at the residuals?

Glen
  • 6,320
  • 4
  • 37
  • 59
  • I ran a Bartlett's test then ran a model (ANOVA) regardless to look at the residuals. Im working in R at the moment (from my understanding, or package I am using) to run a Barlett's test, you have to insert a model in the script. I am using "rcompanion" if you are by chance familiar with it – AyAyRon166 Dec 20 '17 at 06:09
  • So ANOVA (which is a type of regression) assumes equal variances. If the violation is severe (Bartlett's test is one indication it might be a problem, you should visually inspect), then you will want to do a weighted ANOVA to give less importance to the observations with larger variance. – Glen Dec 20 '17 at 06:18
  • When you say weighted, are you refering to a Welch's ANOVA? – AyAyRon166 Dec 20 '17 at 15:52
  • no, I'm referring to weighted least squares regression: https://stats.stackexchange.com/questions/97832/how-do-you-find-weights-for-weighted-least-squares-regression – Glen Dec 20 '17 at 16:50