Are my residuals normal?

  • Thread starter bradyj7
  • Start date
  • #1
122
0

Main Question or Discussion Point

Hi there,

I have performed a multiple regression analysis and I am checking the assumptions. The normal probability plot is not exactly a straight line, but is it okay? If you were looking at the plot would you be satisfied hat the normality assumption is valid?


https://dl.dropbox.com/u/54057365/All/residuLPLOT.jpg [Broken]

I performed an Anderson Darling test on the residuals and the residuals were found not to be normal. What would you do in this case? Would you transform some of the predictors? Would you transform them all or just some?

I'd appreciate any comments.

Thanks

John
 
Last edited by a moderator:

Answers and Replies

  • #2
Stephen Tashi
Science Advisor
6,923
1,182
If you are trying to solve a real life problem, you should consider the realities of the problem and reveal them to your would-be advisors. For example, what are the physics? Is there some physical law that predicts the curve from which you compute the residuals? Is the precision of your measuring equipment limited to a certain number of digits?

There aren't any mathematical laws that can specify a correct answer to your question if it is abstracted to a few graphs. (Of course, there can be lots of suggestions about how to make the graphs look different. I'm just saying that there aren't mathematical theorems that prove one way is better than another.)
 
  • #3
DrDu
Science Advisor
6,015
750
Last not least there are statistical test to decide whether to accept normality or not. I guess they would reject normality in your case. The question is whether this is of any relevance for your problem.
 
  • #4
chiro
Science Advisor
4,790
131
Last not least there are statistical test to decide whether to accept normality or not. I guess they would reject normality in your case. The question is whether this is of any relevance for your problem.
If that first plot is a QQ plot with reference to a standardized normal, I think your assertion will be wrong.

But OP, as others have said, you need to put this into the context of the problem and the data and what both of them actually relate to physically, experimentally, and intuitively.
 
  • #5
DrDu
Science Advisor
6,015
750
  • #6
122
0
Thank you for suggestions. Much appreciated.
 
  • #7
not normal. perform some transformation such as log, reciprocal etc to the variable.
 
  • #8
483
2
Hi there,

I have performed a multiple regression analysis and I am checking the assumptions. The normal probability plot is not exactly a straight line, but is it okay? If you were looking at the plot would you be satisfied hat the normality assumption is valid?


https://dl.dropbox.com/u/54057365/All/residuLPLOT.jpg [Broken]

I performed an Anderson Darling test on the residuals and the residuals were found not to be normal. What would you do in this case? Would you transform some of the predictors? Would you transform them all or just some?

I'd appreciate any comments.

Thanks

John
The residuals are skewed, so they aren't normal. The mode should be around zero but it is actually -0.8.
 
Last edited by a moderator:
  • #9
You can transform dependent variable.

But there is another possibility about data problem. There may be some extreme outliers in the data. Try to remove or adjust the outlier first by some appropriate method such as "Winsoriztion".
 

Related Threads for: Are my residuals normal?

Replies
4
Views
1K
Replies
11
Views
1K
Replies
17
Views
763
Replies
4
Views
6K
Replies
4
Views
3K
  • Last Post
Replies
3
Views
2K
  • Last Post
Replies
1
Views
4K
  • Last Post
Replies
4
Views
3K
Top