Re: Ordinary Least Square Regression: Residual Not Normal

[Oleg Solovyev]

>I found out that my Oridinary Least Square Regression model's residual is
not normal.
>May I know what are the common retractification to this problem?

You can try these:
  1. Look for outliers in your data and delete them if necessary.
  2. Try variables transformations: log, second order etc.
  3. Try to fit other curves to the data.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Swank, Paul R]

You cannot just delete outliers from necessity. It must be clear that
the data are in error. Otherwise, you are cherry-picking your data.
Another solution for certain distributional forms is to use generalized
linear models. It also depends on how non-normal the data are and what
the sample size is. A large sample may indicate that very symmetric and
unimodal distributions are not normal.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Oleg Solovyev
Sent: Sunday, January 25, 2009 1:43 PM
To: [hidden email]
Subject: Re: Ordinary Least Square Regression: Residual Not Normal

>I found out that my Oridinary Least Square Regression model's residual
is
not normal.
>May I know what are the common retractification to this problem?

You can try these:
  1. Look for outliers in your data and delete them if necessary.
  2. Try variables transformations: log, second order etc.
  3. Try to fit other curves to the data.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Peck, Jon]

Indeed.  You could also consider a robust regression technique.  The SPSSINC ROBUST REGR extension command is available from SPSS Developer Central (www.spss.com/devcentral) if you have SPSS Version 17.

But you should look at the patterns in the residuals to see whether there is evidence of a misspecification, a heteroscedasticity problem or other such difficulties.  Also, how much large outliers matter to the results can be gauged by looking at the leverage statistics.

HTH,
Jon


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Swank, Paul R
Sent: Monday, January 26, 2009 8:47 AM
To: [hidden email]
Subject: Re: [SPSSX-L] Ordinary Least Square Regression: Residual Not Normal

You cannot just delete outliers from necessity. It must be clear that
the data are in error. Otherwise, you are cherry-picking your data.
Another solution for certain distributional forms is to use generalized
linear models. It also depends on how non-normal the data are and what
the sample size is. A large sample may indicate that very symmetric and
unimodal distributions are not normal.

Paul R. Swank, Ph.D
Professor and Director of Research
Children's Learning Institute
University of Texas Health Science Center
Houston, TX 77038

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Oleg Solovyev
Sent: Sunday, January 25, 2009 1:43 PM
To: [hidden email]
Subject: Re: Ordinary Least Square Regression: Residual Not Normal

>I found out that my Oridinary Least Square Regression model's residual
is
not normal.
>May I know what are the common retractification to this problem?

You can try these:
  1. Look for outliers in your data and delete them if necessary.
  2. Try variables transformations: log, second order etc.
  3. Try to fit other curves to the data.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[paul wilson-7]

Hi,

I have 20 different samples ranging from about 200 to about 5000 cases.

I'd like to compare their standard deviations and see which ones have more "variation" than others.

However, I can't compare standard deviation of a sample of 200 cases to standard deviation of the one with 5000 because their size is quite different.

I'd appreciate your suggestions on what would be a good strategy to make necessary adjustments so standard

deviations across all 20 samples can be compared and ranked regardless of how many cases are in each sample.

Any ideas on how I could adjust my standard deviations so they take the sample sizes into account and make them

comparable to one another?

Thanks

[David Greenberg]

That the samples are based on different numbers of cases does not preclude your comparing them. David Greenberg, Sociology Department, New York University

----- Original Message -----
From: paul wilson <[hidden email]>
Date: Thursday, September 3, 2009 10:00 pm
Subject: Standard deviation comparison across different samples
To: [hidden email]

> Hi,
>
> I have 20 different samples ranging from about 200 to about 5000 cases.
> I'd like to compare their standard deviations and see which ones have
> more "variation" than others.
> However, I can't compare standard deviation of a sample of 200 cases
> to standard deviation of the one with 5000 because their size is quite
> different.
>
> I'd appreciate your suggestions on what would be a good strategy to
> make necessary adjustments so standard
> deviations across all 20 samples can be compared and ranked regardless
> of how many cases are in each sample.
>
> Any ideas on how I could adjust my standard deviations so they take
> the sample sizes into account and make them
> comparable to one another?
>
>
> Thanks
>
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[paul wilson-7]

Hi David,

Thank you for your quick reply.

I'm glad to hear I may not need to make any adjustments, but before concluding that just wanted to share more detail I failed to mention in my original e mail.

I'm analyzing customer spend from 20 different retail stores.

Each one is different in size and has a different number of registered customers on file.

What I'm reporting on is mean spend for each store and would like to also report on how "spread" is

the spend across each property's customer base and compare those standard deviations.

I guess one can say that we are talking about 20 different populations here as opposed to randomly selecting samples from one population.

I suppose that made me think that I may not be able to compare their standard deviations directly. Can you please confirm that I indeed can regardless of the fact that we

are talking 20 different stores/customer populations as opposed to 20 samples randomly drawn from the same population?

Thanks a lot!

---

From: David Greenberg <[hidden email]>
To: [hidden email]
Sent: Thursday, September 3, 2009 10:19:35 PM
Subject: Re: Standard deviation comparison across different samples

That the samples are based on different numbers of cases does not preclude your comparing them. David Greenberg, Sociology Department, New York University

----- Original Message -----
From: paul wilson <[hidden email]>
Date: Thursday, September 3, 2009 10:00 pm
Subject: Standard deviation comparison across different samples
To: [hidden email]

> Hi,
>
> I have 20 different samples ranging from about 200 to about 5000 cases.
> I'd like to compare their standard deviations and see which ones have
> more "variation" than others.
> However, I can't compare standard deviation of a sample of 200 cases
> to standard deviation of the one with 5000 because their size is quite
> different.
>
> I'd appreciate your suggestions on what would be a good strategy to
> make necessary adjustments so standard
> deviations across all 20 samples can be compared and ranked regardless
> of how many cases are in each sample.
>
> Any ideas on how I could adjust my standard deviations so they take
> the sample sizes into account and make them
> comparable to one another?
>
>
> Thanks
>
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD