RobertNiles.com
About Robert
Reporting Help
Finding Data on the Internet
Questions for Candidates
Stats Lessons
Mean
Median
Percent
Per capita
Standard Deviation
Margin of Error
Data Analysis
Sample Sizes
Stats Tests
Help Board
Bookstore


Statistics for the Utterly Confused

This book offers a super-accessible approach to the much-misunderstood subject of statistics.
More information
-->

Home > Statistics Every Writer Should Know > The Stats Board > Discusssion

more on logistic regression
Message posted by Susan Daniels (via 128.173.184.227) on June 19, 2001 at 3:56 PM (ET)

Thank you JG. Please tell me - is such a result valid (p approaching 1)? Also, what problems might have switched the sign on me...I checked for multicollinearity, but no correlation coefficients are greater than about 0.5.


READERS RESPOND:
(In chronological order. Most recent at the bottom.)

Re: more on logistic regression
Message posted by JG (via 128.8.23.221) on June 19, 2001 at 6:42 PM (ET)

Do we have a 0,1 dependent variable or as an independent variable or both. I am not sure what correlation tells us when we have 0,1 variables.


Re: more on logistic regression
Message posted by JG (via 128.8.23.221) on June 19, 2001 at 6:47 PM (ET)

If you have a P close to 1, it means that your dependent variable is mostly 1's with just a few 0's - this will cause problems with regression or any other curve fitting procedure.


Re: more on logistic regression
Message posted by Jack (via 208.249.113.130) on June 20, 2001 at 4:27 PM (ET)

If one of your coefficients ends up with an unexpected sign, that might arise as a result of an ill-conditioned design matrix. You can overcome that by performing a stepwise logistic regression.

If you have the software, this would involve the computer choosing the "best" independent variable to add to the model at each stage. This would reduce the confounding effects and result in a model with fewer independent variables.


Re: more on logistic regression
Message posted by JG (via 128.8.22.30) on June 21, 2001 at 3:24 AM (ET)

The safest thing to do is to regress against each indepdent variable separately and take it from there.
When the independet variables are not independent - orthogonal - you really have an integer programming problem and stepwise regression is only an approximation. You can also step backwards - start with a full fit and remove variables, etc.



Your $5 contribution helps cover part the $500 annual cost of keeping this site online.

Niles Online Sites:RobertNiles.comTheme Park InsiderViolinist.com

RobertNiles.com™, the site, content and services 咀opyright 1996-2002, Robert Niles.
All rights reserved. Questions? Comments? Read my Privacy Policy, or E-mail me!