Hi guys,
Can you please tell me which technique i use to measure if a non-significant variable changed the other variables in the model by 10% when removed. This is in a logistic regression. A professor told me once that if you remove a non significant variable from a model, it shouldn't change the other variables by more than 10%. How can I test this? Many thanks, Love. |
Love,
Your professor's advice was a bit vague. SPSS lets you add or subtract variables stepwise from your equation, deciding on the basis of the impact of each inclusion or exclusion on the overall fit of the equation. There are two ways: forward stepwise inclusion and backwards elimination. There is a PIN criterion for inclusion of a new variable in the equation, and a POUT criterion for exclusion of a variable already in the equation, the default values of which are 0.05 and 0.10 respectively, but you may change them at will. You can find explanations of their meaning in standard books on logistic regression, such as Hosmer and Lemeshow's Applied Logistic Regression, and in SPSS documentation such as the Syntax Reference manual and the procedure Algorithms (both in your installation CD) or Marija Norusis book on advanced procedures. On the other hand, besides being vague, the advice was a bit contradictory. That a variable is non-significant usually means that its impact is so small that you cannot reliably tell whether it is different from zero. The minimum effect size for a variable effect to be significant depends, of course, on sample size and the level of significance you choose. So a 10% impact may be too much or too little, depending on circumstances. But in the case of logistic regression a 10% impact means a 10% impact on log odds, and that is in most circumstances a pretty significant impact, so the variable causing it is most probably significant. If it is significant (even if the effect is below 10%) you must leave it in the equation. If it is not, you cannot tell whether the effect is for real or just a fluke (even if above 10%, which may happen if your sample is too small or your significance requirements are too high). Luv. Hector -----Mensaje original----- De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de Love (sent by Nabble.com) Enviado el: Tuesday, August 01, 2006 11:57 PM Para: [hidden email] Asunto: Insignificant variable removed in logistic regression Hi guys, Can you please tell me which technique i use to measure if a non-significant variable changed the other variables in the model by 10% when removed. This is in a logistic regression. A professor told me once that if you remove a non significant variable from a model, it shouldn't change the other variables by more than 10%. How can I test this? Many thanks, Love. -- View this message in context: http://www.nabble.com/Insignificant-variable-removed-in-logistic-regression- tf2037395.html#a5606801 Sent from the SPSSX Discussion forum at Nabble.com. |
Free forum by Nabble | Edit this page |