[讨论]Multivarate Regression With Interaction Terms

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

I'm trying to test the following regression model: Y = b0 + b1X1 + b2X2 + b3X3 + b4X4 +b5X5 + b6X5*X1 + b7X5*X2 + b8X5*X3 + b9X5*X4 + b10X6 + b11X7 + b12X8 + b13X9 + e As you can see, this model has interaction terms. First, I plan to use the forward step wise method to include significantly associated variables. Then, I plan to use the backward method to end with the most parsimonious model. I'm not sure how I'm supposed to test this model in SPSS. Would there be a way I can enter interaction terms in a regression model using both forward and backward selection? Any advice would be greatly appreciated. Thank you.

Colleen

[此贴子已经被作者于2005-10-26 10:26:45编辑过]

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏0 回帖

关键词：Multivarate interaction regression regressio interact regression With interaction Terms Multivarate

Hi Colleen Paul is absolutely right. Hosmer & Lemeshow (talking about logistic regression models, but the principles are perfectly right for other regression models), said that you had to select your main effects variables one by one, carefully, basing your decision not only in the significance, but also in your knowledge (confusing variables needn't be significant to be included in a model). Once you had your main effects model, you started selecting the interactions terms trying to take your model easy to interpret. They gave the following guidelines about the interaction terms to be tested: Relevant (clinically/biologically/scientificaly meaningful). DON'T test any interaction term you won't be able to explain in the context of your investigation.

Significant This example (sorry, I did not center the interaction terms) can be useful: * Example dataset (from Campbells' "Statistics at Square Two") *. DATA LIST LIST/deadspac height age asthma (4 F8.0). BEGIN DATA 44 110 5 1 31 116 5 0 43 124 6 1 45 129 7 1 56 131 7 1 79 138 6 0 57 142 6 1 56 150 8 1 58 153 8 1 92 155 9 0 78 156 7 0 64 159 8 1 88 164 10 0 112 168 11 0 101 174 14 0 END DATA. VAR LABEL deadspac'Pulmonary anatomical deadspace (ml)' /height'Height (cm)' /age'Age (years)' /asthma'Presence of Asthma'. VALUE LABEL asthma 0'No' 1'Yes'. * Main effects model *. REGRESSION /STATISTICS COEFF OUTS CI R ANOVA /CRITERIA=PIN(.05) POUT(.10) /DEPENDENT deadspac /METHOD=ENTER height age asthma . * Although age is non significant, it is kept in the model until all the interaction terms are analyzed; Also, it could be a confusing factor *. * Checking interactions (better centered, but this is a fast example) *. COMPUTE heixage = height*age . COMPUTE heixast = height*asthma . COMPUTE astxage = asthma*age . EXECUTE . REGRESSION /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL /CRITERIA=PIN(.05) POUT(.10) /DEPENDENT deadspac /METHOD=ENTER height age asthma /METHOD=STEPWISE heixage heixast astxage. /* Testing interactions one by one *. * We can see that age is not involved in any interaction term What happen if we eliminate it from the model? (see adjusted R²...) *. REGRESSION /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL /CRITERIA=PIN(.05) POUT(.10) /DEPENDENT deadspac /METHOD=ENTER height asthma astxage . HTH Marta

Hi everybody My last model should have been: We can see that age is not involved in any interaction term. What happens if we eliminate it from the model? (see adjusted R²...) *. REGRESSION /STATISTICS COEFF OUTS CI R ANOVA COLLIN TOL /CRITERIA=PIN(.05) POUT(.10) /DEPENDENT deadspac /METHOD=ENTER height asthma heixast . Sorry... Marta

biostatistics@terra.es

Surely there is also another perhaps more important issue in variable selection and that is using stepwise selection techniques on a 'fishing' expedition

What theoretical concepts or models inform your selection? What does your theory, previous research etc. suggest are important variables in this context

Best
Muir

Muir Houston
Research Fellow
Institute of Education
University of Stirling
FK9 4LA

Selection procedures are rarely a good idea for regression, most especially when you have interaction terms since it is possible for the interaction to be added without the main variables. You should run the full model and examine each interaction, one at a time. If the interaction is not significant, drop it and go on to the next interaction. For any interaction that is significant, always keep the individual variables that make up the interaction in the model with the interaction. It is also helpful from an interpretation standpoint to center the variables going into the interactions. Paul R. Swank, Ph.D. Professor, Developmental Pediatrics Director of Research, Center for Improving the Readiness of Children for Learning and Education (C.I.R.C.L.E.) Medical School UT Health Science Center at Houston