Avril Coghlan
2004-Jun-09 10:34 UTC
[R] testing effects of quantitative predictors on a categorical response variable
Hello, I have a small statistics question, and as I'm quite new to statistics and R, I'm not sure if I'm doing things correctly. I am looking at two quantitative variables (x,y) that are correlated. When I divide the data set according to a categorical variable z, then x and y are more poorly correlated when z = A than when z = B (see attached figure). In fact x and y are two (correlated) predictor variables and z is a categorical response variable that x and y affect. I would like to use R to make some statistical test to show that you seem to get z = A when the value of x is much less than y, while you tend to get z = B when x is approximately the same as y. Can anybody tell me what I should be doing? I tried a logistic regression:> glm1 <- glm(z ~ y + x,family=binomial(),trace=T)which gives Pr(>|z|) < 0.01 for both x and y, but I'm not sure if this is valid to do, since x and y are correlated? As well this test does not show that it is for values of x << y that we tend to get z = A, and that for values of x approx = y, that we tend to get z = B. I'm not sure how to show this? I'll be very grateful if anyone can help. Avril -------------- next part -------------- A non-text attachment was scrubbed... Name: avril.ps Type: application/postscript Size: 8001 bytes Desc: not available Url : https://stat.ethz.ch/pipermail/r-help/attachments/20040609/ea3dee5c/avril.ps
jferrer@ivic.ve
2004-Jun-09 11:18 UTC
[R] testing effects of quantitative predictors on a categorical response variable
Hi Avril, I'm not sure what you want to show. Do you want to know the effects of each variable? or just predict when you get z=A and when z=B? In the latter case, I think that, if x and y are in the same units, you could simply try w <- y-x glm1 <- glm(z ~ w,family=binomial(),trace=T) Hope it helps, JR En respuesta a / Antwort zu / Reply to: ~~Hello, ~~ ~~ I have a small statistics question, and ~~as I'm quite new to statistics and R, I'm not ~~sure if I'm doing things correctly. ~~ ~~ I am looking at two quantitative ~~variables (x,y) that are correlated. ~~When I divide the data set according to a categorical ~~variable z, then x and y are more poorly correlated ~~when z = A than when z = B (see attached figure). ~~In fact x and y are two (correlated) predictor ~~variables and z is a categorical response variable that ~~x and y affect. ~~ ~~ I would like to use R to make some statistical ~~test to show that you seem to get z = A when ~~the value of x is much less than y, while you ~~tend to get z = B when x is approximately the same as y. ~~Can anybody tell me what I should be doing? ~~I tried a logistic regression: ~~> glm1 <- glm(z ~ y + x,family=binomial(),trace=T) ~~which gives Pr(>|z|) < 0.01 for both x and y, but ~~I'm not sure if this is valid to do, since x and y are correlated? ~~ ~~As well this test does not show that it is for values of ~~x << y that we tend to get z = A, and that for ~~values of x approx = y, that we tend to get z = B. I'm ~~not sure how to show this? ~~ ~~I'll be very grateful if anyone can help. ~~ ~~Avril ~~ Dipl.-Biol. J.R. Ferrer Paris ~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Laboratorio de Biolog??a de Organismos - Centro de Ecolog??a Instituto Venezolano de Investigaciones Cient??ficas Apartado 21827 - Caracas 1020A REPUBLICA BOLIVARIANA DE VENEZUELA Tel:00-58-212-5041452 --- Fax: 00-58-212-5041088 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ jferrer at ivic.ve