how to use predict()

Question

Want to predict a value but this is clearly not the solution. I am doing a multiple choice test and 0.304... is not an answer.How to use predict() correctly?

library(glm2)
data(crabs)
fit= glm(Satellites~Width,data=crabs, family="poisson")
plot(Satellites~Width,data=crabs)
abline(fit)
predict(fit, newdata=data.frame(Width=c(22)))
1 
0.3042347

I don't think the value is correct since I am doing a multiple choice thest and 0.304.. is not an answer. — Roland Kofler, Feb 19 '13 at 18:38
Ummm... `22 * fit$coef[2] + fit$coef[1]` gives what you have posted. It is right. But clearly, it is not what you expected! — Simon O'Hanlon, Feb 19 '13 at 18:42
Value you got with predict() is log() of expected value because you made a Poisson regression. Just make exp(value) to get the same scale as original values. — Didzis Elferts, Feb 19 '13 at 18:47
Wow how can I imagine that! @DidzisElferts you are right! It fits — Roland Kofler, Feb 19 '13 at 18:49
@Roo yes this shouls be an answer bu I suggest that you modify a little bit your question...to add what you said in the comments for example. — agstudy, Feb 19 '13 at 18:52

score 12 · Accepted Answer · edited Feb 19 '13 at 19:20

12

Function predict() for Poisson regression (for GLM in general) by default will calculate the values on the scale of the linear predictors, i.e. the log scale in this case (see help file for predict.glm).

predict(fit, newdata=data.frame(Width=c(22)))
        1 
0.3042347

To get the predicted values on the scale of the response variable, you should add argument type="response" to function predict().

predict(fit, newdata=data.frame(Width=c(22)),type="response")
       1 
1.355587

edited Feb 19 '13 at 19:20

Ben Bolker

211,554
25
370
453

answered Feb 19 '13 at 19:01

Didzis Elferts

95,661
14
264
201

+1!`preidct` looks sometimes clunky to use! but it is really a great workhorse! – agstudy Feb 19 '13 at 19:10
I am not sure what the scale linear predictors have to do with log() though... even more confused. – Roland Kofler Feb 19 '13 at 19:17
4

You fitted a poisson model. As `?poisson` will tell you, the default link function of the Poisson is the log (that shows up as `poisson(link = "log")` in the help file). "on the scale of the linear predictor" means on the log scale in this context. If this is totally unfamiliar to you, it might be helpful to do some more reading about Poisson GLMs ... – Ben Bolker Feb 19 '13 at 19:19
1

@Roo, Venables & Ripley's MASS book has an entire chapter explaining this: http://www.amazon.com/Modern-Applied-Statistics-W-N-Venables/dp/0387954570 – Ricardo Saporta Feb 19 '13 at 19:30

how to use predict()

1 Answers1