Generating predicted values for levels of factor variable

问题

I am regressing a number of factor variables on a continuous outcome variable using lm(). For example,

fit<-lm(dv~factor(hour)+factor(weekday)+factor(month)+factor(year)+count, data=df)

I would like to generate predicted values (yhat) for different levels of a factor variable while holding the other variables at their median or modal value. For example, how would I generate the yhat for different weekdays while holding other factors constant?

回答1:

I may be able to assist based on @Roland's comments. I think you want plain old ANOVA, which helps determine if factors are important or not. There's no need to factor here, integers or numbers (class: numeric) work fine. I put together the following code as example:

#creates df
(df <- data.frame(h=c(1,3,4,0,2, 3),d=c(2*1:3), m=c(-1, 0, 3, 4, 7, 8), y=c(30,28,27,26,22, 21)))

#creates linear model, gives output
(fit<-lm(df$d~ df$h + df$m+ df$y))

#runs ANOVA on linear model
anova(fit)

#creates predictions from lm based on different values of df$h
predict.lm(fit)

ANOVA is a special case of a regression. The output will tell you whether or not the factor is significant by the P value.

> anova(fit)
Analysis of Variance Table

Response: df$d
          Df  Sum Sq Mean Sq F value  Pr(>F)  
df$h       1 13.2923 13.2923 89.5846 0.01098 *
df$m       1  2.2832  2.2832 15.3879 0.05927 .
df$y       1  0.1277  0.1277  0.8608 0.45147  
Residuals  2  0.2968  0.1484

In this example hours are very highly correlated with your dependent variable days, while months shows the next highest correlation.

Please see the link for a background-

http://www.cookbook-r.com/Statistical_analysis/ANOVA/

FYI - I recommend you include some source code to create your example. In this manner people who attempt to answer your question can all refer to the same example.

FYI2 - I recommend you add the tag "regression"

HTH.

来源：https://stackoverflow.com/questions/14630056/generating-predicted-values-for-levels-of-factor-variable

标签

regression