categorizing a variable turns it from insignificant to significant I have a numeric variable which turns out not significant in a multivariate logistic regression model.
However, when I categorize it into groups, suddenly it becomes significant.
This is very counter-intuitive to me: when categorizing a variable, we give some information up.
How can this be?
 A: One possible way is if the relationship is distinctly nonlinear. It's not possible to tell (given the lack of detail) whether this really explains what's going on.
You can check for yourself. First, you could do an added variable plot for the variable as itself, and you could also plot the fitted effects in the factor-version of the model. If the explanation is right, both should see a distinctly nonlinear pattern.
A: One possible explanation would be nonlinearities in the relationship between your outcome and the predictor.
Here is a little example. We use a predictor that is uniform on $[-1,1]$. The outcome, however, does not linearly depend on the predictor, but on the square of the predictor: TRUE is more likely for both $x\approx-1$ and $x\approx 1$, but less likely for $x\approx 0$. In this case, a linear model will come up insignificant, but cutting the predictor into intervals makes it significant.
> set.seed(1)
> nn <- 1e3
> xx <- runif(nn,-1,1)
> yy <- runif(nn)<1/(1+exp(-xx^2))
> 
> library(lmtest)
> 
> model_0 <- glm(yy~1,family="binomial")
> model_1 <- glm(yy~xx,family="binomial")
> lrtest(model_1,model_0)
Likelihood ratio test

Model 1: yy ~ xx
Model 2: yy ~ 1
  #Df  LogLik Df  Chisq Pr(>Chisq)
1   2 -676.72                     
2   1 -677.22 -1 0.9914     0.3194
> 
> xx_cut <- cut(xx,c(-1,-0.3,0.3,1))
> model_2 <- glm(yy~xx_cut,family="binomial")
> lrtest(model_2,model_0)
Likelihood ratio test

Model 1: yy ~ xx_cut
Model 2: yy ~ 1
  #Df  LogLik Df  Chisq Pr(>Chisq)  
1   3 -673.65                       
2   1 -677.22 -2 7.1362    0.02821 *
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

However, this does not mean that discretizing the predictor is the best approach. (It almost never is.) Much better to model the nonlinearity using splines or similar.
