Thanks for your quick answer. 
I think you might prefer I speak english so I'm gonna try to be clear in english.
For the reprex :
I've done before an analysis of these lost customers, and I identified these variables are discriminating :
I do get these variables for the active people too.
So I think my plan would be :
I merge these 2 dataframes with a new column 'lost' to 0 or 1 .
For the quantitatives variables, I test the linearity and I merge them into slices if necessary.
(4h)
Then I apply the linear regression and try to get a significant model.
customers$mtt_cl<- fct_recode(customers$mtt_cl,
"(5,15]"="(5,10]",
"(5,15]"="(10,15]")
reg2c <- glm(lost~ mtt_cl + nb_depots + age_first_sign ,
titanic,
family="binomial")
pairs(emmeans(reg2c,"mtt_cl ")) %>%
as.data.frame() %>%
arrange(desc(p.value)) %>%
kable()
(4h)
Based on the result of regression, I apply the function predict, and adjust the model until the risk of error is correct.
customers$p_surv <- predict(reg2g,
titanic,
type="response")
confusion <- table(customers$survived,
ifelse(customers$p_surv < 0.382,
"prédit \"non\"",
"prédit \"oui\""))
(6h)
Graphics : an expected graphic is to show the score of each active customer of 'could be lost or not'.
I guess I'll do that with :
reg2g %>%
blr_gains_table() %>%
plot()
(2h)
And of course analysis for each step and interpretation.
So my question is : does that sound great ? logic and timing ? I have to do an estimation before I really do the prediction, but I haven't done that for the last 10 years so I try to get into to give the best estimation (not easy at all !)
For the immediate pb : solved with 'library(pander)' 
