Checking model assumptions and making predictions

Machine Learning for Marketing Analytics in R

Verena Pflieger

Data Scientist at INWT Statistics

Test of PH assumption

testCPH1 <- cox.zph(fitCPH1)
print(testCPH1)

                              rho   chisq        p
gender=Male               0.0317   1.884 1.70e-01
SeniorCitizen=Yes         0.0587   6.507 1.07e-02
Partner=Yes               0.0752  10.116 1.47e-03
Dependents=Yes            0.0131   0.314 5.75e-01
StreamMov=NoIntServ      -0.0448   3.588 5.82e-02
StreamMov=Yes             0.0827  12.174 4.85e-04
PaperlessBilling=Yes      0.0180   0.611 4.34e-01
PayMeth=CreditCard(auto)  0.0253   1.198 2.74e-01
PayMeth=ElektCheck       -0.0427   3.427 6.41e-02
PayMeth=MailedCheck      -0.0851  13.069 3.00e-04
MonthlyCharges            0.1268  25.778 3.83e-07
GLOBAL                        NA 217.172 0.00e+00

Proportional hazards for Partner

plot(testCPH1, var = "Partner")

Proportional hazards for MonthlyCharges

plot(testCPH1, var = "MonthlyCharges")

General remarks on tests

cox.zph()-test conservative
Sensitive to number of observations
Different gravity of violations

What if PH assumption is violated?

Stratified analysis

fitCPH2 <- cph(Surv(tenure, churn) ~ MonthlyCharges +
                 SeniorCitizen + Partner + Dependents + 
                 StreamMov + Contract,
               stratum = "gender = Male",
               data = dataSurv, x = TRUE, y = TRUE, surv = TRUE)

Time-dependent coefficients

Validating the model

validate(fitCPH1, 
         method = "crossvalidation", 
         B = 10, pr = FALSE)

      index.orig training   test optimism index.corrected  n
R2        0.2277   0.2279 0.2277   0.0002          0.2276 10
                            ...

Probability not to churn at certain timepoint

oneNewData <- data.frame(gender = "Female",
                             SeniorCitizen = "Yes",
                             Partner = "No",
                             Dependents = "Yes",
                             StreamMov = "Yes",
                             PaperlessBilling = "Yes",
                             PayMeth = "BankTrans(auto)",
                             MonthlyCharges = 37.12)
str(survest(fitCPH1, newdata = oneNewData, times = 3))

List of 5
 $ time   : num 3
 $ surv   : num 0.905
 $ std.err: num 0.0136
 $ lower  : num 0.881
 $ upper  : num 0.93

Survival curve for new customer

plot(survfit(fitCPH1, newdata = oneNewData))

Predicting expected time until churn

print(survfit(fitCPH1, newdata = oneNewData))

Call: survfit(formula = fitCPH1, newdata = oneNewData)

      n  events  median 0.95LCL 0.95UCL 
   5311    1869      65      53      72

Learnings

	Learnings about survival analyis
You have learned...	to visualize the tenure times of customers
	to model the time to an event and extract factors influencing it
	how to validate the model
	how to make predictions

	Learnings from the model
You have learned...	that being senior citizen increases the probability to churn by 23%
	that a one-unit increase in monthly charges decreases the hazard of churning by about 1%

It is up to you now!

Machine Learning for Marketing Analytics in R