glm(Histology ~ Cigarette + med1 * med2 * med3 * med4 + Age

It looks like the primal test fail. The dual program works: <div class="snippet-cl

See ?detect_separation_control and <code class="notra

"detect_separation" failed about brglm2 HOT 3 CLOSED

ikosmidis commented on June 7, 2024

"detect_separation" failed

from brglm2.

Comments (3)

ikosmidis commented on June 7, 2024

It looks like the primal test fail. The dual program works:

R> library(brglm2)
R> load("data2.rda")
R> form2 <- Histology ~ Cigarette + med1 * med2 * med3 * med4 + Age + Gender + Stage + Race
R> glm(form2, family = binomial("logit"), data = data2, method = "detect_separation", linear_program = "dual")
Separation: FALSE 
Existence of maximum likelihood estimates
          (Intercept) Cigarettenot reported          CigaretteYes 
                    0                     0                     0 
                 med1                  med2                  med3 
                    0                     0                     0 
                 med4               Age>=75              Age55-64 
                    0                     0                     0 
             Age65-74       Agenot reported            Gendermale 
                    0                     0                     0 
              StageII              StageIII               StageIV 
                    0                     0                     0 
    Stagenot reported                RaceNo      Racenot reported 
                    0                     0                     0 
            med1:med2             med1:med3             med2:med3 
                    0                     0                     0 
            med1:med4             med2:med4             med3:med4 
                    0                     0                     0 
       med1:med2:med3        med1:med2:med4        med1:med3:med4 
                    0                     0                     0 
       med2:med3:med4   med1:med2:med3:med4 
                    0                     0 
0: finite value, Inf: infinity, -Inf: -infinity

A post-fit check for infinite estimates agrees with the above. Specifically, there are no diverging standard errors across IWLS iterations, and hence no evidence for infinite maximum likelihood estimates

R> m_glm <- glm(form2, family = binomial("logit"), data = data2)
R> matplot(check_infinite_estimates(m_glm), type = "l", col = "black", lty = 1)

Does this help? Please re-open issue if not

from brglm2.

dasiav7 commented on June 7, 2024

Dear Dr. ikosmidis,

Thank you! What's the difference between "primal" and "dual" in the program?

from brglm2.

ikosmidis commented on June 7, 2024

See ?detect_separation_control and ?detect_separation. For details on the linear programs for detecting separation, pre-fit, see

Kjell Konis (2007). Linear Programming Algorithms for Detecting Separated Data in Binary Logistic Regression Models. DPhil. University of Oxford. <URL: https://ora.ox.ac.uk/objects/uuid:8f9ee0d0-d78e-4101-9ab4-f9cbceed2a2a>

from brglm2.

Recommend Projects