Generally I have trained a logistic regression model with:
glm_model <- glm(response ~ x, data = original_data, family = "binomial")
data_new <- tibble(response = original_data$response,
score_glm = glm_model$linear.predictors)
After that I will again make a non-linear optimization with the minpack.lm package from R
in order to get better results.
a_coef = 0
b_coef = 1
c_coef = 4
d_coef = 7
minpack.lm::nlsLM(response ~ a + b/(1 + exp(-(c + d * score_glm))),
data = data_new,
start = c(a = a_coef,
b = b_coef,
c = c_coef,
d = d_coef),
control = list(maxiter = 1000))
Could this be a valid procedure in order to get better results?
