The [logitr] package fits logit regression models using maximum likelihood estimation. The model assumes an underlying latent Binomial variable.
For a more detailed and complete model with more available and extractable information use stats:::glm. This package was made for learning purposes only and using the standard glm functions instead of this is recommended.
As usual in many other regression packages for R [@R], the main model fitting function logitr() uses a formula-based interface and returns an (S3) object of class logitr:
logitr(formula, data, subset, na.action,
model = TRUE, x = FALSE, y = TRUE,
control = logitr_control(...), ...)The underlying workhorse function is logitr_fit() which has a matrix interface and returns an unclassed list.
A number of standard S3 methods are provided:
| Method | Description |
|---|---|
print() |
Simple printed display with coefficients |
summary() |
Standard regression summary; returns summary.logitr object (with print() method) |
coef() |
Extract coefficients |
vcov() |
Associated covariance matrix |
predict() |
Predictions for new data |
fitted() |
Fitted values for observed data |
residuals() |
Extract the Pearson Residuals |
terms() |
Extract terms |
model.matrix() |
Extract model matrix (or matrices) |
nobs() |
Extract number of observations |
logLik() |
Extract fitted log-likelihood |
Due to these methods some standard utilities also work automatically, e.g., AIC() and BIC()
As this package was created for learning purposes only and merely replicates an already existing function in R the goal was to also get results that match those output by stats:::glm as closely as possible. Because I am not working with any specific data sets I created a dataset arbitrarily using some of the randomization available in R.
## generating data
set.seed(123)
x1 <- rnorm(30,3,2) + 0.1 * c(1:30)
x2 <- rbinom(30,1,0.3)
x3 <- rpois(n=30,lambda = 4)
x3[16:30] <- x3[16:30] - rpois(n=15, lambda = 2)
xdat <- cbind(x1,x2,x3)
ydat <- c(rbinom(5,1,0.1), rbinom(10,1,0.25), rbinom(10,1,0.75), rbinom(5,1,0.9))
## comparison of model outputs
(m0 <- logitr(ydat~xdat))## Probit model
##
## Coefficients:
## (Intercept) xdatx1 xdatx2 xdatx3
## 1.489967 -0.004592 -0.700656 -0.345386
##
## Log-likelihood: -18.69 on 4 Df(m1 <- glm(ydat~xdat, family = "binomial"))##
## Call: glm(formula = ydat ~ xdat, family = "binomial")
##
## Coefficients:
## (Intercept) xdatx1 xdatx2 xdatx3
## 1.489967 -0.004593 -0.700655 -0.345385
##
## Degrees of Freedom: 29 Total (i.e. Null); 26 Residual
## Null Deviance: 41.46
## Residual Deviance: 37.37 AIC: 45.37## comparing AIC and BIC
AIC(m0)## [1] 45.37385AIC(m1)## [1] 45.37385BIC(m0)## [1] 50.97864BIC(m1)## [1] 50.97864While there are slight deviations on the last digit of the coefficients, the results are otherwise mostly identical with both the AIC and BIC also being a perfect match. With this I am fairly satisfied with the result of this project and although I won’t be publishing this I learned an incredible ammount of new things.