For Univariate Linear Regression I can calculate the parameters (And most everything else) from simple sum of squares. Is there a corresponding method for Logistic Regression? Any pointers to code (in any language) would be helpful.

(Yes there are many solvers, but I want to implement as a simple Map-Reduce algorithm, as I have for Univariate Linear Regression)


In general you would need a solver. One exception would be a fully saturated model, that is, a model with only categorical explanatory/right-hand-side/x-variables and one parameter for each combination of groups. In that case the coefficients are just a function of the group means. In case of a bivariate model, it would mean that your only explanatory/right-hand-side/x-variable is a binary variable. Below is an example of how to recreate the results obtained by a solver of a fully saturated bivariate model by just transforming means in Stata.

// load some example data
sysuse nlsw88, clear

// use a "solver" 
logit union collgrad

// collect the means
tempname noncoll coll
sum union if e(sample) & collgrad == 0, meanonly
scalar `noncoll' = r(mean)
sum union if e(sample) & collgrad == 1, meanonly
scalar `coll' = r(mean)

// recreate the coefficients using just means
di as txt    "the constant"                   ///
   as result logit(`noncoll')

di as txt    "coefficient of collgrad: "      ///
   as result logit(`coll') - logit(`noncoll')

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.