I have a dataset of historical tennis matches, one row per match. Each row has the ELO points of each player, and a calculated ELO difference. And of course a column to indicate if Player 1 won as a 1 or 0.

There are other predictor variables involved but for the sake of this example I just want to know how would I go building this kind of model where there can be both numerical and categorical independent variables, and a 1 or 0 dependent variable. (I would be using SAS).

What kind of distribution function would I be using, and what kind of link function would I be using?

Ideally the result would give the probability of winning as a function of ELO difference. Is this possible using a GLM?


You should be using a Binomial GLM, a.k.a. Binomial/Binary logistic regression. Categorical independent variables will be converted to dummies without affecting the generic model structure. For $N$ Binomial input samples and $M$ independent variables (including dummified factors), your GLM components should be the following...

1. Stochastic component:

$ y_i \sim Binomial(n_i, p_i), ~~i=1,...,N$

(NB: For binary response, $n_i=1, \forall i$)

2. Systematic component:

$\eta_i = \beta_0 + \beta_1 x_1 + ... + \beta_Mx_M $

3. Link function:

$g(p_i) = logit(p_i) = \eta_i $

With for final model being:

$logit(p_i) = \beta_0 + \beta_1 x_1 + ... + \beta_kx_k $

| cite | improve this answer | |

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.