Chapter 12 Logistic Regression

Purpose Logistic regression models the probability of a binary outcome \(Y \in \{0,1\}\) as a function of predictors \(X\). It is central in causal inference, especially for estimating propensity scores.

12.1 Model

\[ P(Y=1|X) = \frac{1}{1 + e^{-X^\top \beta}} \]

12.2 Estimation

  • Objective Function: Maximize the likelihood of observing the data given parameters \(\beta\).

  • Loss Function: Equivalent to minimizing the negative log-likelihood.

\[ \ell(\beta) = \sum_{i=1}^n \left[ y_i \log p_i + (1-y_i)\log(1-p_i) \right], \quad p_i = \frac{1}{1+e^{-x_i^\top \beta}} \]

  • Maximum Likelihood Estimation (MLE): Estimates \(\hat{\beta}\) are obtained by maximizing \(\ell(\beta)\).

12.3 Hypothesis Testing

  • Test whether coefficients differ from zero using the z-statistic:

\[ z = \frac{\hat{\beta}_j}{SE(\hat{\beta}_j)} \]

  • Null hypothesis: \(H_0: \beta_j = 0\).

12.4 Role in Causal Inference

  • Logistic regression estimated by MLE is the standard method for estimating propensity scores.

  • These scores underpin Propensity Score Matching (PSM) and Inverse Probability Weighting (IPW), enabling adjustment for confounding.

Takeaway: Logistic regression uses MLE to estimate parameters by maximizing the likelihood (minimizing negative log-likelihood). Its most important role in causal inference is providing propensity scores for adjusting treatment selection bias.