Predictive Modeling

Regression models are the foundation of predictive analytics in R. This lesson covers linear regression for continuous outcomes and logistic regression for binary classification, using R's lm() and glm() functions.

1. Simple Linear Regression with lm()

model <- lm(salary ~ experience, data = df)
summary(model)

The formula salary ~ experience reads as "salary predicted by experience." summary() reveals coefficients, p-values, and model fit.

2. Reading lm() Output

3. Multiple Linear Regression

model <- lm(salary ~ experience + age + education_years, data = df)
summary(model)

# Coefficients with confidence intervals
confint(model)

4. Making Predictions

new_employee <- data.frame(experience = 5, age = 30, education_years = 16)
predict(model, newdata = new_employee)
predict(model, newdata = new_employee, interval = "confidence")

5. Checking Model Assumptions

par(mfrow = c(2, 2))
plot(model)   # residuals vs fitted, Q-Q plot, scale-location, leverage
par(mfrow = c(1, 1))

6. Logistic Regression with glm()

For binary outcomes (yes/no, churn/retain, pass/fail), use glm() with a binomial family:

model_logit <- glm(churned ~ tenure + monthly_charges + support_calls,
                   data = df, family = binomial)
summary(model_logit)

7. Interpreting Logistic Coefficients — Odds Ratios

exp(coef(model_logit))   # convert log-odds coefficients to odds ratios
exp(confint(model_logit)) # confidence intervals for odds ratios

An odds ratio above 1 means the predictor increases the probability of the outcome; below 1 means it decreases it.

8. Making Classification Predictions

predicted_probs <- predict(model_logit, type = "response")
predicted_class <- ifelse(predicted_probs > 0.5, "Yes", "No")

table(Predicted = predicted_class, Actual = df$churned)  # confusion matrix

9. Linear vs. Logistic — When to Use Which

10. Production-Ready Checklist

• ✅ Use lm() for continuous outcomes — always inspect summary() and diagnostic plots.
• ✅ Use glm(family = binomial) for classification — interpret via odds ratios.
• ✅ Check assumptions — linearity, normal residuals, no severe multicollinearity.
• ✅ Validate on new data — training accuracy alone can be misleading.