关于R中lm输出的多元线性回归系数估计的解读疑问

Great question—this is one of the most common pitfalls when moving from simple to multiple linear regression in R, so let’s break it down clearly.

First: Simple vs. Multiple Regression #

• In simple linear regression (only one predictor variable, e.g., lm(y ~ x1)), the coefficient for x1 is exactly the slope you see when plotting y against x1. It represents the average change in y for every 1-unit increase in x1, with no other variables to account for.
• In multiple regression (two or more predictors, e.g., lm(y ~ x1 + x2 + x3)), the coefficients are called partial regression coefficients. They do not match the simple slope you’d get from plotting y against a single predictor alone.

What Do Partial Coefficients Actually Mean? #

Each partial coefficient (the B₁, B₂, etc. from your output) represents the average change in the dependent variable y when that predictor increases by 1 unit, while all other predictors in the model are held constant. It’s the "net effect" of that variable, after controlling for the influence of every other predictor in the equation.

Why the Apparent Contradiction? #

The mismatch between your scatterplot slope and the regression coefficient happens when your predictor variables are correlated with each other (a common scenario called multicollinearity, even if it’s mild). Here’s a concrete example to illustrate:

Suppose we have two correlated predictors, x1 and x2, where x1 tends to increase when x2 increases. The true relationship with y is:
y = 2*x1 - 3*x2 + random noise

• If you plot y vs. x1 alone, the slope will look negative! Because whenever x1 is high, x2 is also high—and x2 has a strong negative effect on y. The scatterplot captures the total relationship (confounded by x2), not the net effect of x1.
• But when you run lm(y ~ x1 + x2), the coefficient for x1 will be close to +2. This is the true net effect: when x2 is held fixed, increasing x1 by 1 unit increases y by ~2 units on average.

Let’s Prove This With R Code #

# Create correlated predictors and a true linear relationship
set.seed(123) # For reproducibility
x2 <- rnorm(100)
x1 <- x2 + rnorm(100, 0, 0.5) # x1 and x2 are positively correlated
y <- 2*x1 - 3*x2 + rnorm(100) # True coefficients: x1=2, x2=-3

# Simple regression (only x1)
simple_lm <- lm(y ~ x1)
cat("Simple regression coefficient for x1:", round(coef(simple_lm)[2], 2), "\n")
# Output: ~-0.8 (negative, matching the scatterplot slope)

# Multiple regression (x1 + x2)
multi_lm <- lm(y ~ x1 + x2)
cat("Multiple regression coefficient for x1:", round(coef(multi_lm)[2], 2), "\n")
cat("Multiple regression coefficient for x2:", round(coef(multi_lm)[3], 2), "\n")
# Output: x1~2.0, x2~-3.0 (matches our true model)

# Plot to visualize the difference
plot(x1, y, main="y vs. x1: Simple vs. Partial Slope")
abline(simple_lm, col="red", lwd=2)
# Add the partial slope (holding x2 constant at its mean)
x2_mean <- mean(x2)
new_data <- data.frame(x1 = seq(min(x1), max(x1), length.out=100), x2 = x2_mean)
y_pred <- predict(multi_lm, newdata=new_data)
lines(new_data$x1, y_pred, col="blue", lwd=2)
legend("topright", legend=c("Simple Slope (Confounded)", "Partial Slope (Controlled)"), col=c("red", "blue"), lwd=2)

Key Takeaways #

• Always interpret multiple regression coefficients as partial effects, not simple slopes.
• Correlation between predictors is the main reason for the mismatch between scatterplot slopes and regression coefficients.
• To visualize a partial effect, you need to hold other predictors constant (like we did by fixing x2 at its mean in the example) instead of plotting the raw y vs. single predictor.

内容的提问来源于stack exchange，提问作者alejandro