2.4 Coefficient interpretation

Coefficient interpretation overview

Coefficient interpretation is how you extract meaning from regression output. A regression gives you numbers, but those numbers only become useful once you can clearly state what they mean in real-world terms. This section covers how to interpret coefficients across different model specifications: linear, logarithmic, interaction terms, and categorical variables.

Coefficient definition

Slope coefficient meaning

The slope coefficient, denoted $\beta_1$, represents the change in the dependent variable ($Y$) for a one-unit increase in the independent variable ($X$), holding all other variables constant. It tells you both the direction (positive or negative) and the magnitude of the relationship between $X$ and $Y$.

This "holding all other variables constant" part is critical. In a multiple regression, $\beta_1$ isolates the marginal effect of $X_1$ on $Y$, net of the other included variables. The slope coefficient is estimated using ordinary least squares (OLS), which minimizes the sum of squared residuals.

Intercept coefficient meaning

The intercept, $\beta_0$, is the predicted value of $Y$ when every independent variable equals zero. Think of it as the baseline: what $Y$ would be before any of the $X$ variables "kick in."

In practice, the intercept often lacks a meaningful real-world interpretation. If $X$ is years of work experience, then $\beta_0$ would be the predicted salary for someone with zero experience, which might be reasonable. But if $X$ is age, then $\beta_0$ would be the prediction at age zero, which is nonsensical. Whether the intercept is interpretable depends on whether $X = 0$ is a plausible value in your data.

Coefficient units

Dependent variable units

The intercept $\beta_0$ is always measured in the same units as $Y$. If $Y$ is measured in dollars, $\beta_0$ is in dollars. If $Y$ is measured in test score points, $\beta_0$ is in test score points.

Independent variable units

The slope coefficient $\beta_1$ is measured in units of $Y$ per unit of $X$. If $Y$ is salary in dollars and $X$ is experience in years, then $\beta_1$ is in dollars per year. Keeping track of units is one of the simplest ways to check whether your interpretation makes sense.

Interpreting linear regression coefficients

Interpreting the slope coefficient

The standard interpretation: $\beta_1$ is the change in $Y$ for a one-unit increase in $X$, holding other variables constant.

• A positive $\beta_1$ means $Y$ increases as $X$ increases.
• A negative $\beta_1$ means $Y$ decreases as $X$ increases.

For example, suppose you regress salary (in dollars) on years of experience and estimate $\hat{\beta}_1 = 1{,}000$. The interpretation: on average, each additional year of experience is associated with a $1,000 increase in salary, holding other factors constant.

Interpreting the intercept coefficient

$\beta_0$ is the predicted value of $Y$ when all $X$ variables equal zero. In the salary example, if $\hat{\beta}_0 = 30{,}000$, the model predicts a starting salary of $30,000 for someone with zero years of experience.

But be careful. If zero isn't a realistic value for your independent variables, the intercept is just a mathematical anchor for the regression line, not something you should interpret substantively.

Putting it together

Consider a model estimating the effect of advertising expenditure (in thousands of dollars) on sales revenue (in thousands of dollars):

$\hat{Y} = 10 + 3.5X$

• $\hat{\beta}_0 = 10$: predicted sales revenue is $10,000 when advertising expenditure is zero.
• $\hat{\beta}_1 = 3.5$: each additional $1,000 in advertising is associated with $3,500 more in sales revenue, on average.

Interpreting interaction terms

What interaction terms are

An interaction term is created by multiplying two independent variables together (e.g., $X_1 \times X_2$). You include one when you suspect the effect of $X_1$ on $Y$ depends on the level of $X_2$.

For example, the return to an additional year of education might differ depending on gender. An interaction term lets you model that.

Main effects vs. interaction effects

Consider this model:

$Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \beta_3 (X_1 \times X_2) + u$

• $\beta_1$ is the effect of $X_1$ on $Y$ when $X_2 = 0$.
• $\beta_2$ is the effect of $X_2$ on $Y$ when $X_1 = 0$.
• $\beta_3$ tells you how the effect of $X_1$ on $Y$ changes for each one-unit increase in $X_2$ (and vice versa).

The total marginal effect of $X_1$ on $Y$ is $\beta_1 + \beta_3 X_2$. You can't interpret $\beta_1$ in isolation once an interaction is in the model. You need to consider both the main effect and the interaction coefficient together.

Interpreting interaction coefficients

• A positive $\beta_3$ means the effect of $X_1$ on $Y$ becomes stronger (more positive or less negative) as $X_2$ increases.
• A negative $\beta_3$ means the effect of $X_1$ on $Y$ becomes weaker (less positive or more negative) as $X_2$ increases.

Interpreting logarithmic coefficients

Logarithmic transformations change how you interpret coefficients. There are four possible specifications, and each has a distinct interpretation. The table below summarizes them:

Log-level model

The dependent variable is logged; the independent variable is not. If $\hat{\beta}_1 = 0.05$, a one-unit increase in $X$ is associated with an approximately 5% increase in $Y$. (Multiply the coefficient by 100 to get the percentage.)

Level-log model

The independent variable is logged; the dependent variable is not. If $\hat{\beta}_1 = 500$, a 1% increase in $X$ is associated with an approximately $500/100 = 5$ unit increase in $Y$. (Divide the coefficient by 100.)

Log-log model

Both variables are logged. The coefficient is interpreted directly as an elasticity: the percentage change in $Y$ for a 1% change in $X$. If $\hat{\beta}_1 = 0.8$, a 1% increase in $X$ is associated with a 0.8% increase in $Y$.

Interpreting categorical variable coefficients

Binary variable coefficients

A binary (or indicator) variable takes values of 0 or 1. Its coefficient represents the average difference in $Y$ between the group coded as 1 and the group coded as 0, holding other variables constant.

For example, if you include a binary variable for gender (0 = male, 1 = female) in a wage regression and estimate its coefficient as $-5{,}000$, the interpretation is: on average, women earn $5,000 less than men, all else equal.

Dummy variables for multi-category variables

When a categorical variable has more than two categories (e.g., education level: high school, bachelor's, master's), you create a set of dummy variables and omit one category as the reference group.

If high school is the reference category:

• The coefficient on the bachelor's dummy is the average difference in $Y$ between bachelor's degree holders and high school graduates, holding other factors constant.
• The coefficient on the master's dummy is the average difference in $Y$ between master's degree holders and high school graduates.

Reference category interpretation

The choice of reference category affects the coefficients but not the model's predictions. Changing the reference category changes what comparison each coefficient represents. If you switch the reference to bachelor's, then the high school and master's coefficients would both be measured relative to bachelor's degree holders. Pick whichever reference category makes the most intuitive sense for your research question.

Coefficient interpretation challenges

Omitted variable bias

Omitted variable bias (OVB) occurs when a variable that belongs in the model is left out, and that variable is correlated with both $Y$ and at least one included $X$. When this happens, the coefficient on the included variable absorbs part of the omitted variable's effect, producing a biased estimate.

For example, if you regress wages on education but omit ability, and ability is positively correlated with both education and wages, then $\hat{\beta}_{education}$ will be biased upward. It will overstate the true effect of education because it's partly picking up the effect of ability.

Multicollinearity effects

Multicollinearity means the independent variables are highly correlated with each other. It doesn't bias your coefficients, but it inflates their standard errors. That makes it harder to find statistically significant results and harder to disentangle the individual effect of each variable.

Signs of multicollinearity include large standard errors on coefficients that theory says should matter, and coefficients that swing wildly when you add or drop a variable. Common remedies include dropping one of the correlated variables or combining them.

Extrapolation dangers

Your regression coefficients are estimated from the data you have. Applying them outside the range of your data (extrapolation) is risky because the linear relationship may not hold in regions where you have no observations. Always note the range of your data when interpreting results, and be cautious about predictions at extreme values of $X$.

Communicating coefficient interpretations

Presenting coefficient estimates

When reporting results, include the point estimate, standard error (or confidence interval), and a measure of statistical significance (p-value or significance stars). But don't stop at statistical significance. Discuss practical significance too: is the estimated effect large enough to matter in the real world?

Visualizing coefficient relationships

Scatter plots with fitted regression lines are a straightforward way to show the relationship your model captures. For multiple regression, partial regression plots (also called added-variable plots) can illustrate the relationship between $Y$ and one $X$ after controlling for the others.

Explaining results to non-experts

Translate your coefficients into plain language. Instead of saying "$\hat{\beta}_1 = 0.03$," say "a one-year increase in education is associated with about a 3% increase in wages." Avoid jargon where possible, use concrete examples, and be upfront about what the model can and cannot tell you. Acknowledging limitations builds credibility.