Why Test an Individual Coefficient?
A fitted regression model gives estimates of how its predictors relate to the response, but an estimate alone does not show how much uncertainty accompanies it. A test of an individual coefficient asks whether the population coefficient for one specified model term differs from zero, using the estimate together with its standard error.
The question is specific to the model and term being tested. For a slope, the test concerns the modeled change in the mean response associated with a one-unit increase in that predictor, with other included predictors held fixed. For an indicator coefficient, it concerns the modeled difference from the reference group, again with other included predictors held fixed. Neither result, by itself, establishes a causal effect.
Hypotheses and the Test Statistic
Write the regression model as \(Y_i=\beta_0+\beta_1x_{i1}+\cdots+\beta_kx_{ik}+\varepsilon_i\). Here, \(\beta_j\) is the population coefficient for the \(j\)th predictor term. The fitted model provides an estimate \(b_j\) and its standard error \(SE(b_j)\). A two-sided test of whether this coefficient differs from zero is:
The null hypothesis says that the population coefficient for this term is zero in the specified model. The alternative allows either a positive or a negative coefficient. The test statistic measures how many estimated standard errors separate the observed coefficient estimate from zero.
The absolute values matter because departures in either direction count against the null. The p-value is not the probability that the null hypothesis is true. It measures how unusual the observed statistic, or a more extreme one, would be if the null hypothesis and the model assumptions held.
For the conventional t test to have its stated exact distribution, the regression model must give the relevant conditional mean, errors must be independent with mean zero and constant variance, and errors must be normally distributed conditional on the predictors. Inference is understood conditional on the observed predictor values. These conditions support the standard error and reference distribution; they do not guarantee that the model is substantively appropriate. In applications, study design and residual diagnostics can help assess whether the assumptions are reasonable, though they cannot prove them.
Why the Two-Tailed Probability Uses the Absolute Statistic
The t distribution under the null is symmetric about zero. If an observed coefficient estimate is far below zero, an equally far estimate above zero is also evidence against \(H_0\). Thus the two-sided p-value adds the probabilities in both tails. For example, an observed statistic of \(-3\) is assessed by looking at outcomes at or below \(-3\) and at or above \(3\).
This is a short justification for the two-sided calculation: symmetry makes both tail areas equal, so the p-value is twice the upper-tail probability beyond \(|t|\). The degrees of freedom determine which t distribution supplies those tail areas; use the residual degrees of freedom reported or otherwise established for the fitted model.
Worked Example: Testing a Regression Slope
Suppose a hypothetical sample of 20 independent clinics is used to model a clinic’s average weekly change in a health score, in points, from the number of exercise-program sessions offered per week. A simple linear regression estimates the slope as \(b_1=-1.20\) points per additional weekly session, with \(SE(b_1)=0.40\) points per session and \(df=18\). Assume the model’s conditional mean is linear and its errors are independent, normally distributed, and have constant variance; hypothetical residual checks show no marked pattern or extreme influential point.
We test \(H_0:\beta_1=0\) against \(H_a:\beta_1\ne 0\). The estimate is three standard errors below zero:
For \(df=18\), the two-sided p-value is \(P(|T_{18}|\ge 3.00)\approx 0.007685\). At the significance level \(\alpha=0.05\), this p-value is smaller than \(\alpha\), so we reject \(H_0\). The hypothetical data provide evidence that the population slope in this model differs from zero. The estimated association is a decrease of 1.20 score points per additional weekly session, but the test does not show that offering sessions caused the change.
Indicator Coefficients and Context
An indicator variable records membership in a category, often using 1 for one group and 0 for a reference group. In a model with an indicator coefficient, the coefficient represents a modeled difference from that reference group, conditional on any other predictors in the model. Testing the indicator coefficient against zero therefore tests whether that modeled difference is zero. Be explicit about the reference group and any adjustment variables when stating the conclusion.
Worked Example: Testing an Indicator Coefficient
In a hypothetical regression of appointment waiting time on a service-category indicator and several specified adjustment variables, let the reference category be standard service and let the indicator equal 1 for expedited service. The estimated indicator coefficient is \(b_j=2.50\) minutes, with \(SE(b_j)=1.00\) minute and \(df=46\). Suppose the model and error assumptions are reasonable for this hypothetical analysis, including independence, constant error variance, and approximately normal errors.
The hypotheses are \(H_0:\beta_j=0\) and \(H_a:\beta_j\ne 0\). The test statistic is:
The two-sided p-value is \(P(|T_{46}|\ge 2.50)\approx 0.01605\). Since \(0.01605<0.05\), we reject \(H_0\) at \(\alpha=0.05\). There is evidence that the modeled mean waiting time differs between expedited and standard service, conditional on the adjustment variables. The estimated difference is 2.50 minutes, with expedited service higher in this model. This is an adjusted association, not proof that service category caused a difference.
Making and Reporting the Decision
Choose the significance level \(\alpha\) before evaluating the test result. Compare the two-sided p-value with that level: reject \(H_0\) when the p-value is less than or equal to \(\alpha\); otherwise, do not reject \(H_0\). A decision should include the coefficient’s context and the evidence conveyed by the test, not just “significant” or “not significant.”
Name the coefficient, write \(H_0:\beta_j=0\), and use \(H_a:\beta_j\ne 0\) for a two-sided question.
Subtract zero from the estimate and divide by its standard error: \(t=b_j/SE(b_j)\).
Use the absolute value of the statistic and the appropriate residual degrees of freedom to calculate \(P(|T_{df}|\ge|t|)\).
Compare the p-value with \(\alpha\), then describe the evidence about the specified coefficient in the units and context of the model.
A p-value above \(\alpha\) means the test did not find sufficiently strong evidence against \(H_0\) at that chosen threshold. It does not demonstrate that the coefficient is exactly zero, nor does it show that any possible difference is practically unimportant. The estimate and its standard error still describe the fitted result and its uncertainty.
Worked Example: A Result That Does Not Cross the Significance Threshold
Suppose a hypothetical simple regression relates daily study time to a quiz score, measured in points. The estimated slope is \(0.12\) points per additional study hour, its standard error is \(0.12\) points per hour, and the residual degrees of freedom are 20. Assume independent observations and a linear conditional mean, with constant-variance, approximately normal errors and no serious residual pattern.
For \(H_0:\beta_1=0\) versus \(H_a:\beta_1\ne 0\), the test statistic is:
With 20 degrees of freedom, \(P(|T_{20}|\ge 1.00)\approx 0.3293\). Because \(0.3293>0.05\), we do not reject the null hypothesis at the 5% significance level. These hypothetical data do not provide strong evidence that the population slope differs from zero. The estimate is positive, but its direction alone is not enough to establish a nonzero population slope; nor does this result establish that the slope is zero.
Common Mistakes and Why This Test Matters
- Using a one-sided p-value for a two-sided question. A test of “differs from zero” counts evidence in both directions. Use \(P(|T_{df}|\ge |t|)\), not only the tail in the observed direction.
- Ignoring the sign when interpreting the estimate. The sign of \(b_j\) gives the estimated direction, while the two-sided p-value assesses evidence against zero in either direction. Report both appropriately.
- Calling a large p-value proof of no effect. “Do not reject” is not the same as “accept \(H_0\).” The test may simply lack enough precision to distinguish zero from plausible nonzero values.
- Interpreting an indicator coefficient without its reference group. State which category is compared with which, and whether the interpretation is conditional on other model predictors.
- Treating statistical evidence as practical importance or causation. A small p-value does not measure the size or importance of a difference and does not establish a causal relationship.
Individual coefficient tests are useful when a research question concerns one specified term in a fitted regression model. Their conclusions are strongest when the model, assumptions, units, and comparison are stated clearly. Always report the estimate and standard error alongside the test result so that the direction and scale of the modeled association remain visible.
Check Your Understanding
Use the coefficient estimate, standard error, two-sided alternative, and relevant degrees of freedom in each response.
- A coefficient estimate is \(-0.80\), with standard error \(0.20\). What is the t statistic for testing whether its population value differs from zero?
- Why does a two-sided test use the absolute value of the observed t statistic?
- An indicator coefficient compares a treatment category with a reference category while adjusting for other predictors. What modeled difference does testing that coefficient against zero address?
- If a test gives \(p=0.12\) at \(\alpha=0.05\), what is the appropriate decision, and what does it not establish?
- Why does rejecting a zero-slope hypothesis not, on its own, establish that changing the predictor causes a change in the response?