Interpreting a Slope Confidence Interval in Context

What a Slope Interval Describes

A regression-slope confidence interval is most useful when its numbers are translated back into the variables they describe. An interval such as \((-3.0,-0.6)\) is not self-explanatory: the reader needs to know what the predictor and response measure, which population is being described, and how the interval’s units are formed.

The population slope, denoted \(\beta_1\), describes the linear change in the population mean response associated with a one-unit increase in the predictor. A confidence interval uses sample data to estimate a range of plausible values for this population quantity. Its endpoints must therefore be interpreted as possible values of the same slope, in the same units, for the same target population.

Definition: In the population linear regression model \(E(Y\mid X=x)=\beta_0+\beta_1x\), \(\beta_1\) is the population slope. It is the change in the population mean response associated with a one-unit increase in the predictor. If \(Y\) is measured in response units and \(X\) in predictor units, then \(\beta_1\) and every endpoint of its confidence interval have units of response units per predictor unit.

This describes a change in the mean response, not a promise about what will happen to every individual. Nor does a slope, by itself, establish that changing the predictor causes a change in the response. A causal conclusion requires an appropriate study design and additional justification.

Model and Conditions for Interpretation

A slope interval is meaningful only relative to a model and a target population. The standard simple linear regression model for inference assumes a linear conditional mean, independent observations or errors, errors with mean zero at each predictor value, constant error variance, and normally distributed errors for the usual exact \(t\)-based interval. In addition, the way the data were collected must support inference to the population being named. For example, a random sample can support inference to the population from which it was drawn, subject to the model conditions.

Under these conditions, a \(100(1-\alpha)\%\) confidence interval for the population slope has the form

\hat{\beta}_1 \pm t^* SE(\hat{\beta}_1)

Here \(\hat{\beta}_1\) is the sample slope, \(SE(\hat{\beta}_1)\) is its standard error, and \(t^*\) is the appropriate critical value with \(n-2\) degrees of freedom. The interval’s lower endpoint is \(\hat{\beta}_1-t^*SE(\hat{\beta}_1)\); its upper endpoint is \(\hat{\beta}_1+t^*SE(\hat{\beta}_1)\). The earlier material on confidence intervals for the population slope develops this calculation. The present task is to explain what the resulting endpoints mean.

The confidence level describes the long-run performance of the interval procedure. If the same sampling and interval-building process were repeated many times under the model, about \(100(1-\alpha)\%\) of the resulting intervals would contain the fixed population slope. It does not mean that, after an interval has been computed, there is a \(100(1-\alpha)\%\) probability that this particular interval contains \(\beta_1\).

Why the Confidence Statement Is Correct

The repeated-sampling interpretation follows from the pivot used to construct the interval. Under the stated model conditions, the standardized difference between the sample slope and the population slope follows a \(t\) distribution with \(n-2\) degrees of freedom. For a two-sided confidence level \(100(1-\alpha)\%\), the critical value is chosen so that the standardized quantity falls between \(-t^*\) and \(t^*\) with probability \(1-\alpha\).

P\left(-t^* \leq \frac{\hat{\beta}_1-\beta_1}{SE(\hat{\beta}_1)} \leq t^*\right)=1-\alpha

Multiplying through by the positive standard error and rearranging places the fixed \(\beta_1\) between the two random endpoints:

P\left( \hat{\beta}_1-t^*SE(\hat{\beta}_1) \leq \beta_1 \leq \hat{\beta}_1+t^*SE(\hat{\beta}_1) \right)=1-\alpha

Before data are collected, the interval endpoints vary from sample to sample, so the coverage probability refers to the interval-building procedure. Once one interval has been calculated, its endpoints are fixed; the correct interpretation describes the population slope and the repeated-sampling coverage, not a probability assigned to that fixed interval.

Reading the Endpoints in Context

To interpret a slope interval, identify four things: the target population, the response, the predictor and its unit, and the interval’s confidence level. Then translate each endpoint into the same sentence, changing only the value of the slope. Preserve the sign and the units. A negative endpoint describes a decrease in the population mean response per predictor unit; a positive endpoint describes an increase.

Name the target.
Say which population’s linear mean response relationship the interval estimates, based on the data collection and model.

Keep the units attached.
Express both bounds in response units per one predictor unit, including any measurement scale such as hours, kilograms, or tens of units.

Interpret both bounds.
State the range of plausible population mean changes associated with a one-unit predictor increase, at the stated confidence level.

Worked Example: A Negative Interval for Resting Heart Rate

Suppose a hypothetical random sample of \(24\) adults is used to relate weekly exercise time \(X\), in hours, to resting heart rate \(Y\), in beats per minute. The estimated slope is \(\hat{\beta}_1=-1.80\) beats per minute per additional weekly exercise hour, and its standard error is \(0.60\) in the same units. A 95% interval uses \(t^*=2.074\) with \(22\) degrees of freedom. The sample scatterplot is roughly linear; the residuals show no clear pattern or funnel, are approximately symmetric, and the adults were sampled independently. These checks support, but cannot prove, the model conditions. Interpret the interval.

First calculate the margin of error:

$$ 2.074(0.60)=1.2444 $$

The endpoints are:

-1.80-1.2444=-3.0444,\qquad -1.80+1.2444=-0.5556

Thus the 95% confidence interval is approximately \((-3.04,-0.556)\) beats per minute per weekly exercise hour. For the population of adults represented by the sampling design, the interval estimates that each additional hour of weekly exercise is associated with a change in mean resting heart rate between a decrease of about \(0.556\) and a decrease of about \(3.04\) beats per minute. The negative signs matter: both endpoints describe decreases. This is an association in the stated population, not evidence by itself that exercise causes the heart-rate difference.

In repeated samples taken in the same way and analyzed under the model, about 95% of intervals constructed by this method would contain that population’s slope. The statement does not assign a 95% probability to this particular fixed interval.

Worked Example: A Positive Interval and Its Units

In a hypothetical study, a random sample of \(18\) students from a specified school population is used to examine the association between hours of tutoring and quiz score in points. The estimated slope is \(3.20\) points per tutoring hour, and \(SE(\hat{\beta}_1)=1.10\) points per tutoring hour. Suppose the linear pattern is reasonable, observations are independent, residual spread is roughly constant, and residuals are approximately normal. Use \(t^*=2.120\) with \(16\) degrees of freedom for a 95% interval.

The margin of error is:

$$ 2.120(1.10)=2.332 $$

The lower and upper endpoints are \(3.20-2.332=0.868\) and \(3.20+2.332=5.532\). Therefore the 95% interval is \((0.868,5.532)\) quiz points per tutoring hour.

For the population represented by this sample, the interval estimates that an additional tutoring hour is associated with an increase of about \(0.868\) to \(5.53\) points in the mean quiz score. These are possible values of the population slope, not predictions that each student’s score will rise by that amount. Since the example is observational, the interval alone does not establish a causal effect of tutoring.

Worked Example: An Interval That Includes Zero

A hypothetical sample of \(30\) patients is used to study the linear association between a measured activity score \(X\), in points, and a laboratory response \(Y\), in milligrams per liter. The estimated slope is \(-0.70\) milligrams per liter per activity-score point, with standard error \(0.50\) in those units. Assume the patients were sampled independently from the target population, the relationship is reasonably linear, and residual checks support constant spread and approximate normality. With \(t^*=2.048\) and \(28\) degrees of freedom, find and interpret the 95% interval.

The margin of error is:

$$ 2.048(0.50)=1.024 $$

The lower endpoint is \(-0.70-1.024=-1.724\), and the upper endpoint is \(-0.70+1.024=0.324\). The interval is \((-1.724,0.324)\) milligrams per liter per activity-score point.

For the target patient population, the interval estimates that each additional activity-score point is associated with a change in mean laboratory response ranging from a decrease of about \(1.72\) milligrams per liter to an increase of about \(0.324\) milligrams per liter. Zero is among the plausible slope values represented by the interval, so the interval does not establish a single direction for the population linear association. The correct interpretation is not that the slope is zero, nor that there is no relationship of any kind; it is that this interval includes both zero and nonzero values.

Common Mistakes

Leaving out the units. Saying “the slope is between \(-3.04\) and \(-0.56\)” is incomplete. State response units per predictor unit and name what both variables measure.
Interpreting the bounds as individual outcomes. A slope describes change in the population mean response per predictor unit, not a guaranteed change for each person or a prediction interval for an individual.
Changing the population mid-interpretation. A sample from one school or clinic does not automatically support claims about all students or all patients. Name the population supported by the sampling design.
Claiming a causal effect from an association. A confidence interval quantifies uncertainty about a model slope. It does not, by itself, show that changing the predictor causes the response to change.
Misstating the confidence level. “There is a 95% chance the true slope is in this interval” treats the fixed parameter as random after the interval is observed. Instead, describe the long-run coverage of the procedure.

Key takeaway: A contextual slope interval gives plausible values for a population mean-response change per one predictor unit. Name the target population, retain response-per-predictor units, interpret both endpoints, and explain confidence as repeated-sampling coverage.

Check Your Understanding

For each question, make the population and units explicit where they are relevant.

A 90% interval for a slope is \((1.2, 4.5)\) kilograms per centimeter. What does each endpoint mean for the population mean response?
A slope interval for a population of clinic patients is \((-2.1, 0.4)\) milligrams per liter per score point. What direction(s) of change in mean response are represented?
In your own words, explain the repeated-sampling meaning of a 95% confidence level.
Why is it incorrect to interpret a slope of \(3\) points per hour as saying every individual gains exactly 3 points for each additional hour?
What extra kind of evidence is needed before describing an estimated slope as a causal effect?