Why Confidence Intervals Are Easy to Misread
A regression coefficient interval gives a range of plausible values for a particular model coefficient, reflecting uncertainty in its estimate. It is easy to give that range more meaning than it has: a 95% interval is not a 95% probability statement about a fixed coefficient, it is not a range containing 95% of individual outcomes, and an interval that includes zero does not prove that an association is absent.
The first safeguard is to identify exactly what is being estimated. In a regression model, a coefficient describes a feature of the model’s relationship between a response and one or more predictors. A confidence interval concerns uncertainty about that coefficient under the model and its assumptions. It does not, by itself, settle whether the model is appropriate, whether a relationship is causal, or what will happen to any particular individual.
What “95% Confidence” Means
Imagine repeatedly collecting samples in the same way and constructing an interval each time using the same method. If the method’s conditions hold, about 95% of those intervals will contain the fixed coefficient in the long run. The particular interval calculated from the data at hand either contains that coefficient or it does not; the repeated-sampling statement describes the method, not a probability assigned to the coefficient after the interval has been observed.
This is the same distinction between a fixed parameter and a random interval that underlies confidence-interval coverage. Before the sample is collected, the interval endpoints depend on which data happen to be observed. Once the sample is collected, those endpoints are fixed numbers. The coefficient does not become random simply because its value is not known.
Worked Example: Interpreting a 95% Interval for a Slope
Suppose a hypothetical study relates students’ weekly hours of practice to a quiz score in points. A simple regression estimates a slope of \(b=2.4\) points per practice hour, with \(SE=0.8\). For the stated degrees of freedom, use \(t^*=2.10\) for a 95% interval. Assume the observations are independent and that checks of the hypothetical data support a linear mean relationship, roughly constant error variance, and approximately normal errors. These conditions support using the conventional interval here.
First calculate the margin of error: $$ t^*SE=2.10(0.8)=1.68 \text{ points per practice hour}. $$ Then calculate both endpoints: $$ 2.4 \pm 1.68 = (2.4-1.68,\;2.4+1.68)=(0.72,\;4.08). $$
The 95% confidence interval is 0.72 to 4.08 quiz points per additional weekly practice hour. In this simple regression, the interval estimates the slope: the change in the model’s fitted mean quiz score associated with one more practice hour. It does not give a 95% probability that the fixed slope is between 0.72 and 4.08. Rather, the procedure used to construct such intervals would cover the fixed slope in about 95% of repeated samples under the conditions described.
A Coefficient Interval Is Not an Outcome Range
A coefficient interval describes uncertainty about a coefficient, in that coefficient’s units. It is not a range of likely response values for individual observations. For example, a slope measured in minutes per training hour cannot directly be read as a range of task-completion times in minutes. Individual outcomes can differ from the model’s fitted mean because of residual variation, even when the coefficient is known precisely.
A useful check is to name the quantity and its units before interpreting the endpoints. An interval for a slope in a multiple regression describes the fitted relationship with that predictor while the other included predictors are held fixed. The interval still concerns the slope—not the outcome values of people or objects in the data.
Worked Example: A Slope Interval Is Not a Range of Completion Times
Suppose a hypothetical multiple regression predicts task-completion time in minutes from training hours and a baseline skill score. The estimated training coefficient is \(b=-1.5\) minutes per training hour, with \(SE=0.4\). Use \(t^*=2.00\) for a 95% interval. Assume the data collection provides independent observations and that residual checks support the specified linear mean model, approximately constant variance, and approximately normal errors. Under these conditions, the conventional interval is appropriate for this illustration.
The margin of error is $$ 2.00(0.4)=0.8 \text{ minutes per training hour}. $$ Thus the interval is $$ -1.5 \pm 0.8=(-2.3,\;-0.7) \text{ minutes per training hour}. $$
In this model, the estimated fitted mean completion time is associated with a decrease of between 0.7 and 2.3 minutes for each additional training hour, comparing observations with the same baseline skill score. The endpoints are not a range of completion times for an individual. Nor do they say that 95% of individuals will finish 0.7 to 2.3 minutes faster per training hour. They describe uncertainty in the model’s training coefficient.
For a ten-hour difference in training, multiplying the slope interval endpoints by 10 gives an estimated fitted-mean difference of \(-23\) to \(-7\) minutes, with baseline skill held fixed. This is still an interval for a modelled difference in fitted means—not a range of individual completion times. The calculation also assumes that interpreting a ten-hour difference within the model is sensible for the data being considered.
Why Including Zero Does Not Prove No Association
For a slope, zero represents no linear change in the model’s fitted mean response per unit increase in that predictor, holding any other predictors in the model fixed. If a confidence interval includes zero, it means that zero is among the coefficient values the interval procedure has not distinguished from the data at its chosen level of precision. It does not show that the true coefficient equals zero.
An interval containing zero may also contain negative and positive values. Those values can describe different directions of association, and some may represent effects that matter in practice. The interval’s width helps show how much uncertainty remains. A wide interval can include zero along with values that would be important in either direction; a narrow interval around zero would express greater precision but would still not prove exact equality.
Worked Example: A Wide Interval That Includes Zero
Suppose a hypothetical regression predicts exam score in points from weekly study hours, adjusting for a prior assessment score. The estimated study-hours coefficient is \(b=0.7\) points per hour, with \(SE=0.6\). Use \(t^*=2.10\) for a 95% interval. Assume the sample was collected independently and checks support the model’s linear mean relationship, approximately constant error variance, and approximately normal errors. Those assumptions support interpreting this conventional interval.
The margin of error is $$ 2.10(0.6)=1.26 \text{ points per study hour}. $$ The resulting interval is $$ 0.7 \pm 1.26=(0.7-1.26,\;0.7+1.26)=(-0.56,\;1.96) \text{ points per study hour}. $$
Zero is inside the interval. In the specified model, the interval includes a small negative slope, no linear change, and positive slopes as large as about 1.96 points per hour. It therefore shows considerable uncertainty about the direction and size of the conditional linear association. It does not establish that the study-hours coefficient is zero or that no association exists.
The interval also does not prove a causal effect of study hours. The coefficient describes an association in the specified regression model, adjusted for the included prior assessment score. Other features of the data or model may matter to a causal claim.
A Short Justification: The Interval Moves, the Coefficient Does Not
The coverage interpretation follows from what varies under repeated sampling. Let \(\beta\) be the fixed population coefficient and let \(B\) denote the interval calculated from a random sample. If the interval procedure has 95% coverage under its assumptions, then before observing the data, $$ P(\beta \in B)=0.95. $$ Here the probability refers to the random interval \(B\), generated by the sampling process. After observing the sample, the interval becomes a particular fixed interval \(B_{\text{obs}}\). The fixed coefficient is either inside \(B_{\text{obs}}\) or outside it; the coverage statement does not turn into \(P(\beta \in B_{\text{obs}})=0.95\).
This distinction is why “95% confidence” is the appropriate frequentist wording. It describes how well the method works over repeated use, not a probability that can be attached to the unknown fixed coefficient after one interval is calculated.
What an Interval Can and Cannot Establish
A coefficient interval can communicate an estimated direction, a plausible range of coefficient values under the model, and the degree of precision in the estimate. The units matter: a slope’s units are response units per predictor unit, while an intercept has response units. The interval does not automatically tell whether the coefficient is practically important; that depends on the context and the range of predictor values that matter.
It also does not repair a poorly specified model or make an association causal. Conventional intervals rely on conditions such as an appropriate linear mean model, independent errors, and constant error variance; the usual \(t\)-based procedure also relies on normal errors for exact small-sample coverage. If those conditions are not reasonable, the stated coverage may not be achieved. A coefficient interval also leaves out sources of uncertainty not represented by the model and procedure, such as omitted variables or uncertainty about model choice.
Identify the response, predictor, any adjustment variables, and the coefficient’s units.
Describe the coefficient values they estimate, not the range of individual response values.
If zero is included, report that the interval includes zero; do not conclude that the coefficient equals zero.
Interpret the interval under the specified model and its conditions, without turning an association into a causal claim.
Common Mistakes
- Wrong: “There is a 95% chance the fixed slope is in this observed interval.” Correction: The 95% describes the long-run coverage of the interval procedure, when its conditions hold.
- Wrong: “The slope interval gives the range containing 95% of students’ scores.” Correction: The interval concerns uncertainty about the slope; it is not an outcome range for individuals.
- Wrong: “Because zero is in the interval, the predictor has no association with the response.” Correction: Zero is one coefficient value included in the interval; the interval may also contain negative or positive slopes.
- Wrong: “The interval proves that more training causes faster completion.” Correction: A regression coefficient interval describes a modelled association. Causation requires additional justification beyond the interval.
Check Your Understanding
For each question, state what the interval does and does not justify.
- A 95% interval for a slope is 0.3 to 1.1 points per hour. What does the 95% refer to?
- A coefficient interval has endpoints of \(-4\) and \(-1\) minutes per training hour. Is this a range of completion times for individuals? Explain.
- A 95% coefficient interval runs from \(-0.2\) to \(1.4\) units per predictor unit. What does including zero permit you to say, and what does it not prove?
- Why should a slope interval in a multiple regression be interpreted with the other included predictors named?
- Can an interval for a regression coefficient alone establish that changing the predictor causes a change in the response? Why or why not?