Reading Confidence Intervals on the Outcome Scale

Why Read an Interval in the Outcome’s Units?

A regression coefficient gives an estimated change in an outcome for a specified change in a predictor. But a single estimate does not show how uncertain that estimate is. A confidence interval adds that information: its endpoints describe a range of coefficient values compatible with the data and the assumptions used for the analysis. Reading those endpoints in the outcome’s original units makes the range concrete. An interval measured in exam points, minutes, or millimeters of mercury can be related directly to the size of changes those units represent.

The aim is not just to notice whether an interval is “wide” or “narrow.” Its location tells us which effect values are represented, while its width tells us how broad a range of values remains compatible with the estimate. Both matter when considering practical consequences.

What a Coefficient Interval Estimates

For a multiple linear regression with \(n\) observations, write the model as

Here \(Y_i\) is the outcome for observation \(i\), \(x_{ij}\) are predictor values, \(\beta_j\) are population regression coefficients, and \(\varepsilon_i\) is the error. The coefficient \(\beta_j\) describes the change in the conditional mean of the outcome for a one-unit increase in predictor \(j\), with the other predictors held fixed. Its units are outcome units per unit of predictor \(j\). The estimate from the data is written \(b_j\).

A commonly used \(100(1-\alpha)\%\) confidence interval for \(\beta_j\) has the form below. For the usual \(t\)-based interval, \(SE(b_j)\) is the standard error of the coefficient estimate, \(df=n-k-1\), and \(t^*_{1-\alpha/2,df}\) is the corresponding \(t\) critical value.

The interval has a confidence level because the method has a repeated-sampling property: if the model and its assumptions hold, and the same procedure is used on many comparable samples, approximately \(100(1-\alpha)\%\) of the resulting intervals will contain the true coefficient. This does not mean there is a \(100(1-\alpha)\%\) probability that the fixed coefficient lies in this particular interval.

Conditions for Interpreting the Usual Interval

A confidence interval is only as trustworthy as the model and procedure that produced it. For the usual \(t\)-based regression interval, the data should be suitable for the fitted linear model: the conditional mean should be adequately represented by the chosen predictors and their specified form; observations should be independent under the sampling or study design; and the error variance should be approximately constant across predictor values. For exact \(t\)-based inference in a small sample, the errors should also be normally distributed. The predictors must not be perfectly redundant, so that the coefficient can be estimated.

In practice, model checks and knowledge of how the data were collected help assess these conditions. The examples below stipulate that these checks are reasonable for the hypothetical data. If important assumptions are doubtful, the numerical interval may not have its stated coverage. Also, this interval concerns a population coefficient, not the outcome for a particular individual: individual outcomes vary for reasons beyond the fitted conditional mean.

Worked Examples: Reading Endpoints and Width

Why Interval Width Matters

The width of a symmetric \(t\)-interval follows directly from its endpoints. If the lower endpoint is \(L=b_j-t^*SE(b_j)\) and the upper endpoint is \(U=b_j+t^*SE(b_j)\), then

Thus, at a fixed confidence level and degrees of freedom, a larger standard error produces a wider interval. A wider interval leaves a broader range of coefficient values compatible with the data and model; on the outcome scale, that means a broader range of possible changes in the conditional mean per predictor unit. A narrower interval concentrates that range more tightly. Width is not a measure of whether an effect matters: practical importance depends on what the endpoint values mean in context and on relevant outcome-based considerations.

Width must also be read alongside location. The same width can describe a range centered on a large positive coefficient, a large negative coefficient, or a value near zero. Consequently, “the interval is narrow” is not a complete interpretation. State the endpoints, their units, and what changes in the outcome they represent. If a practical benchmark has already been defined for the outcome, the endpoints can be described relative to that benchmark; the interval’s width alone cannot settle that comparison.

A Short Guide to Interpreting the Interval

Common Mistakes

• Giving endpoints without units. An interval such as \([-1.0,0.2]\) is difficult to interpret alone. Say, for example, “rating points per minute of waiting.”
• Treating the point estimate as the only plausible effect. The estimate is the center of the interval, not a certainty. Interpret the full range of coefficient values.
• Calling the interval a range for individual outcomes. A coefficient interval concerns a population regression coefficient. Individual outcomes can vary much more because of person-to-person differences and random error.
• Reading width as practical importance. Width describes uncertainty in coefficient units. It does not decide whether the represented effects matter; that requires interpreting the values in context.
• Claiming a probability for the fixed coefficient. A 95% confidence level describes the long-run performance of the interval procedure under its assumptions, not a 95% probability that this particular interval contains the coefficient.
• Turning association into causation. A regression coefficient describes a modeled relationship. A confidence interval does not, on its own, justify a causal conclusion.