Interpreting Crossover Patterns and Uncertain Slopes

Why a Visible Crossover Needs More Than a Plot

Two fitted lines may cross in a graph, or the estimated slope of a predictor may change sign across moderator values. These patterns are useful descriptions of the fitted model, but they are not, by themselves, evidence that the underlying effects differ meaningfully. The coefficient estimates are uncertain. A slope that looks positive at one moderator value and negative at another may have confidence intervals broad enough to include zero, and two fitted lines that intersect at one estimated point may still differ substantially elsewhere.

The key is to match the uncertainty calculation to the quantity being interpreted. A conditional slope is a linear combination of regression coefficients. A difference between two fitted lines at a chosen predictor value is another linear combination. Each has a standard error and can be summarized with a confidence interval. The estimated location where lines cross is a different quantity, and simply reporting its point estimate does not tell us how precisely it is located.

Model, Conditional Slopes, and Crossovers

For a continuous predictor \(x\) and a moderator \(z\), consider the interaction model

The fitted conditional mean at given \(x\) and \(z\) is \(\hat{Y}=\hat{\beta}_0+\hat{\beta}_1x+\hat{\beta}_2z+\hat{\beta}_3xz\). Holding \(z\) fixed, the fitted slope for a one-unit increase in \(x\) is \(\hat{\beta}_1+\hat{\beta}_3z\). Thus, the slope depends on the moderator value; \(\hat{\beta}_1\) alone is the slope only when \(z=0\).

For example, if \(z\) is a binary group indicator, the fitted difference between the group coded 1 and the group coded 0 at a given \(x\) is \(\hat{\beta}_2+\hat{\beta}_3x\). If \(\hat{\beta}_3\ne0\), the two fitted lines intersect at the \(x\)-value that makes this difference zero: \(\hat{x}_{\mathrm{cross}}=-\hat{\beta}_2/\hat{\beta}_3\). This algebra identifies an estimated intersection; it does not quantify its uncertainty.

Before using standard errors and \(t\)-based confidence intervals for these quantities, assume the interaction model is an appropriate description of the conditional mean, observations have independent errors with mean zero and constant variance, and the model matrix has full rank. For exact \(t\)-based inference, also assume normally distributed errors. In applications where these assumptions are doubtful, the reported standard errors and intervals may not have their stated properties. The examples below treat the assumptions as plausible for the hypothetical data and inference shown.

Uncertainty for a Conditional Slope

A conditional slope combines two estimated coefficients, so its uncertainty depends on both coefficient variances and their covariance. For estimates \(\hat{\beta}_1\) and \(\hat{\beta}_3\), the estimated variance of the slope at moderator value \(z\) is

This follows from the variance rule \(\operatorname{Var}(A+B)=\operatorname{Var}(A)+\operatorname{Var}(B)+2\operatorname{Cov}(A,B)\): take \(A=\hat{\beta}_1\) and \(B=z\hat{\beta}_3\), noting that \(z\) is fixed. The standard error is the square root of this variance. A confidence interval is the estimated slope plus or minus a \(t\) critical value times its standard error, using the model's residual degrees of freedom.

Contrasts at the Estimated Intersection

When two fitted lines cross, their estimated contrast is zero at the estimated intersection by construction. That fact is about the fitted lines, not a test result. Uncertainty in the contrast at that \(x\)-value is still needed to assess whether the data are compatible with a nonzero difference in the conditional means.

For the binary-group model, let \(G=1\) identify one group and \(G=0\) the other:

Subtracting the fitted mean for \(G=0\) from the fitted mean for \(G=1\) gives \(\hat{\beta}_2+\hat{\beta}_3x\). Its estimated variance is \(\widehat{\operatorname{Var}}(\hat{\beta}_2)+x^2\widehat{\operatorname{Var}}(\hat{\beta}_3)+2x\widehat{\operatorname{Cov}}(\hat{\beta}_2,\hat{\beta}_3)\). The covariance matters: ignoring it can produce an incorrect standard error and interval.

Sign Reversals in Conditional Slopes

A plot may suggest that the association between \(x\) and \(Y\) changes direction as \(z\) changes. To assess that pattern, calculate the conditional slope at relevant moderator values and examine its uncertainty at each value. A positive estimate at one value and a negative estimate at another are not sufficient evidence that the underlying slopes have opposite signs.

What a Crossover Does—and Does Not—Establish

A crossover can be a useful summary of an interaction: it identifies where the fitted contrast changes sign or where the fitted conditional means coincide. But several questions should remain separate:

• Where do the fitted lines cross? Solve for the predictor value at which the estimated contrast is zero.
• Is a contrast distinguishable from zero at a chosen value? Calculate its standard error and confidence interval at that value.
• Is a difference large enough to matter? Interpret the interval against a context-appropriate threshold for practical importance, if one has been specified.

The first question concerns a point estimate; the second concerns statistical uncertainty; the third concerns practical meaning. A confidence interval that includes zero does not prove equality. A narrow interval around zero may rule out differences large enough to matter, while a wide interval that includes zero may simply be inconclusive. Conversely, an interval excluding zero does not automatically imply an important difference.

The estimated intersection \(-\hat{\beta}_2/\hat{\beta}_3\) is a ratio of estimates. Its location can be highly uncertain, especially when the estimated interaction \(\hat{\beta}_3\) is close to zero. Do not treat the calculated point as a precisely known boundary, or infer its uncertainty by separately inserting the endpoints of coefficient confidence intervals into the ratio. The contrast interval at a specified \(x\) answers a different question from an interval for the crossover location. Also check that the relevant predictor and moderator values are supported by the observed data; an intersection far outside that support is not a reliable basis for substantive interpretation.

Common Mistake

Wrong reasoning: “The fitted lines cross at 3 hours, so the groups are equal at 3 hours; and because they are on opposite sides of the crossover, the groups must differ meaningfully on either side.”

Correction: The fitted contrast is zero at the estimated intersection, but the population contrast there remains uncertain. Calculate its standard error and interval. At other predictor values, calculate the contrast and interval at those values too. Then distinguish evidence that a contrast differs from zero from evidence that its size is practically important. A plotted intersection alone answers none of these uncertainty questions.