Choose Values for Probing an Interaction Within the Data

Why Probe Values Matter

In a model with an interaction between two quantitative predictors, the slope for one predictor depends on the value of the other. A conditional slope is therefore not one universal number: it describes the fitted change in the outcome for a one-unit increase in one predictor at a specified value of the other. To communicate or estimate that slope, we must choose the value at which to evaluate it.

A useful probe value should serve two purposes. It should correspond to a scientifically interpretable question, such as the slope at a clinically recognized benchmark, and it should be supported by the observed data. A value may be meaningful in principle but poorly represented in the data. Conversely, a value may occur frequently but have no particular scientific relevance. Careful probing considers both.

Model and Meaning of a Probe

Let \(x\) and \(z\) be quantitative predictors and \(Y\) the outcome. For a linear interaction model, the fitted conditional mean is

Here \(b_0,b_1,b_2,b_3\) are estimated regression coefficients. At a chosen value \(z=z^*\), the fitted slope of \(Y\) with respect to \(x\) is \(b_1+b_3z^*\). Thus, \(z^*\) is the probe value: it identifies the condition under which the slope for \(x\) is evaluated. One can instead probe the slope for \(z\) at a chosen \(x\); the same selection principles apply.

The model assumes that the conditional mean of \(Y\) follows the stated linear expression over the region being interpreted. In the usual regression formulation, the errors have mean zero conditional on the predictors; independence and constant error variance are used for standard regression inference, and normally distributed errors support exact small-sample \(t\) procedures. These assumptions do not make unsupported predictor combinations informative. In particular, choosing probe values does not validate the model or justify a causal interpretation.

To see why a probe matters, hold \(z\) fixed and compare fitted outcomes at \(x\) and \(x+\Delta x\):

Dividing by \(\Delta x\), for a nonzero change, gives the fitted change per unit of \(x\), \(b_1+b_3z\). The probe value does not change the model; it specifies where within the interaction pattern we summarize its slope.

Choosing Values with Scientific Meaning and Data Support

Start with the scientific question, not with a formula for generating probe values. Ask which levels of \(z\) represent meaningful conditions or comparisons. A threshold defined by practice, a standard dose, or a value that separates recognizable conditions may be more useful than an arbitrary decimal. If no established reference points exist, values such as the observed median and selected percentiles can give interpretable summaries of the sample.

Then check support in the data. The minimum and maximum of \(z\) provide a first screen: a probe outside that interval asks the model to describe a condition beyond the observed values. But being within the marginal range is not enough. The interaction concerns combinations of \(x\) and \(z\). If observations at a proposed \(z\) occur only alongside a narrow set of \(x\) values, a slope in \(x\) at that \(z\) may be weakly supported. The joint distribution, not merely each predictor's separate range, matters.

A common descriptive choice is the mean and values one standard deviation below and above it. This can be convenient when those values are meaningful and well supported, but it is not a rule. The resulting values may fall outside the observed range, land in sparse regions, or have no useful scientific interpretation. Observed quantiles, natural reference points, or a narrower set of values may be better.

Preplanned Probes and Responsible Interpretation

When possible, identify probe values before examining which conditional slopes appear strongest or most interesting. Choosing values only after seeing the results can encourage selective reporting: a reader may see a compelling slope at one value without learning that many alternatives were considered. A small set tied to scientific reference points or a clearly stated descriptive plan is easier to interpret and report transparently.

If a scientific benchmark falls outside the observed range, it can still be named as a question of interest, but the data at hand do not directly support an answer there. The appropriate response is to acknowledge the extrapolation and, when possible, obtain data covering that condition. Moving the probe to a nearby observed value may answer a related question, but it does not answer the original benchmark question.

Likewise, a value inside the range should not automatically be described as well supported. A predictor can have observations at both ends of its range while the combinations needed to estimate a conditional slope are rare. Inspecting the data jointly—for example, by tabulating ranges within bands of the probe predictor—helps reveal this limitation. Where support is thin, describe the slope cautiously and avoid implying that the estimate is equally reliable everywhere.

Common Mistakes

• Using an appealing value without checking the data. A scientific threshold outside the observed range may produce a model calculation, but not a data-supported conclusion at that threshold.
• Assuming any in-range value is safe. The other predictor may not vary near that probe value, so the slope can depend on a poorly represented combination.
• Choosing mean plus or minus one or two standard deviations automatically. These values can be out of range or scientifically unhelpful; inspect them rather than treating them as mandatory.
• Reporting only the most striking probe. Selective choice after examining results can conceal other examined values. State the basis for selecting probes and report the comparison transparently.
• Interpreting a fitted conditional slope causally. A regression interaction describes an adjusted association under the model; probe selection alone does not establish a causal effect.