Why Check Data Support?
A marginal effect can be calculated at a selected predictor value, but the calculation alone does not show whether the data contain observations relevant to that value. For example, an interaction model may report how the expected outcome changes with \(X\) when a moderator \(Z\) is fixed at a chosen level. Before interpreting that result, ask whether the observed data include cases with \(X\) and \(Z\) together near the values in question. Looking at each predictor separately is not enough.
This check is about the evidence available in the observed predictor values. It does not establish that the regression model is correct, that an estimated effect is precise, or that a relationship is causal. It helps readers understand which observed parts of the predictor space are well represented and which are sparse.
Model and Meaning of Support
For two continuous predictors, consider the linear interaction model
Here \(Y_i\) is the outcome for observation \(i\), \(X_i\) and \(Z_i\) are predictors, and \(\varepsilon_i\) is the error. The conditional mean is
Under this model, the marginal effect of \(X\) at \(Z=z\) is \(\beta_1+\beta_3z\). This interpretation relies on the stated conditional-mean form. Standard inference also relies on appropriate assumptions for the reported standard errors, including independent observations and a suitable variance calculation. Checking data support does not verify those assumptions.
For a categorical moderator, the same principle applies within each category. If a conditional slope is reported for group \(G\), inspect the values of the continuous predictor among observations in that group. A total sample count can conceal that one group has few observations, or that its observations cover only a limited part of the predictor range.
Why Marginal Ranges Are Not Enough
The range of a predictor is the interval from its observed minimum to its observed maximum. A target value lying within that interval is not necessarily close to any observed value. More importantly, separate ranges do not reveal which values occurred together on the same observation.
A short justification makes the distinction precise. Let \(S\) be the set of observed pairs \((X_i,Z_i)\). The marginal ranges describe the separate projections of \(S\) onto the \(X\)- and \(Z\)-axes. A pair can lie within both projected ranges without belonging to, or being near, \(S\). Thus, membership of \(x\) in the \(X\) range and \(z\) in the \(Z\) range does not imply that the pair \((x,z)\) has joint support.
Ranges are still useful: they quickly show whether target values fall outside the observed marginal spans. But a range cannot tell whether values inside it are concentrated at a few separated points, whether a group is represented near the target, or whether two target predictor values occur together. Counts and joint displays answer different parts of those questions.
Worked Example: Ranges Contain a Target but Nearby Pairs Do Not
Suppose a hypothetical set of four clinics records a staffing measure \(X\), in staff-hours per day, and a workload measure \(Z\), in cases per day. The observed pairs are \((0,0)\), \((0,10)\), \((10,0)\), and \((10,10)\). A report considers the marginal effect of staffing at \(Z=5\), and a reader asks about the predictor profile \(X=5,\ Z=5\).
| Clinic | \(X\) (staff-hours per day) | \(Z\) (cases per day) |
|---|---|---|
| 1 | 0 | 0 |
| 2 | 0 | 10 |
| 3 | 10 | 0 |
| 4 | 10 | 10 |
The separate ranges are \(0\) to \(10\) staff-hours per day for \(X\) and \(0\) to \(10\) cases per day for \(Z\). Therefore, both target values fall within their respective ranges. That fact does not show that there are observations near either target, or that the two target values occur together.
For an explicit local check, define “near” here as within 1 staff-hour per day of \(X=5\) and within 1 case per day of \(Z=5\). Each observed \(X\) value is either 0 or 10, so every observation is 5 staff-hours from the target \(X\). Each observed \(Z\) value is also either 0 or 10, so every observation is 5 cases from the target \(Z\). No observation meets the stated joint criterion. The joint count is \(0\) out of \(4\), or \(0/4=0\%\).
The ranges place both target values between the observed minima and maxima, but the observations provide no nearby joint pairs under the chosen criterion. The criterion is part of the diagnostic description, not a universal definition of proximity.
Check Counts Within Groups
When a moderator is categorical, examine the continuous predictor separately in every group for which a conditional effect is reported. Record each group’s sample size and range, then count observations in meaningful intervals around the value or region of interest. A group-wide range can include a target even when no observations are close to it. If the group is small, even a nonzero count may warrant caution.
Worked Example: A Conditional Slope in a Sparse Part of One Group
In a hypothetical training program, \(X\) is practice time in hours per week and \(G\) identifies one of two training groups, A or B. A fitted interaction model reports a conditional slope for practice time in group B. Before describing how well the data represent a practice time of 8 hours per week in that group, inspect the observed values:
Group A has \(X=0,2,4,6,8,10\) hours per week, so \(n_A=6\). Group B has \(X=0,2,4,10,10\) hours per week, so \(n_B=5\). The full data set has \(6+5=11\) observations.
Both group ranges are 0 to 10 hours per week, so the target value 8 lies within each range. Now count observations within 1 hour of 8, meaning from 7 through 9 hours per week. Group A has one such observation, \(X=8\), so its local count is \(1/6=0.167\), or about \(16.7\%\). Group B has no observations in that interval, so its local count is \(0/5=0\%\). The two observations at \(X=10\) in group B are each 2 hours from the target, not within the defined interval.
Thus, group B’s range alone would give an incomplete picture: it includes 8, but the group has no observation within the specified local interval. The conditional slope may still be computed from the fitted model, but a description focused on \(X=8\) in group B should disclose the sparse local coverage. The check is diagnostic, not an automatic rule that a model estimate must or must not be reported.
Inspect the Joint Distribution
For two continuous predictors, a scatterplot is often the most direct display of their joint values. When many points overlap, or when a concise summary is needed, divide each predictor into scientifically meaningful bins and tabulate the number of observations in every joint cell. Binning simplifies the display, so the bin boundaries and units should be reported. A cell count describes how many observations fall in that region; it does not show their exact positions within the cell.
For a categorical moderator and a continuous predictor, group-specific summaries are a form of joint display: each group is one category of the moderator, and the predictor distribution is shown within it. With two categorical predictors, a cross-tabulation of their combinations provides the corresponding check.
Worked Example: Marginal Counts Hide a Sparse Joint Cell
A hypothetical data set contains 30 students. Let \(X\) be weekly study time, from 0 to 8 hours, and let \(Z\) be weekly tutoring time, from 0 to 6 hours. For a compact display, classify study time as low (0 to 4 hours) or high (more than 4 to 8 hours), and tutoring time as low (0 to 2 hours) or high (more than 2 to 6 hours). The observed joint counts are:
| Study time | Low tutoring time | High tutoring time | Total |
|---|---|---|---|
| Low (0 to 4 hours) | 12 | 2 | 14 |
| High (more than 4 to 8 hours) | 1 | 15 | 16 |
| Total | 13 | 17 | 30 |
The row totals show 14 students with low study time and 16 with high study time. The column totals show 13 with low tutoring time and 17 with high tutoring time. These marginal totals suggest that each broad category has substantial representation considered separately. But the low-study, high-tutoring cell contains only 2 students. Its share of the full sample is \(2/30=0.0667\), or about \(6.67\%\).
For a report about the combination of low study time and high tutoring time, the joint cell is more relevant than either marginal total. The table identifies this combination as sparse relative to the others, although it does not show the exact predictor values of the two students or whether they lie near a particular target within the bin. A scatterplot or narrower, justified bins could provide more detail. The conclusion should remain about observed coverage, not about the reason for the pattern or the correctness of an interaction model.
A Practical Support Check
Before reporting a conditional effect, predicted profile, or average over predictor profiles, make the support assessment specific to the quantity being reported. A broad statement such as “the predictors are within range” omits group membership, proximity, and joint occurrence. A clear report can state the relevant counts and their limits without turning the diagnostic into a formal test.
State the predictor values, categories, and units at which the marginal effect or profile is being discussed.
Compare each target value with that predictor’s observed minimum and maximum. Treat this only as a first check, not evidence of joint coverage.
For a categorical moderator, report group sizes, ranges, and counts near the relevant predictor values within each group.
Use a scatterplot or a clearly defined joint table to see whether observations occur together near the target combination.
State where coverage is sparse or uneven. Keep this observation distinct from uncertainty in the estimated effect and from assumptions about the model.
Common Mistake and Key Takeaway
A common mistake is to reason, “The target \(X\) is inside the observed \(X\) range, and the target \(Z\) is inside the observed \(Z\) range, so the data support the target combination.” This is not justified. A range records only the observed minimum and maximum; it does not establish that observations are near a target within that range. Separate ranges also do not show whether the two target values occur together. The correction is to inspect the group-specific counts and the joint distribution, and to describe what those diagnostics do and do not reveal.
Check Your Understanding
Answer each question by referring to the relevant ranges, groups, or joint counts.
- Why does a target value lying inside a predictor’s observed range not establish that observations are near that value?
- A continuous predictor has a range that includes a chosen value in both moderator groups. What additional group-specific information would help assess support?
- In the four-clinic example, why do the two marginal ranges fail to establish support near \(X=5,\ Z=5\)?
- What does the low-study, high-tutoring cell count of 2 establish, and what does it leave unknown?
- Explain why a support check does not verify that the interaction model is correctly specified.