Why the Choice of Moderator Values Matters
An interaction model says that an effect may vary with a moderator. But a model can produce conditional effects at many possible moderator values, and a report usually shows only a few. Those values shape what readers see. If they are meaningful and grounded in the data, the reported effects can help answer a practical question. If they are chosen mechanically or without context, the results may be hard to interpret or may emphasize values that are not useful to the intended audience.
The aim is not to find a universally correct set of moderator values. Instead, select a small set that reflects the question, the moderator’s units and distribution, and the values actually observed. Explain the choice so readers can tell what the reported effects describe.
Conditional Effects and the Values at Which They Are Reported
For a continuous predictor \(X\) and a continuous moderator \(Z\), consider the linear interaction model below. Here \(Y\) is the outcome, \(\beta_0,\beta_1,\beta_2,\beta_3\) are regression coefficients, and \(\varepsilon\) is the error term.
The model’s conditional mean is linear in \(X\) and \(Z\), with the effect of \(X\) allowed to change linearly with \(Z\). In this model, the conditional slope of \(Y\) with respect to \(X\), at a specified moderator value \(z\), is:
This expression is the quantity being reported when we describe the conditional effect of \(X\) at \(Z=z\). The choice of \(z\) does not change the fitted model; it determines which part of the model’s pattern is being summarized.
The model’s form must be suitable for the analysis: the conditional mean should be reasonably represented by the specified linear interaction, and the variables and units should be defined appropriately. The usual assumptions needed for estimation and inference also apply when reporting estimates or confidence intervals. Choosing meaningful moderator values does not make an unsuitable model appropriate.
Three Useful Bases for Choosing Values
A strong choice often combines more than one consideration. A value may be directly relevant to a decision, representative of a part of the observed distribution, and easy to explain in its original units.
- Subject-matter relevance. Choose values tied to a meaningful threshold, policy, clinical category, or practical scenario. Explain why that value matters; do not treat an arbitrary round number as important merely because it is convenient.
- Observed quantiles. Quantiles such as the 25th, 50th, and 75th percentiles can describe lower, middle, and higher parts of the observed moderator distribution. They adapt to the data’s scale and may be more informative than a fixed set of round numbers.
- The observed range and units. Report values in units readers recognize and keep them within the range represented in the data. The overall range is a useful initial check, but it does not by itself establish that every combination of predictors and moderators is well represented.
The last point is deliberately limited: a moderator value can lie within the overall range while being uncommon for a particular predictor value. A separate check of data support across predictor combinations is needed for that question.
Worked Example: Choosing Ages That Answer a Practical Question
Suppose a hypothetical program studies whether the association between monthly coaching sessions, \(X\), and weekly exercise minutes, \(Y\), differs by age, \(Z\), measured in years. The fitted conditional mean uses age centered at 50:
Assume the estimated coefficients are \(\hat{\beta}_1=18\) minutes per session and \(\hat{\beta}_3=-0.20\) minutes per session per year. The estimated slope at age \(z\) is therefore \(18-0.20(z-50)\). In this hypothetical data set, ages range from 20 to 80 years. Program staff specifically want to compare effects at ages 30, 50, and 70, each of which is within that range.
At age 30, the estimated conditional slope is:
At age 50, it is \(18-0.20(50-50)=18\). At age 70, it is:
The selected ages have a clear program-specific purpose, are stated in years, and are within the observed age range. The corresponding estimates describe associations: one additional monthly coaching session is associated with an estimated 22, 18, or 14 additional weekly exercise minutes at ages 30, 50, or 70, respectively. This interpretation does not establish that coaching causes a change in exercise.
Quantiles as Distribution-Based Reference Points
When there is no single decision threshold, observed quantiles provide a systematic way to choose values. For example, reporting conditional effects at the lower quartile, median, and upper quartile describes the model’s pattern across the central portion of the observed moderator distribution. These are descriptive reference points, not special values at which the model becomes more valid.
Use the moderator’s original units in the final report. It is fine to calculate quantiles with software, but readers should be able to understand what each value means. Also check whether many observations share a value, whether the distribution is highly skewed, and whether rounding would materially change the chosen value. Quantiles do not remove the need to explain the substantive interpretation.
Worked Example: Reporting Effects at Observed Experience Quantiles
In a hypothetical workforce analysis, \(X\) is hours of a training course, \(Y\) is a later assessment score, and \(Z\) is employees’ years of prior experience. The fitted model gives the estimated conditional slope:
Experience ranges from 0 to 30 years. The observed 25th percentile is 2 years, the median is 8 years, and the 75th percentile is 20 years. These values provide interpretable reference points across the observed distribution. Substituting each value gives:
Thus, for employees with 2, 8, and 20 years of prior experience, respectively, one additional training hour is associated with estimated assessment-score differences of 3.36, 3.84, and 4.80 points. These are estimated conditional slopes, not predictions that every individual’s score will change by those amounts. The quantiles summarize selected parts of the experience distribution; they do not imply that employees occur only at those three experience levels.
The choice is transparent because the values are identified as observed quantiles, expressed in years, and linked to a clear summary of the distribution. If a substantive question instead concerned a particular experience threshold, that threshold could be reported as an additional value, with its rationale stated.
Why Mean Plus or Minus One Standard Deviation Is Not Automatic
A familiar shortcut is to report effects at the moderator’s mean and one standard deviation below and above it. This can be reasonable when those values are substantively interpretable and fall within the useful range of the data. It should not be applied automatically. For a skewed moderator, the mean and standard deviation may produce an unrepresentative or impossible value. Even when all three values fall within the overall range, they may not correspond to useful real-world distinctions.
A safer approach is to calculate proposed values, inspect them in the moderator’s units, and ask whether each helps answer the question. If the standard-deviation values are awkward, use quantiles or meaningful thresholds instead. This is not a rule against using the mean or standard deviation; it is a reminder that a convention is not a substitute for interpretation.
Worked Example: Replacing Awkward Values for a Skewed Moderator
Suppose a hypothetical service analysis relates \(X\), the number of staff-training hours, to \(Y\), a service-quality score, while allowing the training slope to vary with the number of monthly cases handled, \(Z\). Case volume is right-skewed. Its mean is 10 cases, its standard deviation is 18 cases, its median is 6, and its 25th and 75th percentiles are 2 and 14 cases. The observed range is 0.5 to 70 cases. For this program, 10 cases is also an operationally meaningful reference value.
The mean-plus-or-minus-one-standard-deviation rule proposes \(10-18=-8\), \(10\), and \(10+18=28\) cases. A negative monthly case count is impossible and outside the observed range, so \(-8\) is not a sensible value to report. The value 28 is within the overall range, but it is not, by itself, as directly interpretable as the selected quantiles or the operational threshold.
Assume the estimated conditional slope is \(1.50+0.06z\) quality-score points per training hour. Choose 2 and 14 cases to represent the 25th and 75th percentiles, 6 cases for the median, and 10 cases for the operational threshold. The calculations are:
These values serve two distinct purposes: 2, 6, and 14 describe the observed distribution, while 10 answers a program-specific question. The estimates are associated increases in score per additional training hour at the stated case volumes. It is important to label the two rationales rather than imply that all four values were selected in the same way.
A Short Justification for Using Quantiles
Quantiles are useful because they define positions in the observed distribution rather than relying on a fixed scale. For instance, the median divides an ordered sample so that at least half the observations are at or below it and at least half are at or above it. The 25th and 75th percentiles similarly locate the lower and upper quarters, subject to the sample’s quantile convention and ties. Thus, the three values locate parts of the data’s distribution even when the moderator is measured in a different unit or has a skewed shape.
This justification is descriptive, not inferential. Choosing quantiles does not guarantee precise estimates at those values, does not validate the interaction model, and does not show that every relevant combination of predictors has adequate representation. It explains why the selected values summarize locations in the observed moderator distribution.
How to Make the Choice and Report It
Decide which comparison, decision, or description the conditional effects should address.
Record the observed range and useful summaries such as the median and quartiles; identify relevant subject-matter thresholds.
Use values that answer the question, and explain whether they are thresholds, quantiles, or another justified choice.
Substitute each value into the conditional-effect expression and report the value, units, estimate, and uncertainty when available.
Do not imply that the selected values describe everyone or establish support for every predictor–moderator combination.
For a continuous moderator, a concise report might say: “Conditional slopes are shown at the observed 25th percentile, median, and 75th percentile of prior experience (2, 8, and 20 years).” If one value is selected for a practical reason, state that reason explicitly. If values have been rounded for readability, retain enough precision that the conditional-effect calculations remain meaningful.
Common Mistakes
- Choosing round numbers because they look neat. A value such as 10 may be useful if it represents a meaningful threshold; otherwise, its convenience alone is not a rationale.
- Using mean plus or minus one standard deviation by habit. Check whether those values are possible, interpretable, and useful for the question.
- Calling quantiles “typical cases” without explanation. State which quantiles were used and what part of the observed distribution they represent.
- Reporting values without units or context. “At \(Z=8\)” is incomplete if readers do not know whether 8 means years, kilograms, or another unit.
- Assuming an in-range value is automatically well supported. The moderator’s overall range is only an initial check; representation across combinations of predictors is a separate issue.
- Treating conditional associations as causal effects. A meaningful moderator value improves interpretation, but it does not by itself justify a causal conclusion.
Check Your Understanding
Try these questions before moving on.
- In a continuous interaction model, what does it mean to report the conditional slope at a specified moderator value?
- Why might the 25th percentile, median, and 75th percentile be more informative than fixed round numbers for a skewed moderator?
- A moderator has a mean of 12 and standard deviation of 20, measured in nonnegative units. What should you check before reporting effects at the mean minus and plus one standard deviation?
- Give one example of a subject-matter reason for choosing a moderator value and one reason for choosing a quantile.
- Why does a moderator value’s presence within the overall observed range not, on its own, establish support for every predictor–moderator combination?