03 · What You Need to Know
What Interaction Effects Mean and Why AI May Misinterpret Them
What Is a Statistical Interaction?
A statistical interaction occurs when the relationship between a predictor and an outcome varies according to the value or category of another predictor, on the scale specified by the statistical model.
In regression analysis, an interaction is commonly represented by a product term involving two predictors.
For example, researchers may examine whether the association between weekly study hours and examination scores differs between students who use an AI tutoring system and those who do not.
If the association between study hours and examination performance differs across the two groups, the regression model may contain an interaction between study time and tutoring-system usage.
Interaction is closely related to moderation analysis. A moderator is a variable across whose values the association or effect of another variable changes.
However, an interaction in a statistical model does not automatically establish a causal moderation effect. The interpretation remains dependent on the research design and assumptions.
How an Interaction Appears in Linear Regression
Consider a linear regression model with study time, AI tutoring usage, and their interaction.
The interpretation depends on the units and coding of the predictors. If the tutoring variable were coded differently, the numerical meaning of the lower-order coefficients would change.
Why the Main Effects Cannot Be Interpreted Independently
In a conventional linear interaction model using reference coding, the coefficient for X represents its slope when Z equals zero. The coefficient for Z represents its relationship with the outcome when X equals zero.
These are conditional effects rather than universal effects that apply across every value of the interacting predictor.
For example, if β₁ = 2 and β₃ = 1.5, it would be misleading to state that every additional hour of study is associated with exactly two additional examination points for all students. The estimated slope differs according to tutoring usage.
The UCLA Statistical Consulting Group's guidance on interaction models emphasizes that lower-order coefficients must be interpreted in relation to the values of the interacting variables. Its analysis and visualization tutorial demonstrates how conditional slopes change across moderator values.
What Is a Simple Slope?
A simple slope is the estimated relationship between a continuous predictor and the outcome at a specified value of another predictor.
For the linear interaction model, the conditional slope of X is:
Simple slopes help researchers explain the direction and magnitude of an association at substantively relevant moderator values.
For continuous moderators, researchers may examine values such as meaningful observed levels, prespecified benchmarks, or selected points within the data range. The choice should serve the research question rather than follow an arbitrary convention.
Why a Significant Interaction Does Not Mean Every Simple Slope Is Significant
An interaction test evaluates whether the modeled relationship changes across moderator values. It does not automatically establish whether each conditional slope differs from zero.
For example, the estimated slope may be positive in both groups, but the interaction test evaluates the difference between those slopes. Conversely, a slope may be statistically distinguishable from zero in one group and not in another without the difference between slopes being statistically significant.
This distinction is particularly important because AI may compare the significance labels of separate group analyses and incorrectly conclude that the groups differ.
The appropriate question is whether the difference between the conditional effects is supported by the relevant statistical test and uncertainty estimate.
Why Interaction Plots Are Often More Informative Than Coefficient Tables
An interaction plot displays predicted outcomes or estimated conditional effects across values of the interacting predictors.
In a simple linear model with two groups, parallel fitted regression lines indicate equal slopes. Nonparallel lines indicate different fitted slopes and therefore an interaction on the modeled outcome scale.
However, the visual separation of lines does not establish statistical significance. Confidence intervals or appropriate tests are needed to evaluate uncertainty.
Interaction plots can also reveal whether a relationship changes direction, becomes stronger, or weakens across moderator values.
When AI interprets an interaction from a table alone, asking it to explain the predicted outcomes at meaningful predictor values may expose errors that are less obvious in a purely verbal summary.
Why Centering Variables Changes Some Coefficients but Not the Underlying Interaction
Researchers sometimes center continuous predictors by subtracting a meaningful reference value, such as the sample mean.
In a conventional linear model containing the relevant lower-order terms and interaction, centering changes the interpretation of the intercept and lower-order coefficients. For example, a coefficient previously interpreted at zero study hours may instead be interpreted at the average number of study hours.
Under an algebraically equivalent reparameterization, centering does not change the model's fitted predictions or the underlying interaction between the variables.
Centering also does not automatically eliminate problematic multicollinearity or correct model misspecification.
An AI explanation should therefore identify whether variables were centered and what the new zero values represent.
Why Logistic Regression Interactions Are More Complicated
In logistic regression, coefficients describe relationships on the log-odds scale. An interaction coefficient indicates how the conditional log-odds association of one predictor changes with another predictor.
However, the relationship between log odds and predicted probability is nonlinear. Consequently, the interaction coefficient does not directly equal the interaction effect on the probability scale.
Ai and Norton (2003) demonstrated that interaction effects in logit and probit models require particular care. The magnitude and even direction of an interaction on the probability scale may differ from what a researcher expects by inspecting the product-term coefficient alone.
For a binary-by-binary interaction, exponentiating the product-term coefficient gives a ratio of conditional odds ratios, not a probability difference-in-differences.
Researchers may need predicted probabilities, marginal effects, or appropriately defined contrasts to answer their substantive question. The relevant calculation depends on whether the interaction is being interpreted on the log-odds, odds-ratio, or probability scale.
This is especially important when AI is asked to interpret odds ratios from models containing interaction terms.
Does an Interaction Establish That One Variable Causally Moderates Another?
No. A statistical interaction describes variation in a modeled association or effect across values of another variable.
In observational research, that variation may reflect confounding, selection, measurement differences, or other features of the data.
A causal interpretation requires a suitable design and defensible assumptions. Even when a randomized intervention is involved, causal claims about a moderator may require care if the moderator is measured after treatment or is affected by the intervention.
Researchers should therefore distinguish statistical association from causation before describing an interaction as proof that one factor changes another factor's causal effect.
04 · A Practical Example
When AI Treats an Interaction Coefficient as an Ordinary Effect
Hypothetical Example
Does AI Tutoring Change the Relationship Between Study Time and Examination Performance?
A researcher investigates whether the association between weekly study hours and examination scores differs between students who use an AI tutoring system and those who do not.
The researcher fits a linear regression model with study hours (X), tutoring usage (Z, coded 0 for nonusers and 1 for users), and their interaction.
The following hypothetical coefficients are reported:
| Model Term |
Coefficient |
Interpretation |
| Intercept |
50.0 |
Predicted score for a nonuser studying zero hours |
| Study hours |
2.0 |
Study-time slope among nonusers |
| AI tutoring usage |
3.0 |
Predicted difference between users and nonusers at zero study hours |
| Study hours × AI tutoring |
1.5 |
Difference in study-time slopes between users and nonusers |
For illustration, assume the interaction coefficient has p = 0.01 under an appropriate model and testing procedure. The coefficient values are hypothetical and do not represent findings from an actual study.
Calculate the Predicted Scores
| Weekly Study Hours |
Nonusers |
AI Tutoring Users |
Predicted Difference |
| 0 |
50 |
53 |
3 |
| 4 |
58 |
67 |
9 |
| 8 |
66 |
81 |
15 |
These predictions illustrate the interaction within the fitted model. They should only be used substantively for predictor values supported by the observed data and appropriate model assumptions.
An Incorrect AI Interpretation
Problematic AI Interpretation
"Study time increases examination scores by 2 points per hour, AI tutoring increases scores by 3 points, and the interaction contributes an additional 1.5 points. All three factors independently improve academic performance."
This interpretation treats the lower-order coefficients as universal effects and the interaction coefficient as an independent improvement. It overlooks that the estimated study-time slope differs between groups and that the tutoring-group difference changes with study hours.
A More Defensible Interpretation
Revised Interpretation
"The interaction between weekly study hours and AI tutoring usage was statistically significant, B = 1.5, p = 0.01, indicating that the fitted association between study time and examination scores differed between users and nonusers. Among nonusers, each additional study hour was associated with an estimated 2-point higher examination score. Among users, the corresponding association was 3.5 points per hour. The estimated difference in study-time slopes was therefore 1.5 points per hour. Because the analysis was observational, these findings do not establish that AI tutoring caused the stronger association."
How the Researcher Should Verify the Interaction
1. Confirm the model and coding.
Identify which predictors interact and what zero and one represent.
2. Calculate the conditional slopes.
Use the study-hours coefficient and interaction coefficient to obtain the fitted slope for each group.
3. Examine predicted outcomes.
Calculate scores at meaningful observed study-hour values to illustrate the interaction.
4. Verify the statistical evidence.
Check the interaction test and calculate appropriate uncertainty estimates for any conditional slopes or contrasts being reported.
5. Limit the conclusion.
Describe the differing associations without automatically attributing the difference to a causal effect of tutoring.
The essential finding is not that an interaction term adds another independent benefit. It is that the modeled relationship between study time and examination performance differs according to tutoring usage.
06 · What This Means for You
How to Verify AI Interpretations of Interaction Effects
When using AI to explain an interaction, provide the complete regression equation, coefficient estimates, variable definitions, coding scheme, relevant uncertainty measures, and research design.
The interpretation should explain the relationship at meaningful values of the interacting predictors rather than merely describing whether the product-term coefficient is statistically significant.
A Practical Interpretation Framework
If the model is linear with a continuous-by-binary interaction
Calculate and compare the conditional slopes for the two categories.
If both interacting predictors are continuous
Evaluate conditional slopes at meaningful values within the observed moderator range.
If the model uses logistic regression
Specify whether the interaction concerns log odds, odds ratios, or probabilities, then calculate the relevant contrasts.
If AI compares separate significance labels
Check the direct test of the interaction or difference between conditional effects.
If AI claims causal moderation
Examine whether the study design and causal assumptions justify that interpretation.
Ask AI to Calculate Conditional Effects
A useful prompt is:
"Interpret this interaction term using the complete regression model and variable coding. Explain the lower-order coefficients conditionally, calculate the estimated effect of each predictor at meaningful values of the other predictor, and distinguish the interaction coefficient from those conditional effects. Identify what uncertainty estimates are needed, avoid comparing significance labels as evidence of a difference, and do not infer causation from the interaction alone."
For nonlinear models, the prompt should additionally specify the scale on which the substantive interaction is to be interpreted.
Use Predicted Outcomes to Communicate the Interaction
Interaction plots and predicted-value tables can make conditional relationships easier to understand. They may also help researchers identify whether an AI-generated verbal explanation matches the fitted model.
However, plots should use relevant predictor values and communicate uncertainty where possible. Extrapolation beyond the observed range can make an otherwise correct statistical calculation substantively misleading.
Researchers should also distinguish statistical significance from practical importance. A statistically detectable difference between slopes may be too small to matter for the research question.
Finally, AI assistance does not remove the researcher's responsibility for the final statistical interpretation.
