Interaction Terms and Effect Modification

When a comparison changes across another variable, population, or effect scale

Statistics

Regression

Interaction

Effect Modification

Regression interactions, conditional effects, scale dependence, causal effect modification, uncertainty, diagnostics, and interpretation.

Author

Jiahong Sun

Published

October 14, 2026

Modified

October 14, 2026

1 An Effect Can Depend on Context

An interaction term represents nonadditivity on the scale of a statistical model: the association between one predictor and the outcome changes with another predictor. Effect modification is the corresponding scientific idea that a specified effect differs across levels of a modifier or across populations.

These statements are incomplete until the outcome scale, effect measure, covariates, and target population are specified. The same data can exhibit no interaction in mean differences but interaction in ratios, or vice versa. Statistical interaction also does not by itself establish a physical or causal mechanism.

2 The Product-term Model

For predictors $X$ and $Z$, the linear conditional-mean model

\[ \operatorname{E}(Y\mid X,Z) =\beta_0+\beta_1X+\beta_2Z+\beta_3XZ \]

has conditional slope

\[ \frac{\partial\operatorname{E}(Y\mid X,Z)}{\partial X} =\beta_1+\beta_3Z. \]

Thus $\beta_3$ is the change in the $X$ slope for a one-unit increase in $Z$. The “main effect” $\beta_1$ is the $X$ slope specifically at $Z=0$, while $\beta_2$ is the $Z$ slope at $X=0$. They are not average effects unless the coding and design make them so.

Centering $Z$ at a meaningful value $z_0$ makes $\beta_1$ the $X$ slope at $Z=z_0$. It changes the lower-order coefficients but not fitted values or the interaction coefficient when the same linear scale is used. Centering can improve interpretability; it cannot create overlap or remove genuine collinearity between the scientific quantities of interest.

3 Categorical Interactions Are Differences in Differences

Let $A$ and $B$ be binary indicators. The model

\[ \operatorname{E}(Y\mid A,B) =\beta_0+\beta_AA+\beta_BB+\beta_{AB}AB \]

implies four conditional means:

$A$	$B$	Conditional mean
0	0	$\beta_0$
1	0	$\beta_0+\beta_A$
0	1	$\beta_0+\beta_B$
1	1	$\beta_0+\beta_A+\beta_B+\beta_{AB}$

The interaction is the difference in differences:

\[ \boxed{ \beta_{AB} =\{\mu_{11}-\mu_{01}\}-\{\mu_{10}-\mu_{00}\} }. \]

It asks whether the A difference at $B=1$ differs from the A difference at $B=0$. Reversing the order gives the same additive interaction. With factors having more than two levels, an interaction is a set of contrasts and should usually be assessed through scientifically chosen comparisons or an omnibus multi-degree-of-freedom test rather than one arbitrarily selected coefficient.

4 Scale Determines the Interaction

In a generalized linear model,

\[ g\{\operatorname{E}(Y\mid X,Z)\} =\beta_0+\beta_1X+\beta_2Z+\beta_3XZ, \]

$\beta_3$ describes interaction on the link scale.

In a linear model with identity link, it is departure from additivity of means.
In a log-link model, it is departure from multiplicativity of means or risks.
In logistic regression, it is departure from multiplicativity of odds.

A zero product coefficient in logistic regression does not imply constant risk differences or risk ratios. Conversely, heterogeneous odds ratios can coexist with nearly constant risk differences. For binary outcomes, report predicted risks for joint exposure groups and contrasts on scientifically relevant scales; do not let the default link silently define “no interaction.”

For two binary exposures, additive interaction in risks can be summarized by

\[ \operatorname{RERI} =RR_{11}-RR_{10}-RR_{01}+1, \]

when risk ratios use the unexposed group as reference. $\operatorname{RERI}=0$ means no departure from additivity of risk relative to baseline. This is a different null from a product-term coefficient of zero in a log or logit model.

5 Statistical Interaction and Causal Effect Modification

An observed conditional association is not automatically a causal effect. A causal effect modifier is defined relative to an intervention contrast, outcome scale, covariate set, and target population. For example, an additive conditional average treatment effect at $Z=z$ is

\[ \tau(z)= \operatorname{E}\{Y(1)-Y(0)\mid Z=z\}, \]

where $Y(a)$ is the potential outcome under exposure $a$. Identifying $\tau(z)$ from observational data requires consistency, conditional exchangeability, positivity within modifier strata, and an adequate measurement and sampling process. A regression product term supplies a functional form; it does not supply those assumptions.

Effect modification is population dependent. Stratum-specific effects may be stable while a marginal effect changes because the distribution of baseline risk or other covariates changes. Noncollapsible measures such as the odds ratio can also change between conditional and marginal analyses without confounding. Specify whether the target is conditional or marginal and how standardization to the target population is performed.

6 Estimation and Uncertainty

Suppose the fitted covariance matrix of $(\hat\beta_1,\hat\beta_3)$ is $\widehat{\mathbf V}_{13}$. The estimated $X$ slope at $Z=z$ is

\[ \hat s(z)=\hat\beta_1+z\hat\beta_3, \]

with variance

\[ \widehat{\operatorname{Var}}\{\hat s(z)\} = \begin{pmatrix}1&z\end{pmatrix} \widehat{\mathbf V}_{13} \begin{pmatrix}1\\z\end{pmatrix}. \]

Plotting $\hat s(z)$ or predicted outcomes with pointwise or simultaneous uncertainty is usually clearer than reporting only the p-value for $\beta_3$. Inference must still respect heteroskedasticity, clustering, repeated measures, and the number of independent experimental units.

Interaction tests often have low power because the modifier splits the available information and the product term is correlated with lower-order terms. Absence of statistical significance is not evidence of homogeneous effects. Report effect estimates and intervals in prespecified strata, and design studies around the interaction contrast rather than the easier main-effect calculation.

7 The Hierarchy Principle and Model Specification

When $XZ$ is included, retain $X$ and $Z$ unless a scientifically justified parameterization deliberately omits them. Dropping lower-order terms imposes constraints such as equality of groups at $Z=0$; it does not merely simplify the model. Keeping hierarchy also makes the model invariant to shifts of the predictor origin.

Product terms encode a linear change in a linear slope. They may be inadequate when either predictor has a nonlinear effect. A flexible model can include spline-by-group interactions, but flexibility increases data requirements and the risk of discovering unstable subgroups. Choices of cutpoints, modifiers, transformations, and outcome scales should be prespecified or transparently treated as exploratory.

8 Worked Bioinformatics Example

Suppose normalized expression of a pathway score is measured in independent patient biopsies. Treatment $A$ is randomized, tissue type $Z$ is measured at baseline, and the model includes treatment, tissue, their interaction, and prespecified batch covariates.

If fitted treatment differences are $1.2$ log-expression units in tissue 0 and $0.3$ in tissue 1, then the additive interaction estimate is

\[ 0.3-1.2=-0.9. \]

Statistically, the data suggest that the treatment contrast on this scale is smaller in tissue 1. Biologically, that pattern might motivate a tissue-specific mechanism, but the interaction alone does not prove one. Tissue-specific assay behavior, batch imbalance, compositional shifts, or sparse treatment-by-tissue cells may produce the same pattern.

The patient is the experimental unit. Multiple cells, spots, images, or technical runs from one biopsy refine its measurement but do not become independent replicates of the treatment-by-tissue interaction. A cell-level model must retain patient clustering or aggregate to justified sample-level summaries. If tissue was not randomized, causal claims about tissue and mechanistic synergy require additional assumptions beyond treatment randomization.

9 Diagnostics and Sensitivity

An interaction analysis should examine more than the product-term p-value:

Question	Diagnostic or sensitivity analysis
Is there support for every comparison?	Cell counts, predictor overlap, joint-density plots, leverage
Is the chosen functional form adequate?	Predicted-versus-observed patterns, residuals by modifier, splines or prespecified transformations
Does one unit determine the interaction?	Interaction-specific influence and leave-one-independent-unit-out fits
Does the conclusion depend on scale?	Report predictions and contrasts on additive and multiplicative scales when relevant
Is uncertainty calibrated?	Cluster-aware covariance or model, finite-sample corrections, design-respecting bootstrap
Was the modifier selected after inspection?	Distinguish confirmatory from exploratory analysis; validate in new samples
Is a causal interpretation defensible?	Check exchangeability, positivity, consistency, measurement, and target-population assumptions

Dichotomizing a continuous modifier discards information, creates an arbitrary threshold, and can make subgroup conclusions unstable. Prefer a continuous functional form and show predictions over regions with adequate data. Avoid claims that an effect exists in one subgroup but not another merely because one subgroup p-value crosses 0.05 and the other does not; the relevant test compares the subgroup effects directly.

10 Reporting the Scientific Result

Report:

the outcome scale and effect measure;
the modifier, coding, reference values, and target population;
joint cell counts or the range of continuous-modifier support;
predicted outcomes and conditional effects with intervals at meaningful values;
the interaction contrast with uncertainty, not two separate significance labels;
adjustment, clustering, multiplicity, and exploratory selection procedures;
sensitivity to functional form, influential units, and alternative effect scales.

The result should say “the estimated treatment difference changed by … on the … scale,” rather than vaguely claiming that two variables “interact.”

11 Connections and Take-home Messages

Categorical predictors and contrasts shows how group comparisons become estimable linear functions.
Multiple linear regression provides the conditional-mean and design-matrix foundation.
Regression uncertainty and coefficient interpretation derives uncertainty for linear combinations of coefficients.
Linear-model assumptions and diagnostics covers functional-form, leverage, dependence, and extrapolation checks.
Statistical power and sample-size calculation explains why the intended interaction must drive design planning.

An interaction is a departure from additivity on a declared model or effect scale; it is not an intrinsic property of two variables.
Lower-order coefficients in a product-term model are conditional effects at the other predictor’s reference value.
Interpret and plot conditional effects or predicted outcomes, with uncertainty, rather than relying on a product-term p-value.
Causal effect modification requires design and identification assumptions that regression algebra cannot provide.
Sparse joint cells, data-driven subgroup searches, and technical pseudoreplication make interaction estimates especially unstable.

12 References

Cox, D. R. (1984). Interaction. International Statistical Review, 52(1), 1–31. https://doi.org/10.2307/1403235
VanderWeele, T. J. (2009). On the distinction between interaction and effect modification. Epidemiology, 20(6), 863–871. https://doi.org/10.1097/EDE.0b013e3181ba333c
Knol, M. J., and VanderWeele, T. J. (2012). Recommendations for presenting analyses of effect modification and interaction. International Journal of Epidemiology, 41(2), 514–520. https://doi.org/10.1093/ije/dyr218
VanderWeele, T. J., and Knol, M. J. (2014). A tutorial on interaction. Epidemiologic Methods, 3(1), 33–72. https://doi.org/10.1515/em-2013-0005
Gelman, A., Hill, J., and Vehtari, A. (2020). Regression and Other Stories. Cambridge University Press. https://doi.org/10.1017/9781139161879
R Core Team. formula: model formulae and model.matrix: construct design matrices. R statistical software documentation. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/formula.html

--- title: "Interaction Terms and Effect Modification" subtitle: "When a comparison changes across another variable, population, or effect scale" description: "Regression interactions, conditional effects, scale dependence, causal effect modification, uncertainty, diagnostics, and interpretation." author: "Jiahong Sun" date: "2026-10-14" date-modified: "2026-10-14" categories: [Statistics, Regression, Interaction, Effect Modification] toc: true number-sections: true code-fold: show draft: false --- # An Effect Can Depend on Context An interaction term represents nonadditivity on the scale of a statistical model: the association between one predictor and the outcome changes with another predictor. Effect modification is the corresponding scientific idea that a specified effect differs across levels of a modifier or across populations. These statements are incomplete until the outcome scale, effect measure, covariates, and target population are specified. The same data can exhibit no interaction in mean differences but interaction in ratios, or vice versa. Statistical interaction also does not by itself establish a physical or causal mechanism. # The Product-term Model For predictors $X$ and $Z$, the linear conditional-mean model $$ \operatorname{E}(Y\mid X,Z) =\beta_0+\beta_1X+\beta_2Z+\beta_3XZ $$ has conditional slope $$ \frac{\partial\operatorname{E}(Y\mid X,Z)}{\partial X} =\beta_1+\beta_3Z. $$ Thus $\beta_3$ is the change in the $X$ slope for a one-unit increase in $Z$. The “main effect” $\beta_1$ is the $X$ slope specifically at $Z=0$, while $\beta_2$ is the $Z$ slope at $X=0$. They are not average effects unless the coding and design make them so. Centering $Z$ at a meaningful value $z_0$ makes $\beta_1$ the $X$ slope at $Z=z_0$. It changes the lower-order coefficients but not fitted values or the interaction coefficient when the same linear scale is used. Centering can improve interpretability; it cannot create overlap or remove genuine collinearity between the scientific quantities of interest. # Categorical Interactions Are Differences in Differences Let $A$ and $B$ be binary indicators. The model $$ \operatorname{E}(Y\mid A,B) =\beta_0+\beta_AA+\beta_BB+\beta_{AB}AB $$ implies four conditional means: | $A$ | $B$ | Conditional mean | |---:|---:|---| | 0 | 0 | $\beta_0$ | | 1 | 0 | $\beta_0+\beta_A$ | | 0 | 1 | $\beta_0+\beta_B$ | | 1 | 1 | $\beta_0+\beta_A+\beta_B+\beta_{AB}$ | The interaction is the difference in differences: $$ \boxed{ \beta_{AB} =\{\mu_{11}-\mu_{01}\}-\{\mu_{10}-\mu_{00}\} }. $$ It asks whether the A difference at $B=1$ differs from the A difference at $B=0$. Reversing the order gives the same additive interaction. With factors having more than two levels, an interaction is a set of contrasts and should usually be assessed through scientifically chosen comparisons or an omnibus multi-degree-of-freedom test rather than one arbitrarily selected coefficient. # Scale Determines the Interaction In a generalized linear model, $$ g\{\operatorname{E}(Y\mid X,Z)\} =\beta_0+\beta_1X+\beta_2Z+\beta_3XZ, $$ $\beta_3$ describes interaction on the link scale. - In a linear model with identity link, it is departure from additivity of means. - In a log-link model, it is departure from multiplicativity of means or risks. - In logistic regression, it is departure from multiplicativity of odds. A zero product coefficient in logistic regression does not imply constant risk differences or risk ratios. Conversely, heterogeneous odds ratios can coexist with nearly constant risk differences. For binary outcomes, report predicted risks for joint exposure groups and contrasts on scientifically relevant scales; do not let the default link silently define “no interaction.” For two binary exposures, additive interaction in risks can be summarized by $$ \operatorname{RERI} =RR_{11}-RR_{10}-RR_{01}+1, $$ when risk ratios use the unexposed group as reference. $\operatorname{RERI}=0$ means no departure from additivity of risk relative to baseline. This is a different null from a product-term coefficient of zero in a log or logit model. # Statistical Interaction and Causal Effect Modification An observed conditional association is not automatically a causal effect. A causal effect modifier is defined relative to an intervention contrast, outcome scale, covariate set, and target population. For example, an additive conditional average treatment effect at $Z=z$ is $$ \tau(z)= \operatorname{E}\{Y(1)-Y(0)\mid Z=z\}, $$ where $Y(a)$ is the potential outcome under exposure $a$. Identifying $\tau(z)$ from observational data requires consistency, conditional exchangeability, positivity within modifier strata, and an adequate measurement and sampling process. A regression product term supplies a functional form; it does not supply those assumptions. Effect modification is population dependent. Stratum-specific effects may be stable while a marginal effect changes because the distribution of baseline risk or other covariates changes. Noncollapsible measures such as the odds ratio can also change between conditional and marginal analyses without confounding. Specify whether the target is conditional or marginal and how standardization to the target population is performed. # Estimation and Uncertainty Suppose the fitted covariance matrix of $(\hat\beta_1,\hat\beta_3)$ is $\widehat{\mathbf V}_{13}$. The estimated $X$ slope at $Z=z$ is $$ \hat s(z)=\hat\beta_1+z\hat\beta_3, $$ with variance $$ \widehat{\operatorname{Var}}\{\hat s(z)\} = \begin{pmatrix}1&z\end{pmatrix} \widehat{\mathbf V}_{13} \begin{pmatrix}1\\z\end{pmatrix}. $$ Plotting $\hat s(z)$ or predicted outcomes with pointwise or simultaneous uncertainty is usually clearer than reporting only the p-value for $\beta_3$. Inference must still respect heteroskedasticity, clustering, repeated measures, and the number of independent experimental units. Interaction tests often have low power because the modifier splits the available information and the product term is correlated with lower-order terms. Absence of statistical significance is not evidence of homogeneous effects. Report effect estimates and intervals in prespecified strata, and design studies around the interaction contrast rather than the easier main-effect calculation. # The Hierarchy Principle and Model Specification When $XZ$ is included, retain $X$ and $Z$ unless a scientifically justified parameterization deliberately omits them. Dropping lower-order terms imposes constraints such as equality of groups at $Z=0$; it does not merely simplify the model. Keeping hierarchy also makes the model invariant to shifts of the predictor origin. Product terms encode a linear change in a linear slope. They may be inadequate when either predictor has a nonlinear effect. A flexible model can include spline-by-group interactions, but flexibility increases data requirements and the risk of discovering unstable subgroups. Choices of cutpoints, modifiers, transformations, and outcome scales should be prespecified or transparently treated as exploratory. # Worked Bioinformatics Example Suppose normalized expression of a pathway score is measured in independent patient biopsies. Treatment $A$ is randomized, tissue type $Z$ is measured at baseline, and the model includes treatment, tissue, their interaction, and prespecified batch covariates. If fitted treatment differences are $1.2$ log-expression units in tissue 0 and $0.3$ in tissue 1, then the additive interaction estimate is $$ 0.3-1.2=-0.9. $$ Statistically, the data suggest that the treatment contrast on this scale is smaller in tissue 1. Biologically, that pattern might motivate a tissue-specific mechanism, but the interaction alone does not prove one. Tissue-specific assay behavior, batch imbalance, compositional shifts, or sparse treatment-by-tissue cells may produce the same pattern. The patient is the experimental unit. Multiple cells, spots, images, or technical runs from one biopsy refine its measurement but do not become independent replicates of the treatment-by-tissue interaction. A cell-level model must retain patient clustering or aggregate to justified sample-level summaries. If tissue was not randomized, causal claims about tissue and mechanistic synergy require additional assumptions beyond treatment randomization. # Diagnostics and Sensitivity An interaction analysis should examine more than the product-term p-value: | Question | Diagnostic or sensitivity analysis | |---|---| | Is there support for every comparison? | Cell counts, predictor overlap, joint-density plots, leverage | | Is the chosen functional form adequate? | Predicted-versus-observed patterns, residuals by modifier, splines or prespecified transformations | | Does one unit determine the interaction? | Interaction-specific influence and leave-one-independent-unit-out fits | | Does the conclusion depend on scale? | Report predictions and contrasts on additive and multiplicative scales when relevant | | Is uncertainty calibrated? | Cluster-aware covariance or model, finite-sample corrections, design-respecting bootstrap | | Was the modifier selected after inspection? | Distinguish confirmatory from exploratory analysis; validate in new samples | | Is a causal interpretation defensible? | Check exchangeability, positivity, consistency, measurement, and target-population assumptions | Dichotomizing a continuous modifier discards information, creates an arbitrary threshold, and can make subgroup conclusions unstable. Prefer a continuous functional form and show predictions over regions with adequate data. Avoid claims that an effect exists in one subgroup but not another merely because one subgroup p-value crosses 0.05 and the other does not; the relevant test compares the subgroup effects directly. # Reporting the Scientific Result Report: 1. the outcome scale and effect measure; 2. the modifier, coding, reference values, and target population; 3. joint cell counts or the range of continuous-modifier support; 4. predicted outcomes and conditional effects with intervals at meaningful values; 5. the interaction contrast with uncertainty, not two separate significance labels; 6. adjustment, clustering, multiplicity, and exploratory selection procedures; 7. sensitivity to functional form, influential units, and alternative effect scales. The result should say “the estimated treatment difference changed by ... on the ... scale,” rather than vaguely claiming that two variables “interact.” # Connections and Take-home Messages - [Categorical predictors and contrasts](categorical-predictors-and-contrasts.qmd) shows how group comparisons become estimable linear functions. - [Multiple linear regression](multiple-linear-regression.qmd) provides the conditional-mean and design-matrix foundation. - [Regression uncertainty and coefficient interpretation](regression-uncertainty-and-coefficient-interpretation.qmd) derives uncertainty for linear combinations of coefficients. - [Linear-model assumptions and diagnostics](linear-model-assumptions-and-diagnostics.qmd) covers functional-form, leverage, dependence, and extrapolation checks. - [Statistical power and sample-size calculation](../statistical-inference/statistical-power-and-sample-size-calculation.qmd) explains why the intended interaction must drive design planning. 1. An interaction is a departure from additivity on a declared model or effect scale; it is not an intrinsic property of two variables. 2. Lower-order coefficients in a product-term model are conditional effects at the other predictor's reference value. 3. Interpret and plot conditional effects or predicted outcomes, with uncertainty, rather than relying on a product-term p-value. 4. Causal effect modification requires design and identification assumptions that regression algebra cannot provide. 5. Sparse joint cells, data-driven subgroup searches, and technical pseudoreplication make interaction estimates especially unstable. # References - Cox, D. R. (1984). Interaction. *International Statistical Review*, 52(1), 1--31. <https://doi.org/10.2307/1403235> - VanderWeele, T. J. (2009). On the distinction between interaction and effect modification. *Epidemiology*, 20(6), 863--871. <https://doi.org/10.1097/EDE.0b013e3181ba333c> - Knol, M. J., and VanderWeele, T. J. (2012). Recommendations for presenting analyses of effect modification and interaction. *International Journal of Epidemiology*, 41(2), 514--520. <https://doi.org/10.1093/ije/dyr218> - VanderWeele, T. J., and Knol, M. J. (2014). A tutorial on interaction. *Epidemiologic Methods*, 3(1), 33--72. <https://doi.org/10.1515/em-2013-0005> - Gelman, A., Hill, J., and Vehtari, A. (2020). *Regression and Other Stories*. Cambridge University Press. <https://doi.org/10.1017/9781139161879> - R Core Team. `formula`: model formulae and `model.matrix`: construct design matrices. *R statistical software documentation*. <https://stat.ethz.ch/R-manual/R-devel/library/stats/html/formula.html>

\(A\)	\(B\)	Conditional mean
0	0	\(\beta_0\)
1	0	\(\beta_0+\beta_A\)
0	1	\(\beta_0+\beta_B\)
1	1	\(\beta_0+\beta_A+\beta_B+\beta_{AB}\)