2027 Interpreting Interaction on Additive and Multiplicative Scales for MPH and MS Students
Co-Founder and Chief Data Scientist
An interaction result can change depending on whether you compare risk differences or risk ratios. For MPH and MS students, that distinction matters when deciding whether a finding supports a population-health intervention, a research claim, or neither.
This guide explains both scales with one hypothetical example, shows what regression interaction terms do and do not mean, and identifies the coursework and practice needed to analyze interactions responsibly. You can use it to interpret a paper now and to judge whether a program offers enough methods training for your goals.
Key Things You Should Know
- Additive interaction compares risk differences; multiplicative interaction compares relative effects. One can be present when the other is absent.
- In the hypothetical example below, the joint risk is 20% versus 5% with neither exposure: the excess risk beyond additivity is 5 percentage points, while the multiplicative interaction ratio is 1.
- A product term is not a universal interaction test. Identify the outcome scale, estimate all four exposure-group risks when possible, and check uncertainty before drawing a public health conclusion.
What Are Additive and Multiplicative Interaction Scales in Public Health?
Interaction, also called effect modification in this context, means the association between an exposure and an outcome differs across levels of another variable on a specified scale. The scale is essential: a difference in absolute risk answers a different question from a difference in relative risk.
Suppose a study compares a binary environmental exposure A and a binary occupational exposure B with the risk of a respiratory outcome during a defined follow-up period. The four risks are R00 for neither exposure, R10 for A alone, R01 for B alone, and R11 for both. These are risks in a stated population and time window, not interchangeable with odds, rates, or counts.
Additive interaction asks whether the joint risk exceeds what would be expected by adding the two separate risk differences to the unexposed risk. Multiplicative interaction asks whether the joint risk ratio differs from the product of the two separate risk ratios.
Neither scale is inherently the "correct" one: absolute excess risk can inform prevention planning, while relative effects may be useful for etiologic or model-based comparisons.
Students entering graduate study from another field can learn the foundations through statistics and epidemiology coursework; online bachelor's study in public health is one possible foundation, not a prerequisite for every MPH or MS.
How Do Additive and Multiplicative Interaction Measures Differ?
The following entirely hypothetical risks hold the population and follow-up period constant. They show why the two scales can support different, simultaneously valid descriptions of the same four groups.
| Exposure group | Illustrative outcome risk | Risk relative to neither exposure |
| Neither A nor B | 5% | 1 |
| A only | 10% | 2 |
| B only | 10% | 2 |
| Both A and B | 20% | 4 |
On the additive scale, each exposure alone adds 5 percentage points to the 5% baseline. Without additive interaction, the expected joint risk would be 15%. The observed illustrative joint risk is 20%, so the excess risk due to interaction is R11 − R10 − R01 + R00 = 5 percentage points.
Expressed using risk ratios, the relative excess risk due to interaction, or RERI, is RR11 − RR10 − RR01 + 1 = 1. These calculations describe the hypothetical risks; they are not empirical findings.
On the multiplicative risk-ratio scale, the expected joint ratio is 2 × 2 = 4, exactly the illustrative joint ratio. The multiplicative interaction ratio, RR11 ÷ (RR10 × RR01), is therefore 1: no multiplicative interaction in this example. A value above or below 1 would indicate departure from multiplication in the corresponding direction.
A faster undergraduate schedule, including shorter course terms for a public health BS, does not change the mathematics. What matters for this skill is whether the sequence gives students time to master risks, ratios, and regression before advanced epidemiology.
How Do You Interpret Interaction Terms in Epidemiologic Models?
For binary A and B, a regression may include A, B, and their product A × B. The product coefficient describes departure from the model's link-function scale; its presence alone does not establish interaction on every public health scale.
In a linear model for risk with an identity link, the product coefficient represents an additive difference-in-differences in predicted risks, provided the predictions and model are appropriate. In a log-link risk model, exponentiating the product coefficient gives a multiplicative interaction ratio for risk ratios.
In logistic regression, exponentiating it gives a ratio of odds ratios, not a ratio of risk ratios. Odds can differ substantially from risks when an outcome is common, and odds ratios can change with adjustment even without confounding.
To interpret a fitted model, first specify the population, outcome period, reference group, covariates, and link function. Then estimate the four adjusted risks for a common target population, when the design and data permit.
From those risks, calculate the risk-difference contrast and risk-ratio contrast, with confidence intervals that reflect the estimation method. A nonsignificant product-term test on the log-odds scale does not rule out a meaningful additive contrast.
Admission logistics are separate from this analytical requirement. MPH application options without academic references may help an applicant enter a program, but the applicant should still check whether its coursework teaches interpretation beyond logistic product terms.
Interaction analysis usually builds across courses rather than appearing as a guaranteed standalone class. Introductory epidemiology establishes populations, risk, confounding, and study design; biostatistics covers estimation and uncertainty; and intermediate or advanced epidemiologic methods addresses regression, effect modification, and sometimes causal contrasts. Course titles alone cannot establish how deeply either scale is taught.
For an MPH, look for a sequence that moves from foundational biostatistics and epidemiology to an applied regression or advanced methods elective with a reproducible project. For a research-oriented MS, examine the depth of statistical modeling, study design, computing, and thesis or research supervision.
These are typical distinctions, not universal degree requirements. Ask for recent syllabi and check whether students calculate adjusted risks and additive contrasts, not merely fit an interaction term.
If prior grades make graduate entry uncertain, public health master's entry with a modest GPA is a distinct admissions question. Strong preparation in statistics can help demonstrate readiness, but only the individual program can define its requirements or exceptions.
Students need more than the ability to run a regression command. They should be able to define exposure categories and a defensible reference group; distinguish risks, rates, and odds; assess whether each joint-exposure group has enough observations; and explain why a chosen scale fits the decision.
For credible estimates, they also need to examine confounding, selection, measurement error, missing data, and the plausibility of model predictions. Sparse joint-exposure cells can make contrasts unstable.
When data come from a complex survey or clustered sample, the variance calculation and any required weights must reflect that design. Reproducible code should show how adjusted predictions and uncertainty intervals were produced.
These skills can be developed through supervised projects as well as degrees. A student considering doctoral-level specialization should first establish a research question and methods foundation; online epidemiology doctoral funding options matter only if doctoral research, rather than targeted training, serves that goal.
An MPH may emphasize communicating whether an intervention's absolute benefit differs across communities or exposure groups. An MS may devote more time to model specification, estimation, and research design.
Either can offer rigorous interaction training, and either can leave gaps; the decisive evidence is the actual course sequence, assignments, faculty expertise, and access to supervised data analysis.
Compare what students produce. A useful assignment requires an explicit question, a justified scale, four interpretable group estimates, uncertainty intervals, sensitivity checks, and a plain-language account of limitations.
A thesis or applied project can deepen that work if supervision and suitable data are available. Neither an MPH applied practice experience nor an MS thesis should be assumed to include interaction analysis without checking its approved scope.
When choosing among public health master's degree formats, favor the route that supplies the needed quantitative depth and project access over the degree label alone.
Start with a clearly documented teaching dataset or a public dataset whose variables, sampling design, and permitted uses you understand.
A manageable first question uses two well-defined exposures, a binary outcome, a stated follow-up period, and an adequately represented joint-exposure group. Public availability does not remove the need to follow dataset terms or institutional project rules.
A practical analysis proceeds in this order:
- Write the target population, exposure definitions, outcome period, reference group, and reason for reporting absolute and relative contrasts.
- Tabulate the four groups; inspect missing values, cell sizes, and crude risks before fitting a model.
- Fit a defensible adjusted model, check its predictions, and estimate standardized risks for the same target population where appropriate.
- Calculate additive and multiplicative contrasts with uncertainty intervals; document confounder choices, design adjustments, and sensitivity analyses.
- Report the risks and contrasts together, separating observed associations from any causal interpretation.
For restricted, identifiable, or protected data, obtain the necessary data-use permissions and consult the institution about privacy and human-subjects oversight before access or analysis. If a single methods course meets the goal, compare its cost with affordable distance learning public health degrees rather than assuming a full degree is necessary.
Institutional accreditation concerns the institution. CEPH accreditation concerns an accredited school or public health program and its defined scope; it should not be assumed to cover every degree at the institution. MPH applicants can use CEPH status as one program-quality check, while MS applicants should verify whether their exact degree is included rather than inferring coverage from a related MPH.
Accreditation does not certify that a particular course teaches RERI, marginal standardization, or advanced interaction modeling. Check the accreditor's listing for the school or program and degree scope, then ask the department for current syllabi, elective schedules, and examples of assessed methods work.
That distinction matters when comparing public health degree affordability and completion planning: the quickest available route may not include the methods sequence a student needs.
Epidemiologists, biostatisticians, and some health-services or environmental-health researchers may analyze whether associations differ across exposure groups. Public health program evaluators may use absolute-risk contrasts when considering where an intervention could have the greatest population benefit.
In each role, responsible work also includes study design, data quality checks, uncertainty, and clear communication, not simply producing an interaction p-value.
Early-career staff may first clean data, prepare descriptive tables, or reproduce an analysis under supervision. Independently specifying complex models and defending causal claims generally requires deeper training and experience. Job descriptions, not degree titles alone, indicate the expected technical level.
Likewise, comparing MPH graduate earning opportunities should not be mistaken for a salary forecast for interaction-analysis roles; pay depends on the occupation, employer, location, and experience.
How Should Students Choose a Program for Advanced Epidemiologic Methods?
Choose backward from the work you want to perform. If the immediate goal is to interpret published interaction results or add a defensible contrast to a current project, a focused course plus supervised analysis may be enough.
If the goal is sustained applied population-health work, an MPH with substantial epidemiology electives may fit. If the goal is method-intensive research, compare MS programs on modeling depth, thesis expectations, faculty fit, and data access.
Before enrolling, request syllabi and ask whether students estimate both adjusted risk differences and relative contrasts, how uncertainty is taught, and what software and prior statistics are expected. Confirm when electives run and who supervises the relevant project. Compare total tuition and fees, time away from work, and any required software or research expenses against the specific skills you will gain; do not treat a low advertised price as proof of value.
A doctorate is a separate decision tied to a defined research or practice objective. If it becomes relevant later, examine public health doctorate financial planning and outcomes alongside supervision and research feasibility. It is not required simply to interpret additive and multiplicative interaction well.
Other Things You Should Know About Public Health Programs
Yes. An interaction term can involve a continuous exposure, but its interpretation depends on the exposure's units and the model. Check whether a straight-line relationship is plausible before treating one coefficient as describing the full exposure range.
Recoding can change the sign or presentation of a coefficient. The underlying group predictions do not change when the same model is re-expressed correctly. State the reference group so readers can interpret the reported contrast.
No. The joint-exposure group still needs adequate representation, reliable measurements, and suitable comparison groups. Overall sample size alone does not establish that an interaction estimate will be precise.
No. Prioritize interactions motivated by a substantive question and specify them before examining results when possible. Searching many combinations increases the chance of misleading findings and makes interpretation harder.
- Online Master of Public Health Program https://publichealth.berkeley.edu/academics/online
- Measuring additive interaction using odds ratios | springer.com https://link.springer.com/article/10.1186/1742-5573-3-5
- Interactions in statistical models: Three things to know - Duncan https://besjournals.onlinelibrary.wiley.com/doi/10.1111/2041-210X.13714
- Interpreting Odds Ratios in Logistic Regression | MetricGate https://metricgate.com/blogs/odds-ratios-logistic-regression/
- epi.interaction function | RDocumentation https://www.rdocumentation.org/packages/epiR/versions/2.0.99/topics/epi.interaction
- A Tutorial on Interaction | degruyterbrill.com https://www.degruyterbrill.com/document/doi/10.1515/em-2013-0005/html
- RERI: Relative Excess Risk Due to Interaction | CASRAI https://casrai.org/guides/reri-relative-excess-risk-due-to-interaction
