Why Separate the Effects?
A combination response can differ from baseline for several reasons: drug A may change the response, drug B may change it, and the joint presence of both drugs may add a further change. A model that combines these contributions into one undifferentiated term cannot show which part is attributed to each drug and which part is attributed to their joint action.
An explicit interaction term makes that distinction visible. Its meaning is conditional on the chosen response scale and model form: it quantifies the departure from the particular additive prediction encoded by the model. It does not, by itself, establish a molecular mechanism or prove that a combination is beneficial.
A Response Model with Separate Terms
Let \(Y\) be a response measured at a specified time, and let \(x_A\) and \(x_B\) be dimensionless exposure functions for drugs A and B. For example, each could be a saturating function of concentration, \(x_A=C_A/(EC_{50,A}+C_A)\), ranging from zero toward one. Such a function is a modeling choice; its parameters and suitability must be supported by the application.
A simple model that separates the contributions is:
Here, \(\beta_0\) is the predicted response when both exposures are zero. The terms \(\beta_Ax_A\) and \(\beta_Bx_B\) represent the separate drug contributions when the other exposure is zero. The cross-term \(\beta_{AB}x_Ax_B\) is zero if either exposure is zero and contributes when both are present. If \(Y\) is measured in log-burden units, all four coefficients are in log-burden units because the exposures are dimensionless.
The model’s additive-effects prediction is \(\beta_0+\beta_Ax_A+\beta_Bx_B\). At a particular pair of exposures, the interaction changes that prediction by \(\beta_{AB}x_Ax_B\). The coefficient \(\beta_{AB}\) is therefore not generally the interaction effect at every exposure pair; the contribution depends on the product of the exposures.
The effect of changing \(x_A\) also depends on \(x_B\). Holding \(x_B\) fixed, the model’s slope with respect to \(x_A\) is:
Thus \(\beta_A\) is the slope at \(x_B=0\); with drug B present, the interaction changes that slope. The corresponding expression for drug B is \(\partial Y/\partial x_B=\beta_B+\beta_{AB}x_A\).
Assumptions and Interpretation
Before using this model, specify the response, its scale, the observation time, and how each exposure is calculated. The model assumes that these choices are meaningful and consistently applied. It also assumes that the selected functions \(x_A\) and \(x_B\) adequately represent each drug’s exposure dependence over the range being modeled, and that the cross-term \(x_Ax_B\) is a suitable description of the joint contribution. Those are modeling assumptions, not consequences of including the term.
If the response is log viable-cell burden relative to an untreated reference, lower values indicate less burden. A negative \(\beta_A\) then predicts a decrease in log burden as \(x_A\) increases when drug B is absent. A negative \(\beta_{AB}\) predicts an additional decrease when both exposures are present; a positive value predicts an increase relative to the additive-effects prediction. These directional interpretations reverse if the response is defined so that larger values indicate a better outcome.
The scale matters. Additivity on a log-burden scale implies multiplicative combination effects on the original burden scale. A cross-term can therefore be present on one response scale and absent, or different in size, on another. Always state the scale before describing an interaction as positive or negative.
Worked Example: Recovering the Interaction from Four Responses
Suppose a hypothetical experiment measures \(Y=\ln(N/N_0)\), the natural logarithm of viable-cell burden relative to its untreated value at a fixed time. Exposures are either absent (\(x=0\)) or at a specified reference level (\(x=1\)). The observed responses are: neither drug, \(Y_{00}=0.00\); A alone, \(Y_{10}=-0.40\); B alone, \(Y_{01}=-0.30\); and the combination, \(Y_{11}=-0.90\).
At these exposure levels, the model gives \(\beta_0=Y_{00}=0.00\), \(\beta_A=Y_{10}-Y_{00}=-0.40\), and \(\beta_B=Y_{01}-Y_{00}=-0.30\). Substituting the combination response into the model gives:
The additive-effects prediction is \(0.00-0.40-0.30=-0.70\). The observed combination response is \(-0.90\), or \(0.20\) log units lower than that prediction. In this model and on this scale, the fitted interaction contribution at \(x_A=x_B=1\) is \(-0.20\). This calculation describes a departure from the specified additive prediction; it does not establish why the departure occurred.
Why the Factorial Contrast Is an Interaction
The four observations in the example illustrate a general result for the two-level model. At \(x_A,x_B\in\{0,1\}\), the model predicts \(Y_{00}=\beta_0\), \(Y_{10}=\beta_0+\beta_A\), \(Y_{01}=\beta_0+\beta_B\), and \(Y_{11}=\beta_0+\beta_A+\beta_B+\beta_{AB}\). Subtracting the single-drug responses from the combination response and adding back baseline cancels all terms except the cross-term:
This is the two-factor interaction contrast. It equals \(\beta_{AB}\) for this model and these zero-or-one exposure levels. With other exposure levels, the contrast corresponds to \(\beta_{AB}\) times the relevant exposure changes, rather than automatically equaling the coefficient itself. Replicated data also include measurement variability, so the contrast calculated from sample means is an estimate, not an exact parameter value.
State what \(Y\) measures, its units or transformation, and the time point. Decide what direction represents an increase in response.
Define \(x_A\) and \(x_B\), including their units or dimensionless range. The zero-exposure value should correspond to drug absence if the model is to assign baseline and single-drug terms this interpretation.
Choose a cross-term, such as \(x_Ax_B\), and write its coefficient and sign convention explicitly. Interpret it relative to the separate-effects terms.
Evaluate the model at relevant exposures, inspect whether its predictions are plausible on the response scale, and check whether the data include single-drug and combination conditions needed to distinguish the terms.
Interaction Terms in a Dynamic Model
A cross-term can also be placed in a dynamic mechanism. Suppose \(N(t)\) is viable-cell number, and the drugs act on its instantaneous net growth rate. For constant exposures, one possible model is:
Here \(r_0\), \(k_A\), \(k_B\), and \(\gamma\) have units of inverse time. The minus signs define positive \(k_A\), \(k_B\), and \(\gamma\) as growth-rate reductions. In this convention, a positive \(\gamma\) adds a joint reduction in growth rate. This is distinct from the previous response model’s sign convention, where the coefficient is added directly to \(Y\). The model assumes constant exposures over the interval and proportional growth or decline of \(N\); if these assumptions do not fit the setting, this equation is not an adequate description.
Worked Example: Predicting Cell Burden with a Rate Interaction
Consider a hypothetical seven-day interval with constant exposures \(x_A=0.50\) and \(x_B=0.80\). Let \(r_0=0.080\ \text{day}^{-1}\), \(k_A=0.120\ \text{day}^{-1}\), \(k_B=0.090\ \text{day}^{-1}\), and \(\gamma=0.050\ \text{day}^{-1}\). Under the dynamic model, the net rate is:
For a constant net rate, \(N(7)/N(0)=\exp(7r_{\mathrm{net}})\). Therefore:
The model predicts that the cell burden after seven days is about \(0.604\) times its starting value. Without the interaction term, the rate would be \(0.080-0.060-0.072=-0.0520\ \text{day}^{-1}\), giving \(\exp(7(-0.0520))=\exp(-0.364)\approx0.695\). Under the stated model, the interaction term therefore predicts a further reduction in burden. This result follows from the chosen rate equation and constant-exposure assumption; it is not evidence by itself for a particular molecular mechanism.
Model Checks, Alternatives, and Limitations
A product interaction is only one way to represent joint drug effects. It is a useful starting point when each drug has a separate contribution and the joint departure is expected to vary with both exposures. Alternatives include a different joint function, such as a saturating or threshold term, or a dynamic model in which the interaction changes a transition, production, or loss rate. Such alternatives answer different modeling questions and require appropriate data; a more complex term is not automatically more mechanistic.
| Model component | What it represents | Key interpretation check |
|---|---|---|
| Separate additive terms | Each drug’s contribution on the chosen response scale | Is the scale and exposure range appropriate for the assumed separate effects? |
| Product cross-term | A joint contribution that depends on both exposures | Is the interaction zero when either drug is absent, as intended? |
| Dynamic rate cross-term | A joint change in a specified rate governing the system | Are the rate units, sign convention, and exposure assumptions explicit? |
Several diagnostics help reveal when the formulation is failing. Plot observed and predicted responses across the single-drug and combination conditions; systematic combination-only residuals suggest that the joint function may be inadequate. Check that the model reproduces the single-drug responses before interpreting the cross-term. Verify that exposure measurements and response times are comparable across conditions. If predictions violate known response bounds, the chosen linear response form may be unsuitable across the full range. If drug exposures are strongly correlated in the data, separate contributions may be difficult to distinguish; a fitted interaction can then be highly uncertain even when the equation is simple.
An interaction coefficient is also conditional on the model’s response scale, exposure functions, and other included terms. A model that is additive on log burden need not be additive on raw burden. If combination treatment changes the exposure achieved for either drug, a response-model cross-term may capture exposure differences as well as effects downstream of exposure. In that case, interpreting the coefficient as a purely pharmacodynamic interaction would require additional evidence and a model that represents the exposure process.
Common Pitfall
Incorrect reasoning: “The combination response is different from the response to either drug alone, so the interaction coefficient must be nonzero.” A combination will usually differ from either single-drug condition simply because it includes both separate drug contributions.
Correction: Compare the combination response with the model’s additive-effects prediction, which includes baseline and both separate contributions. In the two-level example, the relevant contrast is \(Y_{11}-Y_{10}-Y_{01}+Y_{00}\), not \(Y_{11}-Y_{10}\) alone. Interpret the resulting departure on the stated response scale and within the stated model.
Check Your Understanding
Answer each question using the model and sign conventions stated in the question. Solutions follow.
- For \(Y=\beta_0+\beta_Ax_A+\beta_Bx_B+\beta_{AB}x_Ax_B\), what is the contribution of the interaction term when drug B is absent?
- At \(x_A=0.40\), \(x_B=0.50\), and \(\beta_{AB}=-0.30\), calculate the interaction contribution to \(Y\).
- At zero-or-one exposure levels, hypothetical responses are \(Y_{00}=2.0\), \(Y_{10}=1.5\), \(Y_{01}=1.7\), and \(Y_{11}=1.0\). Calculate the interaction contrast.
- In a model for log burden, what does a negative interaction contribution mean if lower log burden indicates fewer viable cells?
Solutions. (1) It is zero because \(x_Ax_B=0\) when \(x_B=0\). (2) The contribution is \((-0.30)(0.40)(0.50)=-0.060\) response units. (3) The contrast is \(1.0-1.5-1.7+2.0=-0.20\); at these exposure levels, the model’s interaction coefficient is \(-0.20\). (4) The term lowers predicted log burden relative to the additive-effects prediction at that exposure pair, corresponding to a lower predicted burden on the original scale.