Diagnosing Formulation-By-Food Interactions

Why Separate a Shared Food Effect From an Interaction?

A meal can change exposure after both the test and reference formulations. That shared change is a food effect, but it does not by itself indicate that food changes the comparison between products. The distinct question is whether the effect of food differs by formulation. This is a formulation-by-food interaction.

The distinction matters because a product comparison made in the fasted condition does not automatically describe the fed condition. If food affects both formulations by the same proportion, their relative exposure can remain stable even while exposure rises or falls for both. If their food effects differ, the relative exposure changes across conditions. The earlier material on interpreting food-related changes in exposure considers food effects within a condition; here the focus is the difference between those effects for two formulations.

Definition: Let $E_{T,F}$ and $E_{T,S}$ denote exposure for the test formulation in the fed and fasted conditions, and $E_{R,F}$ and $E_{R,S}$ the corresponding reference exposures. The test and reference food-effect ratios are $E_{T,F}/E_{T,S}$ and $E_{R,F}/E_{R,S}$. The multiplicative formulation-by-food interaction is their ratio: $$ IR=\frac{E_{T,F}/E_{T,S}}{E_{R,F}/E_{R,S}} =\frac{E_{T,F}/E_{R,F}}{E_{T,S}/E_{R,S}}. $$ An $IR$ of 1 means the two formulations have the same fed-to-fasted exposure ratio on this multiplicative scale. It does not mean that food has no effect.

The second expression gives a useful equivalent interpretation: $IR$ is the test-to-reference exposure ratio in the fed condition divided by the test-to-reference ratio in the fasted condition. An interaction therefore describes how the relative formulation comparison changes across food conditions. It is not simply the observation that exposure differs between fed and fasted administration.

A Model and Its Interpretation

For formal inference on exposure ratios, a common model represents log exposure as a function of formulation, condition, and their interaction. Let $Z_{ifc}$ be the log exposure for subject $i$, formulation $f$, and condition $c$, where $f=1$ denotes test and $c=1$ denotes fed. A simple mean structure is:

Z_{ifc}=\mu+\alpha f+\beta c+\gamma(fc)+b_i+\varepsilon_{ifc}.

Here, $\mu$ is the reference-formulation fasted mean on the log scale, $\alpha$ is the test-versus-reference contrast when fasted, and $\beta$ is the fed-versus-fasted contrast for reference. The interaction coefficient $\gamma$ is the additional fed-versus-fasted change for test relative to reference. The subject term $b_i$ represents subject-level variation when a design includes repeated observations on the same people; the model must account for the actual study design and within-subject dependence.

In this model, the interaction is a difference of differences on the log scale:

\gamma= \bigl(\text{test fed}-\text{test fasted}\bigr) -\bigl(\text{reference fed}-\text{reference fasted}\bigr).

Exponentiating that difference gives the ratio of food-effect ratios: $IR=\exp(\gamma)$. Thus, testing $H_0:\gamma=0$ corresponds to testing $IR=1$, against an alternative in which the interaction differs from zero. The interaction estimate and its confidence interval should come from a model that accounts for the study design. Repeated measurements, incomplete data, or other design features can affect both the standard error and the valid degrees of freedom.

This model requires a clearly specified endpoint and comparable fed and fasted conditions, including a defined meal and administration procedure. Subjects should be assigned and observed in a way that supports the intended comparisons. If the same subject receives multiple conditions, the analysis must account for that dependence; treating all observations as independent can give an incorrect standard error. Carryover, condition imbalance, missing observations, or deviations in meal administration can also compromise interpretation. For conventional small-sample model-based inference, the residual distribution and variance assumptions used for the confidence interval should be reasonably adequate, or an appropriate alternative method should be used.

Worked Interpretations

Worked Example: A Food Effect Shared by Both Formulations

Suppose a hypothetical study reports the following exposure summaries in matching units. For reference, fasted exposure is 100 and fed exposure is 150. For test, fasted exposure is 105 and fed exposure is 157.5. Treat these values as illustrative summaries for comparing conditions.

The reference food-effect ratio is: $$ \frac{150}{100}=1.50. $$ The test food-effect ratio is: $$ \frac{157.5}{105}=1.50. $$ Therefore: $$ IR=\frac{1.50}{1.50}=1.00. $$ The test-to-reference ratios are also 105/100 = 1.05 when fasted and 157.5/150 = 1.05 when fed.

On these summaries, food is associated with a 50% higher exposure for each formulation, while the relative test-to-reference comparison is unchanged. This is a shared food effect, not evidence that food has no effect. Because the example gives only summaries and no uncertainty estimate, it does not establish that the interaction is statistically negligible in a study population.

Worked Example: A Larger Food Effect for Reference

In a second hypothetical example, reference exposure is 100 fasted and 150 fed. Test exposure is 105 fasted and 126 fed. The units are the same across all four summaries.

The reference food-effect ratio is: $$ \frac{150}{100}=1.50. $$ The test food-effect ratio is: $$ \frac{126}{105}=1.20. $$ The interaction ratio is: $$ IR=\frac{1.20}{1.50}=0.800. $$ Equivalently, the test-to-reference ratio is 105/100 = 1.05 fasted and 126/150 = 0.8400 fed; 0.8400/1.05 = 0.800.

The estimated food-effect ratio for test is 0.800 times the food-effect ratio for reference. In this example, reference exposure rises by 50% with food while test exposure rises by 20%. The test-to-reference ratio is lower in the fed condition than in the fasted condition. These are descriptive contrasts from the hypothetical summaries; without an uncertainty estimate, they do not show whether the interaction is distinguishable from sampling variability.

Worked Example: Interpreting Uncertainty Around an Interaction

Suppose a hypothetical model reports an interaction ratio of 0.920 with a 90% confidence interval from 0.780 to 1.080. This interval is supplied by the model; it is not calculated from separate condition-specific confidence intervals.

The point estimate indicates that the test food-effect ratio is 0.920 times the reference food-effect ratio, or about 8% lower. But the interval includes 1. Values below 1 are compatible with a smaller food effect for test, while values above 1 are compatible with a larger food effect for test. Under the model and its assumptions, the interval gives a range of values compatible with the data and procedure at the stated confidence level.

The interval does not establish that there is no interaction, nor does it establish that the interaction is acceptably small. Those claims require a prespecified criterion for what size of interaction matters and adequate precision to address that criterion. A confidence interval that includes 1 can reflect limited information as well as a small estimated interaction.

Deriving the Ratio-of-Ratios

The interaction formula follows directly from comparing the two food effects. Let $F_T=E_{T,F}/E_{T,S}$ and $F_R=E_{R,F}/E_{R,S}$. Their ratio is:

\frac{F_T}{F_R} = \frac{E_{T,F}/E_{T,S}}{E_{R,F}/E_{R,S}} = \frac{E_{T,F}E_{R,S}}{E_{T,S}E_{R,F}}.

Rearranging the same terms gives:

\frac{F_T}{F_R} = \frac{E_{T,F}/E_{R,F}}{E_{T,S}/E_{R,S}}.

So the ratio of the food effects equals the fed test-to-reference ratio divided by the fasted test-to-reference ratio. If the food effects are identical, $F_T=F_R$, the ratio is 1 and the two test-to-reference ratios are equal. This algebra is the reason interaction is best assessed directly rather than inferred from separate statements about fed and fasted results.

Diagnostics, Alternatives, and Limitations

Start by displaying all four formulation-by-condition estimates, rather than showing only the interaction. This makes it possible to see whether the reported interaction reflects a change in the test food effect, a change in the reference food effect, or both. Check that the fed and fasted comparisons use the same endpoint and compatible summaries, and verify that the meal conditions and dose administration were implemented as intended. Examine missingness, unusual observations, and design features that could change the precision or bias the estimated contrasts.

A common diagnostic error is to compare two separate significance tests: one asking whether food affects test and another asking whether food affects reference. A significant result for one and a nonsignificant result for the other does not demonstrate that the food effects differ. The relevant quantity is their difference on the chosen scale, with its own standard error or confidence interval. Separate estimates may be correlated, especially in a crossover design, so their uncertainty cannot generally be reconstructed by treating them as unrelated.

Interaction depends on scale. The ratio-of-ratios asks whether the food effects are the same proportionally. An additive alternative asks whether the fed-minus-fasted exposure difference is the same for each formulation:

(E_{T,F}-E_{T,S})-(E_{R,F}-E_{R,S}).

That additive contrast can be scientifically relevant for some questions, but it is not interchangeable with the multiplicative interaction. The scale should be selected to match the scientific question and the analysis plan, not chosen after seeing which result appears more favorable. For exposure ratios, the multiplicative comparison is often the natural target.

Finally, absence of a clearly detected interaction is not proof that fed and fasted product comparisons are equivalent. The interaction interval may be wide, and an interaction test addresses a different question from whether a formulation comparison meets a bioequivalence criterion in a particular condition. A study may also have limited ability to estimate an interaction even when its main aim is to compare formulations. Report the estimated interaction, its uncertainty, the scale, and the conditions to which it applies.

Key takeaway: A shared food effect can change exposure substantially while leaving the test-to-reference ratio unchanged. A formulation-by-food interaction is a change in that relative comparison across conditions, quantified here by the ratio of food-effect ratios.

Check Your Understanding

Answer each question, then compare with the concise solutions below.

Reference exposure is 80 fasted and 120 fed. Test exposure is 88 fasted and 132 fed. Calculate the interaction ratio and interpret it.
What does an interaction ratio of 1 mean? Does it mean food has no effect?
A fed-versus-fasted test contrast is statistically significant, while the corresponding reference contrast is not. Does that alone establish an interaction?
A model reports an interaction ratio of 1.10 with a confidence interval from 0.96 to 1.26. What can and cannot be concluded from this interval?

Solutions. (1) Reference food effect is 120/80 = 1.50; test food effect is 132/88 = 1.50; therefore $IR=1.50/1.50=1.00$. The summaries show a shared 50% food effect and an unchanged test-to-reference ratio. (2) The two formulations have equal fed-to-fasted ratios on the multiplicative scale. Food may still affect exposure for both. (3) No. The interaction itself must be estimated with its uncertainty; different significance outcomes for separate tests do not establish that the effects differ. (4) The point estimate suggests a 10% larger food-effect ratio for test than reference, but the interval includes 1. The data are compatible with no multiplicative interaction and with effects on either side of 1; the interval alone does not establish equivalence or rule out a meaningful interaction.