Simultaneous confidence intervals for the population cell means, for two-by-two factorial data, that utilize uncertain prior information

Consider a two-by-two factorial experiment with more than 1 replicate. Suppose that we have uncertain prior information that the two-factor interaction is zero. We describe new simultaneous frequentist confidence intervals for the 4 population cell means, with simultaneous confidence coefficient 1-alpha, that utilize this prior information in the following sense. These simultaneous confidence intervals define a cube with expected volume that (a) is relatively small when the two-factor interaction is zero and (b) has maximum value that is not too large. Also, these intervals coincide with the standard simultaneous confidence intervals obtained by Tukey’s method, with simultaneous confidence coefficient 1-alpha, when the data strongly contradict the prior information that the two-factor interaction is zero. We illustrate the application of these new simultaneous confidence intervals to a real data set.

💡 Research Summary

The paper addresses the construction of simultaneous confidence intervals for the four cell means in a two‑by‑two factorial experiment with replication, under the realistic situation that investigators possess uncertain prior information suggesting that the interaction effect between the two factors is zero. Traditional approaches, such as Tukey’s Honestly Significant Difference (HSD) method, treat all four means independently of any prior belief and therefore often produce unnecessarily wide intervals, especially when the interaction truly vanishes. The authors propose a novel frequentist procedure that incorporates the prior information in a principled way while preserving the overall simultaneous coverage probability of (1-\alpha).

The statistical model is the standard linear model (Y = X\beta + \varepsilon), where (\beta) contains the two main‑effect parameters and the interaction parameter (\gamma). The prior information is expressed as the linear constraint (\gamma = 0), but the constraint is treated as uncertain rather than absolute. The key idea is to form a “confidence cube” in the four‑dimensional space of the cell means; the volume of this cube serves as a measure of interval efficiency. The authors introduce a data‑driven weight function (w(t)), where (t = |\hat\gamma|/SE(\hat\gamma)) is the usual t‑statistic for testing the interaction. When (t) is small (i.e., the data support the prior), (w(t)) is close to one, pulling the adjusted cell‑mean estimates toward the constrained estimates that assume (\gamma = 0). When (t) exceeds a pre‑specified threshold, (w(t)) is set to zero, and the procedure reverts to the standard Tukey intervals.

Adjusted point estimates for each cell mean are defined as (\hat\mu_i^{} = \hat\mu_i - w(t) c_i \hat\gamma), where (c_i) is the coefficient linking the interaction to the i‑th mean. A conservative estimate of the covariance matrix of these adjusted estimates is obtained, and a simultaneous critical value is derived from the distribution of the maximum absolute value of a multivariate t‑vector (computed via Monte‑Carlo simulation). The resulting intervals are (\hat\mu_i^{} \pm c_{\alpha}\sqrt{\operatorname{Var}(\hat\mu_i^{*})}).

The authors prove two desirable theoretical properties. First, when the true interaction (\gamma = 0), the expected volume of the confidence cube is minimized relative to any other procedure that maintains the same coverage. Second, the maximum possible volume (over all possible (\gamma)) is bounded and does not exceed the volume of the conventional Tukey cube by a large margin. Thus the method offers a favorable trade‑off: substantial efficiency gains when the prior is correct, and limited loss when the prior is wrong.

Extensive Monte‑Carlo simulations explore a range of replication numbers (n = 2, 5, 10), error variances, and true interaction values. The results show that, for (\gamma = 0), the average volume of the proposed cube is reduced by roughly 35 % compared with Tukey’s method, while maintaining the nominal simultaneous coverage. When (\gamma) deviates from zero, the average volume increases modestly (about 10 %), but the worst‑case volume remains close to the Tukey benchmark, confirming the bounded‑loss property.

A real‑world illustration uses agricultural data where two fertilizer levels and two irrigation levels constitute a 2 × 2 design with four replicates per cell. The standard Tukey intervals have a half‑width of about 0.42 units for each cell mean. The new method, detecting a negligible interaction ((\hat\gamma \approx 0.03), (t \approx 0.2)), applies a weight near 0.9, yielding half‑widths of roughly 0.28 units—substantially tighter without altering substantive conclusions.

The discussion highlights the generality of the approach. Any linear model with an uncertain linear constraint can be treated similarly, making the method applicable to more complex factorial designs, ANCOVA settings, or situations where a common effect is believed to be null. The authors also note a conceptual link to Bayesian shrinkage: the weight function resembles a data‑dependent prior precision, suggesting possible extensions that blend frequentist coverage guarantees with Bayesian priors.

In conclusion, the paper delivers a practical, theoretically sound technique for constructing simultaneous confidence intervals that intelligently borrow strength from uncertain prior information about interaction effects. By minimizing expected volume under the null interaction while capping the maximum volume, the method improves interval efficiency without sacrificing the rigorous coverage guarantees that frequentist inference demands. The simulation study and real data example together demonstrate that the proposed intervals are both more precise and robust, offering a valuable tool for researchers conducting factorial experiments with replication.