Consider a clinical trial studying a new antihypertensive medication (AntiHyp) across different patient groups. The trial aims to understand how the drug's effectiveness varies between male and female patients, potentially revealing important interaction effects that could inform personalized treatment approaches.
Key Variables:
Parameter | Estimate | Description |
---|
In this model, the treatment effect (β₁) remains constant across patient groups. The lines representing different groups remain parallel, indicating that the treatment works equally well regardless of patient characteristics.
Here, the treatment effect varies by group: \(\frac{\partial Y}{\partial \text{Treatment}} = \beta_1 + \beta_3 \text{Group}\). Non-parallel lines indicate that the treatment's effectiveness differs between groups, suggesting the need for personalized dosing strategies.
This model demonstrates why simply adding terms linearly cannot capture true interaction effects. The treatment effect remains constant: \(\frac{\partial Y}{\partial \text{Treatment}} = \beta_1 + \beta_3\), independent of the patient group. The lines remain parallel because adding terms linearly only shifts the intercept and slope uniformly across groups, unlike true interactions which allow for group-specific treatment effects.
Let's examine how these models look in matrix form for \(n\) observations, where \(x_1\) represents Treatment and \(x_2\) represents Group:
The design matrix \(\mathbf{X}\) in each model reveals important properties:
The optimal coefficients \(\boldsymbol{\beta}\) are found by solving:
\[ (\mathbf{X}^\top\mathbf{X})\boldsymbol{\beta} = \mathbf{X}^\top\mathbf{y} \]For the additive "interaction" model:
In contrast, for the multiplicative interaction model: