---
title: Relative Contrast in Treatment Effects
url: https://www.emergentmind.com/topics/relative-contrast-rc
type: topic
---

# Relative Contrast in Treatment Effects

Relative Contrast (RC) is a framework for comparing conditional treatment effects in a scale-invariant manner, particularly suited for individualized treatment recommendation under resource constraints. The methodology characterizes and estimates contrasts that remain invariant to the scale of outcomes, formalizes a semiparametric single-index model for inference, and achieves semiparametric efficiency bounds under standard causal inference conditions [2010.13904].

## 1. Formal Definition and Theoretical Properties

A relative contrast function $h:\mathbb{R}^+\times\mathbb{R}^+\to\mathbb{R}$ satisfies the following conditions for $a,b > 0$ and $\lambda > 0$:
- $h(a,a)=0$ (self-comparison yields zero contrast);
- For fixed $b$, $h(a,b)$ is strictly increasing in $a$;
- Scale invariance: $h(\lambda a, \lambda b)=h(a,b)$.

Given conditional mean outcomes $\mu_1(X)=\mathbb{E}[Y_1\mid X]$ and $\mu_{-1}(X)=\mathbb{E}[Y_{-1}\mid X]$ for treatments $T=1,-1$, the induced contrast is
\[
R(X) = h\left(\mu_1(X), \mu_{-1}(X)\right).
\]

All scale-invariant contrasts are monotonic transformations of each other—precisely, if $h$ and $\tilde h$ are two relative contrast functions, there exists a strictly increasing $g$ such that
\[
\tilde h(\mu_1(X), \mu_{-1}(X)) = g \circ h(\mu_1(X), \mu_{-1}(X)).
\]
This result ensures that ranking based on any relative contrast function is equivalent up to monotonic transformation and motivates modeling a specific form without loss of generality [2010.13904].

## 2. The Log-Ratio Working Contrast

The canonical operationalization is the log-ratio contrast
\[
R(X) \equiv \theta(X) = \log\frac{\mu_1(X)}{\mu_{-1}(X)},
\]
which is unbounded and particularly convenient for modeling via a single-index structure. The log-ratio contrast notably preserves relative differences and is compatible with scale-invariant requirements. Its range, $(-\infty, \infty)$, aligns naturally with monotonic transformations and ranking procedures necessary for treatment prioritization under constrained resources [2010.13904].

## 3. Semiparametric Single-Index Model and Identifiability

Relative contrast is modeled as
\[
\theta(X) = \psi_0\bigl(X^{\top}\beta_0\bigr),
\]
where $\psi_0:\mathbb{R} \to \mathbb{R}$ is an unknown strictly increasing function and $\beta_0\in \mathbb{R}^p$ is the index parameter. As $\psi_0$ is non-identifiable up to scale, identifiability is enforced by constraining $\|\beta_0\|_2=1$ and, optionally, $\beta_{0,1}>0$. This restriction ensures global identification of the parameter vector $\beta_0$ on the unit sphere [2010.13904].

## 4. Efficient Estimation and Doubly-Robust Loss

Under standard causal inference assumptions (including SUTVA, consistency, and no unmeasured confounding), the observed data likelihood involves nuisance components: the propensity scores $\pi_t(X)=P(T=t\mid X)$, error variances $\sigma_t^2(X)$, and $S(X)=\mu_1(X) + \mu_{-1}(X)$. The efficient score for $\beta$ under the single-index model is
\[
\varphi_{\mathrm{eff}}(Y,T,X) =
\pi_T^{-1} T \left\{ Y \left[1 + e^{-T\psi_0(X^{\top}\beta)} \right] - S(X) \right\}
\nabla \psi_0(X^\top\beta)\;\alpha_*(X),
\]
where
\[
\alpha_*(X) = V(X) S(X) \left\{ X - \frac{\mathbb{E}\left[V(X) S^2(X) X \mid X^\top\beta \right]}{\mathbb{E}\left[ V(X) S^2(X) \mid X^\top\beta \right]} \right\}
\]
and
\[
V^{-1}(X) =
\bigl\{ \pi_1^{-1} e^{-\psi_0(X^\top\beta)} \sigma_1^2(X)
+ \pi_{-1}^{-1} e^{\psi_0(X^\top\beta)} \sigma_{-1}^2(X)
\bigr\}
\bigl(e^{\psi_0/2}+e^{-\psi_0/2}\bigr)^2\,.
\]
Direct solution via the efficient score is operationally intensive. Instead, estimation proceeds by minimizing a doubly-robust loss:
\[
\ell\bigl(f(X);\pi_T,S\bigr) =
\pi_T^{-1} Y \exp\{-T f(X)\}
+
\pi_T^{-1} T (S-Y) f(X).
\]
$\theta(X)$ uniquely minimizes the expectation of this loss over all bounded $f$. Approximation of $\psi_0$ by monotone B-splines allows joint minimization over spline coefficients $\xi$ and $\beta$ under monotonicity and norm constraints [2010.13904].

## 5. One-Step Efficiency Augmentation and Variance Estimation

A one-step procedure refines the doubly-robust pilot estimator. Plug-in estimates $(\hat\xi, \hat\beta)$, together with estimated nuisance parameters, construct $\hat\alpha_*(X)$ and the one-step correction $\tilde\beta$ is defined by the equation
\[
U_n + \bigl\{\hat\Sigma_1 + \lambda_n I\bigr\}(\tilde\beta - \hat\beta) = 0,
\]
where
\[
U_n = \mathbb{E}_n \bigl[\, \nabla_{\beta}\ell(\hat\xi, \hat\beta; \hat\pi, \hat S)\, \nabla \hat\psi (X^\top \hat\beta) \hat\alpha_*(X) \bigr],
\]
\[
\hat\Sigma_1 = \mathbb{E}_n \bigl[
\nabla^2_{\beta}\ell(\hat\xi,\hat\beta; \hat\pi, \hat S)
\{ \nabla\hat\psi(X^\top\hat\beta)\}^2
X \hat\alpha_*(X)
\bigr],
\]
with a small ridge $\lambda_n>0$ to compensate for rank-deficiency from the unit-norm constraint. Explicitly,
\[
\tilde\beta = \hat\beta - \{\hat\Sigma_1 + \lambda_n I\}^{-1}U_n.
\]
Variance is estimated by the sandwich formula,
\[
\text{Var}(\tilde\beta) \approx
\{\hat\Sigma_1 + \lambda_n I\}^{-1}
\mathbb{E}_n\bigl\{ \varphi_n(X)\varphi_n(X)^{T} \bigr\}
\{\hat\Sigma_1 + \lambda_n I\}^{-1},
\]
where
\[
\varphi_n(X)
= \nabla_{\beta}\ell(\hat\xi,\hat\beta; \hat\pi, \hat S)\; \nabla\hat\psi(X^\top\hat\beta)\; \hat\alpha_*(X).
\]
This two-stage procedure achieves the semiparametric lower bound for estimation of $\beta$ under the imposed constraints [2010.13904].

## 6. Theoretical Guarantees

Under regularity assumptions including $\psi_0$ of smoothness order $k>3$, compact covariate support, strict positivity of propensity scores, and convergence rates for nuisance function estimators $\hat\pi$ and $\hat S$ such that $\alpha+\beta > 1/2$, the following properties are established:

- **Consistency and convergence rate (Thm 3.1):**
\[
\|\hat\beta - \beta_0\|_2^2 + \|\hat\psi - \psi_0\|_{L_2}^2
= O_p \bigl(n^{-1+\nu+\gamma}\bigr) + O\bigl(n^{-2k\gamma}\bigr),
\]
where $\gamma<1/10$ indexes B-spline knot count and $\nu\in(0,1/2]$ identifies nuisance estimation rate.

- **Asymptotic normality and efficiency (Thm 3.2):**
\[
\sqrt{n}(\tilde\beta-\beta_0) \xrightarrow[]{d} N(0,\Sigma)
\]
with $\Sigma$ the semiparametric efficiency bound, degenerate along the direction of $\beta_0$ due to unit-norm constraint (rank $p-1$).

These results establish that the two-step procedure is both consistent and achieves the minimax optimal rate [2010.13904].

## 7. Empirical Evaluation and Application

Simulation studies are performed for two outcome-generating models, a continuous outcome with Gaussian noise ("O1") and a Poisson outcome ("O2"), in conjunction with two treatment-assignment mechanisms: a logistic model ("PS1") and a nonlinear model ("PS2"). Sample sizes $n=800, 1000, 1500$ are considered. Competing approaches include Q-learning, a $\Delta$-index model for absolute effect estimation, and an Outcome-Weighted Learning (EARL) variant. Performance metrics are:

- Rank correlation between fitted $\hat\theta(X)$ and true $\theta(X)$.
- Empirical value $\mathbb{E}[Y_{d(X)}]$ under assignment $d(X)=\text{sign}\{\hat\theta(X)\}$.

The relative-contrast methodology uniformly outperforms all alternatives in both ranking accuracy and empirical value. Bootstrap-type intervals for $\beta$ are reported to exhibit close-to-nominal coverage. Further, analysis of a mammography-screening counseling trial with $N \approx 530$ demonstrates practical variable selection and interpretability of the estimated index $\tilde\beta$ within the relative contrast framework [2010.13904].

Source: https://www.emergentmind.com/topics/relative-contrast-rc