Relative Contrast in Treatment Effects

• Relative Contrast is a scale-invariant framework that quantifies differences in treatment effects using a conditional contrast function derived from outcome means.
• The methodology employs a log-ratio contrast along with a semiparametric single-index model to enable reliable ranking and individualized treatment prioritization under resource constraints.
• Efficiency is achieved via a doubly-robust loss minimization and one-step augmentation procedure, satisfying semiparametric efficiency bounds under standard causal inference assumptions.

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 (Liang et al., 2020).

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 $a,b > 0$0, the induced contrast is

$a,b > 0$1

All scale-invariant contrasts are monotonic transformations of each other—precisely, if $a,b > 0$2 and $a,b > 0$3 are two relative contrast functions, there exists a strictly increasing $a,b > 0$4 such that

$a,b > 0$5

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 (Liang et al., 2020).

2. The Log-Ratio Working Contrast

The canonical operationalization is the log-ratio contrast

$a,b > 0$6

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, $a,b > 0$7, aligns naturally with monotonic transformations and ranking procedures necessary for treatment prioritization under constrained resources (Liang et al., 2020).

3. Semiparametric Single-Index Model and Identifiability

Relative contrast is modeled as

$a,b > 0$8

where $a,b > 0$9 is an unknown strictly increasing function and $\lambda > 0$0 is the index parameter. As $\lambda > 0$1 is non-identifiable up to scale, identifiability is enforced by constraining $\lambda > 0$2 and, optionally, $\lambda > 0$3. This restriction ensures global identification of the parameter vector $\lambda > 0$4 on the unit sphere (Liang et al., 2020).

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 $\lambda > 0$5, error variances $\lambda > 0$6, and $\lambda > 0$7. The efficient score for $\lambda > 0$8 under the single-index model is

$\lambda > 0$9

where

$h(a,a)=0$0

and

$h(a,a)=0$1

Direct solution via the efficient score is operationally intensive. Instead, estimation proceeds by minimizing a doubly-robust loss: $h(a,a)=0$2 $h(a,a)=0$3 uniquely minimizes the expectation of this loss over all bounded $h(a,a)=0$4. Approximation of $h(a,a)=0$5 by monotone B-splines allows joint minimization over spline coefficients $h(a,a)=0$6 and $h(a,a)=0$7 under monotonicity and norm constraints (Liang et al., 2020).

5. One-Step Efficiency Augmentation and Variance Estimation

A one-step procedure refines the doubly-robust pilot estimator. Plug-in estimates $h(a,a)=0$8, together with estimated nuisance parameters, construct $h(a,a)=0$9 and the one-step correction $b$0 is defined by the equation

$b$1

where

$b$2

$b$3

with a small ridge $b$4 to compensate for rank-deficiency from the unit-norm constraint. Explicitly,

$b$5

Variance is estimated by the sandwich formula,

$b$6

where

$b$7

This two-stage procedure achieves the semiparametric lower bound for estimation of $b$8 under the imposed constraints (Liang et al., 2020).

6. Theoretical Guarantees

Under regularity assumptions including $b$9 of smoothness order $h(a,b)$0, compact covariate support, strict positivity of propensity scores, and convergence rates for nuisance function estimators $h(a,b)$1 and $h(a,b)$2 such that $h(a,b)$3, the following properties are established:

• Consistency and convergence rate (Thm 3.1):

$h(a,b)$4

where $h(a,b)$5 indexes B-spline knot count and $h(a,b)$6 identifies nuisance estimation rate.

• Asymptotic normality and efficiency (Thm 3.2):

$h(a,b)$7

with $h(a,b)$8 the semiparametric efficiency bound, degenerate along the direction of $h(a,b)$9 due to unit-norm constraint (rank $a$0).

These results establish that the two-step procedure is both consistent and achieves the minimax optimal rate (Liang et al., 2020).

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 $a$1 are considered. Competing approaches include Q-learning, a $a$2-index model for absolute effect estimation, and an Outcome-Weighted Learning (EARL) variant. Performance metrics are:

• Rank correlation between fitted $a$3 and true $a$4.
• Empirical value $a$5 under assignment $a$6.

The relative-contrast methodology uniformly outperforms all alternatives in both ranking accuracy and empirical value. Bootstrap-type intervals for $a$7 are reported to exhibit close-to-nominal coverage. Further, analysis of a mammography-screening counseling trial with $a$8 demonstrates practical variable selection and interpretability of the estimated index $a$9 within the relative contrast framework (Liang et al., 2020).