Maximum Agreement Linear Predictor (MALP)

• MALP is a linear predictor that maximizes Lin’s concordance correlation coefficient by enforcing agreement in both mean and variance while maintaining linear association.
• It transforms the ordinary least-squares predictor via a closed-form rescaling and recentering, offering higher concordance at the expense of increased prediction variance.
• MALP estimation utilizes sample covariances and asymptotic approximations, demonstrating improved agreement performance compared to traditional least-squares linear prediction in simulation and real-data studies.

The Maximum Agreement Linear Predictor (MALP) is the unique linear predictor that maximizes Lin’s concordance correlation coefficient (CCC) between predicted and observed values. Unlike conventional least-squares linear prediction (LSLP), which minimizes mean squared error (MSE), MALP enforces agreement in both mean and variance as well as linear association, offering an alternative optimality criterion for linear prediction when the focus is on concordance rather than error minimization. All explicit definition, properties, estimation procedures, and comparative performance metrics discussed herein are as established in "Maximum Agreement Linear Prediction via the Concordance Correlation Coefficient" (Kim et al., 2023).

1. Formal Definition and Closed-Form Solution

Let $(X, Y)$ be a pair of random variables with $X \in \mathbb{R}^p$, with mean $\mu_X = E[X]$, mean $\mu_Y = E[Y]$, inner covariance $\Sigma_{XX} = \mathrm{Cov}[X, X]$, cross-covariance $\Sigma_{XY} = \mathrm{Cov}[X, Y]$, and response variance $\sigma_Y^2 = \mathrm{Var}[Y]$. For a linear predictor of the form $\tilde Y(x) = \alpha + x^T \beta$, Lin’s CCC with $Y$ is

$$\rho^c(Y, \tilde Y) = \frac{2 \mathrm{Cov}[Y, \tilde Y]}{\mathrm{Var}[Y] + \mathrm{Var}[\tilde Y] + (\mu_Y - \mu_{\tilde Y})^2} \,.$$

MALP is characterized by the optimizer of

$X \in \mathbb{R}^p$0

There exists a unique solution [Theorem 1]: $X \in \mathbb{R}^p$1 with the maximum concordance

$X \in \mathbb{R}^p$2

This solution can be written as a rescaling and recentering of the ordinary least-squares linear predictor $X \in \mathbb{R}^p$3: $X \in \mathbb{R}^p$4 Here, $X \in \mathbb{R}^p$5 is the LSLP.

2. Sample Estimation and Finite-Sample Coefficients

Given data $X \in \mathbb{R}^p$6, define $X \in \mathbb{R}^p$7, $X \in \mathbb{R}^p$8 as sample means,

$X \in \mathbb{R}^p$9

The sample multiple correlation is

$\mu_X = E[X]$0

The estimated MALP is

$\mu_X = E[X]$1

For $\mu_X = E[X]$2, this simplifies to $\mu_X = E[X]$3, where $\mu_X = E[X]$4.

3. Finite-Sample and Asymptotic Distributional Properties

Assuming $\mu_X = E[X]$5, the finite-sample distribution of $\mu_X = E[X]$6 does not have a closed-form due to its dependence on the absolute value and square root of sample covariance matrices. However, approximate normality arises via the multivariate $\mu_X = E[X]$7-statistic and delta method.

For a new $\mu_X = E[X]$8,

$\mu_X = E[X]$9

with asymptotic variance

$\mu_Y = E[Y]$0

Confidence intervals for $\mu_Y = E[Y]$1 can be constructed via

$\mu_Y = E[Y]$2

where $\mu_Y = E[Y]$3 may be estimated by plug-in, jackknife, or bootstrap approaches.

4. Asymptotics for Estimated Coefficients and Construction of Intervals

Applying the delta method to $\mu_Y = E[Y]$4 yields the asymptotic normality

$\mu_Y = E[Y]$5

for a specific closed-form covariance $\mu_Y = E[Y]$6. For $\mu_Y = E[Y]$7,

$\mu_Y = E[Y]$8

and variance

$\mu_Y = E[Y]$9

plus higher-order corrections. Corresponding confidence intervals may be constructed as

$\Sigma_{XX} = \mathrm{Cov}[X, X]$0

5. Comparison with Least-Squares Linear Prediction (LSLP)

Both MALP and LSLP maximize Pearson’s correlation among linear predictors, but only MALP enforces equality of means and variances, leading to $\Sigma_{XX} = \mathrm{Cov}[X, X]$1, whereas for LSLP $\Sigma_{XX} = \mathrm{Cov}[X, X]$2. It is analytically established that $\Sigma_{XX} = \mathrm{Cov}[X, X]$3, so MALP has greater prediction variance than LSLP. LSLP minimizes MSE, whereas MALP maximizes CCC.

Simulation studies (Section 4) under bivariate normality with varying correlation $\Sigma_{XX} = \mathrm{Cov}[X, X]$4 and sample sizes $\Sigma_{XX} = \mathrm{Cov}[X, X]$5 demonstrate:

• Empirical Pearson correlation coefficients (PCC) for MALP and LSLP are identical.
• MALP achieves uniformly higher CCC (by 5–15 points), especially when $\Sigma_{XX} = \mathrm{Cov}[X, X]$6 is moderate or small.
• LSLP nearly always yields lower MSE than MALP.

Application to real data (OCT eye and body fat datasets) confirms that MALP achieves higher CCC and LSLP achieves lower MSE; thus, preference depends on whether agreement or squared-error minimization is the primary goal.

6. Algorithmic Implementation and Practical Considerations

The computation of the MALP in practice proceeds as follows:

1. Fit standard OLS to obtain $\Sigma_{XX} = \mathrm{Cov}[X, X]$7, $\Sigma_{XX} = \mathrm{Cov}[X, X]$8, and $\Sigma_{XX} = \mathrm{Cov}[X, X]$9.
2. Compute $\Sigma_{XY} = \mathrm{Cov}[X, Y]$0.
3. Compute $\Sigma_{XY} = \mathrm{Cov}[X, Y]$1 and $\Sigma_{XY} = \mathrm{Cov}[X, Y]$2.
4. For prediction, use $\Sigma_{XY} = \mathrm{Cov}[X, Y]$3.
5. Estimate prediction variance via the asymptotic formula, jackknife, or bootstrap; construct confidence or prediction intervals as previously described.

MALP consistently outperforms LSLP in concordance (CCC), both in-sample and out-of-sample, provided that the goal is to maximize CCC. The greatest benefit occurs when $\Sigma_{XY} = \mathrm{Cov}[X, Y]$4 and the sample size is large enough for reliable estimation of $\Sigma_{XY} = \mathrm{Cov}[X, Y]$5.

Key caveats include:

• For true $\Sigma_{XY} = \mathrm{Cov}[X, Y]$6, the MALP estimator can be unstable (mixture-like finite-sample distribution for small $\Sigma_{XY} = \mathrm{Cov}[X, Y]$7), and the normal approximation may be inadequate; the jackknife or bootstrap is recommended for standard errors.
• MALP produces higher variance compared to LSLP; for stringent MSE-minimization or small samples, LSLP may be preferred.
• Variable selection strategies for MALP require caution, as CCC does not automatically adjust for model size in the way that adjusted $\Sigma_{XY} = \mathrm{Cov}[X, Y]$8 does for the LSLP context.

7. Summary and Impact

MALP uniquely maximizes Lin’s concordance correlation coefficient among linear predictors by enforcing equality in means, variances, and maximizing linear association. It is readily constructed from the OLS predictor by a deterministic recentering and rescaling. Its finite-sample and asymptotic statistical properties are well understood under normality, with detailed guidance provided for interval estimation. Extensive simulations and real-data illustrations demonstrate the fundamental tradeoff: maximizing agreement (CCC) at the expense of higher prediction variance and MSE. Accordingly, the use of MALP is justified when concordance is the primary objective, while LSLP remains optimal for mean-squared error minimization (Kim et al., 2023).