Residual Coefficient of Variation

• Residual coefficient of variation is a dimensionless statistic that quantifies unexplained dispersion in meta-regression and extreme-value contexts.
• It aids in diagnosing heterogeneity beyond covariate effects or high threshold values, with applications in model checking and tail analysis.
• Estimation methods like REML and simulation-based tests provide actionable insights for selecting thresholds and validating tail behaviors.

The residual coefficient of variation (residual CV, or RCV) is a dimensionless, scale-invariant statistic that quantifies residual heterogeneity in two principal contexts: random-effects meta-regression and extreme-value (tail) modeling. In both regimes, the residual CV provides a direct, interpretable measure of the dispersion in excess of that explained by covariates or beyond a high threshold, and forms the basis for both statistical diagnostics and formal testing procedures.

1. Definition and Fundamental Properties

In random-effects meta-regression, for studies $i = 1, \ldots, k$ with effect estimates $Y_i$ and sampling variances $v_i$, the two-parameter random-effects meta-regression model with moderator(s) $x_i$ is

$$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$

where $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$ models unexplained heterogeneity and $\epsilon_i \sim N(0, v_i)$ models sampling error. The total variance is $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$.

At a given moderator value $x$, the model-implied mean is $\mu(x) = \beta_0 + \beta_1 x$. The residual coefficient of variation is defined as

$Y_i$0

and estimated via

$Y_i$1

In extreme-value analysis, for a nonnegative continuous random variable $Y_i$2 and threshold $Y_i$3, define the threshold-excess variable $Y_i$4, with mean $Y_i$5 and variance $Y_i$6. The residual coefficient of variation is

$Y_i$7

By construction, $Y_i$8 is dimensionless and scale-invariant under positive rescaling of $Y_i$9 (Castillo et al., 2015).

2. Theoretical Justification and Model-Specific Behavior

In classical meta-analysis, the usual CV replaces the between-study standard deviation $v_i$0 for $v_i$1 in $v_i$2. In meta-regression, allowing $v_i$3 to vary adapts this notion to heteroscedastic conditional means (Cairns et al., 2021).

In the context of excess distributions over thresholds,

• For $v_i$4, $v_i$5 for all $v_i$6 (due to the memoryless property) (Castillo et al., 2011, Castillo et al., 2015).
• For $v_i$7 following a generalized Pareto distribution (GPD) with shape $v_i$8 and scale $v_i$9, $x_i$0 is again GPD$x_i$1, yielding

$x_i$2

which is constant in $x_i$3 and depends only on the tail index (Castillo et al., 2015).

A flat residual CV-plot as $x_i$4 increases empirically characterizes a GPD tail and identifies the value of $x_i$5 (Castillo et al., 2011, Castillo et al., 2015).

3. Estimation, Confidence Intervals, and Testing

In meta-regression, $x_i$6 is obtained by weighted least squares, and

$x_i$7

where $x_i$8 and $x_i$9 (Cairns et al., 2021).

Three classes of confidence intervals for $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$0 are:

• Wald-type intervals on the log scale,
• $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$1-adjusted substitution intervals for $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$2 (with nominal level adjustment, then back-transform to CV via $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$3),
• Propagating imprecision intervals using joint bounds of $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$4 and $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$5 (Cairns et al., 2021).

In extreme-value analysis, the empirical RCV is computed at multiple thresholds, and inference is conducted using test statistics such as

$$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$6

where $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$7 is the number of exceedances at threshold $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$8 and $$Y_i = \beta_0 + \beta_1 x_i + \gamma_i + \epsilon_i,$$9. The asymptotic null distribution is a weighted sum of independent $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$0 with analytically tractable weights (Castillo et al., 2015, Castillo et al., 2011).

For unknown $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$1, replace $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$2 by the weighted average estimator

$\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$3

$\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$4-values are obtained by simulation from GPD$\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$5 (Castillo et al., 2015).

4. Diagnostic Plots and Empirical Behavior

The CV-plot or RCV-plot graphs empirical residual CV values against ordered thresholds or exceedances. The key behaviors are:

• Flat RCV-plot: Indicates tail behavior consistent with GPD, with the flat value determining the tail shape parameter $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$6 (Castillo et al., 2015, Castillo et al., 2011).
• Upward trend: Suggests heavier-than-GPD tails.
• Downward trend: Suggests lighter-than-GPD or finite endpoint distributions.

For meta-regression, at a fixed moderator $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$7,

• Small $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$8 indicates little residual heterogeneity relative to the mean.
• Large $\gamma_i \sim N(0, \tau^2_{\mathrm{res}})$9 indicates pronounced heterogeneity, with effects possibly spanning zero (Cairns et al., 2021).

Interpretive benchmarks are: $\epsilon_i \sim N(0, v_i)$0 (modest), $\epsilon_i \sim N(0, v_i)$1–$\epsilon_i \sim N(0, v_i)$2 (moderate), $\epsilon_i \sim N(0, v_i)$3 (large) (Cairns et al., 2021).

5. Applications in Meta-Regression and Extreme-Value Analysis

In random-effects meta-regression, residual CV quantifies unexplained heterogeneity after accounting for moderators, with robust estimation and confidence intervals provided by REML and the outlined interval procedures. Interpretation is grounded in the comparison to the magnitude of the mean effect, allowing cross-study or cross-design comparisons (Cairns et al., 2021).

In extreme-value analysis, the RCV method provides:

• A diagnostic for detecting GPD tails and estimating the shape parameter $\epsilon_i \sim N(0, v_i)$4,
• Formal multiple-threshold testing for GPD conformity,
• An automatic threshold selection algorithm to objectively determine the onset of GPD behavior (Castillo et al., 2015).

Example: Danish fire insurance data fit with RCV yields threshold selection and $\epsilon_i \sim N(0, v_i)$5 estimates in close agreement with MLE, validating both methodology and practical interpretability (Castillo et al., 2015).

6. Practical Implementation and Recommendations

In meta-regression:

• Estimate $\epsilon_i \sim N(0, v_i)$6 using REML,
• Report $\epsilon_i \sim N(0, v_i)$7 with 95% interval (preferably $\epsilon_i \sim N(0, v_i)$8-adjusted or PropImp),
• Use $\epsilon_i \sim N(0, v_i)$9 or $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$0 where $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$1 may be near zero, as these are bounded and avoid unstable CVs,
• Summarize across $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$2 using the geometric mean $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$3 and its CI (Cairns et al., 2021).

In extreme-value settings:

• Plot the RCV against threshold to diagnose tail regime,
• Use the multiple-threshold $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$4 statistic and simulation-based $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$5-values for formal assessment,
• Leverage the threshold selection algorithm outlined above for objective tail modeling (Castillo et al., 2015, Castillo et al., 2011).

7. Interpretation, Limitations, and Relation to Other Measures

The residual coefficient of variation complements widely used heterogeneity indicators such as $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$6 in meta-analysis, providing a scale-invariant and directly interpretable gauge of unexplained dispersion. A principal limitation in both contexts is potential instability when the mean effect approaches zero, in which case one should prefer alternate bounded transforms ($\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$7, $\mathrm{Var}(Y_i) = v_i + \tau^2_{\mathrm{res}}$8) or restrict inference to intervals away from zero. In extreme-value inference, infinite-variance tails can make the ordinary RCV unreliable, but transformation-based stabilization methods extend RCV techniques even to such cases (Castillo et al., 2015).

The RCV and its plot offer both a graphical check and rigorous formal test for model assessment in tail modeling, with mathematically tractable and interpretable properties (Castillo et al., 2015, Castillo et al., 2011). In meta-regression, simulation studies confirm coverage properties of the recommended intervals for moderate to large studies, supporting widespread methodological adoption (Cairns et al., 2021).