Statisticians Justin Dunnam and K. Krishnamoorthy have developed advanced statistical procedures, including an improved modified likelihood ratio test and a parametric bootstrap approach, designed to evaluate the equality of quantiles across several independent normal populations. While experimental analyses routinely evaluate differences among population means, average measurements frequently fail to detect meaningful variations occurring in the upper or lower tails of biological, industrial, and clinical data. Published in the Journal of Statistical Theory and Applications, the investigation provides exact theoretical formulations that enhance earlier generalized variable tests, alongside simplified simultaneous confidence intervals for conducting pairwise comparisons once overall group differences are detected. By replacing computationally intensive simulations with closed-form mathematical moments and derived approximations, the newly formulated testing suite establishes accurate type I error rate control and robust statistical power across varying group counts and sample sizes.
The motivation for testing distribution percentiles emerges from practical analytical scenarios where central tendencies provide an incomplete portrait of experimental outcomes. In clinical pharmacology, determining whether a large majority of patients respond favorably to a novel drug therapy often carries greater relevance than establishing that average physiological responses are identical. Previous investigations, including studies on pathological tremors by Albers and Lohnberg, clinical diagnostics by Rudolfer and Campbell, and wood materials by Huang and Johnson, demonstrated that groups can share identical means while displaying starkly divergent tail behaviors. Despite the broad relevance of percentile comparisons across scientific disciplines, inferential methodology for evaluating several quantiles simultaneously has remained limited.
Early methodological contributions primarily addressed two-sample problems, utilizing nonparametric techniques, generalized variable frameworks, or fiducial methods to evaluate percentile differences and ratios between pairs of distributions. Extending these questions to multiple groups, Li and colleagues introduced the generalized variable test, which relied on fiducial quantities constructed from sample means and standard deviations. However, their formulation required secondary Monte Carlo simulations to approximate the mean and variance of the underlying fiducial quantity, resulting in computational burdens and overly conservative hypothesis testing. Dunnam and Krishnamoorthy resolved this limitation by deriving explicit theoretical expectations and variance expressions using mathematical moments of noncentral t-distributions. By expressing the fiducial quantity directly through noncentral t random variables, the authors eliminated the need for auxiliary simulation loops. This closed-form mathematical enhancement enables researchers to compute test statistics and generalized p-values directly from observed sample statistics without sacrificing analytical rigor.
To provide a more powerful and mathematically elegant alternative, the researchers formulated the DK-modified likelihood ratio test by refining likelihood ratio principles through moment matching techniques. Standard likelihood ratio tests frequently exhibit substantial inaccuracies when applied to small sample sizes, leading to poor type I error control. While Abdollahnezhad and Jafari developed an initial modification based on higher-order asymptotic theory, Dunnam and Krishnamoorthy adapted foundational insights from DiCiccio and colleagues along with Welch’s approximate degrees of freedom methodology. Their approach rescales the likelihood ratio statistic against its expectation and approximates the resulting distribution using a chi-square distribution scaled by estimated degrees of freedom. Because deriving exact theoretical moments for the likelihood statistic under constrained maximum likelihood estimation is analytically intractable, the authors implemented an efficient Monte Carlo routine to estimate the mean and variance, producing an empirical test statistic with superior small-sample calibration.
Complementing their likelihood framework, the authors introduced a parametric bootstrap test that extends earlier mean-testing methodologies established by Krishnamoorthy and colleagues to arbitrary normal distribution percentiles. Under the null hypothesis of identical population quantiles, the true underlying parameters must satisfy a shared quantile constraint. The parametric bootstrap test evaluates group quantile estimates by weighting individual deviations against their estimated sample variances, which incorporate correction factors derived from gamma functions of sample degrees of freedom. Bootstrap samples are iteratively generated from normal distributions governed by constrained maximum likelihood parameter estimates obtained from observed datasets. By computing the test statistic across replicated bootstrap realizations, researchers obtain an empirical p-value representing the probability of observing test statistics as extreme as the empirical sample value. This bootstrap construction ensures that hypothesis tests accommodate unequal variances and heterogeneous group sizes across continuous parameter spaces.
The researchers assessed the performance of the testing procedures through extensive Monte Carlo simulations, generating ten thousand independent sample configurations across diverse parameter combinations, group counts, and sample sizes ranging from five to twenty. The simulation studies demonstrated that the original generalized variable test remains excessively conservative across almost all investigated conditions, exhibiting empirical type I error rates well below the nominal five percent significance threshold, particularly when comparing ten groups. In contrast, both the Abdollahnezhad-Jafari modified likelihood ratio test and the newly proposed DK-modified likelihood ratio test maintained empirical error rates closely centered around five percent. However, the Abdollahnezhad-Jafari modification showed slight conservatism in ten-group comparisons, whereas the DK-modified likelihood ratio test consistently controlled type I error rates across every evaluated sample size and parameter configuration, demonstrating outstanding robustness in high-dimensional group settings.
Statistical power comparisons further highlighted operational distinctions among the candidate tests when evaluating twenty-fifth, fiftieth, and seventy-fifth percentiles across three, five, and ten populations. Due to its inherent conservatism, the generalized variable test consistently exhibited lower power than competing procedures. When comparing up to five groups, neither likelihood modification emerged as a definitive superior choice: the Abdollahnezhad-Jafari method achieved higher power at the lower quartile, whereas the DK-modified test showed superior sensitivity at the upper quartile, with comparable performance at the median. However, when evaluating ten groups, the DK-modified test clearly outperformed the Abdollahnezhad-Jafari test across all percentiles, establishing its superiority for larger group comparisons. Meanwhile, the parametric bootstrap test exhibited the highest overall power for group counts up to ten when evaluating quantiles at or below the median, although its power declined relative to the DK-modified test at the seventy-fifth percentile.
When overall hypothesis tests reject the null hypothesis of quantile equality, practitioners require simultaneous confidence intervals to determine precisely which specific population pairs differ. Prior literature had introduced complex simultaneous intervals based on minimum variance unbiased estimators and noncentral pivotal quantities, requiring repeated numerical evaluations of gamma functions. Building on this equivalence, the authors formulated simplified pairwise fiducial and parametric bootstrap confidence intervals that eliminate repetitive gamma calculations by substituting direct sample estimators. Furthermore, the team developed closed-form simultaneous Bonferroni confidence intervals using an accurate normal approximation to the noncentral t-distribution, providing an analytically tractable closed-form alternative that avoids iterative simulation procedures entirely.
Extensive numerical evaluations of simultaneous confidence interval properties revealed notable differences in coverage probability and interval precision across candidate methodologies. Assessing three-group and four-group scenarios through ten thousand Monte Carlo simulation runs, the researchers evaluated empirical coverage alongside scaled average volume metrics. The parametric bootstrap confidence intervals delivered the strongest performance overall, achieving coverage probabilities exceptionally close to the nominal ninety percent confidence level while producing narrower interval volumes than competing techniques. Pairwise fiducial intervals based on standard point estimators demonstrated performance virtually indistinguishable from their computationally heavier unbiased counterparts, confirming that simplified formulas introduce no loss in coverage accuracy. Although Bonferroni intervals exhibited conservative behavior with coverage probabilities exceeding nominal benchmarks and slightly wider intervals, their computational simplicity and closed-form expressions make them highly suitable for practical applications when sample sizes reach thirty or more.
The investigators illustrated their methodology using clinical data from a Roswell Park Cancer Institute trial that examined vitamin D metabolite concentrations among colorectal cancer patients. Subjects receiving a six-month vitamin D supplementation regimen were categorized into three cohorts based on baseline serum levels: deficient, insufficient, and sufficient. While testing for equality of group means and quantiles for serum 1,25-D3 yielded no statistically significant differences across groups, all evaluation methods detected significant group differences in the median and third quartile of serum 24,25-D3. Consistent with simulation patterns, the DK-modified likelihood ratio test produced smaller p-values than competing likelihood formulations, while the conservative generalized variable test generated higher p-values. Furthermore, simultaneous ninety-five percent confidence intervals constructed for the seventieth percentiles across the three patient cohorts all contained zero, indicating that specific pairwise differences at that percentile remained statistically indistinguishable.
A second empirical application evaluated communication latency among surgical anesthesia providers across four operating room decision-making workflows encompassing 472 recorded messages. The communication events reflected distinct operational contexts, categorized by whether a prior message existed and whether incoming communications were anchored. Every testing procedure identified statistically significant group differences across response latencies for percentiles at or below the thirtieth percentile. Constructing simultaneous ninety-five percent confidence intervals for the tenth percentile revealed that anesthesia providers responding to unanchored messages without prior warning differed significantly from those receiving anchored messages with prior context. The Bonferroni intervals additionally suggested potential differences between other workflow pairings, illustrating how quantile testing uncovers critical operational nuances obscured by aggregate mean summaries.
Subject of Research: Statistical hypothesis tests and simultaneous confidence intervals for comparing normal distribution quantiles
Article Title: Tests for Comparing Several Normal Quantiles and Pairwise Confidence Intervals
Article References: Dunnam, J., & Krishnamoorthy, K. (2026). Tests for Comparing Several Normal Quantiles and Pairwise Confidence Intervals. Journal of Statistical Theory and Applications, 25(1), Article 45. https://doi.org/10.1007/s44199-026-00198-3
Image Credits: AI Generated
DOI: 10.1007/s44199-026-00198-3
Keywords: normal quantiles, likelihood ratio test, parametric bootstrap, generalized variable test, simultaneous confidence intervals, pairwise comparisons, type I error rates, statistical power, noncentral t distribution, Bonferroni intervals, fiducial inference, colorectal cancer study, vitamin D metabolites, anesthesia response latency, lognormal distribution, moment matching, constrained maximum likelihood, multiple comparisons
Cite Scienmag News
Nathaniel Bowman. (November 18, 2026). Modified Likelihood Ratio Tests Improve Quantile Comparisons Across Multiple Normal Populations. Scienmag. https://scienmag.com/modified-likelihood-ratio-tests-improve-quantile-comparisons-across-multiple-normal-populations/
Nathaniel Bowman. "Modified Likelihood Ratio Tests Improve Quantile Comparisons Across Multiple Normal Populations." Scienmag, 18 November 2026, https://scienmag.com/modified-likelihood-ratio-tests-improve-quantile-comparisons-across-multiple-normal-populations/. Accessed 18 November 2026.
Nathaniel Bowman. "Modified Likelihood Ratio Tests Improve Quantile Comparisons Across Multiple Normal Populations." Scienmag. November 18, 2026. https://scienmag.com/modified-likelihood-ratio-tests-improve-quantile-comparisons-across-multiple-normal-populations/








