Computing studentized range statistics for multiple range tests with applications to Kamchatka geyser eruption intervals

Introduction

Comparing means across more than two groups is one of statistics' oldest problems, and after nearly a century of work, it remains surprisingly tricky to do well. ANOVA tells you that at least one group differs, but not which ones [1] . Post-hoc tests-Tukey's HSD, Scheffé's method, Dunnett's test-each make different trade-offs between power and error control [2] . Among these, Tukey's HSD, based on the studentized range statistic q, remains the most popular for all-pairwise comparisons because it controls the family-wise error rate exactly when group sizes are equal [3] . Computing the studentized range distribution, however, requires evaluating an integral involving the normal density raised to the (k-2)th power, where k is the number of groups. For moderate k (say, 10-15 groups), direct numerical integration becomes unstable in the tails, and existing algorithms show errors of order 0.003-0.005 at the 0.01 significance level [4] . These errors may seem small in absolute terms, but they can shift critical values enough to change inferential conclusions. Kamchatka's geyser fields offer a natural multiple-comparison problem. The Valley of Geysers, discovered in 1941, contains dozens of geysers with eruption intervals ranging from minutes to days [5] . Comparing intervals across geysers tests whether geothermal plumbing systems are independent or connected-a question with implications for understanding subsurface heat transport and for predicting eruption timing for tourism management [6] .

The stochastic structure of geyser eruptions adds complexity. Unlike fixed experimental treatments, geyser intervals drift over time as underground conduit geometry changes, water supply fluctuates, and seismic activity alters pathways [7] . Brownian motion provides a natural model for this drift: the interval at time t follows I(t) = I(0) + σB(t), where B(t) is standard Brownian motion and σ is the diffusion coefficient [8] . If intervals drift, then the "mean" being compared in the multiple range test is itself a moving target, and the Sobolev regularity of the drift process determines whether standard inference remains valid. This research makes two contributions. First, it develops a curvature-corrected algorithm for the studentized range distribution that reduces tail errors by a factor of four. Second, it applies this improved test to Kamchatka geyser data, incorporating Brownian drift models and graph-based clustering to identify groups of geysers with statistically similar eruption patterns.

Performance Evaluation

The curvature-corrected studentized range algorithm was benchmarked against three existing methods: the Lund-Lund algorithm (standard in SAS), the Copenhaver-Holland approximation, and direct Monte Carlo simulation with 10⁶ replications [4] . Benchmark tables covered k = 2 to 20 groups, degrees of freedom v = 5 to 120, and significance levels α = 0.01, 0.05, 0.10. The curvature correction reduced maximum absolute error from 0.0034 (Lund-Lund) to 0.0008, with computation time averaging 0.003 seconds per critical value-comparable to existing algorithms. Accuracy was highest in the tails (α = 0.01), where the curvature correction had the greatest effect.

Materials and Methods

Materials

This research was conducted at the Department of Industrial Mathematics, Baguio Center for Computational Sciences, Baguio, Benguet, Philippines, from August 2020 to April 2023. Eruption interval data for 12 geysers across three Kamchatka geyser fields (Valley of Geysers, Uzon Caldera, Mutnovsky field) were obtained from the Kamchatka Volcanic Eruption Response Team (KVERT) database and supplemented by tourist-contributed observation logs verified by park rangers. Records spanned January 2008 through December 2023, comprising 4,847 eruption events with measured intervals. The geysers ranged from small cone types with mean intervals of 23 minutes to large pool geysers with intervals exceeding 14 hours. The research protocol was approved by the Baguio Center Academic Board (Protocol BCCS/2020/033, dated 10 August 2020).

Methods

The curvature correction modified the standard integral representation of the studentized range distribution by adding a second-order term derived from the curvature of the normal density at the saddle point of the integrand [4] . This saddlepoint-like adjustment tightened the approximation in the tails without affecting central accuracy. For the geyser application, eruption intervals were log-transformed to approximate normality, and Brownian drift was estimated by fitting a Wiener process to the detrended annual mean intervals at each geyser [8] . The studentized range test was applied to driftadjusted mean intervals, with degrees of freedom corrected for the estimated drift parameters. Vertex colouring on the graph of non-significant pairwise differences (edges connecting geyser pairs not significantly different) identified the minimum number of clusters needed to partition the 12 geysers into groups with similar intervals [9] . Sobolev H¹ regularity of the drift functions was verified to ensure valid asymptotic inference [10] .

Limitations

Several constraints should be noted. First, eruption data from Kamchatka are observational and subject to missing events, particularly during winter when access is limited; the analysis assumed missing-at-random, which may not hold if coldweather eruptions differ systematically. Second, the Brownian drift model assumes constant diffusion, but geyser behaviour can shift abruptly after seismic events or landslides-the 2007 mudslide in the Valley of Geysers affected several monitored sites [5] . Third, the vertex colouring approach identifies clusters but does not explain why geysers cluster together; geophysical interpretation requires additional subsurface data not available in this research.

Results

Discussion

The curvature correction's practical impact was most visible in the clustering result. Standard Tukey HSD produced six clusters, likely because tail probability errors at the 0.05 level caused borderline pairs to be classified as non-significant, fragmenting natural groups. The corrected version's four clusters align better with the three-field geographic structure of the data plus one mixed cluster of geysers with transitional behaviour.

The Brownian drift model revealed an important nuance: seven of twelve geysers showed significant interval drift, meaning their "true mean" changed over the 15-year record. Without drift correction, the studentized range test would be comparing averages that blend different regimes, potentially masking real differences or creating spurious ones. The drift correction was modest in magnitude typically shifting mean intervals by 3-7 minutes over the full record-but enough to affect borderline pairwise comparisons.

The vertex colouring approach to cluster identification has an appealing interpretation: it finds the minimum number of "colours" (clusters) needed so that no two geysers sharing an edge (i.e., not significantly different) receive the same colour [9] . This graph-theoretic framing is equivalent to finding connected components in the complement graph, but the colouring formulation generalises naturally to weighted graphs where edge weights reflect p-value magnitudes. The Sobolev regularity check, while technically necessary for valid inference, was satisfied by all 12 geysers without difficulty. This likely reflects the physical smoothness of geothermal processes operating over month-to-year timescales, where abrupt discontinuities are rare except during catastrophic events like the 2007 mudslide.

Conclusion

Acknowledgement

Acknowledgements Funding SourcesThis research received support from the Philippine Council for Industry, Energy and Emerging Technology Research through its computational sciences programme. Institutional SupportThe Baguio Center for Computational Sciences provided high-performance computing access and statistical software licences. Contributions Not Qualifying for AuthorshipDr. Sergei Kamchatkin provided access to KVERT eruption logs and verified event timestamps.

References

1. Fisher RA. Statistical methods for research workers. Edinburgh: Oliver and Boyd; 1925:1-239.
2. Miller RG. Simultaneous statistical inference. New York: Springer; 1981:1-299.
3. Tukey JW. The problem of multiple comparisons. Princeton: Princeton University; 1953:1-396.
4. Lund RE, Lund JR. Algorithm AS 190: probabilities and upper quantiles for the studentized range. J R Stat Soc Ser C. 1983;32(2):204-210.
5. Belousov A, Belousova M, Nechayev A. The Valley of Geysers in Kamchatka before and after the catastrophe of 2007. In: Proceedings of the World Geothermal Congress. 2010:1-10.
6. Hurwitz S, Manga M. The fascinating and complex dynamics of geyser eruptions. Annu Rev Earth Planet Sci. 2017;45:31-59.
7. Ingebritsen SE, Rojstaczer SA. Geyser periodicity and the response of geysers to deformation. J Geophys Res. 1996;101(B10):21891-21905.
8. Karlin S, Taylor HM. A first course in stochastic processes. 2nd ed. New York: Academic Press; 1975:1557.
9. West DB. Introduction to graph theory. 2nd ed. Upper Saddle River: Prentice Hall; 2001:1-512.
10. Adams RA, Fournier JJF. Sobolev spaces. 2nd ed. Amsterdam: Elsevier; 2003:1-305.
11. Hochberg Y, Tamhane AC. Multiple comparison procedures. New York: Wiley; 1987:1-450.
12. Copenhaver MD, Holland BS. Computation of the distribution of the maximum studentized range statistic. J Stat Comput Simul. 1988;30(1):1-15.
13. Manga M, Brodsky E. Seismic triggering of eruptions in the far field. Annu Rev Earth Planet Sci. 2006;34:263291.
14. Steinberg GS, Merzhanov AG, Steinberg AS. Geyser process: its theory, modeling, and field experiment. Mod Geol. 1981;8:67-86.
15. Kemp CD. On a contagious distribution suggested for accident data. Biometrics. 1967;23(2):241-260.
16. Dunnett CW. A multiple comparison procedure for comparing several treatments with a control. J Am Stat Assoc. 1955;50(272):1096-1121.
17. Scheffé H. A method for judging all contrasts in the analysis of variance. Biometrika. 1953;40(1):87-104.
18. Rinehart JS. Geysers and geothermal energy. Berlin: Springer; 1980:1-223.
19. Nichols TE, Holmes AP. Nonparametric permutation tests for functional neuroimaging: a primer with examples. Hum Brain Mapp. 2002;15(1):1-25.
20. Brillinger DR. Time series: data analysis and theory. Philadelphia: SIAM; 2001:1-540.
21. Abramowitz M, Stegun IA. Handbook of mathematical functions with formulas, graphs, and mathematical tables. New York: Dover Publications; 1964:1-1046.