The tables of critical values printed in the back of a statistics textbook, for the t, F, studentized range, studentized maximum modulus, Dunnett, and Bryant-Paulson distributions, are not data to be trusted on faith. Each one is the quantile of a known distribution, so each one can be recomputed. This vignette shows that DMAR’s critical-value family reproduces those tables, walks through the parameter conventions that make the reproduction line up, and closes with a case where recomputation caught a typographical error in a printed table.
DMAR’s critical-value functions share a common set of arguments:
alpha_level; the degrees of freedom (df, or
df_numerator and df_denominator in
cv_scheffe(); cv_z() has none); and a count of
groups or comparisons where the distribution needs one. Each returns a
tidy table with one row per critical value:
| Function | Distribution | Appendix table it reproduces |
|---|---|---|
cv_t(), cv_z() |
Student’s t, standard normal | A.1 |
cv_tukey_hsd() |
studentized range | A.4 |
cv_smm() |
studentized maximum modulus | A.5 |
cv_dunnett() |
many-to-one (Dunnett) | A.6, A.7 |
cv_bryant_paulson() |
ANCOVA generalized studentized range | A.8 |
cv_scheffe() |
Scheffe (an F transformation) | derived from A.2 |
The base-R quantile functions qt(), qf(),
qtukey(), and qchisq() cover the t,
F, studentized range, and chi square tables directly; DMAR adds
the multiple-comparison distributions that base R does not ship.
Appendix Table A.1 is one-tailed t critical values at the
Bonferroni-spaced levels .05, .05/2, .05/4, .05/6, .05/8, .05/10. Each
cell is qt(1 - alpha, df):
alpha1 <- 0.05 / c(1, 2, 4, 6, 8, 10)
tab <- sapply(alpha1, function(a) round(qt(1 - a, df = c(5, 10, 20, 60)), 2))
dimnames(tab) <- list(df = c(5, 10, 20, 60), alpha1 = signif(alpha1, 3))
tab
#> alpha1
#> df 0.05 0.025 0.0125 0.00833 0.00625 0.005
#> 5 2.02 2.57 3.16 3.53 3.81 4.03
#> 10 1.81 2.23 2.63 2.87 3.04 3.17
#> 20 1.72 2.09 2.42 2.61 2.74 2.85
#> 60 1.67 2.00 2.30 2.46 2.58 2.66The printed value at df = 1, alpha = .05 is 6.31 = 6.31; the
infinite-df row is the normal quantile cv_z(). The fourth
column’s header prints “.0083”, but the exact level is 1/6 of .05 =
.008333…, and it is the exact fraction that reproduces the printed
cells. A literal .0083 would move a dozen cells by 0.01.
Tukey’s studentized range is qtukey(), used by
cv_tukey_hsd():
# Six groups, 14 error df, alpha = .05: Appendix A.4 prints 4.64.
qtukey(0.95, nmeans = 6, df = 14)
#> [1] 4.638538At very small error df, qtukey() is not accurate,
increasingly so as the number of groups and the confidence level grow,
and below 2 df it returns NaN. There
cv_tukey_hsd() takes the critical value from the exact
distribution instead (its help page gives the region and the
computation). With 20 groups and 2 error df, qtukey() is 17
percent too large:
qtukey(0.99, nmeans = 20, df = 2)
#> [1] 44.52627
sqrt(2) * cv_tukey_hsd(.01, df = 2, n_groups = 20, verbose = FALSE)$value
#> [1] 37.94346The studentized maximum modulus (cv_smm()) needs a word
about its index. Its argument is n_comparisons, the number
m of simultaneous statistics. When the maximum modulus is used
for all pairwise comparisons among a groups, that number is
m = a(a-1)/2, not a. So the “Number of Groups” column
of a printed SMM table maps to
n_comparisons = a * (a - 1) / 2:
# Appendix A.5, alpha = .05, "groups" a = 4, 5, 6 in the large-sample limit.
a <- c(4, 5, 6)
data.frame(groups = a, n_comparisons = a * (a - 1) / 2,
cv = round(sapply(a * (a - 1) / 2, function(m)
cv_smm(.05, df = Inf, n_comparisons = m, verbose = FALSE)$value), 2),
printed = c(2.63, 2.80, 2.93))
#> groups n_comparisons cv printed
#> 1 4 6 2.63 2.63
#> 2 5 10 2.80 2.80
#> 3 6 15 2.93 2.93(cv_smm() evaluates the one-dimensional integral that
defines the maximum modulus and inverts it deterministically, so the
value carries no Monte Carlo error and needs no seed.)
cv_dunnett() compares each of several treatments to a
single control. Its n_comparisons argument is the number of
treatments, that is, a - 1 where a counts all groups
including the control:
# Two-sided, a = 4 groups (3 treatments vs control), 10 error df: A.6 prints 2.76.
cv_dunnett(alpha_level = .05, df = 10, n_comparisons = 3, alternative = "not_equal")| term | value | area_less | area_greater |
|---|---|---|---|
| upper_cv | 2.759 | 0.950 | 0.050 |
With one treatment the many-to-one family collapses to a single two-sided t test, a useful check on the argument mapping:
c(dunnett = cv_dunnett(.05, df = 10, n_comparisons = 1,
alternative = "not_equal")$value,
t = qt(1 - .05/2, df = 10))
#> dunnett t
#> 2.228139 2.228139Appendix A.6 and A.7 reproduce Dunnett’s own tables. Recomputed with
the exact integral above, the printed values agree throughout except for
a scattering of cells at the .01 level that differ by 0.01 in the last
digit, a property of the historical tabulation rather than a
transcription error. (Reaching for a Monte Carlo multivariate t
quantile instead of the deterministic integral would introduce sampling
error of its own, several hundredths at the .01 level, which is why
cv_dunnett() integrates rather than simulates.)
The Bryant-Paulson generalized studentized range is the
analysis-of-covariance analog of Tukey’s range, for comparing
covariate-adjusted means when the covariate is random.
Base R does not compute it; cv_bryant_paulson() and the
underlying qbryant_paulson() supply it. (For its use in a
full ANCOVA analysis see the bryant_paulson_ancova
vignette.)
It reproduces the published anchor exactly:
# Bryant & Paulson (1976) Table 1a: q_{.05; p=1, a=6, nu=14} = 4.83.
cv_bryant_paulson(alpha_level = .05, df = 14, n_groups = 6, n_covariates = 1,
verbose = FALSE)$value
#> [1] 4.829856Recomputing an entire block of Appendix Table A.8 (error df = 20, three covariates, alpha = .01) against the printed values shows a single cell out of step:
a_vals <- c(2, 3, 4, 5, 6, 7, 8, 10, 12, 16, 20)
recomputed <- sapply(a_vals, function(k)
cv_bryant_paulson(.01, df = 20, n_groups = k, n_covariates = 3,
verbose = FALSE)$value)
printed <- c(4.35, 5.03, 5.45, 5.75, 6.09, 6.19, 6.36, 6.63, 6.85, 7.19, 7.45)
data.frame(a = a_vals, printed = printed, recomputed = round(recomputed, 2),
diff = round(recomputed, 2) - printed)
#> a printed recomputed diff
#> 1 2 4.35 4.35 0.0
#> 2 3 5.03 5.03 0.0
#> 3 4 5.45 5.45 0.0
#> 4 5 5.75 5.75 0.0
#> 5 6 6.09 5.99 -0.1
#> 6 7 6.19 6.19 0.0
#> 7 8 6.36 6.36 0.0
#> 8 10 6.63 6.63 0.0
#> 9 12 6.85 6.85 0.0
#> 10 16 7.19 7.19 0.0
#> 11 20 7.45 7.45 0.0Every cell agrees except a = 6, where the table prints 6.09 and the distribution gives 5.99. Two arguments settle which is right. First, the critical value is concave in the number of means, so the increments across the row must decrease; 5.99 keeps them decreasing while 6.09 forces a reversal:
rbind(printed_increments = diff(printed),
recomputed_increments = round(diff(recomputed), 2))
#> [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
#> printed_increments 0.68 0.42 0.3 0.34 0.1 0.17 0.27 0.22 0.34 0.26
#> recomputed_increments 0.68 0.42 0.3 0.24 0.2 0.17 0.27 0.22 0.34 0.26Second, a direct Monte Carlo of the statistic, drawn from its definition as the range of a standard normals divided by an independent scale, agrees with the computed 5.99, not with 6.09:
G <- 2e5; a <- 6; nu <- 20; p <- 3
R <- apply(matrix(rnorm(G * a), ncol = a), 1, function(z) max(z) - min(z))
S <- sqrt(rchisq(G, nu) / nu)
delta <- rbeta(G, (nu + 1) / 2, p / 2) # covariate-shrinkage factor
Q <- R / (S * sqrt(delta)) # Bryant-Paulson statistic
c(simulated_99th_pctile = round(quantile(Q, 0.99, names = FALSE), 2),
cv_bryant_paulson = round(cv_bryant_paulson(.01, df = 20, n_groups = 6,
n_covariates = 3, verbose = FALSE)$value, 2))
#> simulated_99th_pctile cv_bryant_paulson
#> 6.01 5.99The printed 6.09 is a typographical error for 5.99. Recomputation is not just a way to avoid typing a table by hand; it is a way to check the tables one has.
Bryant, J. L., & Paulson, A. S. (1976). An extension of Tukey’s method of multiple comparisons to experimental designs with random concomitant variables. Biometrika, 63, 631–638.
Dunnett, C. W. (1964). New tables for multiple comparisons with a control. Biometrics, 20, 482–491.
Maxwell, S. E., Delaney, H. D., & Kelley, K. (2027). Designing experiments and analyzing data: A model comparison perspective (4th ed.). Routledge.