R: Simultaneous Confidence Bounds Based on Steel's Multiple...
SteelConfInt
R Documentation
Simultaneous Confidence Bounds Based on Steel's Multiple Comparison Wilcoxon Tests
Description
This function inverts pairwise Wilcoxon tests, comparing a common control sample with each of
several treatment samples to provide simultaneous confidence bounds
for the respective shift parameters by which the sampled treatment populations may differ
from the control population. It is assumed that all samples are independent
and
that the sampled distributions are continuous to avoid ties.
The joint coverage probability for all bounds/intervals is calculated, estimated, or approximated, see
Details. For treatment of ties also see Details.
Usage
SteelConfInt(..., data = NULL, conf.level = 0.95,
alternative = c("less", "greater", "two.sided"),
method = c("asymptotic", "exact", "simulated"), Nsim = 10000)
Arguments
...
Either several sample vectors, say
x_1, …, x_k,
with x_i containing n_i sample values.
n_i > 4 is recommended for reasonable asymptotic
P-value calculation. The pooled sample size is denoted
by N=n_1+…+n_k. The first vector serves as control sample,
the others as treatment samples.
or a list of such sample vectors.
or a formula y ~ g, where y contains the pooled sample values
and g (same length as y) is a factor with levels identifying
the samples to which the elements of y belong. The lowest factor level
corresponds to the control sample, the other levels to treatment samples.
data
= an optional data frame providing the variables in formula y ~ g.
conf.level
= 0.95 (default) the target joint confidence level for all bounds/intervals.
0 < conf.level < 1.
alternative
= c("less", "greater", "two.sided"), where "less" results in
simultaneous upper confidence bounds for all shift parameters and any negative upper bound
should lead to the rejection of the null hypothesis of all shift parameters
being zero or positive in favor of at least one being less than zero.
"greater" results in simultaneous lower confidence bounds for all shift parameters
and any positive lower bound should lead to the rejection of the null hypothesis of all
shift parameters being zero or negative in favor of at least one being greater
than zero.
"two.sided" results in simultaneous confidence intervals for all shift parameters
and any interval not straddling zero should lead to the rejection of the null hypothesis
of all shift parameters being zero in favor of at least one being different from zero.
method
= c("asymptotic", "exact", "simulated"), where
"asymptotic" uses only an asymptotic normal approximation
to approximate the achieved joint coverage probability.
This calculation is always done.
"exact" uses full enumeration of all sample splits to obtain the exact achieved
joint coverage probability (see Details).
It is used only when Nsim is at least as large as the number of full enumerations.
Otherwise, method reverts to "simulated" using the given Nsim.
"simulated" uses Nsim simulated random splits of the
pooled samples into samples of sizes n_1, …, n_k,
to estimate the achieved joint coverage probability.
Nsim
= 10000 (default), number of simulated sample splits to use.
It is only used when method = "simulated",
or when method = "exact" reverts to method ="simulated", as previously explained.
Details
The first sample is treated as control sample with sample size n_1. The remaining
s=k-1 samples are treatment samples.
Let W_{1i}, i=2,…,k denote the respective Wilcoxon statistics comparing the common control sample (index 1)
with each of the s treatment samples (indexed by i).
For each comparison of control and treatment i
sample
only the observations of the two samples involved are ranked.
By W_i=W_{1i}-n_i(n_i+1)/2 we denote
the corresponding Mann-Whitney test statistic.
Furthermore, let D_{i(j)} denote the j-th ordered value (ascending order) of the n_1n_i
paired differences between the observations in treatment sample i and those of the control
sample. By simple extension of results in Lehmann (2006), pages 87 and 92, the following equations hold,
relating the null distribution of the
Mann-Whitney statistics and the joint coverage probabilities of the D_{i(j_i)} for any set of
j_1,…,j_s with 1≤ j_i ≤ n_1 n_i.
where P_Δ refers to the distribution under Δ=(Δ_2,…,Δ_k)
and P_0 refers to the joint null distribution of the W_i when all sampled
distributions are the same and continuous. There are k-1 indices j_i that can be manipulated
to affect the achieved confidence level. To limit the computational complexity
standardized versions of the W_i, i.e., (W_i-μ_i)/τ_i with
μ_i and τ_i representing mean and standard deviation of W_i,
are used to choose a common value for (j_i -1-μ_i)/τ_i (satisfying the
γ level) from the multivariate normal approximation
for the W_i (see Miller (1981) and Scholz (2016)), and reduce that
to integer values for j_i, rounding up, rounding down, and rounding to the nearest integer. These
integers j_i are then used in approximating the actual joint probabilities
P_0(W_i≤ j_i -1, i=2,…,k), and from these three coverage probabilities
the one that is closest to the nominal confidence level γ and ≥ γ
and also also the one that is closest without the restriction ≥ γ are chosen.
When method = "exact" or = "simulated" is specified, the same process
is used, using either the fully enumerated exact distribution of W_i, i=2,…,k (based on a recursive
version of Chase's sequence as presented in Knuth (2011)) for all sample splits,
or the simulated distribution of W_i, i=2,…,k. However, since these distributions are discrete
the starting point before rounding up is the smallest quantile such that the proportion of distribution values less
or equal to it is at least γ. The starting point before rounding down is the highest quantile such that
the proportion of distribution values less
or equal to it is at most γ. The third option of rounding to the closest integer is performed using
the average of the first two.
Confidence intervals are constructed by using upper and lower confidence bounds, each with
same confidence level of (1+γ)/2.
When the original sample data appear to be rounded, and especially when there are ties,
one should widen the computed intervals or bounds by the rounding ε, as illustrated
in Lehmann (2006), pages 85 and 94. For example, when all sample values appear to end in one of .0, .2, .4, .6, .8,
the rounding ε would be .2. Ultimately, this is a judgment call for the user. Such widening
of intervals will make the actually achieved confidence level ≥ the stated achieved level.
Value
A list of class kSamples with components
test.name
"Steel.bounds"
n1
the control sample size = n_1
ns
vector (n_2,…,n_k) of the s=k-1 treatment sample sizes
N
size of the pooled sample = n_1+…+n_k
n.ties
number of ties in the pooled sample
bounds
a list of data frames. When method = "asymptotic" is specified, the list
consists of two data frames named conservative.bounds.asymptotic
and closest.bounds.asymptotic. Each data frame consists of s rows
corresponding to the s shift parameters Δ_i and three columns,
the first column giving the lower bound, the second column the upper bound, while
the first row of the third column states the computed confidence level by asymptotic
approximation, applying jointly to all s sets of bounds. For one-sided bounds
the corresponding other bound is set to Inf or -Inf, respectively.
In case of conservative.bounds.asymptotic the achieved asymptotic confidence level is
targeted to be ≥conf.level, but closest to it among three possible choices (see Details).
In the case of closest.bounds.asymptotic the achieved level is targeted to
be closest to conf.level.
When method = "exact" or method = "simulated"
is specified the list consists in addition of two further data frames named either
conservative.bounds.exact and
closest.bounds.exact or
conservative.bounds.simulated and
closest.bounds.simulated.
In either case the structure and meaning
of these data frames parallels that of the "asymptotic" case.
method
the method used.
Nsim
the number of simulations used.
j.LU
an s by 4 matrix giving the indices j_i used for computing the bounds
D_{i(j_i)} for Δ_i, i=1,…, s.
warning
method = "exact" should only be used with caution.
Computation time is proportional to the number of enumerations.
Experiment with system.time and trial values for
Nsim to get a sense of the required computing time.
References
Knuth, D.E. (2011), The Art of Computer Programming, Volume 4A
Combinatorial Algorithms Part 1, Addison-Wesley
Lehmann, E.L. (2006),
Nonparametrics, Statistical Methods Based on Ranks, Revised First Edition,
Springer Verlag.
Miller, Rupert G., Jr. (1981), Simultaneous Statistical Inference, Second Edition,
Springer Verlag, New York.
Scholz, F.W. (2016), "On Steel's Test with Ties", submitted to Journal of Nonparametric Statistics.