Introduction and Overview

Given two paired, continuous variates (denoted as Y₁ and Y₂), the standard frequentist parametric analysis is the paired t-test. This analysis is predicated on the assumption that the scores are independent, identically-distributed data from a normal distribution for each condition. However, the parametric assumptions are not likely to be strictly true. The possibility of mixture processes invalidate the assumption of the scores being independent and identically distributed. Moreover, there are processes where the central limit theorem does not apply because the underlying stochastic process is not based on the mean value of a latent distribution. Processes such as flooding or task completion time have measures that are dependent on the minimum or the maximum of an underlying process rather than a mean value. The parametric assumptions for the paired t -test are not reasonable for this type of research. Consequently, it is prudent to have a powerful alternative method for doing statistical inference that is not based on parametric assumptions. The Wilcoxon signed-rank test is the most powerful frequentist alternative to the paired t-test (Siegel and Castellan, 1988).¹ Because the signed-rank procedure uses rank values, it has the additional advantage of not being susceptible to the undue influence of outlier scores. Consequently, the Wilcoxon signed-rank test is a more robust inferential approach than the standard t-test.

Chechile (2018) developed a Bayesian version for the Wilcoxon signed-rank statistic. The theory for the Bayesian version of the Wilcoxon signed-rank test is described below in the section The Bayesian Wilcoxon Signed-Rank Procedure, and examples for using the dfba_wilcoxon() function are provided in the section Using the dfba_wilcoxon() Function. Sections 2 and 3 involve issues associated with Monte Carlo sampling, approximating the posterior distribution with a beta distribution, the likelihood principle, and the Bayes factor; please see the vignettes for the dfba_beta_bayes_factor(), dfba_binomial() and the dfba_beta_contrast() functions for more information.

The Bayesian Wilcoxon Signed-Rank Procedure

The frequentist Wilcoxon signed-rank procedure is based on the rank of the difference scores d = Y₁ − Y₂ (Wilcoxon, 1945). The ranking is initially done on the absolute value of the nonzero d values; any data pair where d = 0 is discarded. Thus, all the remaining |d| are positive rank values. In a second step, the signs for the d values are assigned to the absolute value d values. The sample T_pos statistic is the sum of the ranks that have a positive sign. The sample T_neg is the positive sum of the ranks that have a negative value; thus T_neg is also a positive value. As an example, consider the data, taken from Chechile (2018), where there are the following eight d values along with their corresponding signed ranks sr:

The resulting T statistics for this example are: T_pos = 23 and T_neg = 13. In general, for n nonzero d scores, T_pos + T_neg = n(n + 1)/2. Note for the above example that T_pos + T_neg = 36. Because the two statistics sum to a constant, statistical inference may be performed with only one of these two statistics. Let us focus on the T_pos statistic.²

The frequentist Wilcoxon signed-rank procedure assumes the null hypothesis that the rate of positive signs is precisely .5, and based on that assumption, it computes the likelihood for the observed T_pos plus more extreme (unobserved) T_pos values. For the above example, the frequentist approach assumes the null hypothesis of a .5 rate for positive signs and computes the likelihood for T_pos ≥ 23 given n = 8. If the summed likelihood (i.e., the p-value) is less than α, then the frequentist signed-rank test rejects the assumed null hypothesis. The inclusion of more extreme likelihoods than the observed likelihood violates the likelihood principle, (see the vignette about the dfba_binomial() function). The likelihood principle (Berger & Wolpert, 1988) is an essential feature of the Bayesian approach.

The likelihood principle can be described as the rule that, upon completion of an experiment, the only likelihood that should be computed is the likelihood of the observed data.

Other possible non-observed outcomes are irrelevant because those outcomes are not part of Bayes theorem. Both parametric and nonparametric frequentist statistics routinely violate the likelihood principle. When there are violations of the likelihood principle, there can be inferential paradoxes (e.g., Lindley & Phillips, 1976; Chechile, 2020).

Chechile (2018) developed the Bayesian model for analyzing the Wilcoxon signed-rank statistic. Because it is a Bayesian approach, it strictly adheres to the likelihood principle. A second major difference of the Bayesian approach from the frequentist analysis is that the null hypothesis is not assumed. Instead there is a sign-bias parameter ϕ_w, which has a probability distribution on the [0, 1] interval. For any value for the ϕ_w parameter, there is a corresponding likelihood for finding the observed T_pos value. The problem is that this likelihood is not known in closed form. However, the likelihood values can be estimated based on Monte Carlo sampling. For example, for any arbitrary ϕ_w value and for n = 8, it is possible to sample a configuration of signs over the integers 1 to 8 and compute the resulting T_pos value, and then to repeat the procedure a number N times. The default value for the number of Monte Carlo samples in the dfba_wilcoxon() function is N = 30000 for each candidate ϕ_w value. The likelihood is estimated by the proportion of the Monte Carlo samples that result in the observed T_pos value. The dfba_wilcoxon() function evaluates 200 candidate values for ϕ_w: .025 to .9975 in steps of .005. Thus, there is a discrete prior and posterior probability distribution over the values .0025, .0075, …, .9975 for the ϕ_w parameter. Unlike the frequentist signed-rank analysis, the Bayesian approach focuses on estimating the sign-bias parameter ϕ_w with point and interval estimates. It also provides interval Bayes factor values.

Chechile (2018) also studied the posterior distribution for ϕ_w as a function of sample size. He reported that the posterior ϕ_w can be accurately approximated by a beta distribution when the number (n) of non-zero d values is greater than 24. The approximation formula closely matches the mean, variance, and quantiles of the discrete distribution of the Monte Carlo sampling approach. Thus the computationally slow Monte Carlo method can be avoided when n > 24. For the beta approximation, the posterior beta shape parameter are a = n_a + a₀ and b = n_b + b₀ where a₀ are b₀ are the shape parameters, and where

Note that the values for the approximate, large-n posterior beta will generally be non-integer values.

Using the dfba_wilcoxon() Function

The dfba_wilcoxon() function has two required arguments, which are the data for the two paired continuous measurements Y1 and Y2. The user should be careful to assure that there is a linkage between the ith score for Y₁ and the ith score for Y₂, such as the data being from the same research participant tested in two different experimental conditions. Besides these two required arguments, there are five other optional arguments (listed with their respective default values): a0 = 1, b0 = 1, prob_interval = .95, samples = 30000, and method = NULL. The a0 and b0 arguments are the shape parameters for a prior beta distribution for the ϕ_w parameter. The default prior is a uniform distribution, but an informed prior can be employed with the selection of different values for a0 and b0 The prob_interval argument is the value for the interval estimate of ϕ_w. The samples argument is the number of Monte Carlo samples that are drawn for each candidate value for ϕ_w. Finally, the method argument is either the string "small" or "large". The input method = "small" specifies to use the small-n Monte Carlo sampling method (described in The Bayesian Wilcoxon Signed-Rank Procedure); the argument method = "large" specifies to use the large-n approximation for the ϕ_w distribution. When method = NULL, the software uses the small-n Monte Carlo approach when n ≤ 24 and uses the large-n approach when n > 24.

set.seed(77)

w1 <- 10 + rweibull(30, .8) 
w2 <- 9.1 + rweibull(30, .7)

A <- dfba_wilcoxon(Y1 = w1,
                   Y2 = w2)

A
#> Descriptive Statistics 
#> ========================
#>   Wilcoxon Signed-Rank Statistics 
#>   n 
#>   30 
#>   T_pos 
#>   316 
#>   T_neg 
#>   149 
#> 
#>   Beta Approximation Model for Phi_W
#>  for n > 24
#> ========================
#>   The posterior beta shape parameters are:
#>   posterior a         posterior b
#>   16.0403             7.959677 
#>   posterior mean      posterior median
#>   0.668347            0.6730931 
#>   95% Equal-tail interval limits: 
#>   Lower Limit     Upper Limit 
#>   0.472595        0.8375461 
#>   95% Highest-density interval limits: 
#>   Lower Limit     Upper Limit 
#>   0.483053        0.8460527 
#> 
#>  
#>   probability that phi_W > 0.5:
#>   prior           posterior
#>   0.5             0.9550913 
#>   Bayes factor BF10 for phi_W > 0.5:
#>   21.26741

B <- dfba_wilcoxon(Y1 = w1,
                   Y2 = w2,
                   method = "small",
                   samples = 100000,
                   hide_progress = TRUE)

#> Descriptive Statistics 
#> ========================
#>   Wilcoxon Signed-Rank Statistics 
#>   n 
#>   30 
#>   T_pos 
#>   316 
#>   T_neg 
#>   149 
#> 
#>   Monte Carlo Sampling with Discrete Probability Values
#> ========================
#>   Number of MC Samples
#>   1e+05 
#>   
#>   Posterior mean of phi_w:
#>   0.6651462 
#>   95% Highest-density interval limits: 
#>   Lower Limit     Upper Limit 
#>   0.471082        0.8344068 
#>   probability that phi_W exceeds 0.5:
#>   prior           posterior
#>   0.5             0.9539032 
#>   Bayes factor BF10 for phi_W > 0.5:
#>   20.69348

plot(A)

plot(B)

References

Berger, J. O., and Wolpert, R. L. (1988). Likelihood Principle (2nd ed.). Hayward, CA: Institute of Mathematical Statistics.

Chechile, R. A. (2018) A Bayesian analysis for the Wilcoxon signed-rank statistic. Communications in Statistics – Theory and Methods, https://doi.org/10.1080/03610926.2017.1388402

Chechile, R. A. (2020). Bayesian Statistics for Experimental Scientists: A General Introduction Using Distribution-Free Methods. Cambridge, MIT Press.

Lindley, D. V., and Phillips, L. D. (1976). Inference for a Bernoulli process (a Bayesian view). The American Statistician, 30, 112-119.

Siegel, S., and Castellan, N. J. (1988). Nonparametric Statistics for the Behavioral Sciences. New York: McGraw Hill.

Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometric Bulletin, 1, 80-83.