Computes the observed value of robust Welch (RW) test, degrees of freedoms (numerator and denominator) and the corresponding p-value for the equality of means of several long-tailed symmetric (LTS) distributions when the variances are unknown and arbitrary.
RW(formula, data, alpha=0.05, verbose = TRUE, p_shape)A list with class "htest" containing the following components:
the observed value of the RW test statistic.
the numerator and the denominator degrees of freedom of the approximate F distribution.
the p-value for the RW test.
the level of significance.
a character string "Robust Welch Test based on MML Estimators" indicating which test is used.
a data frame containing the variables.
a formula of the form left-hand-side(lhs) ~ right-hand-side(rhs). lhs shows the observed values and rhs shows the group corresponding to the observed values.
a formula of the form left-hand-side(lhs) ~ right-hand-side(rhs). lhs shows the observed values and rhs shows the group corresponding to the observed values.
data frame containing the variables in the formula.
the level of significance. Default is set to alpha = 0.05.
a logical for printing output to R console.
shape parameter of the LTS distribution
Gamze Guven <gamzeguven@ogu.edu.tr>
RW test based on modifed maximum likelihood (MML) estimators is proposed as a robust alternative to Welch's F test (Welch, 1951). The test statistic is formulated as follows
$$RW= \frac{T(\hat{\mu}_1, \dots, \hat{\mu}_a;\hat{\sigma}_1^{2},\dots,\hat{\sigma}_a^{2})/(a-1)}{1+(2(a-2)/(3\nu_1))} $$ where $$T(\hat{\mu}_1,\dots,\hat{\mu}_a; \hat{\sigma}_1^{2},\dots,\hat{\sigma}_a^{2})=\sum\limits_{i=1}^a \frac{M_i}{\hat{\sigma}_i^{2}} \hat{\mu}_i^{2}- \frac{(\sum\limits_{i=1}^a M_i\hat{\mu}_i/\hat{\sigma}_i^{2})^2}{\sum\limits_{i=1}^a M_i/\hat{\sigma}_i^{2}},$$ $$\nu_1= [\frac{3}{a^2-1} \sum\limits_{i = 1}^a \frac{1}{n_i-1}(1-( M_i/\hat{\sigma}_i^2)/( \sum\limits_{j= 1}^a M_j/\hat{\sigma}_j^2))^{2}]^{-1},$$ \(\hat{\mu}_{i}\) and \(\hat{\sigma}_{i}\) (i=1,2,...,a) are the MML estimators of the location and scale parameters, respectively, see Tiku (1967, 1968) for the details of MML estimators.
The null hypothesis is rejected if the computed RW statistic is higher than the \((1-\alpha)\)th quantile of the F distribution with a-1 and \(\nu_{1}\) degrees of freedom.
For further details, see Guven et al. (2022).
G. Guven, S. Acitas, and B. Senoglu, B. RobustANOVA: An R Package for one-way ANOVA under heteroscedasticity and nonnormality. Under review, 2022.
M. L. Tiku. Estimating the mean and standard deviation from a censored normal sample. Biometrika, 54:155-165, 1967.
M. L. Tiku. Estimating the parameters of log-normal distribution from censored samples. Journal of the American Statistical Association, 63(321): 134-140, 1968.
B. L. Welch. On the comparison of several mean values: an alternative approach. Biometrika, 38(3): 330-336, 1951.
library(RobustANOVA)
RW(obs ~ methods, data = peak_discharge, alpha = 0.05, verbose = TRUE, p_shape=2.3)
Run the code above in your browser using DataLab