A Power Study of Goodness-of-Fit Tests for Categorical Data

A Power Study of Goodness-of-Fit Tests for Categorical Data Michael Steele School of Mathematical and Physical Sciences, James Cook University, [email protected] Janet Chaseling Australian School of Environmental Studies, Griffith University Cameron Hurst School of Information Technology and Mathematical Sciences, University of Ballarat. 1. Introduction Goodness-of-fit (GOF) tests are used for the analysis of categorical data by applied researchers from many disciplines however studies of their relative powers are limited. Although the Chi-Square ( 2) test is a popular choice for many researchers, power studies show that this may be at the expense of power in some instances. This paper compares the powers of two of the lesser known GOF test statistics based on the empirical distribution function with the 2 test to determine which is the more powerful for the investigated null and alternative distributions. 2. The test statistics used in the power study The test statistics used are 2 (Pearson 1900), the discrete Kolmogorov-Smirnov KS (Pettitt and Stephens 1977) and the discrete Cramér-von Mises W2 (Choulakian et al. 1994). k 2 (1) χ = ( Oi − Ei ) Ei i =1 (2) KS = 2 max Zi 1≤ i ≤ k k (3) W 2 = N −1 2 Z i pi i =1 where k is the number of cells, N is the sample size, pi is the probability for cell i, Oi and Ei and are the observed and expected frequencies for cell i, and Zi is the cummulative sum of the differences between Oi and Ei up to and including cell i. 3. The power study The power for each test statistic is approximated for a uniform null distribution over 10 cells against the increasing trend and triangular ∨ or ’bath-tub’ type alternatives defined in Table 1. The total sample sizes range from 10 to 200 which represents expected frquencies under the uniform null distribution of 1 to 20 per cell. The power of each test statistic is estimated at the 5% significance level from 10000 simulated random samples. The simulated distributions of the test statistic are discrete. To overcome that there may not be a unique test statistic at the required significance level of 5%, linear interpolation of the powers about this level is used for consistency. Table 1. Distributions used in the power study. Cell Probability (2 Decimal Places) Description 1 2 3 4 5 6 7 8 Uniform 0.10 0.10 0.10 0.10 0.10 0.10 0.10 0.10 Increasing 0.03 0.06 0.07 0.09 0.10 0.11 0.12 0.13 Triangular ∨ 0.17 0.13 0.10 0.07 0.03 0.03 0.07 0.10 9 0.10 0.14 0.13 10 0.10 0.15 0.17 4. Results from the power study The powers for the increasing alternative distribution are given in Figure 1. The powers for the triangular ∨ or ’bath-tub’ type alternative are given in Figure 2. A summary of which of the test statistics have the higher power for the two alternatives is given in Table 2. 1 Power 0.8 0.6 KS 0.4 W^2 ^2 0.2 0 10 20 30 50 100 200 S ample S ize Figure 1. Powers for a uniform null and increasing alternative. 1 Power 0.8 0.6 KS 0.4 W^2 ^2 0.2 0 10 20 30 50 100 200 S ample S ize Figure 2. Powers for a uniform null and triangular ∨ or ’bath-tub’ type alternative. Table 2. Summary of the power of the three test statistics. Alternative Distribution General Ranking of Power from Highest to Lowest Increasing W2 > KS > 2 2 > KS W2 Triangular ∨ REFERENCES Choulakian, V., Lockhart, R.A. and Stephens, M.A. (1994). Cramér-von Mises statistics for discrete distributions. The Canadian Journal of Statistics, 22 125-137. Pearson, K. (1900). On the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen from random sampling. Philosophical Magazine, 5(50) 157-175. Pettitt, A.N. and Stephens, M.A. (1977). The Kolmogorov-Smirnov goodness-of-fit statistic with discrete and grouped data. Technometrics, 19 205-210. Steele, M.C. (2002). The power of categorical goodness-of-fit test statistics. Unpublished PhD Thesis, Griffith University.