Two samples: (in)dependent and (non)parametric. Two independent samples, normal model Comparing the means or variances of two (sub)populations. We assume we have: Two independent random samples (sizes m and n) from normally distributed population: X1,..., Xm are independent and Xi ~N(µX, ), Y1,..., Yn are independent and Yj ~N(µY, ). The sample means and sample variances are denoted by: , resp. The observed values as . are unbiased estimators of : , = )= . are unbiased estimators of µX and µY , so µX - µY 1 are independent so = + So ~ N( + ) or: Based on this variable and its distribution we can derive a confidence interval for and a test statistic for a specific value of . But this is only usable if are known. Usually this is not the case. If we use as estimators of and , then we apply the Student’s t-distribution. We distinguish three cases (assumptions): 2 1. and are unknown, but equal: is the best estimator of the common variance ( called the pooled sample variance): is Similar to the one-sample-problem we derive: (1- α)100%-CI( where and c from the t(m+n-2)-table such that P(Tm+n-2 ≥ c) = . 3 The test statistic for the test on H1 can be one-sided: is or or two-sided: 2. and are unknown and different: 4 Test statistic: Where df = number the degrees of freedom: df = between min(m-1, n-1) and m + n – 2 Computer software, like SPSS, computes df without software we use: df = min(m-1, n-1) (1- α)100%-CI( = Where df = min(m-1, n-1) and c such that P(Tdf ≥ c) = α/2 Note: df = min(m-1, n-1) is a “safe” estimate of the real number. That is why this is called “the conservative method”. 3. Large m and n (m > 30 and n > 30): Use the N(0, 1)-distribution as an 5 approximation, even if the populations are not normally distributed (robustness!) Testing the equality of expectations for m > 30 and n > 30: if And (also for large m and n): (1- α)100%-CI( = where c is such that P( Z ≥ c) = α/2 Testing the equality of variances (Levene’s test for the proportion We test H0: H1 can be = > , H0: < or / ): =1 . 6 Test statistic is the proportion of the sample variances which has a so called Fisher distribution with m -1 degrees of freedom in the numerator and n -1 degrees of freedom in the denominator. Notation: Note 1: If H0 is true T will have values near 1. Note 2: The Fisher table contains values c, such that =α c in this equation is the critical value for the right sided alternative H1: > The critical value c for the left sided alternative H1: < can be found using: α 7 Two sided test for H0: , H1: If T ≤ c1 or T ≥ c2 , then reject H0 , where: and When testing the equality of expectations using SPSS always first check the p-value of Levene’s test (α = 0.05) to choose between equal variances assumed and not assumed. Two pairwise dependent random samples: the t-test for the differences. Whenever we have two observation per object (individual in the sample), we can compute 8 the differences and apply the t-procedure for the differences. Usually this is the case when we have observations per object i before (xi) and after treatment (yi). Statistical assumptions for the (dependent) random samples x1,...,xn and y1,...,yn: The differences Zi = Yi - Xi (i =1,...,n) are independent and Zi ~ N(µ, σ2), where expected difference µ and variance σ2 are unknown. t-test on H0: µ = 0 (expected difference is 0): Confidence interval for the expected difference (c from the t(n-1)-table): The relation between two-sided tests and confidence intervals: 9 Suppose θ is the population parameter and we have a (1-α)100%-CI(θ), based on a sample, then H0: θ = θ0 will be rejected in favour of H1: θ θ0 at significance level α if θ0 is in the interval. e.g. A one sample t-test: 95%-CI(µ) = (22.4, 28.1) => reject H0: µ = 30 versus H1: µ 30 at 5%-level (because 30 is not in the interval). A binomial test: 90%-CI(p) = (0.42, 0.51) => Do not reject H0: p = ½ versus H1: p ½ at 10%-level. A two sample t-test: 99%-CI( ) = (-6.8, -1.2), then: reject in favour of at 1%-level, because the difference 0 is not in the interval. 10 Non-parametric or distribution-free tests are used in cases where the assumption of a distribution (having unknown parameters) does not apply. These tests offer for example an alternative method if the assumption of a normal distribution is evidently incorrect and the number of observations is small: this can be the case if the histogram is skewed to the right or the left or when there are outliers. SPSS also provides tests on normality. 1. The Sign Test as an alternative for the ttest on the differences for two pairwise dependent random samples. Statistical assumptions: the differences Zi have an unknown (notnormal) distribution. p = the probability of a positive difference. We test H0: p = ½ versus H1: p > ½ (or p < ½ or p ½) 11 Note: for symmetric distributions of the differences p = ½ is equivalent to µ = 0, but for skewed distributions p = ½ is equivalent to Median = 0. That’s why this test is also called the sign test on the median. The test statistic is X = the number of positive differences: X ~ B(n, ½) if H0: p = ½ is true. n = the number of observed non-zero differences (just cancel the zero differences) The rejection region is: H1 Rejection region p>½ {c, c +1, ....., n} p<½ {0,...., c} p ½ {0,...., c1} {c2, c2+1, ....., n} Determine the critical value(s), using: - For n ≤ 25 the binomial table if available. - Normal approximation is valid for n > 10: X is approximately N(½n, ¼n)-distributed. 12 2. The Wilcoxon Rank Sum Test, as a nonparametric alternative for the two (independent) samples t-test, on the equality of the expectations . The only statistical assumption we need is that we have two independent random samples of numerical variables X1,..., Xm and Y1,..., Yn. We will test H0: the population distributions are the same versus H1: the population distribution of Y is shifted compared to the distribution of X. Example: shift to the left H1 : 13 The test statistic, Wilcoxon’s W, is defined as: W = sum of the ranks of the X-observations if we order all X- and Y-observations. One sided alternatives: Y is shifted to the left H1: If the alternative is true, the X-values will be large and W is large: W ≥ c => reject H0. Y is shifted to the right H1: If the alternative is true, the X-values will be small and W is small: W ≤ c => reject H0. Two sided alternative: Y is shifted to either side H1: If the alternative is true, the X-values will be large or small and W is small or large: if W ≤ c1 or W ≥ c2, then reject H0. The distribution of W We will use a normal approximation of the 14 (discrete) distribution of W and apply continuity correction if m > 5 and n > 5. This is an approximate distribution (N=m+n). The Wilcoxon rank sum test using 1. the critical value c (right sided test): If W ≥ c , then reject . And c such that: P(W ≥ c)= P(W ≥ c+½ ) ≈ P(Z ≥ 2. the p-value p-value ≤ α => reject P(W ≥ w)= P(W ≥ w+½ ) ≈ P(Z ≥ )=α : ) Note: for m ≤ 5 and/or n ≤ 5 the exact distribution of W is computed, using combinatorics (not part of this course). Ties: if two or more observations have the same value, then we call this group of observations a tie. All observations of a tie are 15 assigned the same rank: the mean rank of these observations. We can use the distribution above, if there are just a few ties. 16