Power and Sample Size • IF the null hypothesis H : μ = μ 0 • • • 0 is true, “Null Distribution” then we should expect a random sample mean to lie in its “acceptance region” with probability 1 – α, the “confidence level.” That is, P(Accept H0 | H0 is true) = 1 – α. Therefore, we should expect a random sample mean to lie in its “rejection region” with probability α, the “significance level.” That is, P(Reject H0 | H0 is true) = α. 1 /2 /2 H0: = 0 “Type 1 Error” Rejection Region Acceptance Region for H0 Rejection Region μ0 + zα/2 (σ / n) Power and Sample Size “Null Distribution” “Alternative Distribution” • IF the null hypothesis H : μ = μ is false, 0 0 then the “power” to correctly reject it in favor of a particular alternative HA: μ = μ1 is 1– 1 /2 /2 H0: = 0 Rejection Region Acceptance Region for H0 P(Reject H0 | H0 is false) = 1 – . Thus, P(Accept H0 | H0 is false) = . “Type 2 Error” HA: μ = μ1 Rejection Region μ0 + zα/2 (σ / n) μ1 – z (σ / n) Set them equal to each other, and solve for n… z /2 z | 1 0 | n , where 2 Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis value • HA: μ = μ1 Alternative Hypothesis specific value • significance level (or equivalently, confidence level 1 – ) • 1– power (or equivalently, Type 2 error rate ) Then the minimum required sample size is: N(0, 1) z /2 z | 1 0 | n , where 2 1 z Example: σ = 1.5 yrs, μ0 = 25.4 yrs, = .05 z.025 = 1.96 Suppose it is suspected that currently, μ1 = 26 yrs. Want 90% power of correctly rejecting H0 in favor of HA, if it is false 1 – = .90 = .10 z.10 = 1.28 = |26 – 25.4| / 1.5 = 0.4 2 1.96 + 1.28 n So… minimum sample size required is 65.61 0.4 n 66 Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis value • HA: μ = μ1 Alternative Hypothesis specific value • significance level (or equivalently, confidence level 1 – ) • 1– power (or equivalently, Type 2 error rate ) Then the minimum required sample size is: N(0, 1) z /2 z | 1 0 | n , where 2 1 z Example: σ = 1.5 yrs, μ0 = 25.4 yrs, = .05 z.025 = 1.96 Suppose it is suspected that currently, μ1 = 26 yrs. Want 95% 90% power of correctly rejecting H0 in favor of HA, if it is false 1.28 .90 = .05 .10 z.05 1 – = .95 .10 = 1.645 = |26 – 25.4| / 1.5 = 0.4 2 2 1.645 1.96 + 1.28 n 65.61 81.225 So… minimum sample size required is 0.4 0.4 n 82 66 Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis value • HA: μ = μ1 Alternative Hypothesis specific value • significance level (or equivalently, confidence level 1 – ) • 1– power (or equivalently, Type 2 error rate ) Then the minimum required sample size is: N(0, 1) z /2 z | 1 0 | n , where 2 1 Example: σ = 1.5 yrs, μ0 = 25.4 yrs, = .05 z.025 = 1.96 z Suppose it is suspected that currently, μ1 = 25.7 26 yrs. yrs. Want 95% power of correctly rejecting H0 in favor of HA, if it is false 1 – = .95 = .05 z.05 = 1.645 = |25.7 |26 – – 25.4| 25.4| / 1.5 / 1.5 == 0.4 0.2 2 1.96 + 1.645 81.225 324.9 So… minimum sample size required is n 0.4 0.2 n 325 82 Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis value • HA: μ = μ1 Alternative Hypothesis specific value • significance level (or equivalently, confidence level 1 – ) • 1– power (or equivalently, Type 2 error rate ) Then the minimum required sample size is: N(0, 1) z /2 z | 1 0 | n , where 2 Example: σ = 1.5 yrs, μ0 = 25.4 yrs, = .05 z.025 = 1.96 1 z Suppose it is suspected that currently, μ1 = 25.7 yrs. With n = 400, how much power exists to correctly reject H0 in favor of HA, if it is false? Power = 1 – = P Z z /2 n P Z 1.96 0.2 400 P Z 2.04 = 0.9793, i.e., 98% Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis • HA: μ ≠ μ0 Alternative Hypothesis (2-sided) • significance level (or equivalently, confidence level 1 – ) • n sample size From this, we obtain… s n x1, x2,…, xn But this introduces additional variability from one sample to another… PROBLEM! “standard error” s.e. (estimate) x s sample mean sample standard deviation …with which to test the null hypothesis (via CI, AR, p-value). In practice however, it is far more common that the true population standard deviation σ is unknown. So we must estimate it from the sample! Recall that s2 2 ( x x ) i n 1 SS df Given: • X ~ N(μ , σ ) Normally-distributed population random variable, with unknown mean, but known standard deviation • H0: μ = μ0 Null Hypothesis • HA: μ ≠ μ0 Alternative Hypothesis (2-sided) • significance level (or equivalently, confidence level 1 – ) • n sample size From this, we obtain… s n x1, x2,…, xn “standard error” s.e. (estimate) x s But this introduces additional variability from one sample to another… PROBLEM! sample mean sample standard deviation …with which to test the null hypothesis (via CI, AR, p-value). SOLUTION: X follows a different sampling distribution from before. … is actually a family of distributions, indexed by the degrees of freedom, labeled tdf. Z ~ N(0, 1) t10 tt3 2 t1 William S. Gossett (1876 - 1937) As the sample size n gets large, tdf converges to the standard normal distribution Z ~ N(0, 1). So the T-test is especially useful when n < 30. … is actually a family of distributions, indexed by the degrees of freedom, labeled tdf. Z ~ N(0, 1) t4 .025 William S. Gossett (1876 - 1937) 1.96 As the sample size n gets large, tdf converges to the standard normal distribution Z ~ N(0, 1). So the T-test is especially useful when n < 30. Lecture Notes Appendix… or… qt(.025, 4, lower.tail = F) [1] 2.776445 … is actually a family of distributions, indexed by the degrees of freedom, labeled tdf. Z ~ N(0, 1) t4 .025 William S. Gossett (1876 - 1937) .025 1.96 2.776 Because any t-distribution has heavier tails than the Z-distribution, it follows that for the same right-tailed area value, t-score > z-score. Given: X = Age at first birth ~ N(μ , σ ) • H0: μ = 25.4 yrs Null Hypothesis • HA: μ ≠ 25.4 yrs Alternative Hypothesis Previously… σ = 1.5 yrs, n = 400, x 25.6 yrs statistically significant at = .05 Now suppose that σ is unknown, and n < 30. Example: n = 16, x 25.9 yrs, s = 1.22 yrs • standard error (estimate) = • .025 critical value = t15, .025 s 1.22 yrs 0.305 yrs 16 n Lecture Notes Appendix… Given: X = Age at first birth ~ N(μ , σ ) • H0: μ = 25.4 yrs Null Hypothesis • HA: μ ≠ 25.4 yrs Alternative Hypothesis Previously… σ = 1.5 yrs, n = 400, x 25.6 yrs statistically significant at = .05 Now suppose that σ is unknown, and n < 30. Example: n = 16, x 25.9 yrs, s = 1.22 yrs • standard error (estimate) = s 1.22 yrs 0.305 yrs 16 n • .025 critical value = t15, .025 = 2.131 95% Confidence Interval = (25.9 – 0.65, 25.9 + 0.65) = (25.25, 26.55) yrs p-value = 2 P ( X 25.9) 2P(T15 1.639) Test Statistic: T15 25.9 - 25.4 0.305 95% margin of error = (2.131)(0.305 yrs) = 0.65 yrs Lecture Notes Appendix… Given: X = Age at first birth ~ N(μ , σ ) • H0: μ = 25.4 yrs Null Hypothesis • HA: μ ≠ 25.4 yrs Alternative Hypothesis Previously… σ = 1.5 yrs, n = 400, x 25.6 yrs statistically significant at = .05 Now suppose that σ is unknown, and n < 30. Example: n = 16, x 25.9 yrs, s = 1.22 yrs • standard error (estimate) = s 1.22 yrs 0.305 yrs 16 n • .025 critical value = t15, .025 = 2.131 95% Confidence Interval = (25.9 – 0.65, 25.9 + 0.65) = (25.25, 26.55) yrs p-value = 2 P ( X 25.9) 2 P(T15 1.639) = 2 (between .05 and .10) = between .10 and .20. 95% margin of error = (2.131)(0.305 yrs) = 0.65 yrs The 95% CI does contain the null value μ = 25.4. The p-value is between .10 and .20, i.e., > .05. (Note: The R command 2 * pt(1.639, 15, lower.tail = F) gives the exact p-value as .122.) Not statistically significant; small n gives low power! Lecture Notes Appendix A3.3… (click for details on this section) To summarize… Assuming X ~ N(, σ), σ) test H0: = 0 vs. HA: ≠ 0, at level α… If the population variance 2 is known, then use it with the Z-distribution, for any n. If the population variance 2 is unknown, then estimate it by the sample variance s 2, and use: • either T-distribution (more accurate), or the Z-distribution (easier), if n 30, • T-distribution only, if n < 30. ALSO SEE PAGE 6.1-28 Lecture Notes Appendix A3.3… (click for details on this section) To summarize… Assuming X ~ N(, σ), σ) test H0: = 0 vs. HA: ≠ 0, at level α… If the population variance 2 is known, then use it with the Z-distribution, for any n. If the population variance 2 is unknown, then estimate it by the sample variance s 2, and use: • either T-distribution (more accurate), or the Z-distribution (easier), if n 30, • T-distribution only, if n < 30. ALSO SEE PAGE 6.1-28 Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? { x1 , x2 , x3 ,…, x24 } And what do we do if it’s not, or we can’t tell? Z ~ N(0, 1) IF our data approximates a bell curve, then its quantiles should “line up” with those of N(0, 1). Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? { x1 , x2 , x3 ,…, x24 } And what do we do if it’s not, or we can’t tell? Sample quantiles Z ~ N(0, 1) IF our data approximates a bell curve, then its quantiles should “line up” with those of N(0, 1). • Q-Q plot • Normal scores plot • Normal probability plot Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? And what do we do if it’s not, or we can’t tell? IF our data approximates a bell curve, then its quantiles should “line up” with those of N(0, 1). • Q-Q plot • Normal scores plot • Normal probability plot qqnorm(mysample) (R uses a slight variation to generate quantiles…) Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? And what do we do if it’s not, or we can’t tell? IF our data approximates a bell curve, then its quantiles should “line up” with those of N(0, 1). • Q-Q plot • Normal scores plot • Normal probability plot qqnorm(mysample) (R uses a slight variation to generate quantiles…) Formal statistical tests exist; see notes. Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? And what do we do if it’s not, or we can’t tell? Use a mathematical “transformation” of the data (e.g., log, square root,…). x = rchisq(1000, 15) hist(x) y = log(x) hist(y) X is said to be “log-normal.” How do we check that this assumption is reasonable, when all we have is a sample? Assuming X ~ N(, σ) And what do we do if it’s not, or we can’t tell? Use a mathematical “transformation” of the data (e.g., log, square root,…). qqnorm(x, pch = 19, cex = .5) qqline(x) qqnorm(y, pch = 19, cex = .5) qqline(y) Assuming X ~ N(, σ) How do we check that this assumption is reasonable, when all we have is a sample? And what do we do if it’s not, or we can’t tell? Use a mathematical “transformation” of the data (e.g., log, square root,…). Use a “nonparametric test” (e.g., Sign Test, Wilcoxon Signed Rank Test). = Mann-Whitney Test • These tests make no assumptions on the underlying population distribution! • Based on “ranks” of the ordered data; tedious by hand… • Has less power than Z-test or T-test (when appropriate)… but not bad. • In R, see ?wilcox.test for details…. SEE LECTURE NOTES, PAGE 6.1-28 FOR FLOWCHART OF METHODS See… http://pages.stat.wisc.edu/~ifischer/Intro_Stat/Lecture_Notes/6__Statistical_Inference/HYPOTHESIS_TESTING_SUMMARY.pdf