This preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: § 5 Hypothesis Testing § 5.1 Introduction 5.1.1 A parametric hypothesis H is an assertion about the unknown parameter θ in a parametric model. Example § 5.1.1 (i) Man’s height ∼ N ( μ M ,σ 2 ) Woman’s height ∼ N ( μ W ,σ 2 ) Parameter θ = ( μ M ,μ W ,σ ) H : μ M > μ W (i.e. men are on average taller than women); (ii) P (new medical treatment cures disease) = p P (old medical treatment cures disease) = q θ = ( p,q ) H : p > q (i.e. new treatment is more effective); (iii) stock market undergoes daily change X ∼ N (0 ,σ 2 ) when a political event occurs θ = σ H : σ > 1 (i.e. political event destabilizes the stock market); (iv) A chemical fertilizer is cheap but is potentially harmful to consumers. Outcome of a safety test = X ∼ N ( θ, 1) Acceptable safety standard requires θ < 1 H : θ < 1 (i.e. chemical is safe). 5.1.2 Formulation of hypotheses is subjective, depends on the practical situation as well as the attitude of the person who hypothesizes. 5.1.3 A statistical test of H is a databased procedure for deciding whether the observed data contain evidence against H . 5.1.4 Sample data are random and may be misleading. Conclusions from hypothesis tests may not always produce the correct answer. Good tests have a bigger chance to give correct answers. 45 § 5.2 Hypotheses 5.2.1 Definition. X (data) ∼ p (probability function) A hypothesis test involves two hypotheses about p : 1. null hypothesis H — • conservative, i.e. not to be rejected unless evidence is strong, • usually takes the form of “no difference”, 2. alternative hypothesis H 1 — • specifies the kind of departure from H of interest to the tester, • preferred to H only if such departure is detected with strong evidence. Based on X , the test helps us assess the “relative” plausibility of H and H 1 . 5.2.2 Example § 5.1.1 (cont’d) (i) H : μ M = μ W (no difference) vs H 1 : μ M > μ W (men being taller). However, if one is only interested in whether men and women have different heights on average, the appropriate hypotheses should be H : μ M = μ W vs H 1 : μ M 6 = μ W . (ii) To protect patients, one might consider H : p = q (no difference) vs H 1 : p > q (new treatment more effective) , i.e. we would not recommend use of new treatment unless strong evidence exists in support of its effectiveness. (iii) H : σ ≤ 1 (small destabilizing effect) vs H 1 : σ > 1 (big destabilizing effect); (iv) To protect the farmer, one may wish to test H : θ < 1 vs H 1 : θ ≥ 1 , so that the cheap fertilizer will be abandoned only if there is strong evidence against its safety. To protect consumers, one may do the opposite, i.e. test H : θ ≥ 1 vs H 1 : θ < 1 , so that the cheap fertilizer will NOT be adopted unless there is strong evidence confirming its safety....
View
Full Document
 Spring '10
 SMSLee
 Statistics, Null hypothesis, Statistical hypothesis testing, H0, likelihood ratio test, λx

Click to edit the document details