13. Consider simple versus simple testing from a Bayesian perspective. Let e have a Bernoulli distribution with P(B == 1) == p and P(B == 0) == 1 - p. Given == 0, X will have density p 0 , and given == 1, X will have densit y e e PI· a) Show that the chance of accepting t he wrong hypothesis in t he Bayesian model using a test function cp is R(cp) = E [I{B = O}cp(X) + I{B = 1}(1 - cp(X))]. b) Use smoothing to relate R(cp) to E 0 cp = E [cp(X) e = OJ and E 1cp = E[cp(X) B=l J. 1 c) Derive t he test function cp* minimizing R( cp) . Show t h at cp* is a likelihood ratio test , identifying the critical value k. 12.8 Problems 18 Let F be a cumulative distribution function t hat is continuous and strictly increasing on [O, CX)) with F(O) == 0, and let qa denote t he upper a quantile for F, so F( qa ) == 1 - a. Suppose we have a single observation X with P0(X < x) == F(x/0) , x E IR, 0 > 0. a) Consider testing Ho : 0 < 0o versus H 1 : 0 > 0o. Derive t he significance level for the test cp == l(c,(X) )· What choice for c will give a specified level a? b) Let cpa denote the test with level a in part Ca . Show t hat t he test s cpa , a E (0 , 1) , are nested in t he sense described in Problem 12.17, and give a formula to compute the p-value P(X). 2 12.8 Problems 30. Let X and Y be independent with X N(µy , 1). Take 11µ11 2 = µ'; + µ~ , and consider testing Ho : µx = µy = 0 versus H1: llµII > 0. For rotational symmetry, a test based on T == X 2 + Y 2 may seem natural. The density of T is rv N(µ x, 1) and Y rv ½Io(v'tllµll)exp{-½(t+ 11µ11 2 ) } , t > O· ' otherwise, 0, where I 0 is a modified Bessel function given by ex cos w dw. 3 12.8 Problems a) Show that Io(x) > Ib(x) and that xI6'(x) + I6(x) == xlo(x). b) Show that xI6(x)/I0 (x) is increasing in x. Use this to show t hat for c > 1, I 0 (cx)/ I 0 (x) is an increasing function of x. Hint: log Jo (ex) - log JO (X) 4 = c l 8 log Io (ux) d a u U· 12.8 Problems c) Show that the densities !11µ11 have monotone likelihood ratios. d) Derive the uniformly most powerful level a test of Ho versus H 1 based on T. e) Derive a level a test of Ho versus H 1 based on X and Y that has power as high as possible if µ x == µy == l. Is this the same test as t he test in part (d ? f) Derive a level a test of Ho : µ x == ex, µy == cy , versus H1 : µ x # Cx 2 2 or µy # cy, based on T == (X - ex) + (Y- cy) . g) Derive a 1 - a confidence region for (µ x, µy) r dual to the family of tests in part (fl . What is the shape of your confidence region? ~ 5 12.8 Problems 4 7. A random angle X has density Pe ( x ) == exp[0 cos x] ( ) , 21rio 0 x E [0, 21r) ; 0, otherwise, where 0 E IR and I 0 is a modified Bessel function (I0 (0) == 1). Derive t he uniformly most powerful unbiased test of Ho : 0 == 0 versus H 1 : 0 -# 0 with level a. 6 12.8 P roblems 49. Because a good test of Ho : 0 E n0 versus H 1 : 0 E n 1 should have high power on n 1 and small power on n0 , a test function ¢ might be chosen to minimize /3¢(0)w( 0) dA( 0) Do + (1 - (3cp (0))w(0) dA(0) , D1 where A is a measure on n == n0 U n 1 and w > 0 is a weight function. (With a natural loss structure, Bayes risks would have this form.) a) Derive a test function ¢ * t hat minimizes this criterion. Assume t hat Pis a dominated family wit h densities Pe, 0 E b) Derive the optimal test function¢* explicitly if w is identically one, A is Lebesgue measure on (0, oo) , Pe is the exponential distribution wit h failure rate 0, n0 == (0, 1] , and n1 == (1 , oo). 12.8 P r o blems n. 7