On the Serfling-Wackerly Alternate Approach by I~yr:~nd J. Carro11* Department of Statistios university of North Carolina Chape l Hi Zl.J North Caro Una 2?514 Institute of Statistics fIi.meo Series Ho. 1052 February~ * This 1976 research was supported. by the: Air Force Office of Scientific Research under Grant no. AFOSR-75-2796. On the Serfl ing-!<!ackerly Alternate· Approach- by P4ymond J. Carroll* Department of Statistics university of North CaroZina at ChapeZ RiZZ ChapeZ RiZZ 3 North CaroZina 27514 .e.BSTRACT For the problem of fixed-length interval estimation of the mean and median, the limiting distributions of the stopping rules recently introduced by Serfling and \vackerly [8J are compared to those of Chow and Robbins [4] as both the non-coverage probability a and interval length d converge to zero. The procedures based on the sample mean differ in the centering constants, while those based on the sample median also differ in the scaling constants. Institute of Statistics filimeo Series #1052. *This research was supported by the Air Force Office of Scientific Research under Grant No. AFOSR-75-2796. 1 1. Introduction The primary purpose of this note is to illustrate the second order differences (where they exist) betl'.reen two cO:r1peting approaches to the problem of fixed length confidence :L'1terva1 estimation of the center of [4J~ syrrnnetry: one due to Chow and Robbins (C-R) by Serfling and Wacker1y (S-V'V) [8]. The intervals of prescribed coverage probability 1 - 2a are of the fom differ; C-R fix a and let d ~ the other recently introduced (~-d, Xn+d) but the stopping rules 0, while S-W fix d and let a show that the C-R approach with stopping time N(a?d) for their problem in the sense that if m(a,d) observations with F c(d) =1 c(d) ~ if F 1 as d as both a,d (1) ~ = ~, ~ O. known~ N(a,d)/m(a,d) ~ ~ o. S-W is lIinappropriate i l is the required number of c(d) ;t 1 as a the normal distribution function. ~ 0, although They also show that 0, "mich suggests that the two approaches are equivalent If the S-W stopping rule is denoted by M(a,d) , N(a,d)/i/l(a,d) ~ 1 (a.s.) as a ~ 0, then d -+- o. TIle consistency property (1) says that the rules approximate the same quantity asymptotically, but they still may be different in terms of a limiting distribution. The asymptotic distribution of N(a,d) is easily found (see [6]), and in Section 3 and the distribution of M(a,d) Contrary to (1), this (2) (a) ~~pendix the asymptotic is obtained, a reslut of some interest in itself. pa~er f:L~ds t~at N(a,d) and I;I(a~d) have limiting distributions with the same scaling constants as a,d ~ 0; for large classes of a,d the C-R(S-W) alJIlr6ach has smaller centering constCl.:"1.ts if thepopulation kurtosis exceeds (i5.,less than) 3. 2 In the normal case with d fixed, the S-W approach is preferable as a ~ 0 if d/o < 2.33 . (b) 111e asymptotic distributions of analogous rules based on the sample median ([5], [8]) are also obtained. The differences are more pronounced here; both the scali.l1g and centering constants are different. As an example of the results, Serfling and Wackerly's modification of Geertsema's rule [5] is found to be infinitely nore variable than their own as a 2. ~ o. The Two f'lethods Let Xl'X2 , •.• be i.i.d. observations from a continuous symmetric distribution F(x-e) with unique median 8 and with variance a2 • Let ~(Aa) ChQ'tlI and = I-a. 111e two approaches are: n - 2 and define Let sn2 = n -1 I (Xi-X n) i=l Robbins (C-R) [4]. N(a,d) = inf{n: n ;?: n'(a,d) = inf{n: n ;?: Serfling and Hackerly (S-W) [8]. lCA) Let (A S /d)2} a n 2 (Aao/d) }. = n- l ~(x) i-'i(a,d) = inf{n: n ;?: ;?: As further notation, let . 1 1= Then tn a/tn ll'h(-d) a.Tld m(a,d) = inf{n: Pr{IYn-el I I{X1·-X :S x}, where n Further, define denotes the indicator of the event A. Inn (-d) = inf exp(dt) ;"[ exp(tx)dHn ex). t:SO -00 n Xi-~+d < 0 for some i:sn} d} :S a}. 110 (t) = f exp(tx)dF(x) be the point at which WoCt) = exp(dt)l:1o (t) since F has a unique median at zero, and let to Cd) achieves its infimum. Then, 3 n = n- l i~l exp(t(Xi-5\~+d)), Define Wn(t) to W~(t) = 0, so that for n = inf IDo(Z) large, exp(-zt) t<O and let tn(d) be the solution = Wn(tn(d)). mn(-d) Finally, let I exp(tx)dF(x). 3.. Sample Means The asymptotic distributions of the nIles N(a,d) and H(a,d) are obtained here. The scaling constants are the same but the centering constants are not. For many sequences a,d -+ 0, the value of the population kurtosis detel1nines the smaller centering'constants. I xdF(x) I = 0, 2 x dF(x) natural logarithm. = 1. Let Let A(a,d) log x both denote the be the variallce of w(X) , where = exp{to(d) (y+d)} $(y) R-n x and In this section', assume - yto(d)E exp{to(d) (X+d)}. The proof of the following Lemna is given in the appendix. idea is to represent log mn(-d)/mo(-d) as a sum of i.i.d. random variables with remainder tern and use the techniques of [6]. the case d Lemma 1. Let If, for same fixed. !:J. = -R-n moC -d). £ > 0, as a,d -+ Then, as ° The basic a -+ 0, Note the result includes 4 then (4b) Lemma 2. ([6]). As a 1d + ° or as a + 0, (5) (see Proposition Now~ 1), so that the centering constants A~/d2 -9.na/t1 in (4) may be replaced by (and the conclusion that the asymptotic distributions are equivalent can be made) if, as a,d + 0, Note that if dAa _ A~ for some £ > o? then since A~/(-Z9.na) (4a) holds but because of Proposition S? H(a?d) + + -00 ° +00 + 1, if EX4 > 3 if EX4 = 3 if f:.X 4 < 3, the result being that the C-R approach has slTl.aller centering constants 4 (and is thus in a sense "prefcrab1ell .but lIinapp~o?tiat~") if, EX >3, giving (Za). Lemrra 3. ci (7a) (7b) St~pose X1 ?XZ' .•• are nOrITally distributed with variance and let n = d/o be fixed. As a + 0, 5 where 2 2 Note that (Jcr!0ST,t1"" 1 as n"" 0 as is predicted by Lemmas 1 and 2? 2 2 but that (Jcp!(JS-iJ"" 0 as n"" 00. In Table 1? the values of this ratio are presented for various values of n. It appears that for 0 < 2 2 (J~if < (JCR while the opposite is true if 2.34 < n < 00. n < 2.33? TABLE 1 2 ,., Values of the quantity (JCR/(J~. n 0.00 0.05 0.25 0.50 1.00 1.50 2.00 2.25 2.33 .. 2.50 3.00 3.50 4. 2 2 (Jm/as:!] 1.00 1.00 1.01 1.04 1.15 1. 25 1.19 1.06 1.00 .87 .45 .16 Sample f'1edian. Serfling and 11ackerly [8] also provide a rule M(a?d) the median. C'J€ertsema [5] has defined an analogue N*(a?d) based on to the C-R approach? while S-W provide a modification of this? call it N(a,d). The results of this section are not as satisfactory technically as those of the previous one? but they do shed insight into the behavior of the three rules. 6 Letting "n.i Y denote the ith order statistic in a sample of size n, define t.." bn = max{l, [n/2 - cn 2 /Z]}, an • 1 Zn(c) = n"2(~,a - ~,b )/c n n Z N(a,d) = first tnne n that Zn (c)/n = n-bn+l Z Z 4d /Aa . N*(a,d) = first time n that ZZ(A )/n ~ 4d Z/AaZ • n a Define M(a,d) as in [8]. ~ = -~ If Tn ~ is the sample median, set 2 109(4(F(d) - F (d))) ~ = -~ 109(4(Fn (Tn+d) - F~(Tn+d))). LelTil!la 4. Asstmle that Zn is unifonnly continuous in probability (see [1]). For the S-W adaptation of Geertsema's rule, as a,d For Geertserna's rule as d Lemma 5. B(F,d) + + 0, 0, Define -2 t = l - 2F (d)) 2 (ZF(d) (I-F(d))) {F(d) (l-F(d))-f(d) (l-F(d))/f(O) " +(f(d)/Zf(O))Z}. Then, as a (9) + 0, 7 Lemma 6. If (j~R denotes the asymptotic variance denotes the asymptotic variance in (9), then as A few points need to be emphasized. is tmifonnly continuous in probability. d as -+ 0, while (9) holds only as a -+ a. -+ O. in (3b) and a,d -+ 07 (j~w a~R/a~'1 -+ First, we do not know if ~. Zn Second, (Bb) holds only as The difficulty with (8b) is that, 0, the necessary appeal to the representation theorem [2] cannot be justified. The difficulty with (9) is that, as d -+ 0, n -+ 00, the asymptotic distribution of the process (1 - 2F(d)) (Yn(t+d) - f(d)Yn(O)/f(O)) (Yn(t) = n~(Fn(t)-F(t)) is tmlmO\vn. TIlese are interesting tedmica1 problems and more work is needed. In surrrnary, LeITIlT'as 4 - 6 make it likely that more variable and N*(a.,d) N(a,d) is infinitely is less variable than H(a,d), although more technical work is needed to confinn these impressions. Note again the difference in the centering constants. It is worthwllile to mention that these results extend readily to the ranking problem ([9]). ERR AT ASH E E T page 8 a) The statement of Proposition 1 should be Proposition 1. Unifonnlyas d -+ 09 n -+ 00, page 9 b) The three lines immediately above Proposition 2 should be deleted. 8 APPENDIX The proof of Lemma 1 is long and tedious~ but the basic idea is as stated in Section 3. lle first begin with a lllJrrlber of propositions. Let 0, a be the customary "big oh, little oh". UnifoITIly as d Proposition 1. n -1 tl tn (d) - t o (d) = n. 1 + 0~ n + oo~ ( ) - -X + O(n -1 10gzn) to(d)X. lIO n (X.+d)e~) 1= Proof: It is a simple calculation to show that tn(d) d + 0, n + + As in Lemma 3. Z of [8] ~ we see that for 00. (a.s.). 0 (a.s.) as n sufficiently large I (x.-x +d)eA~{t 1 n = n- B* ~l = n -1 '1 n Now~ n 1 A* n A* = A(l) On n A(l) = n- 1 °h t '1 1= - +d) 2exp{y (d) (X. -x -+d) }• (X. -X In 1= A(2) On + (d) (lG-X +d)} n o·~ + A(3) n n In where ~ n X. eA~{t (d) (X.-X +d)} 1 0 1 n (d)(d-Y~) -1 n L i=1 =e O n L . 1 {X. exp(to(d)X.) - EXexp(to(d)X)} 1 1= 1 + EXexp(to (d) (X+d)) {exp(-to (d)~) -l} + EXexp (to (d) (X+d)) . A(Z) .n n = _n- 1 On x . L1 exp{t (d) (X.-X +d)} 1= t = -~ e 0 (d)(d-~) 0 n- 1 1 n .L 1=1 n {exp(to(d)Xi ) - Eexp(to(d)X)} - Xu Eexp{to(d) (X+d)}{exp(-t o (d)Xn)-1} - Xn Eexp{to(d) (X+d)}. 9 1 11 = On-\'L i=l = de exn{ t 0 (d) (X-1 -::rn +d)} .~ to(d)(d-~1) n + -1 n 2 - 1 1= {exp (t( (d) X-) - EeXf.l (to (d) x) } ) 1 (lJjeJ~{to (d)(X+d)} (exp (-to (d)X".) -1) =0 Since EXexp(to(d)(X+d)) + dEe}9(to(u)(~C+d)) -}~Eexp{to(d)(X+d-~)} this r:;ea.'1S A* = n n 1 o and 1 + O(dn-~2Clog2n)-'2) = -1\t + dEexp{t (cl)(X+d)}. (a.s.L that -1 n I - 1 )..= (X1 +c1)exp(t o (d)X1-) o - x.. . ," (-~2 1) + J dn (loezn) 2 (a.s.) ~ H ana hence that Now ~ by carrying out t:le Taylor expansion (10) one nore 1]laoo as in Lerrrna and going t;1rough the above steps once ElOre ~ cne completes the proof. Proaf: The key to the expansion of to Cd) lies in the relation a = B(X+d)exp(to(d)(X+G.)). 10 By using the facts EX = EX3 = 0, EX Z = 1, one will obtain to (d) by progressively adding terns to the Taylor eX'pansion of exp(to(d)(X+d)). moe-d) = l'.lo (to (d))exp(dto (d)L where Now, 110 (t (d)) = 1 + t 2 (d)EX2/2 + t 04 (d)EX4/24 + o(d4). 00 Thus, 6 = -~ __4 4) m0 (-d) = -to(d)d - (t 2 (d)/2 + t 4 (d)EX'/Z4 - t 04 (d)/8 + o(d) , 0 0 so that using the expansion for to (d) Proposition 3. Uniformly as d fine-d) - moe-d) = n- 1 n L 0, -+ yields the result. n -+ <Xl, {~(Xo) - E~(X)} + i=l -1 O(n logzn) (a.s.) 1 and m (-d) n 10.g mn ( -d) -- (dZ/m (-d))n- l L o . 1 1= o {(X~-1)(1+d2(1-EX4/3))/4 + d(X1o+X31o/3)/Z 1 + d2(X{-EX4)/24} + o(d4) + O(n- l log 2n) As a -+ 0, log(m (-d)/m (-d)) = nn 0 Proof: (a.s.). By 1 n L1 {~(X1o)-E1HX)}/mo(-d) + O(n -1 10g2n ) (a.s.). 1= 0 a Taylor expansion_ exp(tn(d)Xi ) = exp(to(d)Xi ) + (tn(d)-to(d))Xi exp(to(d)X i ) + (tn(d)-to(d))2x~ exp(Zi(d))/Z, 11 (11) exp{ -tn (d) (d-~)}Inn (-d) = {n- I I . 1 1= {eA~(t0 (d)X.) 1 1 - E exp(to(d)X)} + E exp(to(d)X) n i~l {Xi exp(to(d)Xi ) - EX exp(tO(d)X)} + (tn (d)-t O(d))n- + (tn(dJ-to(d) EX exp(tO(d)X) 1 + O(n- I (10 g 2n))} n = n- . L1 {exp(tO(d)X.) - E exp(tO(d)X)} 1 1= - d(tn(d)-to(d))E exp(to(d)X) + + E exp(t0 (d)X) O(n- 11og2n) (a.s.). NeVI, exp{tn(d)(d-~)} = exp{dto(d)} {d(tn(d)-toCd))-tn(d)~}exp(dto(d)) + + 001-11og2n) (a.s.). }/fuUtip1ying the two sides of (11) by this last expression yields the expansion for Inn(-d). Thus, (m (-d) -TIl (-d)) exp{ -dto Cd)} . o n = n- 1 n .r {exp(to(d)Xi ) 1=1 + 0 (n -1 - E exp(toCd)X) - toCd)Xi E exp(toCd)X)} log 2n). We have 4 4 (d ) L (to (d)X. ) K/K! + 0 M0 (t ) 0 = E exp ( t (d)X) = 1 0 + d2/2 exp (t (d) X.) o 1 = K=O 1 + d4 (~ - EX4 /8) 12 This leads to (~(-d)-mo(-d))exp{-dto(d)} = d2n- 1 I {(X~-l) (1+.(1-EX /3)d )/4 + d(X.+X~/3)/2 + d2(X~-EX4)/24} i=l 4 2 1 + O(n -1 1 log zn) + 1 1 oed4) and completes the proof. If A(a 1 d) Proposition 4. ~ I (1JJ(x) - J1JJ(y)dF(y))2dF (XL then A(a,d) = d4 (EX4-1)/4 + o(d4) as d ~ O. A(a?d) = Var(1JJ(X)) = exp(2to (d)d) (r'1o (2to (d)) - r'l~(toCd)) + t~Cd)I'1~(toCd)) + NO\'J, simple Taylor expansions show that f1 ( t 0 2/2 + d4 (1-EC4/4)/2 + o(d4) Cd)) = 1 + d o p~(toCd)) = I + 2 4 d + d CS-EX4)/4 + oCd4) r1o (2to (d)) = 1 + 2d Z + Zd4 + o(d4) , which? l'lith a few computations, yields the proposition. Proposition 5. . For all £ > 0, Z (Aa - C-2~na))/A~a ~ 0 as a ~ O. 2dtoCd)Il~(toCd))). 13 Proof: Aa It is well known that = _~-l(a)~ Thus, A2/(-2~na) a -+ 1 as ex +,0. by L'Hospital's rule applied to A -£ (H(a)) a. -+ Proof of Lemrna 1. if 1 £ Since A;/(-2~na), > 0, so that By the nature of the stopping rule, M(a. ,d) stops the first time n that It is possible, using ted1niques in [6], page 298, to neglect the excess ten1S so that as a Since 1. -+ 0, from Proposition 3, M(a,d)/(-~na/ll) -+ 1 Letting N (a.s.), this completes the first part of Lemma = H(a.,d), note that (1 so that as a,d -+ - b. -1 log (TIt+ 1 (-d) /mo (-d) ))-1 0, M(a,d)(-~na/b.)-l -+ 1 (a.s.) if (12) Once (12) is shown, the proof will be completed by once again using the 14 technique in (6] together with the expansion of log (llh(-d)/mO(-d)) 2 as well as Proposition 4. Now 6. = O(d ) and log (n71 (-d)/m (-d)) = o -1 o(M 10g 2r,'D, so that it suffices to show d- (13) 2 ,-11 . - -+ 0 11 ogl'; ( a. s. ) . A sufficient condition for (13) to hold is that for some 7,,,-1+£ d - "1, a: ~ d Now, as 0, -+ 0 > 0 ( a.s .). 1,1 (a, d) ~ - tna , so that d-7_ -M(a.,d) -1+£ Proof of Lemma 3. -+ £ ~ d-2 (-Qna) -1+£ By Lemma 1, since 6. = n 2/2 -+ O. and (-2tna)/A~ -+ 1, we have 2 = 4A(a,d)/n 6 where crS1J~ where cr~ = (EX4 -1)/d 2 = 2/n 2 . In the normal case A(a:,d) = 1 - c- n 2 (-2tna - - n 2e -n A~)/Aa: 2 -+ Proof of LeL'Ina 4. L m ...., cr[2/2, as n -+ . By Lerrnna 2, ,cornp~.eting the proof because, by Proposition 5, o. In the notation of Block and Gastwirth [3] with 2 2 00 Zn -+ E, (a.s.) by [2] and IS ~ = 1/£(0). Now, N(a,d)/(~/2ka)2 where Neglecting excess tenns, N(a,d)::::: 1 (a.s.), where + ka = d/Aa . ~(a,d/(2ka)2 so that L Since m'" cN(a,d)'2/d ,.. ct;/4k ' the ill1ifonn continuity in probability a guarantees that as a,d + 0, GeertseJTl.a I S nile basically replaces the above, as d Proof of Lerrrna s. + c by It ; by steps similar to a 0, By once again neglecting excess over the boundary, I:!(a,d) - (-,Q,na/L'.) ::::: '" (-.~na)(L'.L'.r~ -1 '" (L'.-~. Then, by expansions of log (l+x) and F(Tn+d) - F(d), 1 '" n~(L'.-L'.n) =~ 1 ( ~ log 1 + {F (T +d)-F(d)}{l-F (T +d)-F(d)}j n n 2n n F(d)-F Cd) = (1-2F(d)) {2F(d) (l-F(d))}-l r/i(Fn(Tn+d)-F(Tn+d)+Tnf(d)) By the representation theorem [2], if Yn(t) + 0(1) (a.s.). 1 = ~(Fn(t)-F(t)), then ri2(L'.-~n) = (1-2F(d)) {2F(d) (l-F(d))}-l(yn (Tn+d)-f(d)Yn(O)/f(O)) + 0(1) Now, M(a,d)/(-,Q,na/L'.) process Y n + 1 (a.s.); denoting II = M(a,d), (a.s.). since the empirical is weakly convergent tmder ra."'ldom sample sizes [7] and 16 T~1 + 0 (a.s.) ~ this yields so that Proof of Lenna 6. The ratio of the variances is O~R A~/ 2d3(f(O))~ 0sw AaB(F,d)/2~ --r:: Since F(d):: ~ + df(D) + d2fi(O) + d3~i(0) + o (d 3) , we have Also, (1-2F(d))2 :: 4d2f 2(O) 2 + o(d ) F(d)(l-F(d)) = ~ - d2f 2(O) 2 + o(d ). Two more Taylor expansions yield fed) (I-F(d))/f(O) (f(d)/2f(0))2 = =~ + d(~~~~~ !<{l + 2d(f' (O)/f(O)) so that Hence, as d + - f(O)) + A~ 2 I 0SV1 2 oCR + ~ ';<;. 2 d2(~~t~~- + d (2fcW + - 2f' (0)) (~(~~))2) + 2 o(d ) 2 + a (d )}, 17 ACKNOWLEDC.afENf of The author wishes to thank Professor Wacker1y for sending preprints [8J~ [9J. REFERENCES [lJ A.11.scombe~ F.J.~ (2] Bahadur, R.R':I A note on qua,'1tiles in large samples, Annals of Mathematical Statistics 3?~ (~Jne 1966), 577-80. Large sample theory of sequential estimation, Proceedings of the CaJnbridge Philosophical Society, 48, (October 1952), 600-7. J [3] Block, D.A. and Gastwirth, J .L':I On a simple estim.ate of the reciprocal of the density fmction, Annals of Mathematical Statistics, 39, (June 1968)~ 1083-85. [4] ChOW, Y.S. and Robbins~ H., On the asymptotic theory of fixed-width confidence intervals for the mean, Annals of Mathematical Statistics, 36, (April 1965) ~ 457-62. [5] Geertserna, J.e., Sequential confidence intervals based on rank tests, Annals of Mathematical Statistics, 41, (June 1970)~ 1016-26. [6] Gut, A':I On the Doments aTld linit distributions of some first passage tL~es, Annals of Probability, 2 (April 1974), 277-308. J [7] Pyke~ R., The weak convergence of the empirical process with raTldom sample size~ Proceedings of the Cambridge Philosophical Society, 64, (January 1968)~ 155-60. [8] Serfling, R.J. and VlackerlY:l D.D., Asymptotic theory of sequential fixed-width confidence iIltervals for location parameters~ Florida State Thliversity Technical Report I!B16, September, 1974. [9] Wacker1y, D. D., An alternative sequential approach to the problem of selecting the best of k populations, University of Florida Technical Report No. 91 (February, 1975).
© Copyright 2025 Paperzz