© Joma the Aman Sit Anoation Measures of Association for Cross Classifications, IV: Simplification of Asymptotic Variances LEO A, GOODMAN and WILLIAM H. KRUSKAL* The oxymoteie tompting theory snus in our 1963 ert [3] for meaner of oneciton preseted In vores erties [1,2] fr on the derivation of ‘xympllie varoner that may be complex ond fadovsn pace coves. Inte reenter, we snplity ond unify theze derivation by expeling the expres Hon of mecnrer of anocain oration Comment onthe uit of onymptti ‘aoe, enon opin th aetion ore ale given 1, INTRODUCTION AND SUMMARY In our 1963 article [3], we discussed asymptotic sam- pling theory for some of the measures of association presented in our earlier articles [1, 2]. Itwould have been impractical to present in [3] asymptotic results for many measures under many sampling methods, so we gave re~ sults only for some of the more important eases, together with general methods so that others might more readily do their own asymptotics, In the present article, we present « more unified way to derive asymptotic variances (which form the nub of the asymptotic theory) for the following two sampling methods: 4. Multinomial sampling over the entire two-way ross classifeation; ', Independent” multinomial sampling in the rows when total row proportions are knowns that is stratified sam- pling in the roms. ‘We then apply this more unified view to several measures of association and obtain fresh formulas for asymptotic variances, Further to illustrate the method, we rederive & few of the asymptotic variances in [3]. We also take this ‘opportunity to correct (4.4.3) of [3]. Use of the asymp- totic variances in practice is also discussed, and finally wwe give a cautionary note about a trap when working out asymptotic variances. ‘The basic notion of this article isto exploit the expres- sion of most measures of association as ratios, and to do a portion of the manipulations towards asymptotie vari- ances in advanee, 4 Lae A. Goodman is Charan Hutchinson Distinguished Sevie Profanor of ‘aie Meaach Ctr the Unveil H. Real prefer of eae ‘aad airman, Depart of Sutin, Unveity of Clogs, 978 Eas SH ‘Set Chas, iS. Tha reerc wu supported in prt by Nationa Science Foubiatoe Resareh Gusta Nou NOP G9 Bf tod NSF GP I607, Some of the tteondntbar's mark o th ales done while vasa Flow at te Goer for ‘Advanced Stases io the Behavioral Seca aod) sina Seine ‘anatonSenar Postdoctoral Feowship Tha shear qretefl vo Steph Flnberg ire of Chieno) abd to Robert Sema (Uairrty of Coonan, [Berkeley for Depa eget, 2, MULTINOMIAL SAMPLING OVER THE ENTIRE TWO-WAY CROSS CLASSIFICATION 2.1, The Generel Case AAs before, let pa be the probability for the a, 6 cell of the cross classification, and let Ray be the corresponding. observed proportion. The ranges are a=1, 2, ++, aand b=1,2, ++, B. The pa are unrestricted, except of course tha; they are non-negative and add to one. ‘Nearly all the measures of association we presented in our earlier articles were written as ratios. We exploit that, streeture and consider a generic measure in the form 7/4, where » and & are mnemonically chosen to stand for sumerator and denominator; fis the generic measure of association, although it could be any ratio of two fune- tiors, » and 3, of the p's. (We assume that &%0 and that y and 4 are differentiable at the needed values of their arguments.) In this section we assume a multinomial sample of size n over the entire aX cross classification, so that. the Reis are the maximum likelihood estimators of the pu’s. tis well known that the covariance between Vn(Ras— pa) and Vin(Raw— pow) is BebarPes — papers 1) where the 6 here is the Kronecker delta; the K super- script is used to avoid confusion with the denominator delta above. ‘The maximum likelihood estimators of ¢, », and 8 are ‘the same functions of the Ra's as ¢, », and 8 are of the a's; we call these Z, N, and D, respectively, so that ZeN/D, and our concern is with the variance of the asymptotic normal distribution of y/n(Z—r). (See Sec~ tion 4 for comments on the possibility that D=0.) Let us define rhs = 80/80, Bis = 88/Apa, (22) since these partial derivatives enter frequently. It is con- venient also to define ba = rb — bra, B= Los prods Nove that d is a weighted average of the ¢a’s. ANS (23),416 ‘Thea by applying the delta method and the supple- mentary tools presented in (3), its easily calculated that, the asymptotic variance of Vn(Z—f) is ' a F_ Des pales — 8)? “5 1 a) Gl Des patts — 8} “This is a simple general form for the asymptotic varianze. We proceed to three examples, the first a reprise from [3] and the second two new. 2.2. The Measure of Association Gemma In [1] we discussed a measure of association called that might be appropriate when both polytomies of the cross classification are ordered. For convenience we re- capitulate the definition of y; its interpretation is given in [1]. In the present setting, 7 is most readily defined by ‘a short chain of definitions, 0, ~ 1, o=m +m, T= 2 Despallie, Me =2 Dav pallvieny Mas = Larre Dees per Thyies = Deve Dares rv The Roman numeral subseripts are mnemonics for the “quadrants” relative to (a, 8). (Quadrants here refer to the conventional Cartesian system with a’ —a eorrespond ing to the horizontal axis and b’—b to the vertical.) To caleulate the derivatives it helps to look at an ex- ample: what is the derivative of 11, with respect to pa? ‘Think of Tas two times (25) pullin + pullasae + prallyae + pselliae + Then one term of the derivative comes from the fourth summand above: 2Tly.n Another term eomes from the frst summand: 2pu. No other summands contribuse terms, and hence the desired derivative is 2(Ilz2:-+ou) In general, the derivative of T with respect to par is 2(thias+ Hata), where Tata = Laree Doce in accord with our mnemonie.t Similarly, the derivative of Tz with respect to pas i8 2(Hyi00-+Misa), and hence Ya = 2h — Thee + That — Tvs), 8 = 2s + Mtsas + hos + Miya) In terms of the notation of A7 of [3, p. 362], aly = Baus — aU), ‘Thus the general expressidn of 3 One igh at rat be concer at hee lelation of dre ty dicaty woe or none ne coher ne et Stelle gore nacre aac The Journal of the American Statistical Association, June 1972 da = 2(Il, — TLRS + 8) ~ 20, + 11) (@) ~ a) ee 6) Fed Les pollo’ — Meas)) since D pai = and D pase! =, Hence, if @ is the sample analog of y, the asymptotic variance of Vin(@—y) is 16 “ cot ‘+n Lov polos’ — awe] en Except for minor notational differences, and the earrying ut of the square, (2.7) is the same as (3.5.5) of [3]. (To see the equivalence, note that IL+lIlz=1—T, and refer to the manipulations at the bottom of p. 362 of [3].) For some purposes the present form may be simpler than the form in [3]. 2.3, Somers' Asymmetrical See Somers [5] has introduced an asymmetrical modifica- tion* of », n,m -rai together with its mate a, with the denominator replaced by 1— D0 64, (Following our past convention, a dot re- placing a subsubseript means summation over the re~ Placed subscript, e.g, ps. = Ds aa). The denominator of ‘se is the probability that two independently chosen units from a population governed by the [pu} cross classifien- tion probabilities do not lie in the same row (are not tied in a), An interpretation of asy may be given in terms of two such independently chosen units; let us eall their (ran- dom) row and column numbers (a, 6) and (a’, 6”), respec tively, and say that the units are weakly concordant when (a~a')(b~b')20, ie,, when the order of the two columns is the same as that of the tivo rows or when there isa tie. Similarly we may define weak discordance as (a—a’) (6—b)S0. Then the conditional probability of weak concordance, given that there is a difference between the rows, is [T+ 3s e— Des obsl/[L— Le ob], and the conditional probability of weal discordance, given axa’, is (Hat Ds ph— Das pla]/[1— Dy pf]. Hence ay is the difference between these two conditional probabilities It is sometimes useful to write the denoininator of Ay, in the form T+Il+ Los pale2—pa), where the third summand, whieh is equal to So» eh— Dona poy may be (2.8) * Somers und the yb die, Weave hanged tito Sy inorder to wed frAssnciation for Cross Classifications: Asymptotic Variances thought of as the probability that two independent ran- dom units are tied in column but not in row. Some important properties of Aye are: 1. aus is indeterminate if and only if the population is con ‘enteated in a single row; 2. fu is Tif and only if both Tle end Les pu(os—eu) are foro. The second condition says that exch column has at most one non-zero cell: henee, after removing all-ser0 tolummts, Ax i= Lif and only if the non-zero cells descend in staircase fashion, perhaps with treads of unequal width. A similar interpretation holds for 3. = —1) 8, aucis 0 in the ease of independence, but the converse need nat hold except in the 22 ease. ‘The numerator » ig the same for Mis as for y, and the denominator is 1— D> of. Hence vy is the same as for 7 and, more simply, 6% = —2pe. It follows that = a), ay = = Mpa ~ 286 (2.9) B= 2% Dep 2 and that Ta) = — 2 ba — 8 = BL = po) — 28(G — AS’). (2.10) Letting chs denote the sample analogue to As,, the desired asymptotic variance for +/7(de— Ase) is, therefore, w pesle( = pa.) — a(R — 7 ae). 1) So far as we know, this result is newly published, 2.4, The Measure of Asociation rs In [1) we presented an asymmetrical measure of asso- ciation, rs, based on a notion—suggested to us by W. Wallisof reconstructing as best one can the cross classi- fieation probabilities. Interpretative deteils are given in Section 9 of [1]. In the present setting, it is easier to work with I ss (which will not affect the asymptotic variance) and to express 1—n in terms of its numerator and de- nominator, ~ Les (oa/en), = 1— Lop 212) ‘The maximum likelihood estimator of Ln is Ln Ene Ra) 1— Doky and we want to find the asymptotic variance of V/n(lo—r)y which isthe same as that of ¥i[(J 4) — (1—n)}. Fellow ing our general prescription, we find b= 20 va = — Daw F 2 = ~ 2palp. + Ly (oar e0)s where 8* is the Kronecker delta. Thus ba = — 2epat Dpalpe: — 8 Lv (ow /pe dy ppv + 8 Dons (pas/ a.) = — 21 — 8) + a(1 ~ ») =H mtb Hs (2.3) = (ty = 2n)6 = [Zea ob/ood]E + Doe] — 2 Dar (We have expressed in several different ways for con- venience of reference.) Hence the asymptotic variance of yn(l—n) under full aX multinomial sampling is 1 ¥ 2 pales — 3), 4) were 6, @as, and ¢ are defined above. This result is new; we did not discuss the distribution of tsunder full multinomial sampling in [3]. 3. INDEPENDENT MULTINOMIAL SAMPLING IN THE ROWS 31. The Generel Case In Section 4 of [3] we dealt with a stratified sampling rethod that may sometimes arise. In this method there ace separate, independent multinomial samples within each row of the cross classification; further, the sample size in Tow @ is ni. =n, where the a's are supposed known, positive, and summing to one. (In practice, nae ‘ill not in general be an integer, s0 one would take ns. as that integer closest to nw,. For purposes of asymptotic theory, wwe need only assume that the ne./(nes) have the limit one as n grows large.) We also assume that the p.’s are known and positive To be specific, we treat separate sampling within rows; if there is separate sampling within columns, one need cnly interchange the roles of rows and columns. The sampling method under discussion may arise either from stratification of a single population, or in comparing sev- «ral different populations, i a= po that is if ma. is proportional to py, matters simplify a bit because then Ras is still the maximum likeli- hood estimator of pu, just as in the case of full multi- nomial sampling. It was for this reason that we restricted ourselves in [3] to the proportional sampling rate ease, ‘The general ease, however, is not essentially harder—it requires only earrying multiplicative factors along—and, tosave space, we deal with it directly. (This remark qual fies « possibly misleading statement in the last paragraph of Section 4.4 of [3}) It is convenient to deal with conditional row proba- bilities, and accordingly we write Ba = pape, — Ray = Resi Re48 for all relevant values of a and 6. (For asymptotic pur- poses, there is no problem about the positiveness of R., since we have assumed «>0 and n,. is within I of my, but there may be a problem with zero Rs. for small finite sample sizes.) The maximum likelihood estimstor of a is clearly Raj hence that of pa is (ps./Re.)Ras, which is almost («-/v,) Ras. Further, the covariance between ¥/7(Ras— fut) and Vii(Rew Avy) is readily seen to be 1 Holi — Aap) (3) ae Note the 1/, factor, and the first Kronecker delta outside the brackets, because of independence among the multi- nomials. As before, we let f=»/8 be a generic measure of associa tion, but now we regard f, », and & as functions of the a's (and, of course, of the Known p,’s). Then Z, N, D. (Z=N/D) are the corresponding sample quantities, ob- tained by replacing pu with Ry to obtain maximum likeli- hhood estimators. The p,.'s stay unchanged since they are known constants. Next, let »% and 83 be the partial derivatives of » and 4, respectively, with respect to As. We use asterisks in- stead of primes to avoid confusion about the argument of diferentistion, (Nonetheless, we record the relationships Wo= pers and 85 p..8.) Asin Section 2, we introduce = La sat, where 6; is a weighted average of the g's in row a only. (Some symbolic distinction from the notation of (2:3) is necessary, and the ++ superscript is suggestive of fixed row marginals.) The methods of [3] then show easily that the asymptotic variance of ¥/n(Z~f) is FEL Dasaleh ~ 5 = 8 — 50, (32) ay In short, we obtain here a linear combination of the within-row variabilities of the 3’s (weighted by the a's), rather than the overall variability obtained in. Seetion 2 ‘This is a simple general form for the asymptotic vari- ‘ance under independent sampling within rows. We illus trate it next with two examples, the first a generalization of Section 4.2 of [3], and the second a generalization and correction of formula (4.4.3) in [3] 3.2, The Measure of Association hs Tn [1], we presented an asymmetrical measure of a3s0- ciation, \, bused on the concept of optimal prediction, Interpretative details are given in Seetion 5.1 of {1}. Ia the present setting, itis easier to work with 1A. (which Journal of the American Statistical Association, June 1972 will not affect the asymptotic variance) and to express 1-2, in terms of its numerator and denominator, —Zerm 8 @a) where pum is the maximum of pat, “+, eas; and p.nis the maximum of p.,+--, p.2 We shall assume, as in [3] (cee Sec. 3.1 there), that pea equals exactly one of the pu, SAY pana Gnd that p.m equals exactly one of the ps say pa (b= 1, -- +). (Note about symbolism: In [3] we used pin for what is called py here, where b.» is the value of the column subscript index maximizing 1, +, pa; and we did not in [3] need a symbol for what wwe now call bs, the value of the column subscript index maximizing pat, ~-~ a) ‘The maximum likelihood estimator of 1—Ny is Da lou Baa) Max 3 Ge Ra)” 1 pm 1-be= and we want to find the asymptotic variance of Vi(Ls—N), which is the same as that of Vn[(I—Ls) ~(1-2)]. To follow our general prescription,’ first write ve l— Zales §= 1 ~ Max Do (oud). (8.5) Recalling that the pu’s are fixed, we diferentiate with respect to fas to find = panos 8 ms = pads A schematic sketch of the cross classification will aid in seeing why these are the derivatives. It follows that ~ routes + 850 Bay (3.8) 8. = boasy ~ YPay Hence the asymptotic variance of v/7i(Ly—N) under inde- pendent sampling within rows is, from the second form of (3.3), L atk ak ak Fy Le [Da esl 6 hey — 28706 Ba iby en = z {8* De teraa(l — faa 31) = Bel Lr (apadiy) — Le fedsnBaro] Hot Le Bapars(l = pans)}y where é.=ps./ia, and SY denotes summation over those values of @ for which By =b¢. (This summation usage aad been used in {3].) The "Te ihe be eed tha ee iy he cto Atand aot fl tata so dteaty with the styapttie henry coatAssociation tor Cross Clessitications: Asymptone Variances ity sign in (3.7) is, when all 6,=1, exactly the same as (4.2.6) of [3], except for notation changes. Although these expressions appear rebarbative, they are often not diffi- cult to use in specific cases; we illustrated the use of the sample analogue of (8.7) in Section 3.2 of [3]. 3.3 The Measures of Association 15 We return to 1, of Section 2.4, but now under inde~ pendent sampling in the rows, Tt is convenient to work swith In and to express it as »/8, where = Lesmin 95 1- ToC Leeda) G8) ‘The maximum likelihood estimator of | —ry given in See- tion 2.4 should have Ry. replaced by pa., sinee the pa.’s are known, Hence the estimator now is [t — Dosa pe Rad/l — Ss (Ze pe Ra)" Following our general prescription, a Vs = — paar = — pay so that ba = 2pa — 2vp0-p.5 = Ppa [8 Date — » Ds pada eis helpful to let Yas = BP vps, 80 that On = Brudin On = Baar where Jz = Ss fuss. In these terms, the desired asymp- totic variance for independent sampling in rows is FEeatonlts =v 8.10) ‘When the sample sizes by rows are proportional to the e's; ies, when w= py. 80 that all @=1, then (3.10) with 45," deleted gives the asymptotic variance. Formula {4.4.3) of [3] purported to give that asymptotic variance, but in error; the second term of (4.4.8) of [3] is wrong. 4, USE OF THESE RESULTS IN PRACTICE Probably the most common use of these results in practice (see Section 3.2 of [3]) is to treat V/n(Z—D), as approximately unit-normal, where estimator of the asymptotic variance o%. In the setting of our sequence of articles, # is readily taken as the maxi- mum likelihood estimator of 6, as follows. Any o? isa funetion of the pa's, which may for conve- nience be written, perhaps in part, in terms of the Ba's To find the maximum likelihood estimator of ¢2, make the replacements in the arguments of ¢?as listed below. Recall that Ras =Nas/, the proportion of all observations in the (a, b) cell, and that Rus= Nay/ne., the proportion of ob- servations in the (a, 8) eell relative to row a. is a consistent Full multinomial sampling. (See. 2) pa Ras Independent sampling in rows. (See. 3) pat Raslpa./ots) = Ruse = Rasps Ba Res In practice, ¢ may be zero; we disewss this problem in {3] eg, in connection with y on p. 324 of [8]. Provided that ¢ > 0, however, the probability that ¢ = 0 approaches zero 8 1 grovs, 80 for large enough samples the ¢=0 problem disappears, We have no analytic information abcut what “large enough” means, but we have encourag- ing evidence from the simulations reported in [8] and those of Rosenthal [4]. In the next section we consider the meaning of seribed earlier. Tt ean also happen in practice that D=0, and then Z is undefined. This problem was discussed in (3. p. 320] for the case of ty. Since we assume throughout that 5%0, and since D converges to 5in probability, the D=0 prob ler: also vanishes as n gets large. 0 for the measures of association de- 5. WHEN DOES o=0? 5.1, Full Multinomiel Sampling From (24), it is clear thet, under full multinomial sampling, ¢=0 if and only if pa(éa~3) =0 for all a, b, What does this mean for the examples of Section 2? Gamma. Criteria for ¢=0 in the case of G under full multinomial sampling were discussed in Section A7 of [3] The basic condition given there may be rewritten, if T1,>0, as follows: If two individuals, 1 and 2, are drawn independently at random from the population, then, whenever pa>0, ineoncordant with 1 | 1 in (a eel PFD b disordant with 1 Pin @ Deal) does not depend on the choice of (@, 8), except that both numerator and denominator may be zero for sume 0. By our general assumption, both II, and Ty cannot be 0, If either M1, or ly is O (ie., y= +1), then o=0. If at least one corner cell has positive probability, this be- comes an equivalence: ¢=0 if and only if y= £1 A family of cross classifications for whieh o=0 is the balanced cruciform family: all the probability is in single row and column (neither of them borders), and there is equal probability in the two horizontal arms of the “cross” a6 well as equal probability in the two vertical limbs. A specifie numerical example is420 ,, Where cells without numbers have zero probsbilities. In such a balanced eruciform case, My=Hy so -y=0, and G9 =@8 for every céll with pa>0. ‘There are, however, other cross classifications with '=0, but for which y is not —1, 0, or 1. All appear to be very special. For example, consider the 4X4 ease in which, there is probability 0.25 in cells (1,2), (2,1), (8,4), and (4,3); other cells have, of course, zero probability. Here T1,=.5, T=.25 s0 that =1/3, yet it is easy to check that ¢=0 because the coneordance-discordance ratio is 2 for the four eetls with positive probability. The measure &ye. A necessary and sufficient condition that ¢=0 here is that, for all cells with pus>0, BS) — BS) = Aull = 94). We do not have a nent characterization of these eases. The measure 4. This was not discussed in [3] for full multinomial sampling. The major finding here is that if all pa>0, then ¢=0 if and only if independence holds, ie., if and only if pa =p.p.s for all (a, 6). (Note that ths implies =0.) To see this, assume that all p>0, so that to say ¢=0 is to say (see (2.13)) that for all a, by, by = bes, — dary = — Bro, — ps.) + 26(Fan, ~ Bot) Hence, taking the difference of this quantity between rows a; and as, Bad — Baits = Bods ~ Bat Finally, add over by and recall that 4,.= 1, Tt follows that Busy Pray OF that poy ™ pay. Busy Now average both sides over a: to obtain pa, as a product of a factor depending only on a; and another depending only on by. This shows independence. Conversely, if pw=ea-p- for all a, b, sub~ stitution shows immediately that 3=» and that al 4-0. Ti some pus are 0, we do not know a nice way to char acterize ¢=0, 5.2. Independent Sampling in Rowe Here ¢=0 if and only if 4.(¢5—¢1)=0 for all a, t. What does this mean for the examples of Section 3? The measure Ns Tn [3, p. 315] we asserted that for My, =0 if and only if or 1, but we did not there give a proof, In our current notation and approach, a proof may be given relatively easily. We assume, without loss of generality, that all p,.>0; if au.=0, just delete that row. Recall first that, Dy = 0 means pare: for all a, oF, Dew Pay equivalently, Dr M means 3) pm = 3 pate) = 1 Now suppose that #=0, ie. that pu(s—82)=0 for all 2, b, In row a, look at the a, bys cell, for Which fares EUS? be positive, Thus, for all a, iy — FL = Bleu. — pas) — H(deBhemey ~ ns) = 0. Journal of the American Statistical Assoctation, June 1972 Next, add over a, to obtain &—( 57’ ps.—p.n)y=0. Hence, either »=0 (whence 1= 5 pom and d=1) or else Lm pm Di be tam = 1 Di pe = 0. Zr pw, then Dy pu Eber pa 0. 1) 1h Henee #3,—82 =0. Similarly, and éf,,,=J for alla, Hence pay, for all a, #0, 80d ps. —pary)=0. if N=0, 1-N=1, b=, Py ~ BE = Alpes ~ pas — Pee Bas] =O. ‘This completes the proof. ‘The measure ry. As in Section 5.1, the result here is that, if all ps>0, 0 =0 if and only if independence holds, The argument from independence to o =0 is immediate; in the other direction, if all p»>O and ¢=0, then AG, — 83.) = 0, ~ Ba) ~ Los, ~ ps) = 0 for all a, by, by. The corresponding demonstration in See- tion 5.1 then applies. . CAUTIONARY NOTE ABOUT ASYMPTOTIC VARIANCES In working out asymptotic variances of the above ‘kind, there is a trap that stems from the singularity of the FF —hC———— Xs puro... (ee Footnote 1.) Because of these relation= ships, a given function of the pa’s may be expressed in a variety of ways, and sometimes one way is more conve- nient than another. Which way an expression is written makes no difference (except for convenience of eomput: tion) in the final asymptotic variance, provided that the same symbolie functional form is used throughout in find= ing derivatives. If not, incorrect results may be obtained. We illustrate with a very simple ease. Suppose that (Xin, Xia) form a sequence of pairs of random variables (n=1, 2,3, «+ +) such that Xu-+Xin=O, and such that the pair (vin(Xin—2), Vn(Xan+2)) has in the limit as n becomes large the (singular) bivariate normal distribution with means zero, variances 1, and covariance —1. Note that we are treating the singularity consistently: first, 2++(--2) =0; second, the asymptotic variance of Vil (Xin =2)-+(Xie+2)], which should be zero, is indeed 1-241=0. Now let the function of interest be Ya= Xt, Its deriva- tive with respect to Xin (evaluated at Xin=2) is 2X2—4; the corresponding derivative with respect to Xm is zero. Hence the asyinptotic variance of Vnu(¥a~4) is 16(=(4)*x1). But the function might just as well have been written Y,=X3, The evaluated derivatives with respect to Xt, Xr, respectively, are 0 and —4, Hence the asymptotic variance of Vn(¥,—4) is again 16. ‘A more interesting way of writing the funetion for illustrative purposes is Yu=4X},+3X3. The evaluated derivatives now are (3)(2)=4/3 and (§)(—2)=-8/3, respectively. Hence the asymptotic variance isAssociation for Cross Classifications: Asymptotic Varionces @-@Ca+C)- 3, a) 3, 3 1s before. Thus, no matter how we choose to write the function, we get the same asymptotic variance, provided ‘we remain faithful to the same symbolic form during the differentiation process. 1s, however, we do not remain with one symbolic form, incorrect results may occur, In the above example, suppose that we write the function as X3, before getting the Xie derivative, and as X, before getting the Xow derivative. Both evaluated derivatives will then be 2er0,, and see will obtain the grossly wrong asymptotic variance of 0, instead of the foursquare correct value of 16, [Received December 1970, Revised August 1971.] Technieal Reporte (continued from page $87) Gould, PJ. and Howe, Stephen, A New Result on Interpreting Lagrange Multipliers as Dual Variables,” January 1971, 8.55, Mineo Series No. 788. Lindgren, Georg, “Wave-Lengih and Amplitude for « Stationary Process After a High Maximum,” February 1971, $70, Mimeo Sirive No. 742 Skachiman, Richard, “Generation of the Admissible Boundss of a Convex Polyptope,” Pebroary 1071, 8.55, Mica Serie No. 743. Gould, F.J., *Continuously Differentiahle Exact Penalty Pune- tions for Nonlinesr Programs with Tolerances,” April 1971, $.60, Mimeo Series No. 744 Lindgren, Georg, "Wave-Length and Amplitude for a Stationary Process After a High Maximum: Decreasing Covarisnee Func tion,” April 1071, 8.85, Mimic Series No. 745 Smith, Woolleott, “The fafinitely Many Server Queue with Semi- Morkovien Arrivals and Customer Dependent Exponential Service Times," May 1071, $55, Mico Series No. 748 Behboodian, Javad, “Information Matrix for e Mixture of Two Normal Distributions,” April 1971, $.63, Mica Series No. 747. "Information Matrix for & Mixture of Two Exponentish Distributions,” May 1971, $.00, Afinco Series No. 748. Oregon Stote University, Department of Statistis, Corvallis, Ove. 97331 Seymore, Goorge E., “Interval Integer Programming,” May 1971, Technical Report No. 22. Ramsey, Fred L., "Bayesian Biosseny,” May 1071, Technieat Report No, 2. Land, Charles, and Johnson, Bruce K, “A Note on Two-Sided Confidence Intervale for Linear Punetions of the Normal Sean and Variance,” August 1971, Technical Report No. 24. Pon. sania Stote University, Department of Statistics, Uni versity Park, Penn. 16502 © P, ond Boswell, MT, *A Characteristic Property of the MilUvariste Normal Distribution and Some of its Applics- tions." Repart No.1 ‘nut Tnnardnn, K.G., "On Acceptance Sampling without Replacement: Equivalence of Hypergeometrie and Tnverse Hypergeometrie Acceptance Sampling Plans,” Report No. 2. "On Sampling with Replacement from Populations with, 42i REFERENCES: I) Goodmsn, Leo A. and Kruskal, Willian H., Association for Cross Classifications,” Journal of the American, ‘Statistica! Asociation, 49 (December 1954), 732-4 21 Goodman, Leo A. and Kruskal, William H., “Measures of Association for Cross Classfictions, 1: Further Discussion nd References,” Journal of the American Statistica clssocion Hon, 54 (March 1959), 125-83. [8] Goodman, Lea A. and Kruskal, William H., *Messures of Association for Cross Classifeations, III: Approximate. Sampling Theory,” Journal of the American Statistical Assan ciation, 58 June 1903), 310-4. (4) Rosenthal, Irene, “Distribution of the Sample Version of the ‘Measure of Association, Gamma,” Journal of the -merican, Statistical Association, 61 (Sune 1966), 40-83. {5} Somers, Robert H., "A New Asymmetric Measure of Ascocine tion for Ordinal Variables,” Aerican Sociological Review, 27 (Desember 1962), 799-811, Measures of Multiple Charscters," Report No. 3. And Joshi, SW, “Purther Results on Minimuin Vasiance Unbiased Estimation and Additive Number Theory,” Report Ho. 4 Joshi, SW. and Patil, G.P,, *Certsin Structural Properties of ‘the SumSymmetric Power Saries Distribution,” Report No. 8, snd Patil G, P., *Sum-Symmetrie Power Series Distibue ‘ions and Minimum’ Variance Unbiased Estimation,” Report Wo. 8. See eee of Transformations of Distsibution Funetions," Report No. 7 Stiteler, WM. and Patil, G.P., “Variance 10 Mean Ratio and Morisita's Fadex as Measures of Spatial Patterns in Ecclogieal Populations,” Report No. 8. Janardan, K.G. and Patil, G.P., “On the Multivariate Palys Dis. iwibution: A’ Model of Contagion for Data with Multiple Counts,” Report Xo. 3 Haight, F.A., "Group Size Distributions, with Applicstions to ‘Vehicle Occupancy,” Report No. 10. Harkness, W.L., “The Classieal Occupaney Problem Revisited,” Bepart No 17. Jsnsrdan, K.G. and Patil, G.P., “On Acceptance Sampling with ‘out Replacement,” Report No. 12. Bowell, MT. and Patil, G.P., "Characterization of Certain Dise ‘uete Distributions by Differential Equations with Respect 10 ‘Their Parameters," Report No. 18. Janacdan, K.G. and Patil, GP, “The Multivariate Inverse Polya Distribution: A Model of Contagion for Data with Mule tiple Counts in Inverse Sampling,” Report No. 14, Boswell, M-T. and Patil, @.P., "Chance Mechanisms Generating ‘Negative Binomial Disteibutions,” Report No. 15. Janardan, K.G. and Pati, G.P., “A Unified Approach for a Class of Multivariate Hypergeometric Models,” Report Ne. 16 and Patil, G.P., "Multivariate Modified Poly snd n> ‘verse Polya Distributions," Report No. 17. and Patil, G.P., “Location of Modes for Certain Cnivari- ste and Multivariate Discrete Distributions,” Report No, 18 Hettmansperger, T.P. snd Sievers, G.L., “On the Efficiency of Some Median Tests,” Report Yo. 19. Gedambe, A.V, and Patil, G.P., “Infinite Divisbilicy and Addi- ivity of Certain Probability Distributions with an Application “e Mixtures and Randomly Stopped Sums,” Report No. £0. . (Continued an page 447)