(PDF) O-A194 RN AkP WdORKSPACE FOR CONDUCTING … · considered: Sign Test, Wilcoxon Signed-rank Test, Mann-Whitney Test, Kruskal-Wallis Test, Kendall's B, Spearman's R, and Nonparametric

O-A194 971 RN AkP WdORKSPACE FOR CONDUCTING NONPRAHETRIC1/STATISTICAL INFERENCE(U) NAVAL POSTGRADUATE SCHOOL

MONTEREY CA W F YAGTS JUN 87

UMCLRSSIFIED F/G 12/3 HEhhE IlEEEE

1111 1 1080 128

lul ~l ''

11111'' .6IIII1II -8

HI1111- liiil,

MICROCOPY RESOLUTION TEST CHART

NATONAL BUREAU OF STANDAROS 963-A

AV .. &V ' mv ,

'---

, . ON J':FILE ""

NAVAL POSTGRADUATE SCHOOLMonterey, California

DTIC

ELCT

OCT 0 7 198

THESISAN APL WORKSPACE FOR CONDUCTING

NONPARAMETRIC STATISTICAL INFERENCE

Wayne Franz Vagts

June 1987

Thesis Advisor: T. Jayachandran

Approved for public release; distribution is unlimited

. .,'.,,) 183

sEcuairy CILASSIFICATION OF TIS OAGE

REPORT DOCUMENTATION PAGEI# REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS

UNCLASS IFIED)___________________I& SECURITY CLASSIFICA TION AUTHRI401TY 3 RISTRIBUTI0Nd4AAILAILhTY OF .REPORT

_____________________________ pproved for public release;l b OECLASSIF#CATION /DOWN4GRADiJNG SCHEDULE distribution is unlimited.

4 PERFORMING ORGANIZATION REPORT NUMBER(S S MONITORING ORGANIZATION REPORT NUMBER(S)

64 NAME OF PERFORMING ORGANIZATION J60 OFFICE SYMBOL 74 NAME OF MONITORING ORGANIZATION

Naval Postgraduate School 53 Naval Postgraduate School

&c AODRE 55 City. State. and ZIP Cod.; 7b ADDRESS (City, State, and ZIP Coll

Monterey, California 93943-5000 Monterey, California 93943-5000

, NJAME O2F ;UNOINGiSPONSORING Bab OFFICE SYMBOL 9 PROCUREMENT INSTRUMENT IDEN TIFICATION 'juM8RODRGANIZATION (if dO0hCrabAi

8c AODRESS(City. Sfate. and ZiPCod.) 10 SOURCE OF FUNDING NUMBERS d

PROGRAM PROJECT TASK IWORK .JNlrELEMENT NO INO NO ACCESSION NO

1VTLE Incil e 5cui Clawholcarion)

*AN APL .JORKSPACE FOR CONDUCTING NONPAIRAMETRIC STATISTICAL INFERENCE

12 PERSONAL AUTIIOR(S)VAT,1yeFrn

1TYPE OF FIEP iiRT t 3b TIME COVERED 114 POR (Yea, Month. Day) jsPAGE M lUNMases hesis FROM _ TO ue116

'61 SUPPLEMVENTARY NOTATION

17 COSATI CODES ff"%~ f Tj jj t14& ~~ELD GROP SB-GOUP Mann-Whitney Test, Kendall Test, Spearman Test,

Nonparametric simple regression, (continued)'9 ABSTRACT (eontonue on leverseo of nocenary and odentify by block number)

This thesis contains programs written in APL and documentation forperforming certain nonparametric tests and computing nonpara-metricconfidence intervals. These methods of inference are particularly usefulin dealing with Department of Defense related problems as illustrated intaie several nilitari examples worked in Appendix C. The following non-oarametric tests ire considered: Sjign Test, Wqilcoxon Sqned-rank -71stMann-W'hitney Test,. Kruskal--'Jallis Test, KendaL1's B, Soearmnan's a, andNonparametric Linear Regression. The tests are based on the exact -

*distributions of the respective test statistics unless a large sampleapproximation is determined to provide at least a three decimal placeaccuracy. The software consists of two APL workspaces; one, which is

*designed for microcomputers (IBM PC's or compatibles) and is menu driven,

0 DS~RiSUTION /AVAILAILITY OF ABSTRACT 21 STfTS~_IYeSFCTO

(3..NCLASSIFIE0flJNLIMITIEO 0 SAME AS RIPT 03 DTIC USERS MCassi22a NAME OF RESPONSIBLE INDIVIDUAL 22b TFJEPNQNI (Ii.do Area Code) 22c OFFi ILL O

. Jayachandran 4 0-6 620000 FORM 1473.64 MAR 83 APR eolil e~ay be used unil~ Olluted SECURITY CL.ASSIFICATION O r,415 PAGE -All other editions are obsolete

Jill" 11 111 Jil

ISUNUY CLASUIFCAnOW OF T"15 PAQ9 jWtm D O&,

Block 19. Abstract (Continued)

and the other, without menus, is designed for the mainframecomputer (IBM 3033) at the Naval Postgraduate School.

Block 18. Subject Terms (continued)

Exact Distribution, Asymptotic Approximations.

p.a,

SIN 0102- LF- 014-6601

2 SCURITY CLAIMPICAIOW OP IrIS PAG,(Mh, Veta Dnt.e.d)

a ~ ~ ~ : ~ *.- ~ *p~ **,..'-* -**-~'j*~** *~*** !*.':*

Approved for public release; distribution is unlimited.

An APL Workspace for Conducting NonparametricStatistical Inference

Wayne Franz VagtsLieutenant Commander, United States Navy

B.S., University of Notre Dame, 1975

submitted in partial fulfillment of therequirements for the degree of

MASTER OF SCIENCE IN OPERATIONS RESEARCH

from the

NAVAL POSTGRADUATE SCHOOLJune 1987

Author: _ _ _ __ _ _ __ _ _

Wayfie Franz tagt

Approved by:

Toke Jeyadhandran, Thesis Advisor

Robe Recond Reader

Peter urdue, Chairman, Department ofOperations Research

Kneale T. Marshall, Dean rormationand Policy Sciences

ABSTRACT

This thesis contains programs written in APL and

documentation for performing certain nonparametric

tests and computing nonparametric confidence intervals.

These methods of inference are particularly useful in

dealing with Department of Defense related problems as

illustrated in the several military examples worked in

Appendix C. The following nonparametric tests are

considered: Sign Test, Wilcoxon Signed-rank Test, Mann-

Whitney Test, Kruskal-Wallis Test, Kendall's B,

Spearman's R, and Nonparametric Linear Regression. The

tests are based on the exact distributions of the

respective test statistics unless a large sample

approximation is determined to provide at least a three

decimal place accuracy. The software consists of two

APL workspaces; one, which is designed for

microcomputers (IBM PC's or compatibles) and is menu-

driven, and the other, without menus, is designed for

the mainframe computer (IBM 3033) at the Naval

Postgraduate School.

TABLE OF CONTENTS

LIST OF TABLES............................................. 7

ACKNOWLEDGEMENTS........................................... 8

I. INTRODUCTION.......................................... 9

II. WORKSPACE DESIGN ISSUES............................1

III. GENERAL SAMPLE SIZE CONSIDERATIONS ANDASYMPTOTIC APPROXIMATIONS.......................... 13

IV. TESTS FOR LOCATION BASED ON

SINGLE AND PAIRED-SAMPLE DATA..................... 15

A. ORDINARY SIGN TEST............................ 15

B. WILCOXON SIGNED-RANK TEST.................... 18

V. TESTS BASED ON TWO OR MORE SAMPLES................ 24

A. MANN-WHITNEY TEST............................. 24

B. KRUSKAL-WALLIS TEST........................... 3

VI. TESTS FOR ASSOCIATION IN PAIRED-SAMPLES.......... 35

A. KENDALL'S B................................... 35

B. SPEARMAN'SR................................... 38

VII. NONPARAMETRIC SIMPLE~ LINEAR REGRESSION............ 43

A. COMPUTATION OF THEESTIMATED REGRESSION EQUATION................ 43

B. HYPOTHESIS TESTING............................ 44

C. CONFIDENCE INTERVAL ESTIMATION............... 4

VIII.AREAS FOR FURTHER WORK............................. 46

LIST OF REFERENCES....................................... 47

APPENDIX A: DOCUMENTATION FOR THE

MICROCOMPUTER WORKSPACE................... 49

APPENDIX B: DOCUMENTATION FOR THEMAINFRAME COMPUTER WORKSPACE ............ 52

APPENDIX C: WORKSPACE FAMILIARIZATIONTHROUGH PRACTICAL EXAMPLES .............. 54

APPENDIX D: MAIN PROGRAM LISTINGSFOR MICROCOMPUTER WORKSPACE .............. 78

APPENDIX E: MAIN PROGRAM LISTINGSFOR MAINFRAME COMPUTER WORKSPACE ........ 88

APPENDIX F: LISTINGS OF SUBPROGRAMSBASIC TO BOTH WORKSPACES ................ 98

APPENDIX G: LISTINGS OF PROGRAMS USED TOGENERATE C.D.F. COMPARISON TABLES ...... 107

INITIAL DISTRIBUTION LIST ........................... 115

V V ~ V V~~ ~'

LIST OF TABLES

1. C.D.F. COMPARISONS FOR THE SIGN TEST ............. 16

2. C.D.F. COMPARISONS FORTHE WILCOXON SIGNED-RANK TEST ................... 20

3. C.D.F. COMPARISONSFOR THE MANN-WHITNEY TEST ....................... 26

4. C.D.F. COMPARISONS FORTHE KRUSKAL-WALLIS TEST ......................... 32

5. C.D.F. COMPARISONS FOR THE KRUSKAL-WALL TS TEST USING COMPUTER SIMULATION............. 33

6. C.D.F. COMPARISONS FOR KENDALL'S B .............. 37

7. C.D.F. COMPARISONS FOR SPEARMAN'S R .............. 41

Ac.-esion ,For

NTIS CRADTC TABULlmjnirvzwc C]

.i. ' j-VOID

ACKNOWLEDGMENTS

Many thanks to Professor Larson for his APL

workspace STATDIST from which several of the normal

theory based asymptotic approximations are computed.

APL*PLUS/PC System software and APL*PLUS/PC TOOLS are

used in the construction of the microcomputer work-

space.1 IBM's VSAPL is the APL version used for the

mainframe workspace.

IAPL*PLUS is a copyrighted software from STSC,Inc., a CONTEL Company, 2115 East Jefferson Street,Rockville, Maryland 20852.

I.INTRODUCTION

Although nonparametric procedures are powerful

tools to the analyst, they are currently underused and

often avoided by potential users. Perhaps one reason

for this is the difficulty in generating the exact

distributions of the test statistics, even for moderate

sample sizes. Consequently, tables of theseko

distributions are only available for very small samole

sizes and normal theory based approximations must then

be used.

The purpose of this thesis is to make a variety of -'

nonparametric procedures quick, easy and accurate to

apply using menu driven computer programs in APL.1

These programs use enumeration, recursion, or

combinatorial formulas to generate the exact null

distribution of the various nonparametric test

statistics. This allows hypothesis testing and

confidence interval estimation to be based on exact

distributions without the use of tables. For larger

sample sizes, the normal, F, and T distributions are

1 APL was chosen because it is an interactivelanguage that is especially powerful at performingcalculations dealing with rank order statistics andvector arithmetic. Menus are not included in theworkspace designed for the mainframe.

used to approximate the distributions of the test

statistics with three decimal place accuracy.

Section II addresses workspace design issues, 'to

include, workspace requirements and assumptions

regarding its use. Section III discusses the methods

used to assess the accuracy of different asymptotic

approximations, and the sample sizes required for an

approximation to yield three decimal place accuracy.

Section IV gives background information and discusses

programming methodology for nonparametric tests based

on single and paired sample data. In Section V,

nonparametric tests for two or more independent samples

are considered. Section VI discusses nonparametric

tests for association; and, Section VII deals with

nonparametric simple linear regression. Section VIII

recommends other nonparametric tests that may be added

to the workspace and areas for further work.

To show application of nonparametric statistical

methods to Department of Defense problems, several

military examples are worked in Appendix C.

II.WORKSPACE DESIGN ISSUES

This section presents a brief overview of the

design considerations used in developing the APL

workspace for both the mainframe and microcomputer.

A. EQUIPMENT AND SOFTWARE REQUIREMENTS

The microcomputer must be an IBM PC or AT

compatible, equipped with 512 kilobytes of RAM and the

APL*PLUS/PC system software, release 3.0 or later, and

IBM's DOS, version 2.00 or later.1 The 8087 math

coprocessor chip is no required to run this software,

but will increase the computational speed.

B. KNOWLEDGE LEVEL OF THE USER

The user is expected to have had some exposure to

APL and a working knowledge of nonparametric

statistics. Familiarity with microcomputers or the

Naval Postgraduate School mainframe computer is

assumed.

1 The APL system software requires 144 kilobytes ofRAM while the NONPAR workspace requires an additional190 kilobytes.

ii0V

: Z:ell11

C. SELECTION OF TESTS

The nonparametric tests chosen for this workspace

are some of the more widely known, and are considered

basic material for any nonparametric statistics course.

More information about the tests can be found in any of

the textbooks that are referred to in this document.

D. MENU DISPLAYS

The microcomputer's workspace is designed around

the use of menus. This was accomplished using the

software package PC TOOLS from STSC. These menus are

designed to guide the user through the selection of the

tests without an excessive amount of prompting. The

main menu displays the choices available in the

workspace, while the test menus give the background

information and options available for each test. Help

menus to provide additional information about the

tests are also available.

E. ORGANIZATION OF WORKSPACE DOCUMENTATION

Separate documentation is included for the

microcomputer's and mainframe computer's workspaces

(see Appendices A and B. respectivelyY. These

appendices explain the organization and opera:>irn of

the workspaces. Appendix C, which provides example

problems for each nonparametric test, is applicable to

both workspaces.

. . .-. . . . .. -. , . ,... ....... ., .. ,. . . ,.. .... -.. ,.- .-. -. - ,,' -N .

III.GENERAL SAMPLE SIZE CONSIDERATIONS AND ASYMPTOTICAPPROXIMATIONS

In this thesis, the term alpha value is used in

a general sense, and refers to the probability of

rejecting a true null hypothesis. The term P-value

refers to the probability -hat a test statistic 4i''

exceed (or not exceed in the lower-tailed test) the

computed value, when the hypothesis being tested is

true.

For selected values, the exact cumulativedistribution functions (C.D.F.' of the test s.a::s-/cs

are compared with those obtained from normal based

asymptotic approximations. The results of the compari-

sons are used as a basis for assessing the accuracy of

the approximations. In those cases where more than one

asymptotic approximation has been suggested in the

literature, the accuracy of each approximation is

compared over a range of desired C.D.F. values and

sample sizes. From the results, the most consistently

accurate approximation, and the sample size for which

,hat approx.mation provides a! >'as . -_ e iec:=a

place accuracy is determined.

Once the accuracy comparisons were completed for a

specific nonparametric test, microcomputer capabilities

were considered. In some cases, generation of the

exact distribution up to the desired sample size took

too long or was not possible on the PC. When this

occurred, the mainframe computer was used to generate

the required distributions with the results stored in

numerical matrices for quick recall by the

nonparametric test programs.

.1S:

a-... - S

IV.TESTS FOR LOCATION BASED ON SINGLE AND PAIRED-SAMPLE DATA

The tests assume that the data consists of a

single set of independent observations Xi or paired

observations (Xi,Yi), i=1,2,...,N, from a continuous

distribution. For the single and paired-sample cases,

the null hypotheses are concerned with the median of

the Xi and the median of the differences X;- Y-;.

respectively. The tests considered are the Ordinary

Sign Test and the Wilcoxon Signed-Rank Test.

A. ORDINARY SIGN TEST

The Sign test can be used to test various

hypothesis about the population median (or the median

of the population of differences). Confidence

intervals for these parameters can also be constructed.

As a final option, nonparametric confidence intervals

for the quantiles of a continuous distribution are

offered.

1. Computation of the Test StatisticFor single-sample data. the test statisti, K

is computed as the number of observations Xi greater

than the hypothesized median M0 . For the paired-sample

case, K is the number of differences Xi- Yi that exceed

M0 . All observations Xi (or Xi- Yi) that are equal to

MO are ignored and the sample size decreased

accordingly. As long as the number of such ties is

small relative to the size of the sample, the test

results are not greatly affected. Gibbons CRef. 2:pp.

108].

2. The Null and Asymntotic Distribution of K

The null distribution of K is binomial

with p = .5. in Table 1, the exact values of the

C.D.F are compared with the corresponding approximate

values using a normal approximation with and without

continuity correction.

TABLE 1. C.D.F. COMPARISONS FOR THE SIGN TEST

PROBC I kJ; FOR SAMPLI SIZE EQUAL TO 24.

TEST STAT. VALUE I 51 6 1 7 8 I 9 1 01 I 1iI I I I I'

EXACT C.D. . I .00331 I .01±33 I .0396 I .07579 I .15373 I .27003 I .41941I I I I I I [

ERROR; NORMAL I .00117 I .00407 I .01134 I .02456 I .04339 I .06352 I .07786I I I I I I

ERROR; NORM. /CC 1-.00068 1V.00104 1V.00114 1V.00073 1 .00001 I .00043 I .00028I I I ,

PROSEX .k]; FOR SAMPLE SIZE EQUAL TO 25.

TEST STAT. VALUE I 6 1 7 1 8 1 9 I 10 I 11 I 12I I I Ii

EXACT C.D.F. I .00732 I .02164 I .05338 I .11476 I 2121 I .34502 I .50000I I ,

0' 'I21A 0774 2±17?5 .)' 0 5,3 :.o7.)7

!02M. 'L .00,33 011 1 1 00092 .0031 .00032 1 .00044 .3 00 .16i I

16 ,

As can be seen, for sample sizes greater than

25, a normal approximation with continuity correction

is accurate to at least three decimal places.

3. Hypothesis Testing

P-values are computed for three basic

hypotheses comparing the median of the population or

the median of the population of differences M with a

hypothesized median M0 . P-values are taken from the

cumulative distribution of the binomial for the

following tests of hypothesis.

a. One-sided Tests

(1) H0: M = Mo Versus HI: M < Mo. The

P-value equals Pr[K < k], where k is the computed

value of the test statistic.

(2) H0: M = M0 Versus HI: M > M0 . The

P-value equals Pr[K > k].

b. Two-sided Test.

(1) H0: M = M0 Versus Hi: M 0 M0 . The

P-value equals twice the smaller value of a(1 ) or

a(2), but does not exceed the value one.

For sample sizes greater than 25, a normal

approximation with continuity =orrectin is used.

4. Confidence interva. Zs,:imation

Confidence intervals for the population

median are based on the ordered observations in the

sample. For paired-sample data, confidence bounds are

4. P IF IP e

obtained from the ordered differences of the pairs of

data. A 100(1- a )% confidence interval is determined

in the following manner. Let k be the number such that

Pr[K < k] < ( a/2). Then, the (k+i)th and (N-k)th

order statistics constitute the end points of the

confidence interval. Gibbons [Ref. 1: pp.104].

For computing confidence intervals when

sample size N is greater than 25, a normal

approximation with continuity correction is used.

Also included under this test is an option to

generate nonparametric confidence intervals for any

specified quantile given a sample size N from a

continuous distribution. The end points of the

intervals are sample order statistics.

B. WILCOXON SIGNED-RANK TEST

The signed-rank test requires the added assumption

that the underlying distribution is symmetric. This

test uses the ranks of the differences Xi- M0 (or Xi-

Yi- M0 ) together with the signs of these differences to

determine the test statistic. Confidence intervals for

the median can also be constructed.

1. omoutat.Zn cDt' the res- StatestL;

For single-sample data, the test statistic W

is computed as follows.

.'*

• - " ... ' . . - -. :. - -.. 2. .- . .. . .. . . . . '-

L 1 if Xi - MO > 0Let Zj

I 0 if Xi- M0 <0

Nand let ri - rank(IXi- Mo). Then, W = Ziri.

j=1

For paired sample data, W is calculated in

the same manner, except the differences to be ranked

are the paired-differences minus the hypothesized

median. Zero differences are ignored and the sample

size is decreased accordingly. When ties occur between

ranks, the average value of the ranks involved are

assigned to the tied positions. It has been shown that

a moderate number of ties and zero differences has

little effect on the test results.1

2. The Null and Asymptotic Distribution of W

The exact null distribution of W is given by:

Pr[W = w] = UN(w)/2N, w - 0,1,2,...,N(N+1)/2, where

UN(w) is the number of ways to assign plus and minus

signs to the first N integers such that the sur e

positive integers equals w. It can be shown "ee

Gibbons [Ref. 1:pp. 112)) that uN(w), for successive

values of N, can be computed using the recursive

relationship:

UN(w) = UNI(W-N) - UN_1(w)

IFor more information on the effects that zeros andtied ranks have on the Wilcoxon Signed-Rank Test, seePratt and Gibbons [Ref. 3).

Exact C.D.F. values were compared with those

obtained using the following asymtotic approximations:

student's T with (N-i) degrees of freedom (T),

student's T with continuity correction (TC), normal

(Z), normal with a contin ity correction (ZC), the

average of T and Z as suggested by Iman [Ref. 4], and

the average of TC and ZC.

As can be seen in Table 2 below, the average

of TC and ZC gives the most consistently accurate

results with three decimal place accuracy when the

sample size exceeds 9.

TABLE 2. C.D.F. COMPARISONS FOR THE WILCOXON

SIGNED-RANK TEST

PROBEW i w]; FOR SAMPLE SIZE EQUAL TO 9.

TEST STAT. VALUEI 3 5 1 6 8 9 12 1 14

EXACT C.D.F. 1 .00977 I .01953 .02734 1 .04883 1 .06445 1 .12500 1 .17969

ERROR; NORMAL 1".00067 I .00046 1 .00204 1 .00591 1 .00958 1 .01824 1 .02272

ERROR; NORM. W/CC 1".00243 I.00247 1".00167 1 .00023 1 .00269 1 .00693 1 .00806

ER02; T DIST 1 .00518 I .00608 1 00673 1 .00701 1 .00817 1 .00826 1 .00818I i i

ERROR; T WCC 1 .00355 I .00276 1 .00233 I .00012 1 -.00010 1.00437 1 .00725

E2OR; AVE T/2 1 .00225 I .00327 I .00438 1 .00646 1 .00888 1 .0325 .01545

1.0: 56E 02 . .. . 4 .. j033 , 0 17 I -3

10j------------

TABLE 2. (Continued)

PROBEW 5 w]; FOR SAMPLE SI2. EQUAL TO W0.

TEST STAT. VALUEI 5 I 7 I 8 I ±O I 12 I 15 I £7.0097 I .044 .04±9

EXACT C..F. 1 .01355 11 .06!43 .1621 1 .161!3

ERROR: NORMAL I .0Oii5 1 .00023 I .00099 1 .00476 1 .00837 1 .01490 1 .01888

ERROR; NORM. W/CC I-.00270 I .002i9' I .00198 1 .00043 I .00229 I .00558 1 .00710

ERROR: T DI3T I .00399 I .0052 I .00525 .00665 I .00661 I .00661 .00681

ERROR; T U/CC I .00249 I .00243 1 .00182 1.00151 1".00052 1 .00376 I .00571I I I I I

ERROR: AVE T/Z 1.00142 1 .00267 1 .00312 1 .00571 .00749 I .01075 1 .01234

ERP.OR- AVE 1,'C I .000iO .'00i2 .00008 i .00097 .00089 I .)009i .0070I I I

3. Hypothesis Testing

P-values are computed for three basic

hypotheses comparing the median of the population or

the median of the population of differences M with a

hypothesized median M0 as shown below.

a. One-sided Tests

(1) HO: M = M0 Versus HI: M < M0 . The

P-value equals Pr[W < w], where w is the computed

value of the test statistic W.

(2) H0: M = M0 Versus Hi: M > M0 . The

P-value equals PrLW > w].

b. Two-s ided Test

(1) H0: M = M0 Versus Hi: M # M0 . The

P-value equals twice the smaller value of a(l) or

a(2), but not exceeding the value one.

UWWWWTWWWWWJWWN

For sample sizes greater than 9, an average

of the normal and student's T approximations, each with

continuity correction, is used. Computations of the P-

value for each of the alternative hypotheses are:

a. HI: < M

Let PZC = Pr[ Z < (w +.5 - Mw)/ w] and

let PTC = Pr T( - ) <w -___uwl___.5

gaw2 [Lw - jwj - .5] 2 -5N - I

where Z is standard normal, T(N_1) has a student's T

distribution with (N-I) degrees of freedom, Uw =

N(N+1)/4 and aw2 = (N(N+I1)(2N+1)/24.)). Then, the

P-value for the test is (PZC + (1 - PTC))/ 2 if w is

less than mw and (Pzc + PTC)/2, otherwise. The above

formulas are obtained from those given by Iman [Ref. 4]

after inclusion of a continuity correction.

b. Hi: M > M0

The P-value equals ((1 - PZC) + PTC)/2 if

w is less than Pw and ((1 - PZC) + (1 - PTC))/2,

otherwise. The computation of PZC and PTC is similar

to the above excent the sign of the ccn-inu-7

correction is changed.

c. Hi: M * M0

The P-value equals twice the smaller

value of a or b above, but not exceeding the value one.

4. Confidence Interval Estimation

For single-sample data, the confidence

interval for the population median is based on the

ordered averages of all pairs of observations (Xji+

Xj)/2 such that i < J. A 100(1- a)% confidence

interval is determined in the following manner. Let w

be the number such that PrCW < w] < ( a/2). Th'en, the

(w+1)th and (m-w)th order statistics, where m =

N(N+1)/2 or the total number of paired-averages,

constitute the end points of the confidence interval.

A confidence interval for paired-sample data is

computed in the same manner, except the end points are

taken from the paired-averages of the differences X i -

Yi. Gibbons [Ref. 1:pp. 114-118].

For computing confidence intervals when

sample sizes are greater than 9, a normal approximation

with continuity correction is used.

V.TESTS BASED ON TWO OR MORE SAMPLES

The tests assume that the data consists of

independent random samples from two or more continuous

distributions. The general null hypothesis is that the

samples are drawn from identical populations. The

Mann-Whitney and Kruskal-Wallis tests are considered.

A. MANN-WHITNEY TEST

The Mann-Whitney test is based on the distribution

of the test statistic U, which can be used to compare

the equality of the population medians or variances

for two samples.1 The Mann-whitney test with a

modified ranking scheme can be used to test for

equality of variances if the population means or

medians are assumed to be equal (Conover [Ref. 5:pp.

229-230]). If the medians differ by a known amount, the

data can be adjusted before applying the test. A

confidence interval for the difference in the medians

of the two populations can also be estimated.

1. Computation of the Test Statistic

For -he comparison Df Dopulat.on medilans. -he

test statistic U is computed from the combined ordered

1The test statistic U and the method used tocompute it are taken from Gibbons [Ref. 1:pp. 140-141].

! . . . , - -. " " ".... < -:< ...'.'. . ',. ' . . '..''....'..:, '[,:, .-. ' .. ,:.' .' .'.'': .... . .

arrangement of observations Xi and Yj, i =

J = 1,2,...,M. Let r i = rank(X i ) in the combinedN

ordered sample and RX = ri. Then,i=I

U = RX - M(M+1)/2.

For testing the equality of variances, the

computation of U is similar except for the method of

assigning ranks to the ordered sample. This method

ranks the smallest value 1, largest value 2, second

largest value 3, second smallest value 4, and so on

by two's, until the middle of the combined ordered

sample is reached. For either test, tied ranks for the

combined sample are assigned the average value of the

ranks involved. A moderate number of ties has little

effect on the test results

2. The Null and Asymptotic Distribution of U

The exact null distribution of U is

determined using a recursion algorithm due to Harding

[Ref. 6].

Exact C.D.F. values were compared with

approximate values obtained from the following

as7mtotic distributions: student's T 'ith (n-2) degrees

of freedom where n = N + M, the total number of

observations in both samples (T), student's T with

continuity correction (TC), normal (Z), normal with

continuity correction (ZC), the average of T and Z

. . . . . . -* . .

(Iman [Ref. 7]), and the average of TC and ZC. The

results for various sample sizes are given in Table 3.

TABLE 3. C.D.F. COMPARISONS FOR THE MANN-WHITNEYTEST

PROC U i u: FOR SAMPLE SIZES 4 EQUAL n. 9 AN4D M EQUAL 70 9.

TE3T STAT. VALUEI 14 I 7 1 i8 I 2 I 23 I 27 I 29

EXACT C.. F. I .00933 i .0999 I .0255 I .04636 1 .06796 1 .12904 I .17005

ERROR; NORMAL -.00026 1 .0000 1 .00163 .00441 1 .00632 1 .0±243 i .0±51±

ERROR; NORM. W/CC 1".00146 1-.00114 I-.00032 .00026 1 .00±30 .00354 I .00436

12..OR; T )137 1 .0237 .00337 i .00372 .00464 .005-5 .00676 1 .00779

ERROR; T (WCC I .00i19 1 .00±08 1 .00095 1 .00007 1".00073 1".00259 I .00323

122O2; AVE -/: I .00105 1 .0029 .00270 1 .00453 1 .00603 I .00959 1 .01145

E2ROR; AYE 7C,'."C I.00014 1 .00003 1 .00004 1 .0001- 1 .00026 I .00048 I .00054* I I 4

PROBEU i u3; FOR SAMPLE SIZES N EQUAL TO 7 AND M EQUAL TO 12.

TEST STAT. VALUE 14 1 071 19 21 24 1 27 30

XXACT C.D.F. 1 .00853 I .0792 I .02732 1 .04156 1 .07111 .11342 .17012I I ,

E202; NORMAL 1".00045 I .00062 I .00±87 1 .00359 1 .00701 i .0±097 1 .01487I .

ERROR; NORM. W/CC 1".00152 I.00128 I.00079 1".00003 1 .0054 1 .00322 1 .00453

ERROR; T DIST 1 .0099 I .00292 I .00356 1 .00422 1 .00527 1 .00643 1 .00796 %i i. FiIiiI

ERROR; T W/CC 1 .00094 I .00092 I .00068 1 .00026 1".00066 " .00i78 [".002629 '

E220P.; AVE T: 1 .00077 1 .00177 .00 272 .0039 t1 .00614 1 .00870 1 .01±42

RR2: ZYE 7:':: ",- -. 30, . ." ..0044.,307C2

26 .

TABLE 3. (Continued)

?!03tU ~.ul; FOR SAMPLI 31:13 3 EQUAL TO 1. AND I EQUAL TO L7.

T S TAT. VALUE 1 13 1 16 1 13 I 20 i 23 I 27 I 30

E xACT C.. .00963 1 .0933 1 .02916 .04245 .07013 1 ,44O t . 734

E2ROR; OR4IAL 00077 .00039 1 .00101 .00345 .30639 1 .01210 1 .01!64

122OR: .O42. 1'r, .0090 ". i0±0 ".00037 .00007 .30133 .004A4 .00573

1102: 7 3 3 302t0 .3022 .002,99 .00390 1 .)0544 .00774 .00'3

122OR; T W/CC 1 .000? 1 .00023 1 .00023 I .00021 1 .00007 1".00025 -.00051

E2OR: AV! T,: 30027 .301'0 .30225 .30363 .30616 I .3099 .3±IZ0

?20.30AVE ,3 06 i ".30063 )0032 .)0014 .30037 .0209 .30264

?ROEU i u1 FOR SA1PLI SIZES N 1QUAL TO 3 AND M ZGUAL 73 27.

13 STAT. VALUE I 7 1 10 I 12 I5 I 10 I 23 I 26

EXACT C.D.Y. I .00764 1 .01650 1 02512 1 .04236 1 .07734 I .12660 I .17433

ERROR; NORMAL 1V.00265 ".00099 1 .00072 1 .00389 1 .00874 1 .0342 1 01680

ERROR; NORM. W/CC I.00363 1".00254 -.0033 .00089 .00405 1 .00665 1 .00831II I I

ERROR; T DIST I.00121 1 .00031 1 .00173 1 .00414 1 .00741 I .01026 1 .0±253i -i i i

ERROR; T W/CC 1-.00219 1".00129 1".00042 1 .00097 1 .00250 1 .00323 1 .00390

ERROR; AVE T/Z I.00193 1-.00034 1.00122 1 .00402 1 .00808 I .01±84 I .0±466I I

ERROR- AYE TC,/Z I .00291 1.00092 t .00087 .00093 1 .00327 1 .00496 1 .0060

As c-n be seen from the tables. the 3ver2. e

o f ZC and TC g-ves :he most cons.stanz'1 accurate

results. For sample sizes NxM > 80, nearly three

decimal place accuracy is obtained in all cases.

. . . . .. . . . . ./

3. Hypothesis Testing

P-values are computed for three basic

hypotheses comparing the medians or variances of the

two populations as shown below.

a. One-sided Tests

(i) HO: MX = My Versus HI: MX < My or

HO: VX = Vy Versus HI: VX > Vy. The P-value equals

Pr[U < u], where u is the observed value of the test

statistic.

(2) HO: MX = My Versus Hi: MX > My or

HO: VX = Vy Versus HI: VX < Vy. The P-value equals

Pr[u > u].

b. Two-sided Test

(1) HO: MX = My Versus Hi: MX # My or

HO: VX = Vy Versus HI: VX # Vy. The P-value equals

twice the smaller value of a(l) or a(2), but not

exceeding the value one.

For sample sizes NxM greater than 80, the

average of the normal and student's T approximations,

each with continuity correction, is used. Computations

of the P-value for each alternative hypothesis are:

a. HI: MX < My or Vx > Vy

Let PZC Pr[ Z < (u +.5 - pu)/ au] and

le T r_~-)j - ,Uu) - .5

l (N+M-li) a 2 [ iu-g,. ! -5 ] -5N+M-2 N+M-2

where Z is standard normal, T(n_2) has a student's T

distribution with (n-2) degrees of freedom, Pu = NxM/2

and au 2 = (N(M)(N+M+1))/12. Then the P-value for the

test is (Pzc + (1 - PTC))/ 2 for u less than 9u and

(PZC + PTC)/ 2 , otherwise. The above formulas are

obtained from those given by Iman [Ref. 7] after

inclusion of the continuity correction.

b. Hi: MX > My or VX < Vy

The P-value equals ((1 - PZC) + PTC)/2 if

u is less than pu and ((1 - PZC) + (1 - PTC))/2,

otherwise. The computation of PZC and PTC is similar

to the above except the sign of the continuity

correction is changed.

c. Hi: MX 0 My or VX * Vy

The P-value equals twice the smaller

value of a or b above, but not exceeding the value one.

4. Confidence r.n er. a Estma tiorn

Confidence intervals for the difference in

medians, (My - MX), are based on the ordered

arrangement of the differences (Yj- Xi), J = 1,2,...,M;

1 = 1,2,...,N for all I and j. A 100(1- a)%

confidence interval is determined in the following

manner. Let u be the number such that Pr[U < u] <.

( a /2). Then, the (u+1)th and (m-u)th order

statistics, where m = NxM or the total number of

possible differences, constitute the end points of the

confidence interval.

For computing confidence intervals when

sample sizes NxM are greater than 80, a normal

approximation with continuity correction is used.

B. KRUSKAL-WALLIS TEST

The Kruskal-Wallis test is a nonparametric analog

of the one-way classification analysis of variance test

for equality of several population medians. Gibbons

[Ref. l:pp. 99].

1. Computation of the Test Statistic

Calculations of the test statistic H center

around the ordered arrangement of the combined samples

from which the sum of ranks for each sample is derived.

Let Xij, J=1,2,...,ni and i=1,2,...,k, be independent

random samples from k populations. Let rij = rank(Xij),

R - rij, and N = ni. Then,J=1i=

H = (12/(N(N+1)) (R1 2 /n i ) - 3(N+1)

If ties occur in the combined sample, they

are resolved by assigning the average value of the

ranks involved. A correction based on the number of

observations tied at a given rank and the number of

ranks involved, is included in the calculations. A

complete description of the correction factor is given

in Gibbons [Ref. 2:pp. 178-179].

2. The Null and Asymptotic distribution of H

The null distribution of H is generated by

enumeration. Each possible permutation of ranks is

listed for the combined sample, and the corresponding H

value computed. The frequency distribution of H is the

total number of occurrences of each distinct H value.

The H values are arranged in increasing order while

maintaining the frequency pairings. The null

distribution is obtained by dividing the cumulative

frequencies by nl!n2!...nk!/N!.

Due to computer limitations, generation of

the exact distribution of H was only possible for k = 3

populations with n = 4 observations in each, and 4

populations with 3 observations in each. Most of the

distributions were generated on the mainframe computer

and saved in matrices for quick recall by the Kruskal-

Wallis test program.

Exact C.D.F. values were compared with the

corresponding approximate values using the following

- ' * . z

distributions: chi-square with (k-1) degrees of freedom

(C), F distribution with (k-1) and (N-k) degrees of

freedom (F), and F with (k-1) and (N-k-i) degrees of

freedom (Fl). The chi-square distribution uses the

Kruskal-Wallis H statistic, while the F and F1

distributions use a modified H statistic, Hi

((N-k)H)/(k-1)((N-I)-H); see Iman and Davenport CRef.

8]. As can be seen in Table 4, F1 gives the most

consistently accurate estimates.

TABLE 4. C.D.F. COMPARISONS FOR THE KRUSKAL-WALLIS TEST

PROS E I h: ?OR 4 GROUP OF I SAMPLES CONZ5STING OF , 4, AND 3 9S.

TZS!' STAT. 7ALLIE 17.1439 5 .7121 I5.1318 15.5,985 3. -30 '.2121 :9585EXACT C.D.Z. .00970 L .01905 i .0266 1 .3 866 .079±0 I .129i8 .1778U

--------------------------------- --- ---.------

SRROR; CHISQURE V-1.8 0 i-.01582 [.01585 -. 01220 1-.30814 1 .00746 I .01241----------------- t------------.4-----------------

ERROR; F DIST I .00304 I .00736 I .00836 i .01113 I .01820 I .01696 I .00990------- . +------------.--------------------- -------------------

ERROR; F / DF I .00084 I .00U26 I .00403 I .005U4 I .01138 I .00895 I .00172

PROBCR Z hJ; FOR A GROUP OF 3 SAMPLES CONSISTING OF 4, 34, AND % OBS.TEST STAT. VALUE I 7.6538 I 6.9615 I 6.5000 I 5.6923 I 4.g615 I 4.2692 I 3.5769 "

-------4.-----------4.-----------4.-----------4----------.------.------4-----EXACT C.D.P. I .00762 I .01939 I .02996 I .04866 I .08000 I .12190 I .17299

---------------------------------------------4 %.4 4

ERROR; CISQUARE I-.01416 1-.01139 1-.00882 1-.0094 1-.00368 1 .00361 I .005774.------.-------4------------4------------4------------4-----------------------

ERROR; F DIST I .00290 I .00839 I .01204 I .01100 1 .01272 I .01225 1 .00263i.--------4---F;=---------+------------4------------4------------4----------------------ERROR: F I/1 DF I .00149 I .00600 I .00890 I .00647 I .00708 I .00592 1.00384

PROBES k h3; FOR A GROUP OF 4 SAMPLES CONSISTING OF 3. 3, 2, AND 2 0S.TEST STAT. VALUE I 7.6364 7.1818 I 7.0000 I 6.5273 I 6.0182 I 5.3818 I 4.8727

-----4--------4.--------4.--------.--------4---------4------- ------ 4-- --- -----

EXACT C.D.P. 1 .01000 I .01921 1 .02921 I .04921 I .07184 I .12984 I .17952-------4.. ... .. . .------------ .-------------4......------ --------------.----------

ERROR; CRISQUARE 1.04u16 1".04712 1".04269 1-.03939 1-.03089 1".01604 1".00183

ii; ------- 4------------4------------- 4--- 4---.4._+--------------------...EOR: F Drs., I .002SLL -'0250 1 .0073C !.00880 .01094 I12 .00902

ORRCR: . W/- F .J01: -. C -. 02C-.007 - - -5 -JC50 -. 08u3

PROBES I hl. FOR A GROUP OF 4 SAMPLES CONSISTING OF 3, 3, 3, AND 2 OBS.TEST STAT. VALUE I 8.0152 I 7.6364 I 7.1515 I 6.7273 1 6.1970 5.4697 I 4.9697

--~-4------4------4------.------------4------- ------ 4----------------------

EXACT G.D.P. I .00961 I .01831 I .02974 j .0 9 8 I .07805 I .12740 I .17571---- 4-------.-------.------4----- ----- 4._+ ------- -------- 4=------ --- _+:--------------

ERROR; CRISQUARE V.03609 1.03584 1 .037148 1-.0316 1".02436 1.01306 I .00168-----4-------4.---------- .4------------4------------4-------------4--------- -------

ERROR; F DIST I .00215 I .00481 I .00U41 I .00920 1 .01185 I .01036 I .01214------------------- 4------------4.--------- ---- 4-------------4--------- ---- 4---- ------

ERROR; F Y/I OF .00133 1.0001 I.00269 I .00030 1 .00098 1.00230 1.00091

16*

A ifinal accuracy comparison between the C and F1

approximations was conducted by computer simulation for

5 populations with 8 observations each. Initially,

30,000 permutations of the 40 ranks were randomly

generated (no tie ranks allowed), and the H statistic

calculated for each permutation. Then the empirically

determined percentiles Hp for selected values of p

between .01 and .18 were compared with the

approximations given by the C and F1 distributions.

The results are shown in Table 5. It can be seen that

the F1 approximation compares well with the simulated

results, giving three decimal place accuracy, while

the C approximation is less accurate.

TABLE 5. C.D.F. COMPARISONS FOR THE KRUSKAL-WALLIS TEST USING COMPUTER SIMULATION

PROBES ~h : BASED ON 10000 GENERATEfD H'S FOR 5 SAMPLES OF 8 OHS. EACH.

TES STAT 1AU -0 12.2 2I 11.06 8- -0 I -9.2121 --- 8.129 --- 7.030- -- 6.232----------------------------- ++------------4------------4----------------------

C.D.F. VALUE 1 .01000 1 .02000 1 .03000 1 .05000 1 .08000 1 .13000 1 .18000-------------- ---------------4.+4 ------------ 4-------------4----------------------

ERROR; CUISQUARE 1Z-.00573 1-.00584 1-.00646 1-.00601 1-.00696 1-.00432 1-.002464.------------------------ -- -. 4------------4----------------------

IRRR; iii D I.00081 1 .00231 1 .00259 I .00350 1 .00161 1 .00109 1-.00103

PROBES k hl-. BASED ON 20000 GErNERATEPD HIS FOR 5 SAMPLES OF 8 OHS. EACH.

TETSAT-AU--1.1 I15 I184- --- 9.163 --1-- I 7034 6.220--------------------4-------4--------4------------4------------4-------------4----------------------C.D.F.VALUE 1 .01000 1 .02000 1 .03000 1 .05000 1 .08000 1 .13000 1 .18000---------------------------------- 4------------4------------4-------------4-----------------------

ERROR:-- CHS-rsQU-ARE-1-.00516 1-.00596 1-.007u5 1-.00716 1-.00696 1-.00413 1-.00334---------------------.----------- 4---------------------------;----------4. ----------- ;--+=---------

0ROR W1 -IF .00.26 F .0021 .001E6 .0023- 1 .00161 1 .3013O1 .00199

?R0 CH 3ASED ),V ',J,) 0 ;MVERATE: 31 5 ?OR 3 7,;,MPLZS )F 3 .,'8S. ZAC7.

----------------------------4-------.--.--- 4

------------------.------.--------4------------4------------4-------------4-------------4----------

* RROR: CHIrSQARE -1-.00522 1-.00684 1--00802 1-.00677 1I.00563 1-.00213 1-.00020--------------------4-------4--------4------------4------------4------------4-----------------------ERRO: FW/1 ? I.00121 1 .00143 1 .00111 1 .00273 1 .00301 1 .00344 1 .00142

3. Hypothesis Testing

P-values for the test HO: the population

medians are all equal versus Hi: at least two

population medians are not equal, are computed as:

Pr[H > hi, where h is the value of the observed test

statistic.

For three or more populations with at least 4

observations in each, the F1 approximation is used.

VI.TESTS FOR ASSOCIATION IN PAIRED-SAMPLES

The tests described herein assume that the data

consists of independent pairs of observations (Xi,Yi)

from a bivariate distribution. The general null

hypothesis is that of no association between X and Y.

Kendall's B and Spearman's R are considered.

A. KENDALL'S B .v

Computation of the Test Statistic

The test statistic is computed by comparing

each observation (Xi.Yi) with all other observations

(XjYj) in the sample. If the changes in X and Y are -?

of the same sign, sgn(Xj - Xi) = sgn(Yj - Yi), the pair

(Xi,Yi) and (Xj,Yj) is "concordant" and a +1 is scored.

If the signs are different, the pair is "discordant"

and a -1 is scored. Any ties between either the X's or

the Y's scores a zero for that pair. The sum of all

scores divided by the total number of distinguishable

pairs, (N(N-1))/2, gives B. If zeros are scored, the

denominator is reduced by a correction factor which is

basea on 'he number of 'observations tied at a given

rank and the number of ranks involved in each of the X

and Y samples. A complete description of the

correction for ties is given in Gibbons [Ref. 2:pp.

N V-V

289). The value of B ranges between 1, indicating

perfect concordance, and -1, for perfect discordance.

Gibbons [Ref. 1:pp. 209-225].

2. The Null and AsymDtotic Distribution of B

The null distribution of B is derived from

the following recursive formula given in Gibbons [Ref.

1 :pp.216].

u(N+1,P) = u(N,P) + u(N,P-1) + u(N,P-2) + ... + u(N,P-N)

where u(N,P) denotes the number of P concordant

pairings of N ranks. This formula is used to generate

the frequency with which the possible values of P

occur. Division by N! results in the probability

distribution of P. Since, B = (4P/(N(N-1)))-1, the

null distribution of B is easily determined.

Exact C.D.F. values were compared with those

obtained using a normal approximation, with and without

a continuity correction factor (CC = 6/N(N 2 -1), pro-

posed by Pittman [Ref. 11] for the Spearman's R test).

The results for various sample sizes are provided in

Table 6. As can be seen, for sample sizes greater than

13, a normal approximation with continuity orrectin

proviaes three decimal place accuracy.

N. W

TABLE 6. C.D.F. COMPARISONS FOR KENDALL'S B

PR~3 b]; FOR SAMPLE SIZE EQUAL TO ±.3.

IEST STAT. VALUE 0.±28i 0.4615 0.4±03 I 0.3590 1 0.3333 1 0.2564 1 0.2303

EXACT C.D.F. .00748 .0524 1 .02363 1 .04999 .06443 I .2593.i .1!309

ERROR; NORMAL 1.00014 1.00121 1 .00313 I .00620 1 .00808 I .01473 1 .01703

ERROR; 1PRM. W/CC ".00013 1 .00073 .00239 .00497 .00658 .01223 .014:5

PROBCB 2 b]; FOR SAMPLE SIZE EQUAL TO 14.

TEST STAT. VALUE 0.472! 1 0.4236 1 0.4066 1 0.3626 0.2967 0.2527 0.2033

EXACT C.I.F. .00964 i.1773 1 .323 9 .03973 .07353 ,I S! .6541

ERROR: NORMAL I .00035 .00140 .00213 .0043i .00839 .0±2! i .31627

EROR; 4ORM. WCC 1 .00008 ! .10035 .00009 .00345 , 0074 f .310,57 .)1371

3. Hypothesis Testing

P-values for tests of no association between

X and Y are computed for three types of alternative

hypotheses. Because the distribution of B is symmetric,

all probabilities can be taken from the upper tail

using the absolute value of b, the observed value of

the test statistic. Linear interpolation is used when

b lies between tabulated values. The P-values are

'Cmruted as f-L'ows.

a. One-Sided Alternatives

The one-sided alternative tested depends

on the sign of b. A positive b will automatically test

g _ .. ,-..r, . . - .- , . ' - "" "-"- " '. -" "" -' " "-" "'" " " "-". -" . " - - -" 1- ""**" *, ***.": '

I k -V Irv

for direct association or concordance, while a negative

b will test for indirect association or discordance.

The P-value equals Pr[B > IbIl].

b. Two-Sided Alternative

The P-value equals twice the probability

computed for the one-sided hypothesis.

For sample sizes greater than 12, a normal

approximation with continuity correction is used. The

approximate P-values are then:

1 - Pr[ Z < ((fbj - CC) - b)/ ab], where Z

is standard normal, CC is the continuity correction,.

9b = 0, and b2 = (4N + 10)/9N(N-I ), for the one-

sided test and twice this P-value for the two-sided

test.

B. SPEARMAN'S R

The Spearman's R Test requires the added

assumption that the underlying bivariate distribution

is continuous. The test measures the degree of

correspondence between rankings, instead of the actual

variate values, and can be used as a measure of

association between X and Y. Gibbons [Ref. 1:pp. 226].

:o~mDita::on of The Tes Sta:;st/2

The test statistic R is computed in the

following manner. Let ri = rank(Xi) and si = rank(Yi)

and Di = ri - si . Then,

.4~~~~~~~~~~~~~~ N44 . .~~ S* .~.- .-.. S..~-

6 DR = 1 - i:=1R

N(N 2 - - 1)

where N is the size of the sample. If ties occur in X

or Y, they are resolved by assigning the average value

of the ranks involved. A correction factor, based on

the number of observations tied at a given rang and -he

number of ranks involved, is included in the

calculacions. A complete iescriotion f 1he

-:correction factor is given in Gibbons [Ref. 2:pp. 279].

The value of R ranges between 1, indicating perf.ct

direct associat-on, and - , for perfect indirect

association. Gibbons [Ref. 1:pp. 226-235].

2. The Null and Asymptotic Distribution of R

The null distribution of R for a given sample

size N is generated by enumeration. The method, as

presented in Kendall [Ref. 9], involves generation of

an N by N array of all possible squared differences

between any two paired ranks of X and Y. All N!

permutations of N ranks are used to index values from

the array. The sum of these indexed values for each

permutation - 'nves - .se - ,D sum .f squarea1 iiffeences

which are then converted to the R statistic. The

frequency distribution of R is the total number of

occurrences of each distinct value of R divided by N!.

Due to mainframe computer memory limitations

in the APL environment, generation of the distribution

of R was limited to sample sizes of 7 or less.1 Using

tables, provided by Gibbons [Ref. 2:pp. 417-418] to

supplement computer computations, a numerical matrix,

called PMATSP, was created to store the cumulative

distributions of R for sample sizes less than 11. This

matrix allows for quick recall of cumulative

probabilities by the Spearman's R Test program.

Exact C.D.F. values were compared with those

obtained using a student's T approximation with (N-2)

degrees of freedom (see Glasser and Winter [Ref. 10]),

and a normal approximation. Both normal and T

approximations were computed with and without a

continutity correction factor, CC = 6/N(N 2 -1) (Pittman

[Ref. 11]). From the results presented in Table 7, the

most consistently accurate approximation is given by

the T distribution with a correction.

3. Hypothesis Testing

P-values for tests of no association between

X and Y can be computed for three types of alternative

hypotheses. Because the listribution of R is symmetric.

all probabilities are taken from the upper tail using

1 The memory capacity of the mainframe computer inthe APL environment is limited to 2.5 megabytes.

TABLE 7. C.D.F. COMPAR:SONS FOR SPEARMAN'S R

PROBER 1 r]; FOR SAMPLE Sl= EaUAL TO 9.

TEST STAT. VALUE I 0.7333 I 0.7167 I 0.6667 1 0.6000 1 0.5333 1 0.4333 0.3500* ,I060 053 .33I 030

IX AC 2 .Y. .30361 I 01343 .3944 I .)4040 ..07376 .12496 i .179 9

E2202; NORMAL (-.0047 ".00290 1 00022 I .0'35 1 .00305 .0±479 I .0319

ERROR: NORM. Y/CC 1'.00553 1-.0043 I .00095 I .00123 I .00493 1 .0±029 I .0±236

E2RO; ) 31 1 .00235 1 .00352 I .00451 I .00459 .)0415 1 .002938 I .00±38

ERO; T /CC 1 .00153 1 .00203 I .00251 I .00175 I .00042 1-.00212 I.00479* I II

?20 03 2 1 0! RO 3A1,1,LZ 3::- -_,UAL 70 _'O.

7'37 STAT. VALUE I 0.7455 1 0.i727 1 0.6364 0. 5636 ,). 4*3 1).406' 1 0.3333

IXACT 0.3.1. .00370 .0943 .02722 .04314 .7.41 .12374 .17437

12202; A4ORMAL 1 3096 " .0230 I .009± .30 27. i .00700 .)121 .31673

ERROR; NORM. W/CC I .00457 1.00327 1 .00210 1 .00095 1 .00451 1 .00867 1 .0i±23... . I

ERROR; T DIST i .00204 1 .00296 I .00326 I .00323 I .00253 1 .00±59 1 .00107i 1~I

ERROR; T U/CC I .00144 1 .00±35 I .00±34 I .0014 1-.00035 1.00230 1V.00361I I I I I - I

the absolute value of r, the observed value of the test

statistic. The P-values are computed as follows.

a. One-Sided Alternatives

The one-sided alternative tested depends

on the sign of r. A positive r 'ir1! test f:r -

associ~a:z n, w negative r tests for Lnalrecz

association. The P-value equals Pr[R > Irl].

b. Two-Sided Alternative

The P-value equals twice the probability

computed for the one-sided hypothesis.

For sample sizes greater than 10, an

approximation based on the student's T distribution

with (N-2) degrees of freedom and continuity

correction, is used. The P-valuels are:

1 - PrE T(N_2) . ((Ir - CC) - 9r)/ ar],

where T(N_2) denotes the T distribution with (N-2)

degrees of freedom, CC is the continuity correction,

Mr = 0. and ar 2 = (1 -(Irf-CC) 2 )/(N-2), for the one-

sided test,. and twice this P-value for the two-sided

test. Gibbons CRef. I:pp. 218].

42 ,

VII.NONPARAMETRIC SIMPLE LINEAR REGRESSION I

Nonparametric Linear Regression assumes that the

data consists of independent pairs of observations from

a bivariate distribution and that the regression of Y

on X is linear. The program estimates linear

regression parameters based on the data samples. It

then allows the user to input X values to predict the Y

values. Hypothesis testing and confidence interval

estimation for the slope of the regression equation is

offered. If the estimated slope lies outside the

confidence interval, an alternate regression equation

is offered with an opportunity to input X values to

predict the corresponding Y values.

A. COMPUTATION OF THE ESTIMATED REGRESSION EQUATION

The least squares method is used to estimate A and

B in the regression equation Yi = A + BXi + ei

(i=1,2,...N), where ei (unobservable errors) are

assumed to be independent and identically distributed.

A and B are computed from the following equations:

IExcept for program design considerations, theinformation and concepts provided in the section areparaphrased from Conover [Ref. 12 :pp. 263-271].

!* k 43

N N N

B = i=1 i=1 i=IN Xij2 - YX

i=I B

N N 2

Yi - BXi

B. HYPOTHESIS TESTING

P-values for testing hypotheses about the slope of

the regression equation are based on the Spearman's

rank correlation coefficient R between the Xi and U i =

Yi - BoXi, where B0 is the hypothesized slope. The

appropriate one-sided test of hypothesis, HO: B = B0

versus HI: B < B0 or Hi: B > B0 , is automatically

chosen based on the sign of the computed test statistic

r (positive r tests, HI: B > B0 ; negative r tests, HI:

B < B0 ). The P-value is computed as: Pr[R > Irl].

P-values for two-sided tests, HO: B = B0 versus HI: B #

B0 , are also presented.

For sample sizes N greater than 10, P-values are

approximated using a T distribution with (N-2) degrees

of freedom and contlnui-,y correction.

C. CONFIDENCE INTERVAL ESTIMATION

100(- Q )% confidence bounds for the slope

parameter B are determined as follows. The n possible

slopes, Sij = (Yi-Yj)/(Xi-Xj), are computed for all

pairs of data (X i ,Yi) and (Xj,Yj) such that i < J and

X i # Xj and rearranged in increasing order to give S(1)

< S( 2 ) < .• < s(n). Let w be the (1- a/2) percentile

of the distribution of Kendall's statistic with sample

size n• 1 Let d be the largest integer less than or

equal to (n-w)/2 and u the smallest integer greater

than or equal to (n+w)/2 + 1. Then S(d) and S(u) are

the desired lower and upper confidence bounds,

respectively.

For sample sizes larger than 13, a normal

approximation with continuity correction is used to

estimate the confidence intervals.

If the slope of the estimated regression equation

does not lie within the computed confidence interval,

the program automatically calculates a new regression

equation where the slope is the median of the two-point

slopes Sij and the intercept is the difference of the

medians of the X and Y samples, My- MX.2

IKendall's statistic is defined here as Nc - Nd,where Nc is the number of concordant pairs ofobservations and Nd is the number of discordant pairs.Conover rREF. 2:p. 256.

2 This procedure is recommended by Conover [REF.12:pp. 256].

VUII.AREAS FOR FURTHER WORK

To create a more versatile and powerful software

package, the NONPAR workspace could be expanded to

include some or all of the following nonparametric

tests: tests for randomness based on runs. Chisquare

and Kolmogorov-Smirnov(K-S) Goodness-of-fit tests,

Chisquare and K-S general two sample distribution

tests, Chisquare 'est for independence, and -the

Friedman test for association.

LIST OF REFERENCES

1. Gibbons, J. D., NonDarametric StatisticalInference, McGraw-Hill, Inc.. 1971.

2. Gibbons, J. D., Nonparametric Methods forQuantitative Analysis, Holt, Rinehart, andWinston, 1976.

3. Pratt, J. W. and Gibbons, J. D., Conceots ofNonparametric Theory, pp. 160-176, Springer-Verlag, Inc., 1981.

4.. Iman, R. L., "Use of a T-statistic as an Approxi-mation to the Exact Distribution of the WilcoxonSigned-ranks Test Statistic,' Communcazions inStatistics, vol. 3, no. 8, pp. 795-806, 1974.

5. Conover, W. J., Practical Nonparametric statis-

tics. John Wiley and Sons, Inc., 197! .

6. Harding, E. F., "An Efficient, Minimal-storageProcedure for Calculating the Mann-Whitney U,Generalized U and Similar Distributions," AppliedStatistics, vol. 33, no. 1, pp. 1-6, 1984.

7. Iman, R. L., "An Approximation to the ExactDistribution of the Wilcoxon-Mann-Whitney Rank SumTest Statistic, " Communications in Statistics --Theoretical Methods, A5, no. 7, pp. 587-598, 1976.

8. Iman, R. L. and Davenport, J. M., "NewApproximations to the Exact Distribution of theKruskal-Wallis Test Statistic," Communications inStatistics -- Theoretical Methods, A5, no. 14, pp.1335-1348, 1976.

9. Kendall, M. G., Kendall, S. F., and Smith, B. B.,"The Distribution of the Spearman's Coefficient ofRank Correlation in a nniverse n hlch .i

Rankings Occur an Equal Number of Times,'Biometrika, vol. 30, pp. 251-273, 1939.

10. Glasser, G. J. and Winter, R. F., "Critical Valuesof the Coefficient of Rank Correlation for Testingthe Hypothesis of Independence," Biometrika, vol.48, pp. 444-448, 1961.

. . . . " S - " - - - .- ' - - ' '.' ' ' ' " ".' ' '.• '.. " .' . '. %* - '.'% %' ' .'. -.- -

11. Pitman, E. J. G., "Significance Tests Which May beApplied to Samples From any Populations. II. TheCorrelation Coefficient Test," Journal of theRoyal Statistical Society, Supplement 4. pp. 225-232, 1937.

12. Conover, W. J., Practical Nonparametric Statis-tics, 2d ed., John Wiley and Sons, Inc., 1980.

APPENDIX A

DOCUMENTATION FOR THE MICROCOMPUTER WORKSPACE

1. General Information

This appendix describes the organization and

operation of the IBM-PC (or compatible) version of the

workspace. Appendix C continues from where this

appendix leaves off, to walk the user through each test

by working practical examples.

Before proceeding any further, the user should

refer to section II (Workspace Design Issues) for

general information about workspace requirements and

assumptions regarding its use.

To get started, enter the APL environment in the

usual manner and load the NONPAR workspace.

2. Workspace Menus

This workspace is designed around the use of

menus. They guide the user through the selection

process of choosing a nonparametric test and a test

option. Three types of menus are used; the main menu,

test menus, and help menus.

a. The Main Menu

Within moments of loading the NONPAR

workspace, the main menu will appear. It is titled

Nonparametric Statistical Tests. This menu presents

general information about the workspace. Its primary

purpose is to list the choices of nonparametric tests

available and provide an option which allows the user

to exit the main menu into APL to copy data into the

workspace or return to DOS. Each test choice is listed

with some information about the test's area of

application. To make a selection from the menu, move

the cursor (using the cursor keys) to highlight the

desired choice, and press enter. As a reminder to the

user. a footnote at the bottom of the screen describes

the procedure for entering a choice. Once a test has

been selected from the main menu, a sub-menu

appropriate to the test appears. To exit from any menu

back to the main menu, press the Escape key.

b. Test Menus

The title of the test menu is the name of the

nonparametric test chosen. The text portion of the

menu gives a general overview of the test, to include,

the method used to compute the test statistic, and a

description of the various options that may be

exercised. The third section consists of the list of

test options available. These options include

returning to the main menu or choosing the help menu.

Test menus may have options listed in single or

multiple-paged formats. The comment in the final block

of the menu lets the user know if a certain menu is

21L

wwzrwwxw-- --qvtW W

multiple-paged or not. To make a selection from a

multiple-paged menu, use the page-up or page-down key

to locate the desired option. Proceed with the scroll

keys to highlight the choice, and press enter. Once a

test option is entered, the user is prompted to input

the data required to run the test. When the option for

more information is selected, the help menu is

displayed.

c. Help Menus

The title of the help menu usually begins

with the words "More Information About..." followed by

the title of the nonparametric test. The text portion

of the menu explains the test and its options in

greater detail. No choices are offered in the menu.

To return to the test menu, press any key.

APPENDIX B

DOCUMENTATION FOR THE MAINFRAME COMPUTER WORKSPACE

1. General Information

This appendix describes the organization and

operation of the mainframe computer workspace. To load

a copy of the NONPAR workspace from the APL library,

enter the APL environment and type: )LOAD 9 NONPAR.

Within a few moments the variables LIST and DESCRIBE

are displayed on the screen. These variables provide a

description of the workspace.

2. The NONPAR WorksDace

The NONPAR workspace consists of seven

programs which call several subprograms during their

execution. The exact syntax for each test and its

corresponding nonparametric test name is given in the

following format:

SYNTAX: Nonparametric Test and Application.

a. SIGN: Ordinary Sign Test for Location in Singleand Paired-sample Data.

b. WILCOX: Wilcoxon Signed-rank Test for Locationin Single and Paired-samole Data.

c. MANNWHIT: Mann-Whitney Test for Equal Medians orVariances in Two Independent Samples.

d. KRUSKAL: Kruskal-Wallis Test for Equal Medians in

K Independent Samples.

e. KENDALL: Kendall's B: Measure of Association forPaired-sample Data.

f. SPEARMAN: Spearman's R; Measure of AssociationBetween Rankings of Paired Data.

g. NPSLR: Nonparametric Simple Linear Regression;

Least Squares.

The list presented above can be displayed at any

time by typing: LIST.

-or each test program, t here exists a HOW variable

that gives a full description of the test and the

various options that may be exercised. To display any

o the HOW variables, -usm enter the test program's

name with the suffix HOW appended (i.e. SIGNHOW).

A test is run by entering the program's name. The

user is immediately prompted to input data. Enter

numerical data separated by spaces or as a variable to

which the numbers have been previously assigned.

Several of the tests require a considerable amount of

prompting before all the necessary data has been

entered.

* .. .',. . -. -. ~ x-. V.-

APPENDIX C

WORKSPACE FAMILIARIZATION THROUGH PRACTICAL EXAMPLES

1. General Information

This appendix applies to both the mainframe and

microcomputer workspaces. Its purpose is to acquaint

the user with the organization of the programs and the

type of prompts to be expected.

Extensive error checking has been included in the

programs to ensure that the data is of the proper form.

Should a program become suspended, clear the state

indicator by entering: )RESET, check over the data for

errors, and restart the program. 330 kilobytes of

computer memory are needed to load APL and the NONPAR

workspace; to avoid filling up the remaining workspace

area, the user should minimize data storage in the

NONPAR workspace. To exit a program at any time, press

the Control and Escape keys, simultaneously.

2. Practical Examples

a. Sign Test

(1) Descriction of Problem I. A Sinclair mine

is manufactured to have a median explosive weight of

not less than 16 ounces. The explosive weights of 15

mines, randomly selected from the production line, were

recorded as follows: 16.2 15.7 15.9 15.8 15.9 16 16.1

15.8 15.9 16 16.1 15.7 15.8 15.9 15.8.

(a) Is the manufacturing process packing

enough explosives in the mines?

(b) What range of values can be expected

for the median of the explosive weights 90% of the

time.

(2) Solution. ro see if the manufacturing

process is meeting the specifications, we test the

hypothesis Ho: M = 16 versus Hi: M < 16.

(3) Workspace Decision Process.

(a) Microcomputer: Choose the Sign Test

from the main menu, and the option, Single Sample;

Test HO: M = Mo versus Hi: M < Mo, from the test menu.

Skip to the Program Interaction section below.

(b) Mainframe: Enter SIGN at the keyboard

and receive the prompt:

DID YOU ENTER THIS PROGRAM FOR THE SOLE

PURPOSE OF GENERATING CONFIDENCE INTERVALS FOR A

SPECIFIED SAMPLE SIZE AND QUANTILE? (Y/N).

Enter N (If Y is entered, the user will

go directly to this last option of the test). The next

promptIS

THE NULL HYPOTHESIS STATES - THE

POPULATION MEDIAN (M) IS EQUAL TO THE HYPOTHESIZED

MEDIAN (Mo); HO: M = Mo. WHICH ALTERNATIVE DO YOU WISH

TO TEST? ENTER: 1 FOR Hl: M < Mo; 2 FOR Hl: M > Mo:

3 FOR Hi: M # Mo.

Enter 1. The next prompt is:

ENTER: 1 FOR SINGLE-SAMPLE PROBLEM; 2

FOR PAIRED-SAMPLE PROBLEM.

Enter 1.

(4) Program Interaction. The prompt is:

ENTER THE DATA (MORE THAN TWO OBSERVATIONS

ARE REQUIRED).

Enter the data separated by spaces or as a

variable to which the data has been previously

assigned. The next prompt is:

ENTER THE HYPOTHESIZED MEDIAN.

Enter 16. The following is dislayed.

COMPUTATIONS ARE BASED ON A SAMPLE SIZE OF:

13.

THE TOTAL NUMBER OF POSTIVE SIGNS IS: 3.

THE P-VALUE FOR HO: M = 16 Versus Hi:

M < 16 IS: .0461.

Consider a significance level of .05.

Since the P-value of .0461 is less than .05, we reject

HO: M = 16 in favor of Hi: M e 16 and conclude that the

manufacturing process is not packing enough explosives

in the Sinclair mine. The next prompt is:

WOULD YOU LIKE A CONFIDENCE INTERVAL FOR

THE MEDIAN? (Y/N).

.,*.* a . .* *z****** e. .~ .

Enter Y (If N is entered, the progam asks

if confidence intervals for a quantile are desired).

The next prompt is:

ENTER THE DESIRED CONFIDENCE COEFFICIENT;

FOR EXAMPLE: ENTER 95, FOR A 95% CONFIDENCE INTERVAL.

Enter 90. The following is displayed.

A 90% CONFIDENCE INTERVAL FOR THE MEDIAN OF

THE POPULATION IS: ( 15.8 < MEDIAN < 16 ).

The next prompt is:

WOULD YOU LIKE CONFIDENCE INTERVALS FOR A

SPECIFIED QUANTILE? (Y/N).

To see the form of the results, we generate

confidence intervals for the 30th quantile. Sample

size is automaticly set at the number of data points

entered eariler. Enter Y (If N is entered, the

mainframe program ends; or, the Sign test menu

reappears). The next prompt is:

ENTER DESIRED QUANTILE; FOR EXAMPLE: ENTER

20, FOR THE 20TH QUANTILE.

Enter 30. The following is displayed.

ORDER STATISTICS 1 COEFFICIENTS3 8 i .8231602 9 .9494901 .91600

***** THIS TABLE GIVES CONFIDENCE COEF-

FICIENTS FOR VARIOUS INTERVALS WITH ORDER STATISTICS AS

END POINTS FOR THE 30TH QUANTILE.

The mainframe program ends. The menu-

driven microcomputer program pauses for input from the

keyboard by prompting:

PRESS ENTER WHEN READY.

Press Enter and the Sign test menu

reappears.

b. Wilcoxon Signed-rank Test

(1) Description of Problem 2. A special train-

ing program is being considered to replace the regular

training that Radio Telephone Operators receive. In

order to evaluate the effectiveness of the new training

program, proficiency tests were given during the third

week of regular training. Twenty-four trainees were

chosen at random and grouped into twelve pairs based on

proficiency test scores. One member of each pair

received specialized training while the other member

received regular training. Upon graduation, the

proficiency tests were given again with the following

results.

Specially Trained Group (X): 60 50 55 71

43 59 64 49 61 54 47 70

Regularly Trained Group (Y): 40 46 60 53

49 57 51 53 45 59 40 35

(a) Does the special training program

ensure higher scores?

IrU

(b) By what range of values can the

scores of the two groups be expected to differ 95% of

the time?

(2) Solution. To test the hypothesis that the

special training program raises profficiency scores, we

test HO: M(X-Y) = 0 versus Hi: M(X-Y) > 0.

(3) Workspace Decision Process

(a) Microcomputer: Choose the Wilcoxon

Signed-rank Test from the main menu, and the option,

Paired-sample; Test HO: M = Mo versus Hi: M > Mo, from

the test menu. Skip to the Program Interaction section

below.

(b) Mainframe: Enter WILCOX at the key-

board and receive the prompts:

THE NULL HYPOTHESIS STATES - THE

POPULATION MEDIAN (M) IS EQUAL TO THE HYPOTHESIZED

MEDIAN (Mo); HO: M = Mo. WHICH ALTERNATIVE DO YOU WISH

TO TEST? ENTER: 1 FOR Hi: M < Mo; 2 FOR Hi: M > Mo;

3 FOR Hi: M A Mo.

Enter 2. The next prompt is:

ENTER: 1 FOR SINGLE-SAMPLE PROBLEM; 2

FOR PAIRED-SAMPLE PROBLEM.

Enter 2.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN TWO OBSERVATIONS

ARE REQUIRED).

* 59

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA (NUMBER OF Y ENTRIES MUST

EQUAL NUMBER OF X ENTRIES).

Enter the Y data. The next prompt is:

ENTER THE HYPOTHESIZED MEDIAN FOR THE

DIFFERENCES OF THE PAIRED DATA.

Enter 0. The following is dislayed.

COMPUTATIONS ARE BASED ON A SAMPLE SIZE OF:

12.

THE TOTAL SUM OF POSITIVE RANKS IS: 60.5.

THE P-VALUE FOR COMPARING THE MEDIAN OF THE

POPULATION OF DIFFERENCES TO THE HYPOTHESIZED MEDIAN,

HO: M(X-Y) = 0 Versus Hi: M(X-Y) > 0, IS: .0505.

Consider a significance level of .05.

Since the P-value of .0505 is greater than .05, we do

not reject the null hypothesis that the two training

cources are equally effective. However, due to the

closeness in values, the choice of rejecting or not

rejecting the null hypothesis is strictly a judgement

call. The next prompt is:

WOULD YOU LIKE A CONFIDENCE INTERVAL FOR

THE MEDIAN? (Y/N).

Enter Y (If N is entered, the mainframe

progam ends; or, the Wilcoxon test menu reappears).

The next prompt is:

ENTER THE DESIRED CONFIDENCE COEFFICIENT;

FOR EXAMPLE: ENTER 95, FOR A 95% CONFIDENCE INTERVAL.

Enter 95. The following is displayed.

A 95% CONFIDENCE INTERVAL FOR THE MEDIAN OF

THE POPULATION OF DIFFERENCES IS:

( -1 _ MEDIAN(X-Y) < 16.5 ).

The mainframe program ends. The menu-

driven microcomputer program pauses for input from the

keyboard by prompting:

PRESS ENTER WHEN READY.

Press Enter and the Wilcoxon test menu

reappears.

c. Mann-Whitney Test for Equality of Medians

(1) Description of Problem 3. A group of Army

and Navy officers were given the Defense Language

Aptitude test. From the results, 14 Army and 17 Navy

officers' scores were randomly selected. These scores

are listed below.

Army (X): 35 30 55 51 28 25 16 63 60 44 20

42 47 38.

Navy (Y): 54 26 41 43 37 34 39 50 46 49 45

33 29 36 38 42 34.

(a) Is there sufficient evidence to claim

that Navy officers score higher on this test than Army

officers?

(b) By what range of values can the

scores between the two groups be expected to differ 90%

of the time.

(2) Solution. To see if Navy officers score

higher on the exam, we test HO: Mx = My versus

HI: Mx < My.

(3) Workspace Decision Process.

(a) Microcomputer: Choose the Mann-

Whitney Test from the main menu, and the option, Test

HO: Mx = My versus HI: Mx < My, from the test menu.

Skip to the Program Interaction section below.

(b) Mainframe: Enter MANNWHIT at the

keyboard and receive the prompts:

DO YOU WISH TO COMPARE THE MEDIANS OR

VARIANCES OF THE POPULATIONS? ENTER: 1 TO COMPARE

MEDIANS; 2 TO COMPARE VARIANCES.

Enter 1. The next prompt is:

THE NULL HYPOTHESIS STATES - THE

MEDIANS OF X AND Y ARE EQUAL; Mx = My. WHICH

ALTERNATIVE DO YOU WISH TO TEST? ENTER:

1 FOR HI: Mx < My; 2 FOR Mx > My; 3 FOR Mx # My.

Enter 1.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN ONE OBSERVATION IS

REQUIRED).

'Ile

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA.

Enter the Y data. The following is

displayed.

THE SUM OF THE X RANKS IS: 224. THE U

STATISTIC EQUALS: 119.

THE P-VALUE FOR HO: Mx =My versus HI:

Mx < My IS: .5078.

We do not reject the hypothesis of equal

population medians and conclude the median of all Army

scores is equal to the Navy's. The next prompt is:

WOULD YOU LIKE A CONFIDENCE INTERVAL FOR

THE SHIFT IN LOCATION (My - Mx)? (Y/N).

Enter Y (If N is entered, the mainframe

program ends; or, the Mann-Whitney test menu

reappears). The next prompt is:

ENTER THE DESIRED CONFIDENCE COEFFICIENT;

FOR EXAMPLE: ENTER 95, FOR A 95% CONFIDENCE INTERVAL.

Enter 95. The following is displayed.

A 95% CONFIDENCE INTERVAL FOR THE SHIFT IN

LOCATION BETWEEN POPULATIONS X AND Y IS:

-10 < My-Mx < 10 ).

The mainframe program ends. The microcom-

puter program pauses for input from the keyboard by

prompting:

PRESS ENTER WHEN READY.

Press Enter and the Mann-Whitney test menu

reappears.

d. Mann-Whitney Test for Equality of Variances

(1) Description of Problem. Referring to

problem 3 in section c(l). Is there sufficient

evidence to claim that Army scores vary more than Navy

scores?

(2) Solution. To see if Army scores vary

more, we test HO: Vx = Vy versus HI: Vx > Vy.

(3) Workspace Decision Process.

(a) Microcomputer: Choose the Mann-

Whitney Test from the main menu, and the option, Test

HO: Vx = Vy versus HI: Vx > Vy, from the test menu and

receive the prompt:

ENTER THE DIFFERENCE OF THE MEANS OR

MEDIANS (Mx - My).

Because we believe the population

medians to be approximately equal, We enter 0. Skip to

the Program Interaction section below.

(b) Mainframe: Enter MANNWHIT at the

keyboard and receive the prompts:

DO YOU WISH TO COMPARE THE MEDIANS OR

VARIANCES OF THE POPULATIONS? ENTER: 1 TO COMPARE

MEDIANS; 2 TO COMPARE VARIANCES.

Enter 2. The next prompt is:

'~~~.%~ %,9 %. %I

THE TEST TO COMPARE VARIANCES, REQUIRES

THE TWO POPULATION MEANS OR MEDIANS TO BE EQUAL. IF

THEY DIFFER BY A KNOWN AMOUNT, THE DATA CAN BE ADJUSTED

BEFORE APPLYING THE TEST. ENTER THE DIFFERENCE OF

MEDIANS (Mx - My) OR 900 TO QUIT.

We enter 0. The next prompt is:

THE NULL HYPOTHESIS STATES - THE

VARIANCES OF X AND Y ARE EQUAL; Vx = Vy. WHICH

ALTERNATIVE DO YOU WISH TO TEST? ENTER:

1 FOR HI: Vx < Vy; 2 FOR Vx > Vy; 3 FOR Vx # Vy.

Enter 2.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN ONE OBSERVATION IS

REQUIRED).

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA.

Enter the Y data. The following is

displayed.

THE SUM OF THE X RANKS IS: 166. THE U

STATISTIC EQUALS: 61.

THE P-VALUE FOR HO: Vx = Vy versus Hi:

Vx > Vy IS: .0112.

Consider a significance level of .05.

Since a P-value of .0112 is less than .05, we reject

the null hypothesis of equal variances in favor of

LFuvI:A

Vx > Vy and conclude that Army scores do vary more than

Navy scores.

The mainframe program ends. The

microcomputer program pauses for input from the key-

board by prompting:

PRESS ENTER WHEN READY.

Press Enter and the Mann-Whitney test menu

reappears.

e. Kruskal-Wallis Test

(1) Description of Problem 4. During a recent

Monster Mash involving four Navy SEAL Teams, one of the

events consisted of the number of pushups a man could

do in 2 minutes. Eight men were chosen randomly from

each Team. The following scores were recorded.

SEAL 1: 90 96 102 85 65 77 88 70.

SEAL 2: 64 79 99 95 87 74 69 97.

SEAL 3: 101 66 93 89 71 60 76 98.

SEAL 4: 72 78 73 81 83 92 94 86.

Are the different Seal Teams considered to

be equally fit?

(2) Solution. To see if the Seal Teams are

equally fit. we test the hypothesis that all the

population medians are equal.

. .o

(3) Workspace Decision Process.

(a) Microcomputer: Choose the Kruskal-

Wallis Test from the main menu; and, once the test menu

is displayed, press Enter.

(b) Mainframe: Enter KRUSKAL at the

keyboard.

(4) Program Interaction. The prompt is:

ENTER THE NUMBER OF POPULATIONS TO BE

COMPARED (MUST BE GREATER THAN TWO).

Enter 4. The next prompt is:

ENTER YOUR FIRST SAMPLE.

Enter the SEAL 1 data separated by spaces.

The next prompt is:

ENTER YOUR NEXT SAMPLE.

Enter the SEAL 2 data. The next prompt is:

ENTER YOUR NEXT SAMPLE.

Enter the SEAL 3 data. The next prompt is:

ENTER YOUR LAST SAMPLE.

Enter the SEAL 4 data. The following is

displayed.

THE H STATISTIC EQUALS: .1335.

THE P-VALUE FOR HO: THE POPULATION MEDIANS

ARE EQUAL versus HI : AT LEAST TWO POPULATION MEDIANS

ARE NOT EQUAL IS: .98893.

N.N.v.'V,

We do not reject the null hypothesis that

the population medians are equal and conclude that the

SEAL Teams are equally fit.

The mainframe program ends. The microcom-

puter program pauses for input from the keyboard by

prompting:

PRESS ENTER WHEN READY.

Press Enter and the Kruskal-Wallis test

menu reappears.

f. Kendall's B

(1) Description of Problem 5. In order to

determine if cold weather affects target marksmanship,

Naval Special Warfare recorded small arms marksmanship

scores and corresponding air temperatures for a period

of one year. 20 men were chosen at random, and their

scores averaged for different air temperatures. The

average score for each air temperature is shown below.

Air temperature (X): 50 55 20 50 65 55 30

52 40 60.

Average scores (Y): 210 200 165 165 260

215 175 191 180 235.

Can it be said that colder temperatures

have an effect on marksmanship scores? Is that effect

positive or negative?

- -- . .

(2) Solution. We test the null hypothesis

that no association exists between cold temperatures

and marksmanship.

(3) Workspace Decision Process.

(a) Microcomputer: Choose Kendall's B

Test from the main menu; and, once the test menu is

displayed, press Enter.

(b) Mainframe: Enter KENDALL at the

keyboard.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN TWO OBSERVATIONSPP

ARE REQUIRED).

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA (NUMBER OF Y ENTRIES MUST

EQUAL NUMBER OF X ENTRIES).

Enter the Y data. The following is

displayed.

KENDALL'S B EQUALS: .7817.

THE P-VALUE FOR HO: NO ASSOCIATION EXISTS

versus: Hi: DIRECT ASSOCIATION EXISTS IS: .00045.

THE P-VALUE FOR THE TWO-SIDED TEST OF

HYPOTHESIS 1S: .0009.

Since the P-value for the one-sided test

equals .00045, we reject the null hypothesis that no

association exists between temperatures and

A, A'

marksmanship in favor of direct association. We

conclude that colder temperatures tend to cause lower

marksmanship scores.

The mainframe program ends. The microcom-

puter program pauses for input from the keyboard by

prompting:

PRESS ENTER WHEN READY.

Press Enter and Kendall's B test menu

reappears.

g. Spearman's R

(1) Description of Problem 6. When fitness

reports are written, officers of the same grade are

ranked against each other based upon their demonstrated

level of performance. Last marking period, the

Commanding and Executive Officers separately ranked 9

Ensigns as shown below.

EnsignsA B C D E F GH I

CO (x): 6 4 1 5 2 8 3 7 9XO (Y): 5 6 3 4 1 9 7 2 8

Does any association exist between the two

sets of rankings?

(2) Solution. We test the null hypothesis

that no association exists.

(3) Workspace Decision Process.

(a) Microcomputer: Choose Spearman's R

Test from the main menu; and, once the test menu is

displayed, press Enter.

(b) Mainframe: Enter SPEARMAN at the

keyboard.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN TWO OBSERVATIONS

ARE REQUIRED).

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA (NUMBER OF Y ENTRIES MUST

EQUAL NUMBER OF X ENTRIES).

Enter the Y data. The following is

displayed.

SPEARMAN'S R EQUALS: .5500.

THE P-VALUE FOR HO: NO ASSOCIATION EXISTS

versus: HI: DIRECT ASSOCIATION EXISTS IS: .0664.

THE P-VALUE FOR THE TWO-SIDED TEST OF

HYPOTHESIS IS: .1328.

Consider a significance level of .05.

Since a P-value of .0664 exceeds .05, we do not re.ject

the null hypothesis that no correspondence exists

between the two sets of rankings.

The mainframe program ends. The microcom-

puter program pauses for input from the keyboard by

prompting:

PRESS ENTER WHEN READY.

Press Enter and the Spearman's R test menu

reappears.

h. Nonparametric Simple Linear Regression; LeastSquares

(1) Description of Problem 7. Battery-powered

Swimmer Proplusion Units are sometimes used to aide

swimmers during long underwater swims. Recent tests

have shown that a nearly linear relationship exists

between water temperature and battery life for these

units. The following 17 data points were randomly

selected from the test results.

Water temperature (X) Battery life (Y)70 365 2.7550 1.840 1.260 2.455 1.952 1.7550 1.743 1.640 1.172 2.7555 248 1.535 .970 3.368 357 2.3

(a) Find the fitted regression equation.

(b) For the following water temperatures,

predict the battery life of the units: 61 52 46 36.

if the slope of the regression line equals .05.

(d) What range of values could be used

as the slope of the estimated equation line 90% of the

time?

(2) Solution. To determine the estimated

regression equation, we use nonparametric lirear

regression.

(3) Workspace decision process.

(a) Microcomputer: Choose Nonparametric

Simple Linear Regression from the main menu; and, once

the test menu is displayed, press Enter.

(b) Mainframe: Enter NPSLR at the

keyboard.

(4) Program Interaction. The prompt is:

ENTER X DATA (MORE THAN TWO OBSERVATIONS

ARE REQUIRED).

Enter the X data separated by spaces. The

next prompt is:

ENTER Y DATA (NUMBER OF Y ENTRIES MUST

EQUAL NUMBER OF X ENTRIES).

Enter the Y data. The following is

displayed.

THE LEAST SQUARES ESTIMATED REGRESSION

EQUATION IS:

Y -- 1.263 + .060668X.

The next prompt is:

DO YOU WISH TO ENTER SOME X VALUES TO GET

THE PREDICTED Y'S? (Y/N).

Enter Y (If N is entered, the program skips

to hypothesis testing for the slope).

ENTER X VALUES.

Enter 61 52 46 36. The next prompt is:

THE PREDICTED Y VALUES ARE: 2.44 1.89

1.53 .92.

WOULD YOU LIKE TO RUN SOME MORE X VALUES?

Enter N. The next prompt is:

WOULD YOU LIKE TO TEST HYPOTHESIS ON B, THE

SLOPE OF THE EQUATION? (Y/N).

Enter Y (If N is entered, the program skips

to confidence interval estimation). The next prompt

is:

ENTER THE HYPOTHESIZED SLOPE.

Enter .05. The following is displayed.

SPEARMAN'S R EQUALS: .5756.

THE P-VALUE FOR HO: B = .05 versus Hl: B >

.05 IS: .0079.

THE P-VALUE FOR THE TWO-SIDED TEST OF

HYPOTHESIS IS: .0158.

Consider a significance level of .05. Since

a P-value of.0079 is less than .05, we reject the null

hypothesis that B - .05 in favor of B > .05, and

conclude the slope of the regression line is greater

than .05. The next prompt is:

WOULD YOU LIKE A CONFIDENCE INTERVAL FOR

THE SLOPE? (Y/N).

Enter Y (If N is entered, the mainframe

program ends; or, the Nonparametric Regression test

menu reappears). The next prompt is:

ENTER THE DESIRED CONFIDENCE COEFFICIENT;

FOR EXAMPLE: ENTER 95, FOR A 95% CONFIDENCE INTERVAL.

Enter 90. The following is displayed.

A 90% CONFIDENCE INTERVAL FOR B, THE SLOPE

OF THE ESTIMATED REGRESSION LINE, IS:

( .05333 < B < .07 ).

If the estimated slope, does not lie within

the confidence interval, the following would be

displayed.

THE LEAST SQUARES ESTIMATOR OF B LIES

OUTSIDE THE CONFIDENCE INTERVAL. DISCARD THE LEAST

SQUARES EQUATION AND USE:

Y = -1.4458 + .060833X.

THIS EQUATION IS BASED ON THE MEDIANS OF

THE X AND Y DATA, AND THE MEDIAN OF THE TWO-POINT

SLOPES CALCULATED FOR THE CONFIDENCE INTERVAL ON B.

The next prompt is:

DO YOU WISH TO ENTER SOME X VALUES TO GET

THE PREDICTED Y'S FROM THE NEW EQUATION? (Y/N).

To compare results, let us input the

temperatures in the new equation. Enter Y (If N is

entered, the mainframe program ends; or, the

Nonparametric Regression test menu reappears). The

next prompt is:

ENTER X VALUES.

Enter 61 52 46 36. The following is

displayed.

THE PREDICTED Y VALUES ARE: 2.265 1.72

1.35 .74.

The next prompt is:

WOULD YOU LIKE TO RUN SOME MORE X VALUES?

Enter N. The next prompt is:

WOULD YOU LIKE TO TEST HYPOTHESIS ON B, THE

SLOPE OF THE EQUATION? (Y/N).

To compare results once again, we enter Y

(If N is entered, the mainframe program ends; or, the

Nonparametric Regression test menu reappears). The

next prompt is:

ENTER THE HYPOTHESIZED SLOPE.

Enter .05. The following is displayed.

SPEARMAN'S R EQUALS: .8287.

.. .. . . . . .~p* . .. ....... ......... - .. . - : .-.-p. ". ..-s .. .. . : -

I ,<' . < < : , , "- ; ;le"i

THE P-VALUE FOR HO: B = .05 versus Hi: B >

.05 IS: .0000.

THE P-VALUE FOR THE TWO-SIDED TEST OF

HYPOTHESIS IS: .0000.

The mainframe program ends. The microcom-

puter program pauses for input from.the keyboard by

prompting:

PRESS ENTER WHEN READY.

Press Enter and the Nonparametric

regression test menu reappears.

APPENDIX D

MAIN PROGRAM LISTINGS FOR MICROCOMPUTER WORKSPACE

SKENA.AA-B;BX*BY;CCXCY.D;DD;DX;Dr.DXY;S;POS;NEG;XX;Y;N;DEN;NN;NUM;P;PVAL-bU;SV:T'U;V!AT-Z:X;Z:WW;CHA;E:P ' 'I a THIS FUNCTION COMPUTkS THE kENDAfi STATISTIC WHICH IS A MEASURE

2 OF ASSOCIATION BETWEEN SAMPLES. P-VALUES ARE GIVEN FOR TESTING ONE3 AND TWO-SIDED HYPOTHESIS FOR NO ASSOCIATION VERSUS ASSOCIATION.

4 a SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: TIES, TIESK, KENDALP.5 ZNTERP, INPUT AND NORMCDF.

7 . E DISPLAY TEST MENU AND INPUT DATAI :MENU IZN l

1 0

12 BI:RrNPUT 2

Xl + (Q+)+RqORDER I IN INCREASING ORDER OF X.ORDER X IN INCREASING ORDER

'-19 a-x4AXJio20 A COMPUTE CURRENT RANKING OF Y

22 I NOW ORDER Y RANKS IN INCREASING ORDER23 D4-A AJ2 A TIES EX.ST IN EITHER X OR Y RANKED VECTOR USE MID-RANK METHOD25 DD-1 TIES D26 XX-l TIES 327 'FIND ORIGINAL RANKING OF Y WITH TIES RESOLVED28 YY-DDEC329 N-pX30 A COMPUTE NUMBER OF DISTINGUISHABLE PAIRS31 NN (Nx N-i))+232 S-pO33 AAu34 a POSITIVE ONES COME FROM A RUNS UP CONDITION- NEGATIVE I FROM RUNS DOWN35 A ZERO IS SCORED FOR TIES. MULTIPLY THE RESUL±S FOR EVERY ELEMENT AN D SUM36 Li: AA+37 EX(XAA(A+X38 CX4 XX AA <(AA+XXx)39 DX'-DX+CX40 BY.(77A (AA+X))'41 74iYY AA JAA.YY ).(-1)42 DY Y+ CY~43 DX!-DXxDY'4u POS (DXY'0)45 NGS ZO )X46 SSLS.47 -(AAcN-o )x L zl48 A SUM FINAL VECTOR TO DETERMINE S49 S.+/S50 A OBTAIN THE NUMBER OF TIES IN EACH VECTOR USING THE TIES! FUNCTION51 U-TIES B52 VTIE K D53 SU-+/ (2U)54 SV.+/ (2V55 A CALCU^ATE THE B STATISTIC INCLUDING THE CORRECTION FOR TIES56 TS( (NN-SU)x(NN-SV))*0.557 AT58 * (N>13)/NORM59 A CALL XENDALP TO CALCULATE THE RIGHT TAIL OF THE CDF OF B60 P-KENDALP N

6 :ALL :NTERP TO CALCULATE ?- 7ALUE 3Y :NTERPOLAT:ON* * PVAL-AT :NTERP P.0, -qPVALz1,)IL364 PVAL-0.5,65 *L366 CALCULA E P VALUE USINC NORMAL APPROX.67 NORM:NUM (3xAT )x((2XNN)*O68: DEN42x k2 "' N5 )*0.5'70 PVAL-1- NORMCDF Z)71 a IF a 1 POSITIVE PRINT OUT DIRECT ASSOCIATION.72 L3:.(T>0) L5.731 CHA 4'INDIRECT'

ef~f4 e*

7u -L775 LS:CHA.'DIRECT'76 L7:PV-o.2xPVAL77 ;(PVsIL878 PV179 L8:'KZNDALL''S B EQUALS: '§ ('&T TCNLs0 '$E P-VALUE FOR H0: NO AliO~lZAION EXISTS rfHUS'al I ill: ' (u!CHA) I ASSOCIATIONEXSTS IS: I (4vPVAL) OTCNL82, 'TE P-VALUE FOR THE T-SIDEb TEST OF HYPOTHESIS IS: ',t4fPV),T6'NL83 'PRESS ENTER WHEN READY.'8'. WW!Ea85 -NI

V KRWL*NUMDENOMA;CR-D-K;AA:BBbDDE-F:N-OF;P-PVAL-RSOF:SR;TSOR-CHA:B1 TH±S FeICTIOIJ COAP'Tg T X LSKAL!WALLIS'TEST S±AtZSTC& H wHrct :s2 A MEASURE OF TEE EQUALITY OF K INDEPENDENT SAMPLES.

~31 SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: TIES TIESK INDEXPLSAFDISTN. INTERP AND THE VARIABLES PMATKW20, PMATE.V31. PMA±J.33. PMA±"34

Sia?MAXKWUl,* PMATKW42, ANVD PMTKW'.43.~6I71A MENU CHOICES AND ROUTE TO PROPER STATEMET FOR ACTION.8S NI: B-MENU KRWLQgj

I0 MENU MAINQ1AI~11 -01i2 a PP"u1 131 31: 'ENTER THE NUMBER OF POPULATIONS TO BE COMPARED (MUST BE GREATER THAN T

15 .(- '~ (1I)0)E

S17 n INITIALIZE VECTORS E AND F AND VARIABLE C18 E*P'SOFR'p019 C'-020 ATHIS LOOP FACILITATES ENTERING THE SAMPLE VECTORS AND STORING THEM21 CHA" FIRST'

*22 L1;C-C~l23 'ETER TOUR ',(*CHA),' SAMPLE.'

25 ( (0oD -xo)/NEXT26 D-1 D27 a CONCATENATE SAMPLES AS THEY ARE ENTERED AND STORE THEM IN VECTOR E28 NEXT:E-E., D29 A RECORD THE LENGTHS OF THE SAMPLES AS THEY ARE ENTERED30 F*.PD31 CA'ET32 *!;C((XE-1 1)/Ll33 C A-14ASTI34 'C<K)/L135 A RECORD SIZE OF ALL SAMPLES WHEN COMBINED36 N.4/F

38 OF-F[739 A ORDER COMBINED SAMPLE VECTOR TO BE USED BY TIES FUNCTION40 D-E9 E'41 A CALR IND EXPLS TO INCREMENT INDEXES WHEN TIES OCCUR WITHIN ONE SAMPLE42 AA.F* INDEXPLS E43 a CALL TIES TO BREAK TIES BY MIDRANK METHOD414 BB-1 TIES D'45 C-046 a THIS LOOP CALCULATES THE H STATISTIC

* 47 L2:C*C+1*48 A SUM OF RANK S FOR EACH SAMPLE IS CALCULATED

49 SR4-+ BBilFiC)+(AAEChJD50 A CiLCUAT SUM OF 1AK SQUARED DIVIDED BY THE INDIVIDUAL SAMPLE SIZE51 SR. SR*2),FEC52 at STORE EACH CALCULATION53 SOFR.*SOFRSR54 5(< /L2551 A SUM ACROSS ALL SAMPLES*5S TSOR-+ISOFR

57 c3 CALCULATE ?INAL 3 S4TTISTIC

*59 a RECALCULATE H WITH CORRECTION FOR TIES.60 A.TIFSJ E*61 NUM- + (A *3))-(+IA).62 DEPNOM.Nx ((N*2 )1)*63 H*H* (1-(NUM.DENom))*64 A SYSTEM F LOGIC AL STATEMENTS ENSURE PROPER PROD. IS ACCESSED*65 * OF ~ '2 LOUTPUT*66 -jK>u ) IFAPPROX*67 *:3 1I

.6 A/~O 3 333 )v (Ft >)/APO.69 *(A/(F 2 l)o O PUTll)1AP.70 0 ((= )(FL:2 /Pf42

.4%

C713 -((A/(Of= 3-1 1 lflV(A/(OPX 3 2 1 1))V(A/(0F= 3 3 1 1))V(A/(OF= 3 3 2 1))) /P'ui

72 .(K=4 /P4373 r*i(01/(O >4 2 1 ,1(A POF 3 1 1)))OUTPUT75 * ((OL±]L4A(OV3j)76 *( / 0O= 3 2 11)(/(f 3 1))V('./(OF= 3 3 2))V(A/(OF= 3 3 3)))/P3377 * A/ 0OF3 2 2 v~OFL o'j1J4)/P3478 J_ AL P RPRI.T VARIABLE ACCESS CDP79 P23:P.-PMLTK20c CN4);;j80 -PM81 P31:?PXATKW31[(N-5);;382 *?M83 P33:P-PLITXU33r(N-5);;J814 -Pm85 P34:P-PI4ATKU34E(N-6);;]86 *PM87 P41 :P-PM.ATKI41[(N-5): ;]88 *PM89 Pu2:P.PMATKNu2E(N-5):;90 -PM91 P43:P-PATKNu3 E(-792 CALL ITRTO)RILCULAZZ P-VALUE BY INTERPOLATION93 PM:PAL-9. :NTERP ?S941 *(PVALa 1IOU.TPUT95] *L596] a CALCULAT P-VALUE USING TH F DIS rT U/ONE LESS D.F.IN DENOM APPROX.97] FAPPROX: F. (N-KI)*(K x (-98] PVAL1I-(( PK1, N -1)FISTN F)199) 'L5.100~ OUTPUT:2VAL-IGREATER THAN .251101 L5:,THZ 3 STATISTIC EQUALS,: ('aH) "TCNLS1021 'THE P-VALUE FOR HO: THE ?6OP&LATIOIJMqEDrANS ARE .7 UAL 7:,P,2 TCNVL.103] 1 31: AT LEAST TWOC ?0PUATON MEDIANS ARE NOP ZQUZ _ , ":PVALv,!_

COTCVL101 PRESS ENTER WHEN READY.'15 AA-0

E0lZ:'I ERROR: ENTER A SINGLE INTEGER VALUE GREATER THAN 2; TRY AGAIN.' ,TCNL108 -91

7 XANW-N-MPV2ABCG*MM*N;RX*U*NM1 PNU;PVAL-NMNUMZ ;NUMZ1 DEN-DENC;,DENCI A T&-NM GGi;L*A1*PVA PkDb* IPXkC';IAtPHA;BB;CC;UFld;2;P1 ;NN1;-NN

i st 'His 'FuNCfIOI U§S THk sUA OF RANkS AOCEDURE TO CALCULATE THE MANN-2: . WHITNEY u STATSI WHICH is USED IN COMPUTING THE P-VALUE FOR THE TEST3 a OF LOCATION AND SCALE. THE C.I. FOR (MX--!Z) THE SHIFT IN LOCATION IS4. ALSO COMPUTED. SUBPROGRAMS CALLED BY THIS FUNbTIoN INCLUDE: TIES. TIES25 *AINDEXPLS, VARMW, MANWP, * NPUT, CONFMW * NORMCDF * AND NORMPTH.

7. N1:MENL' MANWEL8,a MENU CHOICI5AND ROUTE TO PROPER STATEMENTS FOR ACTION.9.N3:D-CHOICEM PAGEDMENU MA!N1PQ

10 D F.011 (D=2,3,4 52/B112 (D=6 7 83/' 313 * D15lk114 MNU MAINQflJ15 Q16 B3:ENTR THE DIFFERENCE OF TE MEANS OR MEDIANS (Mg - I.17 DZ FF-U18 * ((pDIF)')/E319 a ENTER DATA VECTORS20 B1: 'ENTER X DATA (MORE THAN ONE OBSERVATION IS REQUIRED).'21 11.022 -(N)=0 ) /123 'ENTER Y DATA.24 MO.025 A IF CALCULATIONS INVOLVE VARIANCES ADJUST X B! THE DIFFERENCE IN MEANS26 N-.N- 01FF27 a CONCATENATE X AND Y SAMPLE VECTORS28 A.N,M

29 DETERMINE 3:ZZ OF ", IND 21 7ECTCRS AZVD ASSIGN TO .N AND X!M'3 0 NN-oN-1 mm-om32 AOMPUTE SIZE LIMIT- OF LEFT TAIL OF NULL DISTRIBUTION33 NM- (NNxMME )234 NM14L NM35 a ORDER A AND ASSIGN To B36 BA A37 C.N AM NDEXPLS A38 AB ALkLIE FUNCTION TO BREAK TIES USING MIDRANK METHOD .

39 G.1ri~ TIES40 A F FALSE CALCULATE TEST FOR VARIANCES41 :*fD=2 3 4,51/9542 a' cALVARMW TO GENERATE RANKS REQUIRED FOR VARIANCE TEST43 A C-ALW TIESTORCORD TIES IN TEE DATA AND BREAK TIES IN GO

45 0.0G TIES B46 a CLCUATE SUM OF X RANKS

48 A ONVERT TO MWINWEIT U STATISTIC49g t7RX- ((RNx (NN.1 )5+2)50 Ul.U51 A. IF SIZ OFXTMSSZEO GO TO NORMAL APPROX52 .((NMx2 )~0/L253 N4INM54 NN-INM55 AMANWP FUIJCTICN CALCULATES LEFT TAZL CUMULATIVE PROBS. OF U STATISTIC56 P.NN2 MANWP NNI57 .(D=5)/LIO58 ALOGICAL STATEMENT ENSURES ONLY LEFT SIDE OF NULL DIST IS USED59 -b(UigNMI )/L360 A CON VERT U STAT WHEN GREATER THAN LEFT TAIL VALUES61 LT.(NNxMM)-U12 A IJF gjl.ISA FRACTIONAL, rNTERPOLATE P VALUE63 23: 0(i U) NON

65 .(U2>O)/PiS66 PV.1-((PCU2+12 )+2)67 PF368 ± V.-PU> U2i))~69~ P3:PVI*(PEU2+13+PCU2+2j )+2L70 .J .cHzqxL711 NON:-%U>0)/GOL72] PV.173 -P217u GO:P7.1-P[Uj

~76: CHECE:-(U1-JJM1)/L477, PV2'-pv$73 ?V-?VI~79: PVI*P.V2so 80* L4l , I CONFIDENCE INTERVAL DESIRED GO TO LboL82, M2-(D=5)/L10L83 *a COMPUTE TEE NORMAL APPROXZIATION ,V/CORRECTIOY FACTOR

8u NUMZ-(U+0.5)-NM,85 NUMZle(U-.)-y''86 DEN-((MMxNNxCMM+NN+!))+1.2)*0.S87 Z*NUXk+DENso8 Z1.NUMZI+DEN89 RUM. I U-NM)90 DENC-i. (NN4MM-1)x(DEFN*2) ).(NN*MM-2) i-( ((NUM-0.5)*2)+(NN+MM-2) ))0.591 D9NQ1.( (NN+MM 1)c (DEN*2 ). (NN+xM2 )-((NUM+0.5)*2)+(NN 11*)))*0.592 TC. (NU -0.5 ) DENC93 TC1(NUM+0 )+DENCl94 * ( UNM )/SECOND95 PVI14((NORMCDF Z)+( (NN+MM- 2) TDISTN TC )4+296 PV. (1-(NORMCDF Z1)(1-( (NN*MM-2) TDISTN TCI)))+297 uL99 SECO fD:PVI.((NORMCD1 Z )+(1-((NN+MM-2) TDISTN TC)))+2991 Pv.{(1-(ZJORMDj 1 (N2+MM-2) TDISTN TCl))+2100) L4:2FV3,x(~(VPV

102) .103] NS:PVM-(3, 1)p(PVI PV PV3)104) 'TE SUM OF THE I RNKS IS: ',(*RX),'. THE U STATISTIC EQUALS: ',(oU1),

O2TCNLS105) A* POIA TTMENT FOR VARIANCE OUTPUTL106) .D=6 7 I VAR

L107 'THE P-IU'AUE FOR go: mX = MI U52B:M ,MOI[-;3, ZI:1(6 L& VPVMCD-1;13),CTCNL i ~'(LCCC-:), iI:'

ls09 VARPVMe( 3t1 2pjPV.PVI,PV3)110O 'THE P-AO o V1 YIE59i Hi: VZ ',(*LOGIC[D-5;1J),' Vi IS: '

S (6 4 PVM D-kH±N'RADY.111) 'PRESS ENI'R ED112) B!.2113] N1114) L8:'WOULD YOU LIKE A CONFIDENCE INTERVAL FOR THE SHIFT IN LOCATION (Ml1

Mj) (YIN).',O CNL~115 B41

-' ::a~ LPHA- f':- 0 +v11 ROUTE TO NORMAL APPROX. FOR CONF. lNT. OF LARGER SAMPLE SZZES120 .((NMx2)>80 ) L5121 A COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE122 CDFe-P123 A INDEX POSITION OF VALUE IN CDF :5 ALPHA124 INDEX-(+/(CDF5ALPHA))

N 125 * (INDEX>0)/L61261 NDEX~l127 .*L6

a,129 a COMPUTING CONFIDENCE INTERVALS USING NORMAL APPROX. W/C.F.129 L5:DEN-( (MMxNNx(MM+NN+1))+ 12)0.5130 UALPHA. (DENx(NORMPTH ALPHA)) +NM-0.51311 A ROUND UALPHA DOWN AND INCREMENT BY ONE

132 INDEXLUALPHA+i1331 L6:IPX.NN INDEX134 CI IPX CbNFMW A135. A ',(CC).' CONFIDENCE INTERVAL FOR THE SHIFT IN138 'LOCATION BETWEEN POPULATIONS X AND Y IS:'.OTCNL.37. ' ( ' (sC[),' S M - M2 I ',(*CIE2)),' )',TCNLi38 PRESS ENTER WHEN READY.'139 BB I'140 -N3.... El:'ERROR: SAMPLE CONTAINS LESS THAN TWO ENTRIES; TRY AGAIN.' I.CNL

.143 E3:'ERROR: YOU HAVE ENTERED MORE THAN ONE VALUE. TRY AGAIN.'.CTCNL144 -. 83

7 NPLR-N-SUMX-SUM-XBARTBARSUMX2SUMXY B A 'WW XX'BB;U-D;ALPA;P;CC;CDF;TALPHA;NN:CI:SLOP-S;RFASR ;DENOM.INDtX:FiX:: ;R: CkAE-PV

PROGRAM CONDUCZ' NONPAIAMERIC LINEAR &ERSSON. Hk L2EAST SQUARES22 a ESTIMATED REGRESSION LINE IS COMPUTED WITH HYPOTHESIS TESTING AND,31 a CONFIDENCE INTERVAL AVAILABLE FOR THE SLOPE B. IF B DOES NOT LIE IN THEX-1 m C.I. AN ALTERNATE 3EGRESSION INE :S PROPOSED. SUBPROGRAMS CALLED ARE5 S A SPMANP, KENDAL?, NORMPTH, INPUT. AND CONFLR.6] QPP*57 P DISPLAY MENU AND INPUT DATA.a NI:E MZNU102 MENU MAINQJL.1, -0

~12 JN2:R-INPUT 2F:3 Q.*,1.R

is Y4(Q) R?

16 A ASSIGN THE SIZE OF X (AND Y) TO NSN X COMPUTE THE SUM OF X'S AND Y'S

19 SUMX-/X20 SUMY +/Y

2 COMPUTE THE MEAN OF X AND22 XBAR-SUMX+N3 YBAR.SUMY+N

24 A COMPUTE THE SUM OF THE X'S SQUARED25 SUMX2-+/(X*2)26 A COMPUTE THE Jam OF X TIMES I27 SUMXY-I(XxY)28 A COMPUTE 'B' THE SLOPE OF THE ESTIMATED LEAST SQUARES REGRESSION LINE29 B-((NxSUMXY)-tSUMXxSUMI)),((NxSUMX2)-(SUMX*2))3 ) A COMPUTE 'A', THE I-INTERCEPT31 A-YBAR-(BxXBAR)32 FF 'N'33 'THE LEAST SQUARES ESTIMATED REGRESSION EQUATION rS:',DTCNL34 '7 ' (5A),' + ' (5mB) 'X.' OTCNL35 'DO 70 WISH TO EN±ER SOAE X U'LUES TO GET THE PREDICTED Y''S? (YIN).'36 WW-*M37 *(WW'N')I/l38 L2:'ENTER X VALUES.'39 XX .40 a CALCULATE PREDICTED Y'S41 YY A+BxXX42 'THE PREDICTED Y VALUES ARE, ' (wYY).TCNL43 WOULD Y%;U LIKE TO RUN SOME MORE X ALUES? (YIN).'44 WW45 .'!rI:Y')/L24 LN)IWOULD T LIKE TO TEST HYPOTHESIS ON B, THE SLOPE OF THE EQUATION? (Y/

8--(W=I'N' /L3498"IENZER THE HYPOTHESIZED SLOPE.'

51 ACOMPUTE52, U*Y-(BsxX) CMUEU'

.4 A CALL SPMANP TO COMPUTE R AND ASSOCIATED P-VALUES.-=Z D X SPMANP 11

37 HA-< 1.5 LiI:PVE2xD[239 0 - PV51IL19RoJ PV Z

S612 L19:,SPEARMAN''S R EQUALS: ' (4D[]) JTCNL62 'THE P-VALUE FOR HO: B = I,20BB), 'VERIUS'

'63 HI: B ',(WCHA[I 2)7.3) ' IS1 ( 2D2J,)TCTCNL,64 'THE P-VALUE FOR THE TWO-SIDED TEST OF HYPOTHESIS IS: I, TCN2xD[2)),TCN6 L-5)' IF USING THE NEW REGRESSION EQUATION BASED ON MEDIANS, EXIT HERE.6O L3--wPF'IN' 1L1867 ~SS ENTER WHEN READY.'

69 -NI70 ACOMPUJTE CONFIDENCE INTERVALS ON B71 L18:'WOULD YOU LIKE A CONFIDENCE INTERVAL FOR THE-SLOPE? (YIN).'72 W.73 *(MW=lN')/N1714 LO:CC*INPUT 575 a CHANGE ENTERED VALUE TO ALPHA76 ALPHAe-(100-CC).20077 a ROUTE TO NORMAL APPROX. FOR CONF. INT. OF LARGER SAMPLE SIZES78 , (N13))L579 a COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE80 ?-KENDAL? N81 CDF.PE2-'82 a 1PDEX POSITION OF VALUE IN CDP :9 ALPHA83 INDEX-(+/(CDFSALPHA))SL& *-(INDEX>0)/L685 INDEX 186 -L687 a COMPUTING CONFIDENCE INTERVALS USING NORMAL APPROX. W/C.F.88 L5:DENOM( (Nx(N-ijx((2xN)+5))419)*0.5ao 9 ALPHA-DENOMX 'II (NORMPT3 ALPHA)

91 L6:TALPHA P[3;INDEXI.G21 7ALHA93 LG:C."X CONFLR Z9g4 NN-leCIS95 SLOPES I+CI96 RR+L((NN-TALPHA)+2)

98 *.RRzQV)L209gQ RR.1i0o L20:SR-'-Fl+((NN+TALPH6A+2))101~ SR!021 -&SR:(nSLOPES))IL21i03 SReaSLQPESS104 L21:'A ' (mCC). CONFIDENCE INTERVAL FOR B THE SLOPE OF105 ' TEE ETIMATED REGRESSION LINE, IS:' .,TCNL107 'PES I EDY (SSOS RI, B :51,(5vSLOPESCSR]). ' ,QCV107 'PRESS ENTER HEN READY.T108 WW.3109 m IF 9 OUTSZDE THE C.I. CALCULATE NEW EQUATION BASED ON MEDIANS1o -((B3SLOPES[ZFRA(SLOPSCSR.) /N1111 I ORDER X AND Z112 X X AX)113 ]4- itY1114 A CHECK TO SEE IF THE SIZE OF SS IS EVEN OR ODD FOR FINDING MEDIANS115 *((2 NN)=O)/Si116 A COMPUTE MEDIAN FOR ODD CASE117 B SLOPES[((NN+1)+2))118 -S2,119 a COMPUTE MEDIAN FOR EVEN CASE120 SI:H (SLOPES[(NN 2)]+SLOPESi((NN+2 +2 2121 A VO THE SAME FOR THE X AND Y VECTORS122 S2:W((2 N)=0)/S3123 YBAR L((N+) +2))12U XBARX4((N+1 +2)]125 -OUT.126 S3:BAR-e(IT(N+2)]+YE((N+2).2)J).2127 XBAR (X[(N.2)'+XA'N+2)+2)]2128 A COjMUTE NEW INTERCEPT 'A'129 OUT:A YBAR-(BxXBAR)130 'THE LEAST SQUARES ESTIMATOR OF B LIES OUTSIDE THE CONFIDENCE INTERVAL.'

131 'DISCARD THE LEAST SQUARES EQUATION AND USE:'. TCNL132 ' Y:',(uA).' ' (682.'X',OTCNL.1331' THIS EQUATION IS BASEb ON TAE MEDIANS OF THE X AND Y DATA.AND TEE'C134 'MEDIAN OF THE TWO-POINT SLOPES CALCULATED FOR THE CONFIDENCE INTERVAL 0

N B.' LTCNL[135) A ALLOW USER TO DO SOME ANALYSIS ON NEW EQUATION[136] ' DO YOU WISH TO ENTER SOME X VALUES TO GET PREDICTED 7''S137) 'FROM THE NEW EQUATION? (Y/N).'

[138] FF4-IL139) .(FF=:I')/L2140 3 FF''

SIlGN;A-C-B;D:PVAL-X-MO-N-CDF-ALPHA*CI;Y*AA;BB;CC;DD;PV;PVI;NNN:KPOS;ORDD; XZ.ORDi.kALPHA.Z.Zl. UA.WW.PVM.PV3i.

1 A lHIS PUNCTIOk OSEt T~fOADINIR ri~ TEST TO CALCULATE THE K2 STATISTIC P-VALUE AND CONFIDENCE INTERVAL AS A TEST FOR MEDIANS.

A THE LAST 6PTION WILL DISPLAY A TABLE OF CONFIDENCE INTERVALS OF ORDERED4 a STATISTICS WITH CONFIDENCE COEFFICIENTS.

I SUBPROGRAMS CALLED B! THIS FUNCTION INCLUDE: BINOM, NORMCDF, NORMPTH,*6, INPUT,AND QUANG.

8 NI:MENU SICNGl p9 A MEN -TrICES AND ROUTE FOR PROPER ACTIONS

10] N4:C*CHOICES PAGEDMENU SZGNQIZ

I2: C2,3,4, 5 )/L"2 C=7 8 1 10)/L9

13 11i14 =C11)/L20

is MEU MAINQJ116. .017 aINPUT DATA FOR SINGLE SAMPLE CASEIS' LS:A"e119 X.IrNPUT I20 NNN.QX

22: MO.INPUT 323 D-X-AO24' *L1125 a PAIRED SAMPLE CASE26' Lg:AA4+227: R-INPUT 228~29* X11+ )+30 7.Q+)+

12 NNN-aDD

S3"~ MO-ZNP9TS35 9 ( - )m36 jaCOMPRESS D TO REMOVE ZEROS37 Li1:A.-(D-0)/D38~ A RECORD LENGTH OF A AND ASSIGN TO N3O N.a KEEPING TRACK OF POSITIVE SIGNS

41 ?POS-+/(A > 0)L&2 -(N>25;1NORM

13 PV.4L-31NO Y

*45 -

L47 Pi:PVI-i-PVALEXPOS3L48 P2:PV-PVALE(KPOS+1 ~J

-L6Z0 ? N IS GREATER THAN 30 Uf NORMAL APPROX W1 CONTINUITY CORRECTION

till ?ORM:z1-((KPO$-o.s -( .541),*(0.5x(N*0.5)*52 7-(KPOS+0.5 )-(0.5 XND+(0.5x(N*0.3)).53: PV.NVORMCDFZ.54. PVI.1-(NORMCDF ZI)*55* A IF PAIRED SAMPLE TEST GO TO L17 FOR OUTPUT STATEMENT56 LG:PV3,.2x (L/(PV,P VI))'57' *(V N558a PV3-1.59. N5:PVM-(3 1)o (PV PVI PV3)160, 'CO9PUTA±IONS AhE BASED ON A SAMPLE SIZE OF: '~N'.',1D2CNL

*61 'THE TOTAL NUMBER OF POSITIVE SIGNS IS: ',(fKPciS) , UCNL62. -(C=78,9)/L17 HiN (LOI(C)1), '(

61'THE P-U'LUE FOR HO:. M NO Iqj9:Xl,*OI[Ci;3,6MO),' IS: $.(5PVHEC-15,TCNL

642 *PL1S65 Li 7: 'THE P-VALUE FOR COMPARING THE MEDIAN OF THE POPULATION OF'66 'DIFFERENCES TO "HE HSIE MEDIAN DZCNL1671 'HO: M(X-Y) =I I EFRYPO S M(X-± rLOGIC(C6);13).f ',(fMO),,

69 BB-0f70 * B ~'')L1672 a INPUT SIZE OF CONFIDENCE INTERVAL73 L16:CC*ZNPUT 574 AiPHiA+ -iQ0-C C )200751 . NNN>25 )INORMI76 a COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE77 CDF.HINOM NN78 Aj INDEX POSITION OF CDF FOR ALPHA + 2719 9.+/( CDF.SALPHA)80 *(B'O J/SKIP81 B4-182 -SKIP8a a COMPUTING CONFIDENCE INTERVALS BY NORMAL APPROX.uiL PORMI:AALPHA-(fQ.5x(NNN*0.5))x(NORMPTY ALPHA))'(C.SxNN0-0.5gr- 4 .0UrNo -:;HA dOWN :0 .VEAREST :YVTZCZR! V ,v CTzMENTr 3Y )NE

d7 4 ~ :7 5INGLZ SAMPLE CASE Go TO L

89~ LU CADDC22 ,4,5 LCULATE AND PRINT OUT CONF. INT. FOR PAIRED SAMPLE CASE90 L:ORDD.DDEADD)9I 1 78.ORDD933' A[3ORD .(sCC CONFIDENCE INTERVAL FOR TE MEDIAN OF THE9g' 'POPULAIO lFDPFfRENCES IS: ' OTCNL96 *QN 1,(C L13).1 s MEDIAN(X-I) :5 ',(*CIC2J).' )'.DTCNL

98 7:OD.E CALCULATE AND PRINT OUT CONF. INT. FOR ONE SAMPLE CASE99 77.-oORDX

1001 .;,oD 1ACOPENEITERVAL FOR THE MEDIAN oF THE POPULATION Is:

S102J 1 (TN !9~12, MEDIAN~ S (oCr[212') OTCNLL103 2QUANT:'WOULD YCU LIKE C6NFIrDiNCE INTERVALS FOR ASPECIFIAD QUANTILE? (YIN

104 ). 6 OTC[105] *(UW='7')/B11063 .N41072 L20:1ENTE17 TEE SIZE OF THE SAMPLE.'108g NNN.Q110913 i:1ENTER DESIRED QUANTILE; FOR EXAMPLE: ENTER 20, FOR THE 20TH QUANTILZ.

LIII ~ (QUA5C)v (UA>100))/EIL112 k; QP QUANC UAL1132L1143 a* THIS TABLE GIVES CONFIDENCE COEFFICIENTS FOR VARIOUS INTERVALSLilS)05 'WITH ORDER STATISTICS AS END POINTS FOR THE 1.(vQUA), 'TH QUARTILE. ',QTC

SLS1161 'PRESS ENTER WHEN READY.'1172 BB.M1183 -N4

!1 Z: ERROR: THE QUARTILE 7ALUE MUST LIE BETWEEN 0 AND !00; TRY AGAIN.'

7 SPMAN:X:YA:Q:R:CA3BB3PV] THIS FUNCTIrON CbMPeITES THE SPEARMAN R STATISTC wHICy MEASURES

N , THE DECREE OP CORRESPONDENCE BETWEEN RANKINGS OF TWO SAMPLES. THE P-3 aVALUE :5 G17EN FOR TESTINC ONE AND TWO-SIDED 3YPO THESIS OF A4SSOC.:AT:ON14 A SUBPROGRAMS CALLED By THIS FUNCTION INCLUDE: TIS, TIESK, SPEAR?,

5 A INPUT, SPAPROX, INTERP, AND THE VARIABLE PbL4TSP.67 A DISPLAY MENU AND INPUT DATA.

8 VI:B-MENU SPMANQgJ

.0; 7iNU MANQ5I1

12 BI:R.INPUT 2

15 +1) +R16 A AL SPMANP TO CALCULATE THE STATISTIC AND ASSOCIATED P-VALUES17, A-X SPMANP 718 *(A E1>0 1Li19 CHA.'IND RECT'20 +L221 L1:CHA*'DZRECT'22 L2:PV4.2xA L2J23 * PV:51)/ 324 P .121 L3:'SPEARMAN''S R EUALS: '§ (4vAjC1),0TCNL26 TE P-VALUE FOR HO NO AS.O IATION EXISTS ZSUlI:27 ' Hi: ' 'wC1iA I ASSOCIATION-EXT ErSA2j IS:283 'THE P-VALUE FOR THE T-SID b TEST OF HYPOTHESIS IS: ',.4v(2xAL2 ji,0T CN

C291 'PRESS ENTER WHEN READY.'L30 B.0!

v wIsiGABD-EFPV2Z;ZDEN;NUMZ-NUMZi ;PVALXMO;N-TPLUS-CDF-TALPA;ALPHA.H.C±.±-AABbCCNN!DbPV;POS;±PObS;NM;PVI;TkFS;kN;C;'VM;PV3;R;Q;NUM.T .tC1:TAAP:D~kT-bEN±I

I A THIS AUNION 6SES i'HE WILCOXON SIGNED RANK -TEST TO TALCULATE THE TPLUS2 aSTATISTIC P-VALUE, AND CONFIDENCE INTERVAL AS k~ TEST' FOR MEDIANS.3jaSUBPROGRAAS CALLED'BY THIS FUNCTION INCLUDE: TIES, WILP, NORMCDF,

q NORMPTH, CONFW, INPUT 1tND TEE 7,ARIABLE ?MATRIX.

:~.MA U WIZEN CHOICES AND ROUTE FOR PROPER ACTIONS.U81 N4 :C*CHOICEW PAGEDMENU WILQ~g

92 * 2,3,4,5j)/L810 = 6 7 8,9 )/L9

CC* MEU INQfL9j13 * 0

14A INPUT DATA FOR SINGLE SAMPLE CASE15~ L8:AA.116. X.-INPUT I17~ NNN*gis~ * (C=5ftLI6191 MO.INFUT 3

S20 D-X-NO21 *Li122 aPAIRED SAMPLE CASE23 L9:AA-2

25 Q~ieR26 .+Qi)R

28 DD-.x-Y29 NNN-oDD30 *(C=9)IL1631 MO..TNPUT 1L32 D-(X-I)-MOL33 1:.(0 COMPRESS D TO REMOVE ZEROSL35 A RECORD LENGTH OF A AND ASSIGN TO N36 N.DA37 A KEEPING TRACK OF POSITIVE SIGNS38 aO.Ao TAKE THE ABSOLUTE VALUE OF A; ASSIGN TO B AND ORDER B

*,^', a EORDER POSITIVE SIGNS TO COINCIDE WITH PROPER POSITIONS IN B

'14*4: Ao.~s ~ CALL FUNCTION TO BREAK TIES45. E~l TIES B46 A CALCULA.'E TPOS BY ADDING ACROSS ALL POSITIVE VALUES OF E47: Tpos.+/(POSxE)l8. TPOS1.*TPOSug" a G;IVES SIZE OF LEFT TAIL OF PROBABILITZ DISTRIBUTIONR?~ NM-(LC(+/iN)+2)fl 11

a (N9I3 GO TO STATEMENTS BASED ON LENGTH OF VECTOR E

S53 a GENERATE NULL DISTRIBUTION FOR TPLUSL54. F.WILP NE55] A IF TPOS FALL IN LEFT HALF OF PROB DIST CALCULATE PVALUE AS NORMAL~56 L1:'.(fPOSs(NM-H)/TREGS57 A OTHERWISE USE THE NEGATIVE T STATISTIC58 TPOS- ( +iN)-TPOS159 a IF TPOS IS FRACTIONAL USE BOTH THE INTEGER ABOVE AND BELOW AS TPLUS~60 TREG:-( C: ITOs)=o )/NON62 fP:PV- (('FLTPLUSJ+F~TPLUS+jj+2)63 PVI.(FCTPLUS+iJ+ EFT LUS+23 )$264 *CHEC651 NON:.(TPOS>0)/GO66 PV.167 -P268 GO:PVi-F[C(TPOS) 2169 P2:PVI-F rTPOS+i 270 CHEC:.TFSl (NM-i) )/L671 PV2-PV72 PV*PVI73 PVI+-PV274 *L675 a COMPUTE NORMAL APPROX. WICONTINUITZ CORRECTION FACTOR r76 L3:TRAP-(Nx(N+li) *477 NUMZ-(TPOS+O.5 )-TRAP781 NUMZI.(TPOS-O.5)-TRAP79 DEN-( (Nx(N+1)?c((2xN)+1)),2'4)*0.5so Z.-NUMZ .DEN81 Zi-NUMZI+DEN82 ACOM UTE STUDENT T APPROXIMATION WITH CONTINUITY CORRECTION FACTOR83 NUM IT FTOS-AP84 DENT. ((N(DEN*2) *(N1) -( NUM-C.5)*2)4(N-1) *0.585 DENTI.+ (Nx (DEN*2 )+(N-1 )~(NUM+0.5)*2),(N-1M3 *0.5

86 T.(NM-0.5 )+DENT87 T1.(UM+0.5)+DENTI88 aCOMUTEAVERAGE OF TC AND ZC

a9 *(PS9( (+/1)+2j))<SECOND90 y.NOMD 11Tp (N-i) TDiITN TI)+

Pv1 m (NORMD I T TISTN TC +J2 T1).92 *opL693 SECOND:PV*((1-NRMCDF Zi) )+(.(N-i) TDI.~TN TC ) )+21L&l PVZ..((NORMCDF Z *(I-((N-1. TcDISTN TC) )i)+2

ga8 N5:PVM.(3 i)p(PvI PV PV3)L9j 9] COMPUTA±ION ARA BASED ON A SAMPLE SIZE OF I(wN) ' '02CNL

102 'THE TOTAL SUM OF POSITIVE RANKS IS* I (OTPOR') '*['QCNLLi01.A IF PAIRED SAMPLE TEST GO TO £17 FOR O/JTPU' STATEMENTL102 J AA=22 L17L1031 'THE, P-VALUE FOR HO: M s AMO),l yj5q Hi: M ',(*LOGXCEC-1;1]).' '(

, MO ' S:,(44PVMCC-1;i2),UCNL

105~ L17:1 THE P-VALUE FOR COMPARING THE MEDIAN OF THE POPULATION OF'L106] 'DIFFEjENCFS TO THE 6POTHESIZEDMED fANt~ OTCNL1107 'Ha: M(X7 I ' l~bC (uQ'~i] : M(I-) *(LOGICEC-513),' '.(fMO),'.

IS:,. UP M[4,

86 *

108 L18: 'WOULD YOU LIR2EA'CONFIDENCS INTERVAL FOR THE MEDIAN? (Y/N).'109 BB* zf ')LI110 *(B'7)/6111 090112 L16:CC-INPUT 5113 ALPHA.-(i00-CC)*200114 A ROUTS TO NORMAL APPROX. FOR CONY INT OF LARGE SAMPLE SIZE115 * NNN>9 /L4&16 CF*-WILP NNN117 it INDEX POSITION OF CDP FOR ALPHA + 2118 CDF-(CDFzo)/CDF119 NAPHA.(-I (CDFsALPHA))120 *(TALPHAxO )/JUMP121 TALPHA-1122 .0JUMP123 P COMPUTING CONFIDENCE INTERVALS BY NORMAL APPROX. Ar/C.?,1214 L4:TRAP.(NNNx(NNVN+1 14

15 DEN.l (NN(NN1) (2NNV)+1 ))+24)*0.51126 TS ALPHA)) +TEAP-0.5'27 TALPHAS128 A ROUND TALPHA DOWN ZC IN7TEGER 7ALUE AND .NCREMENT BY ONE129] TALPHA.LTALPHA+l

.130] a IF ONE SAMPLE CASE GO TO L7

L132: a CALCULATE AND PRINT OUT CON?. INT. FOR PAIRED SAMPLE CASEL133, L5:CXeTALPHA CONFW DDL134 ' A I (CC) I CONFIDENCE INTERVAL FOR THE MEDIAN OF THEL135' 'POPULATI6N OF bIFFERENCES IS: ',OTCNL~136 ' ( I (mClIJ),l 59 MEDIAR(X-Y) :9 ',(vCI[23).' )',OTCNL137, 'PRESS ENTER WHEN AEADY.'13 3BW '

139 N4v,luol ACALCULATE AND PRINT OUT CON?. ZNT. FOR ONE SAMPLE CASE

14ij L7:CZ.TALPHA CONFW X142] 'A (*CC),'I CONFIDENCE INTERVAL FOR THE MEDIAN OF THE POPULATION IS:

S143~ C ' (0 CZ C~. I MEDIAN S ',(sC 2]), 0'TCNL14lu] 'PRESS ENTER WHEN AEADZ.

&N % N

APPENDIX E

MAIN PROGRAM LISTINGS FOR MAINFRAME COMPUTER WORKSPACE

V KENDALL-A*AA:BBX;BY-C-CXCY;D;DDDDXYDXY;SPS;NEG;XX;ZY;N;DEN;NN;NUM.P.PV PVAL. UTSV'. V. AT Z'X'Y' HA:Q'r

1 a FH±S .UNC!fON co*kP&TkS'THk kEkDALL b HTATISTIC WHICH IS A MEASUREa OF ASSOCIATION BETWEEN SAMPLES. P-VALUES ARE GIVEN FOR TESTING ONEa AND W0-SIDED BYPOTHESIS --OR NO ASSOCIATION VERSUS ASSOC:AT.:N.

4 A SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: TIES, TIESK, Z4ENDALP.a INTERP INPUT AND NORMCDF.R-INPUT 2

l~:Q+i)R10 a Y4. ORDER Y IN INCREASING ORDER OF X

a ORDER X IN INCREASING ORDERS C. "a COMPUTE CURRENT RANKING OF Z

16 A NOW ORDER 7 RANKS ZN INCREASING ORDER18 . Ii TIES EXIST IN EITHER X OR Z RANKED VECTOR USE MID-RANK METHOD

19 DD-1 TIES D20 XX.1 TIESB20 a FIND ORIGINAL RANKIN: OF Y WITH TIES RESOLVED22 '7-DD C23 NoX24 A COMPUTE NUMBER OF DISTINGUISHABLE PAIRS25 NN-(Nx(N-1)) 226 S-pO27 AA4-028 A POSITIVE ONES COME FROM A RUNS UP CONDITION- NEGATIVE I FROM RUNS DOWN29 A ZERO IS SCORED FOR TIES. MULTIPLY THE RESULTS FOR EVERY ELEMENT AND SUM

0 LI:AA-AA+131 BX (XX AA >(AA+XX))32 CX XX AA AA+X))x(-)33 DX4.BX+ X34 BY.,(77 AA1>AA+Y))x(,35 CI. IAA <~ AA+U)x(136 D7.7+CY37 DXY DXxD"38 POS (DX1>039 NEG4,DXI<O x ()

42 A SUM FINAL VECTOR TO DETERMINE S143 Si.+/544 a+OBTAIN THE NUMBER OF TIES IN EACE VECTOR USING THE TIESK FUNCTION45 U*TIESK B46 V TIESK D47 SU +/ (2 U)4 A CALCULATE HE B STATISTIC INCLUDING THE CORRECTION FOR TIES50 T*S+; (NN-SU )x(NN-SV) )*0.51 AT Ij

53 CALL EENDALP TO CALCULATE THE RIGHT TAIL OF THE CDF OF B54 Pi.KENDALP N55 CALL INTERP TO CALCULATE P-VALUE BY INTERPOLATION56 PVAL-AT INTERP P57 .(PVALz1)/L3

.58 PVAL*O.5

1 2:,LCULATT ? 7ALUE 7C:NC NORMAL APPROX.o , NORM:NUM*',xAT ,xt,2I NN)O.z)

62 DEN (2x((2xN)+5))*0.563 Z NUM+DEN64 PVAL1- (NORMCDF Z)65 IF B IS POSITIVE PRINT OUT DIRECT ASSOCIATION.66 L3:*(T>O) /L567 CHA-' INDIRECT'68 *L769 L5: CHA'DIRECT'70 L7:PV.2xPVAL71 (PVSI )IL S ,72 PV. I73 L8:'KZENDALL''S B EQUALS #,(4T)

-N .,

75) :THE P-VALUE FOR Ho: NO ASSOCIATION EXISTS VERSUS'76 Hi: ',(vcHA).'ASSOCIATION EXISTS IS: '.(L4.PVAL)77] 1 178 'THE P-VALUE FOR THE TWO-SIDED TEST OF HYPOTHESIS IS:'.(4IPV)[79] ' '

7 KRUSKAL NUM DENORA C HDK*AA8BbDD-E;F:NOF-P PVAL;R;SOFR ;SR-TSOP;CA~1 A THIS PUNelION C.WHU§'E tgkA K&SkAL-i ALL±S TE.§T'STAT±S±I!C H WH±CH tt2 A A MEASURE OF THE EQUALITY OF K INDEPENDENT SAMPLES.3 a SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: TIES TIESE INDEXPLS.

I4 a PDISTN INTERP AND THE VARIABLES PMATKW20 PMATKW31. PMA±KW33,5 a PMATFJ34, PMATfKA41, PMATKWJ42, AND PMATK;4i.

7 BI: 'ENTER THE NUMBER OF POPULA.IONS TO BE COMPARED (MUST BE GREATER THAN T

.91 -((K<3)>l)IE1 )1n

11 pK'INITIALIZE VECTORS E AND F AND VARIABLE C:I E-F-SOFPRooC-014 ATHIS LOOP FACILITATES ENTERING THE SAMPLE VECTORS AND STORING THEM15 CHA,,'FZRST'16 LI:C*C+l

' ENTER YOUR 1.(*cHA),' SAMPLE.'

, -. "9ooD)z01/NEXT

2t Q CONCATENATE SAMPLES AS THEZ ARE ENTERED AND STORE THEM IN VECTOR E2 NEXT:E.E.D23 A RECORD HE LENGTHS OF THE SAMPLES AS THEY ARE ENTEREDL24) FDL A25 H 'NEXTi26 (C(K-))ILL27 CRA-'LAST'28 (C<K)1L129 RECORD SIZE OF ALL SAMPLES WHEN COMBINED30 NV-/F31 A ORDER SAMPLE SIZES LARGEST TO SMALLEST32 OF FC'F-33 ORDER COMBINED SAMPLE VECTOR TO BE USED BY TIES FUNCTION34 DEW*E35 CALL INDEXPLS TO INCREMENT INDEXES WHEN TIES OCCUR WITHIN ONE SAMPLE36 AA-F INDEXPLS E37 a CALL TIES TO BREAK TIES BY MIDRANK METHOD38 BB.1 TIES D39 C-040 a THIS LOOP CALCULATES THE H STATISTIC41 L2:C C+1

42 A SUM OF RANKS FOR EACH SAMPLE IS CALCULATED43 SR-+/BB (F C)+(AA C)

44 LCU A OF AANKi SQUARED DIVIDED B THE INDIVIDUAL SAMPLE SIZE46 i SR*2 S[CiTORE EACH CALCULATION47 SOPR.$OFRSR48 (C' )/L250 S SUM ACROSS ALL SAMPLES50 TSOR .+/SOFR5152 H*KTSORx(12+ NX(N+1)))-(3x(N+l53 a ECALCULATE H WIT CORRECTION FOR TIES

q 54 A-?IESKEZ55 NUM., +/(*S3 ) )- (+/A)56 DE x( (N*2)-1 ,)57 H* 1 -(NUMDENOM))58 A SYSTEM OF LOGICAL STATEMENTS ENSURE PROPER PROB. IS ACCESSED59 OF[ 3 <2 )LOUTPUT60 (0. ) IFAPPROX

1 * f3) IF62 * A A/(OF= 3 3 3 3))v(OFE1)>3))/FAPPROX63 r(OF: 2 1. 1 /OUTPUT

--- ,., U',,05.j (OF: -k :v t,(F: 3 2 .-)v(A/(OF: 3 3 :;)vA,(0F 3 3 2: "k/ F -(A/(OF=

67 "F: (OF[1>4)PAPPROX68 1 (AJ/( OF= 3 1 1)))IOUTPUT69 0 ".%4 (OF3) I IP3170 * / (OF: 3 2~ 1) v (A : ) )v(A/(OF= 3 3 2))v(A/(OF= 3 3 3)))/P33-71 ^ OF= 3 2 2 iL JP372 A AL APPROPRIATE VARIABLE ACCESS CDF73 P23:PPMATKW20C - );;J74 .OPM75 P31:P..PMATXW31C(N-5)1;;76 *1PM77 P3 :P.PMATKW33 (N-5):;)7011

ftC CC .. . . . . . C. -

79 P34:P-.PMAfXW34[(N-6); 3]80 PM81 P1i:P-PMATKW i(N-5);;382 -PM83 P42:P-PMATX242[(N-5)•;;84 _PM85 P43:P'PMATXW43 C (N-7) ,86 a CALL INTEAP TO CULATE P-VALUE BY INTERPOLATION87 PM:PVAL-9 ZNTERP P88 *(PVAL- I)/OUTPUT89

iPAL

go a CALCULA-7 ?-VALUE USING THE F DIST WIONE LESS D.F.IN DENOM APPROX.gi FAPPROX:? -Kx ) ((-1(1-))92 PVALi-(((K-),( (N-K)-i)) DISW F) -

93 -L594 OUTPUT:PVAL.'GREATER THAN .25'95 L5:'THE B STATISTIC EQUALS: ',(4*H)96 '97 'THE P-VALUE FOR Ho: THE POPULATION MEDIANS ARE EQUAL VERSUS98 199 ' HI: AT LEAST TWO POPULATION MEDIANS ARE NOT EQUAL IS: ', CPVAL)

,101 .40102, El:'ERROR: YOU MUST ENTER A SINGLE INTEGER VALUE GREATER THAN 2; TRY AGAI) N.1

) 1"103] '

7 XANNWBIT-N.M-PV2-A;BLCG;MM-NN'RX;U:NM1:P;NU:PVAL:N:NUMZ:NUMZI;DENC :Z-- LPHA : NDF INEX - '.k-i; C ;ALPHA, 3B; CC; U ;U2; PV; NNI; NN2 . 2V ; DZVI; ;AA ;O C

;P N PV3;D- Z2R;DEN;bENCTCTC1 :NUM.] THIS FUJNCTION USE'S TE' SUJM OF RANKS PROCEDURE TO CALCULATE THE2 q MANN-WHITNEY U STATISTIC USED IN COMPUTING THE P-VALUE FOR THE TEST3 A OF LOCATION AND SCALE. THE C.I. FOR 2(Y)-M(X), THE SHIFT IN4 a LOCATION, IS ALSO COMPUTED. SUBPROGRAMS CALLED INCLUDE: TIES, TIES27 AINDEXPLS, VARM , MANWP. INPUT, CONFMW, NORMCDF , AND NORMPTH.6 DIFF-O

'DO YOU WISH TO COMPARE THE MEDIANS OR VARIANCES OF THE POPULATIONS?'

0 B2:' ENTER: I TO COMPARE MEDIANS: 2 TO COMPARE VARIANCES.'

12 ( )A(AA=2))/E213 AA=i/Ni14 ' THE TEST TO COMPARE VARIANCES REQUIRES THE TWO POPULATION MEANS'Is 'OR MEDIANS TO BE EQUAL. IF THEY DIFFER BY A KNOWN AMOUNT,16 'THE DATA CAN BE ADJUSTED BEFORE APPLYING THE TEST.'1710833:' ENTER THE DIFFERENCE OF MEDIANS (M(X) - M(Y)) OR 900 TO QUIT).'19 DFF.Q20 DFF )>1 )/E321 F=900)/0221' HE NULL HYPOTHESIS STATES - THE POPULATION VARIANCES ARE EQUAL; V(X) :V(Y).S23224 ' WHICH ALTERNATIVE DO YOU WISH TO TEST?'25 '261 N2:'EZNTR: 1 FOR Hl: V(X) < V(Y); 2 FOR Hi: V(X) > V(Z); 3 FOR Hl: V(X)*_V(z). '

287 '(D-l)A(D-2)A(D*13))/E4291 ,31301 Ni:' THEs NULL HYPOTHESIS STATES - THE MEDIANS OF X AND Y ARE EQUAL; M(X): M(Y).31232 WHICH ALTERNATIVE DO YOU WISH TO TEST?'3334] B6:'ENTMR: I FOR Hl: M(X) < M(Y); 2 FOR Hi: M(X) > M(Y); 3 FOR Hi: M(X)

[3 M(Yb'l

36 '(?Dsi)A(Dz2)A(D:3,)/Er.37 a LNTR DATA 7CTORS

38 31: ENTER ; DATA (MORE THAN LNE OBSERVATION IS REQUIRED).'L4 N>) =0)/El

41 'ENTER I DATA.'

413 it ICALCULATIONS INVOLVE VARIANCES ADJUST X BY THE DIFFERENCE IN MEANS44 N.-N-DIFFL15 A CONCATENATE X AND I SAMPLE VECTORS46 A'N.M47 A DETERMINE SIZE OF X AND Y VECTORS AND ASSIGN TO NN AND MM148 NN'pNSMM M COMPUTE SIZE LIMIT OF LEFT TAIL OF NULL DISTRIBUTION

51 NM-(NNXMM).2521 NMI*LNM

53 a ORDER A AND ASSIGN TO B514 8.AAA55 C. (NNMM)INDEXPLS A56 A CALL TIES FUNCTION TO BREAK TIES USING MIDRAJK METHOD57 C-1 TIES B58 a IF FALSE CALCULATE TEST FOR VARIANCES59 * (AA1 )/8560 a C Z; VARMW TO GENERATE RANKS REQUIRED FOR VARIANCE TEST61 GG.VAP.MW NN+MM)62 A CALL TIES TO RECORD TIES IN THE DATA AND BREAK TIES IN CG63 G.GG TIES B614 A C'LCULATE SUM OF X RANKS65 BS:RX*+/(G[(C[1;(INN)1166 A CONfVERTTO MANNUHIT U STATISTIC67 U.RX-((NNx(NN+1) 2.2)68 U1.U69 A IFqSIZE OF X TIMES SIZE OF Y > 80; GO TO NORMAL APPROX70 (NMX2 )~0)/L2

7.1 ! IN.LNN,MM721 XN2. INN MM73 amANwp FuIJCTIoN CALCULATES LEFT TAIL CUMULATIVE PROBS. OF U STATISTIC714 P-NN2 MANWP NNI75 A LOtqICAL STATEMENT ENSURES ONLY LEFT SIDE OF NULL DIST IS USED

*76 . UgNMIA/L377 A CONVERT U STAT W'HEN GREATER THAN LEFT TAIL VALUES:79 U*(NXlgjl A FRACTIONAL, INTERPOLATE P VALUE

82 4- 20 P'83 PV.-.-( (PEU2+1,J)42)

:851 PM:V.-( P1U23 +PrU2 +l ; +2)86] ?3:PVI.( U2+1 -'PU2+2 +873 -CHECK881 NON:*(U>0)/GO

91 PV2i-

95J PV:PVIPU+.

962 PVI.PV297 -L14982 A COMPUT THE NORMAL APPROXIMATION U/CORRECTION FACTOR9 L2:NUMZ. (0+0.5)-NM100 NUMZ + U-0.5 -NM101 DEN- (MMxNNX (MM+NN+1))+12)*0.5102 Z+NUMZ+DEN103 Zb.N(/ 1N

*1014 NUM. (U-NM )105 DENC.( ( (NN+MM-)x(DEN*2) +N+M-)- NUM-.5)*2)G(NN.MM-2) *0.5106 DENC1.( (NN+MM- )x (DEN*))(NN+MM2)(NUM+0.5)*2),(NN M2)*0.5107 TC*(NUM-0.5)+DENC108 TC1.(NUM+0.5)+DENCI

109 * IN)SECOND110 PVI- (NORMCDF Z)+( (NX+MM-2) TDISTN TC) )42ill PV* ( (1-(NORMCDF Z1 )+(l-( NN+MM-2) TDISTN TC1)))*2112 PL14113 SECOND:PVI.( NORMCDj' Z)+(1-((NN+MM-2) TDISTN TC)))*2114 PV~( (NOM Z J)+ (GvN+MM 2) TDITN TC1))*2115 L4YPV3x(I (P, V)116 .PV3 1)IN117 PV3.1118 N5:PVM*(3 ,lp PVI PV PV3)119 'THE SUM OF THE Ik s KSis: .,(*RX).,. THE U STATISTIC EQUALS: ',(mU1)120 1121 A.(A LOGICAL STATEMENT FOR VARIANCE OUTPUT122 *(BA2 ) IVAR123 'TE, PVALUE FOR NO0: M(X) M(Y) VERSUS Hl: M(X) ',(mLOGCD;1J),' M(Y)

IS: ',(4*PVMED;13)S1241)2125 *L8~126~'~ VA P-VAlUF PR'PVC'7(X (7:7RSUS 71: 7(X) '(LGC:2'77

:129 -0L1301 LS:' UQULD YOU LIKE A CONFIDENCE INTERVAL FOR THE SHIFT IN LOCATION(M()-

M(.1))? (YIN).'

132 *(BBIN' )I0133 L1O:CC*ZNFUT 5134 ALPHA. (100-CC) *200135 1% ROUTE TO NORMAL APPROX. FOR CONF. INT. OF LARGER SAMPLE SIZES136 .((NMx2)>80) 1L5137 A COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE138 CDF.P139 A INDEX POSITION OF VALUE IN CDF S ALPHA140 INDEX*(+I(CDFSALPHA ))141 .(INDEX>0) 1L61142 rNDZX4-1

91i

143 -L6144 COMPUTING CONFIDENCE INTERVALS USING NORMAL APPROX. W/C.F.1'45 L5:UALPHA-(DENOMZx(NORMPTH ALPHA ) NM-0.5146 A ROUND UALPHA DOWN AND INCREMENT 81 ONE147 INDEXeLUALPHA I148 L6:IPX NN INDEX149 CI.IPX CON5MW A150 ' A ' (DCC) ' PERCENT CONFIDENCE INTERVAL FOR TEE SHIFT IN151 ' LOCATION hETWEN POPULATIONS X AND Y IS:'152 '

153 I ( ',(*CIE1]),' S M(Y) - M(X) S '.(CI[23).' )'154155 -0156 E:'ERROR: THE SIZE OF YOUR SAMPLE IS LESS THAN TWO; TRY AGAIN.'157 1158 -81159 E3:'ERROR: YOU HAVE ENTERED MORE THAN ONE VALUE. TRI AGAIN.'160

161 -83162 E2:'ERROR: YOU DAVE NOT ENTERED A VALUE OF 1 OR 2; TRY AGAIN.'163

164 -82165 E4:'ERROR: YOU HAVE NOT ENTERED A VALUE OF 1, 2, OR 3; TRY AGAIN.'166] '167 *(A=2)/N2~168 .86

7 NPSLR-N.SUMXSUMXYXBAR:YBAR;SUMX2:SUMX!:B;A;WW-XX;BB:UOD:ALPHA:P:CC;CDF;TALPkA.NN-CI.SLOPES-RA.SR ....DENOM..NDE: X .... 4.w.... A..V- A PROGRAM CbNDIUCIS NOANPAAAMk'tR±C LINEAR REGRESSION. T LE EAST SQUARES3: 2 ESTIMATED RECRESSION LINE :S COMPUTED hITH HYPOTHESIS TESTINC AND3 CONFIDENCE INTERVAL AVAILABLE FOR THE SLOPE B. IF B DOES NOT LIE IN14 A THE C.I. AN ALTERNATE REGRESSION LINE IS PROPOSED. SUBPROGRAMS CALLED

5 ARE: SPMANP, K2NDALP, NORMPTH, INPUT, AND CONFLA.6 CPP.57 A INPUT DATAa R INPUT 2

10 X~~( 1*11 Q+i12 a ASSIGN TE SIZE OF X (AND 7) TO N3 N pX13 n COMPUTE THE SUM OF X'S AND I'S

15 SUMX-+/X16 SUMX +/Y17 A COMPUTE THE MEAN OF X AND Y18 XBAR.*SUMX#N19 YBAR SUN 4N20 a COMPUTE THE SUM OF THE X'S SQUARED21 SUMX2 +/(X*2)22 a COMPUTE TEE Sam OF X TIMES I23 SUMXY.+ (XxY)24 A COMPUTE '', THE SLOPE OF THE ESTIMATED LEAST SQUARES REGRESSION LINE25 B.((NxSUMX7)-(SUMXxSUMI ))((NxSUMX2)-(SUMX*2))261 aCOMPUTE 'A, TE 7-INTERCEPT27 A YBAR'(BxXBAR528 FF-'N'29 ' THE LEAST SQUARES ESTIMATED REGRESSION EQUATION IS:30 '31 ' z = ',(*A),' + ',(wB),'X.'3 2 '33 'DO YOU WISH TO ENTER SOME X VALUES TO GET THE PREDICTED Y''S? (Y/N).'34 WW -35 (WAW='N')/LI36 L2:'ENTER X VALUES.'37 XX.I938 a CALCULATE PREDICTED 7'S39 Y7.A+BxXX40 'THE PREDICTED Y VALUES ARE: ',(O77)41 ' )42 ' WOULD YOU LIKE TO RUN SOME MORE X VALUES? (7IN).'43 -*J-3.4 .WW=1Yl)/L2451 I:',WOULD IOU LIT. TO TEST HYPOTHESIS ON 3. THE SLOPE OF THE EQUATION? /

IV).46, WW-47 (WW 'N')IL348 'ENTER THE HYPOTHESIZED SLOPE.'49 88.050 A COMPUTE U'S51 UI-(BBxX)52 CHA.'> 153 a CALL SPMANP TO COMPUTE RHO AND ASSOCIATED P-VALUES.54 D.X SPMANP U55 *(D ~l) ')/Lul56 CA "57 Ll:PV 2xD[23

V -, r-*'2, , , , ,, .....

59 aPv-l6 60 L19:' SPEARMAN''S R EQUALS: 1,(4*DE13)61] 1621 'THE P-VALUE FOR HO; B '(B)'VERSUS HI: B ',(vCHAEI 2]).E0B).' IS:

63 1 1('4D[2J64 'THE P-VALUE FOR THE TWO SIDED TEST OF HYPOTHESIS IS: ', (4*PV)65 '66 a IF USING THE NEW REGRESSION EQUATION BASED ON MEDIANS, EXIT HERE.67 L3:.(FF='Y1)/068 A COMPUTE CONFIDENCE INTERVALS ON B69 'WOULD YOU LIKE A CONFIDENCE INTERVAL FOR THE SLOPE? (TIN).'70 W71 (W=fI)/072 L1O:CC-INPUT 573 A CHANGE ENTERED VALUE TO ALPHA714 ALPHAe(100-CC)+20075 4 ROUTE TO NORMAL APPROX. FOR CONI. IN?. OF LARGER SAMPLE SIZES76 *'(N>12)/LS77 A COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE[78 P-RENDALP NF7 7 ,) - [ -S80 A I;VDEX POSITION OF VALUE IN CD! :5 ALPHA81 INDEX-(+/ (CDFSALPHA))82 *(INDEX>0)/L683 INDEX-*-84 *L6185 a COMPUTINU CONFIDENCE INTERVALS USING NORMAL APPROX. WIC.F.

* ".86 L5:D9NOM-(N(- )x((2xN)+5))+18)*0.587 TALPHAeDENOMx (I (NORMPTH ALPHA))88 -L989 L6:TALP9A-Pr3. 'NDZX]~90 Lg:CI+X CON'A I[91 NN-+CI[92 SLOPES.1+C1[93 RR-L ((NN-TALPHA)+2)Lg :o *(RR=0)/L2095 RR-2.96 L20 :SR-.rC1.<( NNI+TALPHA )+2))9g7 -<SR:5(pSLOPESJ)/L2I

~98 LSR.:SLOPES9 j21:'A ' (vCC ) I PERCENT CONFIDENCE INTERVAL FOR B, THE SLOPE OF100 z TH ETIAfkD REGRESSION LINE, IS:'101 1

102 C ,(vSLOPE SERR]),' < B < ',(vSLOPESESR]),' ).1104 IF B OUTSIDE TSE(C.I. CALCULATE9 NEW EQUATION BASED ON MEDIANS105 (.BZSLOPES[RRI)A(B5SLOPESJSRk)J106 A ORDER XN1071 X.-X [109 CH HCK TO SEEF IF THE SIZE OF SS IS EVEN OR ODD FOR FINDING MEDIANS110 *((21INN)=0)/S1111 a COXPI7TE MEDIAN FOR ODD CASE

113 -S2114 a COMPUTE MEDIAN FOR EVYEN CA SE115 S1:Be(SLOPESt (N2 ))+SLOPESf ((NN+2)+2) ]942116' 12: DO TH SAME F OR THE X AND Y VECTORS17 A2.( IN )0)/5

118 IBARI. -Y N+1 +2)1 19 XBAR*XL (N+1).21]

* 120 -OUT*121 S3:YBAR+( [ (N42 )J+Z[((N42)*2 ))2

* *~122 X ++XLN2)] +E (N+2 )#2)] 4.123 A COMPUTE NEW I NTE-RCEPT I A'*124 OUr:A-YBAR-(HBcXBAR),25 'THE LEAST SQUARES ESTIMATOR OF B LIES OUTSIDE THE CONFIDENCE INTERVAL.'

1126 'DISCARD THE LEAST SQUARES EQUATION AND USE:'

[128 + ,w)'4 ',(*B),'X[129[130 ' THIS EQUATION IS BASED ON THE MEDIANS OF THE X AND Y DATA AND~1311. THE 'IEDIANi OF THE IWO-POINT SLOPES CALCULATED FOR THE CONFIDENCE INTER?

* ~ALLOW 9SER :0 O 0SOME ;NAL.-S:S IN NVEW EQ UAZ'N .li34J 'DO YOU WISHj TO ENTER SOME X VALUES TO GET PREDICTED Y''S FROM THE NEW E

131QUATION? (Z/N).'

136] 4.(FF:'I')/L2137] FF+'7

*** % ~ > ~*~~..

7 SIGN;A.CBDPVALXMO.NVCDF;ALPHACIY-AABBCC;DD;PV;PVI;NNN;KPOS;ORD

I a HIS PLINCTIOk OSEb THk CADI----iTS'O ACLT H2 a STATISTIC P-VALUE AND CONFIDENCE INTERVAL AS A TEST FOR MEDIANS.

E3 a THE LAST bPTIORV OILL DISPLAY A TABLE OF CONFIDENCE INTERVALS OF'4] A ORDERED STATISTICS wITH CONFIDENCE COEFFICIENTS.5S. A SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: BINOM, NORMCDF, NOEMPTE,~6. A INPUT SAND QUANC.

7 'I

le I DID YOU ENTER THIS PROGRAM FOR THE SOLE PURPOSE OF GENERATING CONFIDENCE INTERVALS FOR A SPECIFIED SAMPLE SIZE AND QUARTILE? (YIN).'

H 3 WW-m101 7' W1Y)/B'4i I.(~= THE NULL HYPOTHESIS STATES - THE POPULATION MEDIAN (M) IS EQUAL T

0 THE HYPOTHESIZED MEDIAN (MO); HO: M =MO.2

~13 WHICH ALTERNATIVE DO YOU WISH TO TEST?'14I15B3:' ENTER: I FOR Hi: M < MO; 2 FOR Hi: M > MO; 3 FOR HI: M = MO.'

E B2: 'ENTER: 1 FOR SINGLE-SAMPLE PROBLEM; 2 FOR PAIRED-SAMPLE PROBLEM.'H9 AA-O20 (.q )A (,U=2 ))I E2

S21 * (AA2)ILg22.A INPUT DATA FOR SINGLE SAMPLE CASE23] X.INPUT 124 NNN.OX

25 MO+.INPUT 3.21D-X-MO

2 PAIRED SAMPLE CASE

301 i -e-R

S33 DD-X -73U NNN-oDD

35 MO.INPUT361 D-(X-Y)-MON37 ACOMPRESS D TO REMOVE ZEROS358 LU4:A-(D:O)/D39 A RECORD LENGTH OF A AND ASSIGN TO N'40 N-pAL41 $I KEEPING TRACK OF POSITIVE SIGNS42 K.POS.+I (A0O)43 -(NZ30 )INORM'44 PVAL-BINQM N45 ;(PS>O )IFi

47 -P248 P1:PVI-l-PVAL[KPOS)49 P2:PV-PVALE (K.POS+i150 *L651 a IF N I .GREATER THAN 30 U E NORMAL APPR9X WFI-CONTINUITY CORRECTION

52NR(Zi (P -0 5)-(O 5XN) 4 0.5x(N*0.5)j53 Z (KPO +0.5 -(6.5xN)5.*(0.5x(N*O.5))155 PVIi- (NORMCDF Zi)56 AIF PAIRED SAMPLE TEST GO TO L17 FOR OUTPUT STATEMENT57 LPV2x(/(PVP VI))58 (V3)/559 PV3-160 N5:PVM-(3 1)p PV PVI PV3)61 'COMPUTA±IO S AE BASED ON A SAMPLE SIZE OF: ',(VN)6263, 'THE TOT-AL NUMBER OF POSITIVE SIGNS IS: '.(*KPOS)64 ' '

65 -(AA=2 )IL1766 'THE P-VALUE FOR Ho: M =',(*MO),' VERSUS HI: M ',(*LOGIC[C;12),1 ',(vMO)

L69] L17:' THE P-VALUE FOR COMPARING THE MEDIAN OF THE POPULATION OF1701 'DIFFERENCES TO THE HYPOTHESIZED MEDIAN,'

'90q: M(X-71= ,ZMO).' 7ERSUS 31: M(X-Y) ,, c.L0CZCC*.::: . :SO)

~73 1 ,

74 L18: 'WOULD YOU LIKE A CONFIDENCE INTERVAL FOR THE MEDIAN? (YIN).'75 B.76 (BB'Y)L677 :*UANT78 a INPUT SIZE OF CONFIDENCE INTERVAL79 L16:CC*INPUT 580 A4PHA*(100-CC)*20081 * (NNNZ30)INORM182 aCOMPUTING CONFIDENCE INTERVALS By EXACT P-VALUE83 CDF*BINOM NNNa" aINDEX POSITION OF CDF FOR ALPHA + 285 B.+I( CDF!5ALPHA)86 -(B> ) /SFJP

87 3.1ias *SxxP89 a COMPUTING CONFIDENCE INTERVALS BY NORMAL AJPROX.90 NORM1:KALPHA-((0.5x(NNN*0.5))x(NORMPTH ALPHA) )+(0.5xNNN)-O .591 A BOUND KALPHA DOWN IO NEAREST INTEGER AND INCREMENT BY ONE92' B.LEALPBA+l93 A IF SINGLE SAMPLE CASE GO TO L794 SKIP: (AA:I)/L795 A CALCULATE AND PRINT OUT CONF. INT. FOR PAIRED SAMPLE CASE96 L5:ORDD.DD [,DD]'7 7Z+PORDD91 CI.ORDDBJ ,ORDDr(YY-(B-I))'99 'A ' (uC PERCENT CONFIDENCE INTERVAL FOR THE MEDIAN OF THE100 :POPULATION OF DIFFERENCES IS:'102 ',(ocl[lJ),' s MEDIAN(X-Y) i ',(fCIC23),' )f103] ' '

104] QUANT105 a CALCULATE AND PRINT OUT CONF. INT. FOR ONE SAMPLE CASE,1061 L7:ORDX.-X'X"L10T7 II ORDX1I08] CZ-ORDX[RjORD!HUY-'(8"i))]Mg 091 ,mC.! P CENT :ONF:DENCE N.RVT7AL FOR THE MEDIAN OF THE POPULATIO110]1,111 ' ( '.(wCI12),' f. MEDIAN < ',(VCI[2]).' )'112]113 QqANT: 'WOULD YOU LIKE CONFIDENCE INTERVALS FOR A SPECIFIED QUANTILE? (Y/N

nt. : -WW='Y)I!

,117' N'ENTER DESIRED SAMPLE SIZE (SINGLE INTEGER 7ALJE).'

L120 Ij ((1 (NNN)=o0)/Eu121 1:'ENTER DESIRED QUANTILE; FOR EXAMPLE: ENTER 20, FOR THE 20T QUANTIZ.

R23 (QUA 50,v(QUA>1OO))/Z

12D NNN QUANC QU AS1251[116] '***** THIS TABLE GIVES CONFIDENCE COEFFICIENTS FOR VARIOUS INTERVALS WITH 1127 'ORDER STATISTICS AS THE END POINTS FOR THE ',(mQUA),'TE QUANTILE.1128 1129 -0130 E1:'ERROR: THE QUANTILE VALUE MUST LIE BETWEEN 0 AND 100; TRY AGAIN.'131 4132 -BI133 E2:'ERROR: YOU HAVE NOT ENTERED A VALUE OF I OR 2; TRY AGAIN.'134 11.35 -B2136 E3:'ERROR: YOU HAVE NOT ENTERED A VALUE OF 1, 2, OR 3; TRY AGAIN.'137 1138 -3139 E4:'ERROR: YOU HAVE NOT ENTERED A SINGLE, INTEGER VALUE: TRY AGAIN.'140 1141 *34

V SPEARMAN:X;Y'A' Q R;CHA;PV;I A THIS MNT1A CO*MP&JTES THE SPEARMAN R STATISTIC WH::S XEAS:'FF-2 A THE DEGREE OF CORRESPONDENCE BETWEEN RANKINGS OF TWC s;MF:F t

3 VALUE IS GIVEN FOR TESTING ONE AND TWO-SIDED HYPOTH.S: :FA SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: T:ES. T:ESK. SF7.EZ.

5 A INPUT SPAPROX, INTERP, AND THE VARIABLE PMATS:.,6, R.INPUt 27 M +Rq ll +

7,H'L SPMANP TO CALCV '_' 'F ST4=s-~ Z.r r-.~

,L * :L215 L1:CHA-'DIRfCT'16 L2:PV-2xA 2J17 *(PVSI)IL3Is8 PV-119 L3:'SPEARMAN''S R EQUALS '.(uA:11")20,21 'THE P-VALUE FOR HO: NO ASSC-.. :..221 1 81: ',(*CBA),' ASSCC:4T$ rl -

23 ! '2u 'THE P-VALUE FOR THE TWO-SlDFR TEST YF FFr25 '

76 -t14 971 FIN APL WORKSPACE FOR CONDUCTING NONPRRRlETRIC 12/2STATISTICAIL INFERENCE(U) NAVAL POSTGRDUATE SCHOOL

UNCASIFI~tMONTEREY CR N F YROTS JUN 677 F/IFIE F/L2/3 NL

*, *%, %

1 1164

.. ~p.

,,,, q ~' W~~ii~W'U~i li~~u~z T~aP~~ V~A~'IM MDIAV (N) IS EQUAL r'1 MY.-OF AZrnuLTIVE DO YOU VISE ?0 TSf?E1Y wU: I FORII x NcRa 2 FOR 11: X b X01 3 FOR1III a NO.,

0A(C-2)A CS)E; SZ*' L5W-SAM~rj PRODLEM: 2 F05 P&XIUD-SJJIFL PROBLEM.'

a -Lp * A 2 z ' X:PUr DATA FOR SINCZ9 SAJIPE CASE

Al!:.@T PAIRED SAMPLE CASE

L11:.(Dn)/DCOMPRESS D TO REMOVE ZEROS2 NaA RECORD LENCTOF A AND ASSIGN TO N

'PS(-)ZPN RC OF POSITIV SIGNS:AEE I'll ASOLUTIT 7A.L0Z OF A. ASSIGN ?0 3 AND ORDER 3

pos.voinul; osrrrys SIGNS TO COINCIDE wITE PROPER POSITIOS IN I

s.- Eel CI UCO OALL POSITIVE VAUE 0REA EX2W; BY ADDINGE I ACROSS ALPSfT AUSO

N~~e(L(( Of LI??r TAIL OF PROD AJILIT' DISTRIUTION

(#2,16) /L3 SrArEMEJTS UASED ON LENCTI OF VECTOR 8ao-N GENERATEf NULL DISTRINUTION FOR TPLUS

"' FA LT EAL? 0F PROD DIST CALCULATE PVALUX AS NORMALai E E NE sGATvz r' STATISTIC

S~( USE DOTE YEE INTEGER ABOVE AND BELOW AS TPLOS

So )/GO

.r2

: j Ihhi?1111 VOA PPROCXMATIN:TE! -COTINUITY FCRCINFTO

)(NuM-0.5 *2 )(N-1 ).O.s

i Dr96

"S SC AN b5 S %% -S%.

gS;,~S /N2)$C~

4~~ ~~ :V( TDSTOY5*

r, 3:1

o0 2ry '1N OTAL SON OF POSITIVE RANKS IS: '.(VTPOSl)

1I012 a, F PAIRED SAMPLE TEST GO TO Li7 FOR OUTPUT STATEMENT

8.041 J-J"2 i4 K M q ,(WNO),' VERSUS HI: M ',(vL0GICcC;i3).' ',(G4O;15 ,:I. 6', ~ j

P107 1i7: '"99 -7ALJ9 FOR COMPAR:NG THE VFEDrAN OF THE POPULAZION OF108 : DiFJEENCES TO THE BIPOTESkIZED MEDIAN,f,.... '3C: N(X-Yj2 '690).' 7ERSUS 3i. N(X-I) ,cLCC;:.''.(*NaO.'

Li12 L1 U:'VOULD YOU LIKE A CONFIDENCE INTERVAL FOR TEE MEDIAN? (Y/N).'

.. , ALPSA-':OC-CC)#200iiS. aROUTE TO NORMAL APPROX. FOR CON? IN? OF LARGE SAMPLE SIZZ

119. -(NNN>16VLu~120, a COMPUTING CONFIDENCE INTERVALS BY EXACT P-VALUE121~ CDF.YILP tN122 a INDEX POSITION OF CjPF FOR ALPHA + 2123: TALPIA (+/(CD?!CALPEA))

(TJUMP WIUM

a COMPUTING CONFIDENCE INTERVALS BY NORMAL APPROX. W/C.F.126 L.:UOZ(2xNNN NNli)x((2wNNN)+1)),3 *! .129 TxPE. ( DNM NORMPTB ALPHA)(Nx 5NN1) )4tL) BYON130 a rN MAWI 1A Di131 TALPE*LTALP PEA DOWN TO IrNTEGER V LUE AN -NCRH~T3 N132 TALB LZLAi133 aUP.Az)L IF ONE SAMPLE CASE 0O TO0 L7134 a CALCULATE AND PRINT OUT CONP. INT. FOR PAIRED SAMPLE CASE135 L3:CI. ALPA1136 'A 'jV (!C RETOmNDENMCE INTERVAL FOR TEE MEDIAN OF THE137 'POP LXI OF DFFERENCES IS:'136 ''139 ' '(*C13),, f, NEDIAN(X-7) & ',(vC1E23).' )l140 '141 .001042 aCALCULATE AND PRINT OUT CONY. IN?. FOR ONE SAMPLE CASE143 L7:CI.TALPIA CONFW X144 'A ',(uCC).' PERCENT CONFIDENCE INTERVAL FOR T BE MEDIAN OF THE POPULATIG

N IS:'*145~ 1 1

:146' l '(vCI[I3).I S9 MEDIAN :9 ',(fCIE23),' ),147 1,140 *W0,149 22:'ERROR: YOU HAVE NOT ENTERED A VALUE OF 1 OR 2: TRY AGAIN.'

.151' -821152 E3:'RROR: YOU HAVE NOT ENTERED A VALUE OF 1, 2, OR 3; TRY AGAIN.',153 ' '

.1 sk: .33

97 a

d' . w . .~~ ~ .**a ***>**~~ *~*~* ~~U ,L

APPENDIX F

LISTINGS OF SUBPROGRAMS BASIC TO BOTH WORKSPACES

T .0# N0ON N iN P:X;'CDF

II .Tf S UN ri N -IS A S5PROGRAM OF THEP SIGN TEST (S IGN. I? CALCULATES2~ * 55 CD? OF THE DINONAL WHEN PROBABILITY = .5. N=SAA4PLE SIZE.

V O+L CBIN R

2 HIS ?r UNYTION IS A SUBPROGRAM OF CONFIDENCE INTERVAL GENERATOR FORI T153 Q f QUATILE. IT RETURNS THE VALUE OF THE BINOMIAL CDF AT R,WITH N.P=L WHERE N=SAMPLE SIZE AND P=PROBABILITY.

V CON*XX CONFLR 71'BBSS;AA'A'XR'YR;BC-S1I l HIS FUNCTIOk It A SUShaOGAAMO kONPARAM ERIC LINEAR REGRESSION22 a (NPFLR. I CALCULATES THE TWO-POINT SLOPE FOR EACH PAIR OF POINTS3- (XIII) AND (X 4 JJ .)ALL<< AND XI-XJ. ALL SLOPES ARE ORDERED AND4 USETO FIND TH INTERL OR 8, THE SLOPE OF ESTIMATED5" A EQ ATION. XX= X DATA SAMPLE AND I= Y DATA SAMPLE.

7 a RECORD THE SIZE OF XX AND INITIALIZE VARIABLES

11 a THIS LOOP CoMPRESSZS ZAND 7 DOWN TO WHERE THE XX < ALL OTER XXS12 L2: A.Ai13 A.(X%iAA.,CXX)14 XA A <X

15 YR.A/YI16 XR17 .(*B=0)/L318, C*O19 a THIS LOOP CALCULATES THE SLOPE OF EACH PAIR OF PAIRED DATA.20 LI: C+C+121 5. (71 CAAJ-YR C) )(XXCAA2-XRCCJ)22 SSTSS S23 (C<B ILI24 L3:(AA<BE )/L225 JON (pSSl, CSS[ 4SS3)

A'7 V CONFM-AA CONFMW RBA;B;C*D-EF'GHNTI a THIS FUNCTION IS A'SUBPROGA OF THE xANs-WRITNsY TEST (MANW).

2 A IT COMPUTES CONFIDENCE INTERVAL ENPOINTS FOR THETA, THE SHIFT IN3 A LOCATION BETWEEN X AND Y. AA=INDEX POSITION OF C.I. ENPOINT.4, BB=COMBINED DATA SAMPLES.

A AqSIGN SIZE OF X VECTOR TO A; INDEX POSITION FOR CONF INT TO B.

Q SH ASIN X 79CTOR TO ;C; V ECTOR TO Di DA + 38

11 a REORDER X AND Y VECTOR VALUES TO ASCENDING ORDER

14 a INITIALIZE VECTOR 5 AND VARIABLE F15 Mis F4-17 mINNER AND OUTER LOOPS CALCULATE ALL POSSIBLE DIFFERENCES; EVERY 718 a ELEMENT MINUS EVERY X ELEMENT. B VECTOR STORES THESE DIIFERENCES.19 12:G-120 Li:E.D (F)-CEG3211 H-0B.E

p~jV ~ s ~ *~. Pb ' - ~ - r ~ .>. % -. *~~=~ -.. !

ob~ (F( VODI sAhri Ir vicTOiR VALUES TO ASCINDING ORDER

L20 UE 08ZTECN INT VAL"E FROM H

v GONr CINFW B-C;DEZG3N~ FAI S5 '*(fl,9Ar COIN f ~ "" IT RANK TEST

2 . ®r~k C NR A 55 E~ah AS D ON THE3 im E~ Lf.l fA PRID goFIf A Oj 5 RA j1 Z A s X . A x INDEX

aSTART I ACCUMULATION VICTOR OF? MITE ORCINAL VICTOR VALUES7 U00

id' OUTSIDE LOOP INCREMEfNTS D AND RE SETS f TO DiiL2:D-D+i

12 I-D13 j a NSID LOOP CENERATES NEXT SIT OF AVERAGES AND CONCATEPNATES TO ORIGINAL

VCR

*17 aCONTINUE INNER LOOP MNT= I EQUALS THE SZE OF C

19 CONTINUE OUTER LOOP UNTIL D EQUALS THE SZE OF C LESS ONE20 (Da p)I)L:21 ORDER FINAL ACCUMULATED VECTOR H*23 aNDXQFIT VALUES OUT OF H

V FeD? FDI STN X-A:M-NRM-RNLM-LNSM SNM2N21 YETAI FuNcrI6N It~ tus6ROfAI of 'fDkUtAL-FJALLIS TEST IKRWL) AND

2 EYZSU DENT TDIST(DSN.I jXfCA MCLAT PP O UMULATI VE PROBS.

6 A xX+Sm.x-lDF5 .14DF7 * (M'>2 )vNp2is a REATl THE 1ST. FOUR SPECIAL CASES.

"I'L' 1 0sUN1 Lll:Pet-IOA*0.3)+04211 -012 L12:P.A*0.g13 -014 L21:P41-(1-A)*0.g15 -0161 L22:P-A17 .0018 a BEGIN THE GENERAL CASES. INITIALIZE THE QUANTITIES.1.g L:R MIN2) ORNe1 I P220 S =0(R :O )xjLM-LM~e *.5xM-221 S4.(RN=0 iRNILN-LN2*0.5xN-222 a R SN CSE N-EVEN FIRST.23 .NODD-i1LM214 11 I~ O2

25 SIe+,/x\(1-A)e.x1+(1+0.5xM)%N226 .M VN27 NODD:eMEPVENx N28 SN. (1-A )*0.;5=29 :.REAT TH PORTION OF SN THAT DOSN'T CHANCE M/ M ODD OR EVEN.30 ,~+eL ) O=LN

'4.31 S e-SNXI+4&/X\(1-A).x142x~M241+2x1LN2 I REAT :%9 q'--;7EN 3 UBC-SE13:-( (242).!=12M),MEVEJV3.-LC

35 efMZVEN~36 a TRAT THE N-ODD SUBCASE37 S JNX2+ol38 eENDE=M39 SN.Nxx/l++*1+2xi0.5.N240 EV9N :*MODDxxU:21M'41 -ENDx%2:N42 SN4.%++/x\[emailprotected] *IND44 a NEXT TREAT TitE $PECIA4 CASES FOR ODD-M.'45 MODD:SMN (0xM:1)4.( A*0 *5)+0+2) x3=M'46 .ENDxt(1:M)v3=

'46 END: .(SNXAH.)+4xRNx.N-RM)+( (RN"RNx loA*0.S)4o+S)-2xRNxSNx(-A)*0.5

~.0*

a SO I G SDI z LATA ATPLI I CURSN wl

RECORD SIZE OF N TO DETERMINE NUMBER OF SAMPLES IN BAA'.oNa LOCATE LARGEST SAMPLE SIZE

m* ORDER B SMALLEST TO LARGESTI DD-BCAD

aD -(AiET UP MATRIX OF SIZE REQUIRED TO STORE SAMPLE RANKING.Xx1(Al ) X0a CONCATENA!E 0 AND N1 FNINO';RD CUMULATIVE SUM$ OF SAMPLE SIZES.

1 2 HIS LOOP INDEXES OuT ORGINAL SAMPLE VALUES FOR FURTHER CALCULATIONS.2 LI :CC*CC I.3. ,OCTZ. C'.j RI rYAL SAMPLe 7ALUES ZN Bi4 Zl*BC(NCC t "C )

• a 4x1 Os(CIoNS OF ELEMENTS OF X VECTOR IN B AND ASSIGN TO C~ DD X 111

221 * TkIS LOOP DtErCT$ TIED INDEXED POSITIONS AND INCREMENTS TEE INDEXING OFEACH SUCCZSS. . .. ED POS:T:ON BY ONE

9 a TES± DTH ELEMENT OF C AGAINST REST OF C FOR :!iES30 L2:E.(C:D]=D+C-

a SET P EQUA. TO ITE APPROPRIATE SIZE 'DTH SIZE) ZERO 7ETOR

33, CQN.ATENATE F AND vECTOR OF 0'S AND 's (1'S APEAR WHEN TIE OCCURED)

3S ADD RESULTANT F VECTOR TO C

38 a CONTINUE FOR ENTIRE C VECTOR

1 1 J DjES OSF TE INPUT PROMPTING AND ERROR CEECKUNG.*(L=TR2 L3P O d PROTDREA TW:BEVAIN'R RQIE)3 X? IS 1 SUBPROGAM OF SIGN, VISIG, MANYI, AEN, SPMAN, AND NPLR.

it -O 2',~ I

2:'ENTER X DATA (MORE TEAN TWO OBSERVATIONS ARE REQUIRED).'

,1 '14zA r bAA (NUMBER OF 7 ENTRI ES MOST EQUAL NUMBER OF X ENTRIES).'."

15 r~a

" 00 L2:'rTER T AE H MPOTHESIZED MEDIAN.' R

2 !IN.U Z49

2 *(pIN))I )/E3 "-

23 .0 2

24 l' ENTER THE HPOTYESIZED MEDIAN FOR T LE DIFFERENCES OF THE PAIRED DATA.'

26 *U(IN)>I)/E3

11-0

2. 3: THENTER THE DESIRED CONFIDENCE COEFFICENT:

30 ' FOR EXAMPLE: ENTER 95, FOR A 95 PERCENT CONFIDENCE INTERVAL.'31 I

32 *(IN0 )v (ZN>100 ))/Ei&3 332 *(( 1/IN )2O)/ES

23324 *Ql35 E1:' RR THE SIZE OF OURSME IS LES TH AN THFF REE; TrY AG AI.' DAA.

39 E :'ERROR: SAMPLE SIZES ARE NOT EQUAL; WANT To TRY AGAILN? (I/N).':1 *{B:'Y' )/L2

2 'ENTER RIGHT ARRO 5, TO QUIT.'310

32 Yj0J-lJ~ll0))1a

33 M I X0 /9

34~ ,0

49ARO: TIE 17P0ESIZZD MEDIAN MUST BE A SINGLE VALUE: TRI AGAIN.'

:3: .)/L34:1EROR: THIS VALUE MUST LIN BETWEEN 0 AND 100; TRY AGAIN.'

E*L5E5: ERROR: TEIS VALUE MUST BE AN INTEGER; TRY AGAIN.'

.c114 *L S

R ~ D i A~fj UNN I'L.U TEE

£PASS D AS :50 RE GIKENT.

aSjPAR4TE TEE CD? TABLE AND STATS. INTO SINGLE VARIABLES.

a j .C 3;a HERE A FIRST EXCEEDS OR EQUALS ONE OF TEE TABLE VALUES.

'l -(+Cj80STf DOES NOT EQUAL ANY TABLE VALUES SET NTR zz-.Ju nD ~ lcfomoCFrs ocjz2. F AC

~o a I? :NDgXED POSITION EQUALS ONE INDEX P-7ALUE OUT OF CC.

~20) a TEWS ONDUCT INTERPOLATION TO GET PROPER P-VALUE.21 L), ; lsDpifj-E2j4 gL~ D2lj.D_ 1

r s rNT* 0CR-1J.+PL

2~~~~~~ ~ L N.lDAPNB D N (KEN) AND N N-

PARAMETRIC IN R ~ "SlfON ( NlR. IT CALCULWES THE CUMULAIEISexTRIU, W 0O Ba MPL SIZE N.

5 a NN2I ITALXZE FREQUENCIES FOR X FOR SAMPLE SIZE MN.7 a D TERMjNE $IZE OF RIGHT PROD. TAIL VECTOR

91 acoi OUERJ gO INCREMENTS THROUGH TEE N SAMPLE SIZES TILL THE DESIRED ONE IS

10 Jl1:i+0011 C012 f:00z13 NN.tN+114 *.((NNX(NN-1)).2)+215 A41.1N a INNER LOOP GENERATES NN+1 FREQUENCIES FROM TEE VECTOR OF mN FREQUENCIES17 L2:A*A-1Is C4:C4XC (B-A))19 DDC20 AA-d

21 a ISIZE Q D EQALS NN AND INDEXES OF X STILL REMAIN GO TO L422 *(( UN1 (-)F1/L23 aTE S8S CON TINosE To INCREMENT TERU OLD FREQS

25 a WREN LFT HALF OF NN+l VECTOR IS COMPLETE GO TO L5

34q "+XiUB-A)J-XZAAJ

31 E.1.31] a OM TE N VC O REQ D BND ONCTENINGDWT39 L3:9*D

'40 CONTINUE UNTIL SIZE OF SAMPLE N IS REACHED

'42] :NGNEATE VECTOR OF CORRESPONDING P STATS OF PROPER SIZE C

101

"3 P.0 .1.44 1 1 I~S 4 RO mghP VICTOR

~ '~r4~j'?%~) l*miA±E.z1 FROM Tar FREQUENCY VCTOR

14 C ANC FRZQS TO)Cf C JDALE AND OUTPUT B S'TATS WFIAPPRO. CD! VALUES

V MAN-N MANWP M-F.Q P-ST;U-VB -NNMUUMM[1r F UNJTfON IS 'A5Bjb)MOP TOE MANN-WHITNEY TEST MXAYW).2 mJ l RAUJ r E STYIir f. FPOR HEU STATSTIC. N LSZ OF

51Qa~ U~NUMB9j)OF TERMS TO BE INCLUDED IN LEFT TAIL DISTRIBUTION LESS 1S M. NM+ +7) a F~,M0 SET F VECTOR EQUAL TO 1 CONCATENATED WITH MM -ZRO'S

aI SET P EQUAL TO THE MINIMUM OF NeM OR MM

11ii a SET Q EQUAL TO THE MINIMUM OF N OR MM122 Q-L/M MM13) a GOTO LINE DEOM IF MM IS LESS THAN N+1(SIZE OF X+1'5

15 IFNmMN+ llENRATE FIRST BLOCK OF RECURSIVE RESULTS USING NUM LOOP

.9 PRIMAR)' ?ORMULA USED IN GENERATION OF FIRST BLOCK OF REcURS:'7F RESULTSt19 prj@?g 7r-"202 As S .V-W' tEC;?MZNTD VALUE TO U! AND TESTS I? Z THIS NEW U

23 aGENERATE FINAL RECURSIVE RESULTS USING DENOM Loop1 2 u4 DENOM:S.1~28a P4:V4-RT OMA USED IN GENERATION OF FINAL RECURSIVE RESULTS

[292 *L4x't(QkS*S+1[30) ot C NR FREQUENCY TABLE TO CDF VALUES FOR FINAL OUTPUT[312 MAN.(+\F )4(NL(N+M)

V Z.NORMCDF X-AB;C-D1pl mEAUA- bO~A . FF 26.2.11 fA AM qN

~.11 1A S.lAo4 SNJS V XDPW ~a~:f 5 &UI~~ POTDT hRM ~;ir t7H CHANE 0

52 1VI FDRiW PROBLEMS WITU x\

/§4)/Z4-X .X)=P.X )13+C

*A.5BJ.*A*2)X2)*05S)X+/X\A*2).12(C0r0r/)+

12 210 B *C 26~ 89790 56295540 52050600 19934640 3690160 341952 15232 256133 D- 2227021(32432400)7567360 6I540480*2162i0 3843840)314440 15360 256

V Z.NORMPTH P ABC-D T.S.RS_ V API DS 12977

11 S (p)T.0.42< L )/Z ~2.L12. A- 2.5066282N884 -18.6150006252 '1.39119773534 -25.4tslO6OL963713 Ba .735109309 23.083 6743743 21.06224101826 3.13082909833

17, Ef. 2.78189113 -2.29796479134 4.85014127135 -2.321212768581is D4- 3. 43?6892476231.637 Q678j1897193 S-(xS X( R.*0.t x ) +.X)1+.(R.(I o*5-)0.5- )..* 1 2)e.XD)

102]

j(0 O.424 IQ 31100-1Sir: 'ONE 05 8 ROSE I EANI Afi QQr Qf sffaE.'

V FePERK NK:X::a THS FN±bI$' CALLED 3Y SPEARP. IT GENERATES ALL POSSIBLE

a PERUTAT IONS OF N RANKS. NSAMPLE SIZE.3 0X.N=P" I1 I pl

5 r

1:1:. Ox N<X*+1+7 (~(m)eX)\Z

9 tiL1 XIN-1),N)p (.P) "Y

V M.N QUANC Q1JL;:I THIS FUNCTION vGkzvs A CHOICE OF NONPARAMETRIC CONFIDENCE INTERVALS2 FOR THE QT QUARTILE OF A CONTINUOUS POPULATION. N=SAMPLE SIZE,3,,~=AxQUATIL8. IT CALLS ER SUBPROGRAM CBIN.4]52 a CENTERS CHOICES AT "THE ORDER STATISTIC NEAREST ESTIMATE.!6 I'L 0 •5+N/xQ 1 00

07 Is i. NX±ODERD DEVIATIONS (APPROX.)

I r-! x 1u1 ±2 K!R N 1; BOT il7 '12 CONF DE COEFFICES13 L-I(2 pK)o(N 0.01xQ) CBIN(CI l)+K-1),(IE2]-K))

I & M- -ORDEA STATIHT1CS I COEFFICIENTS15 a PART o F iOR T -ING OUTp 2t

-7 U( 0 6 0 2,18 U-U, 4( , 1L19 MM,Ll. U, 12 6 c((pE),1)pLr.

v SPN-N SPAPROX X'YI n THISF UNCTION tS A SUBPROGRAM OF SPEARMAN'S R (SPHANP)

aIf APPROXIMATES THE CUMMULATIVE PROB FOR R WHEN PASSED THE SAMPLE3 n I IN THE LEFT ARGUMENT AND TE ABSOLUTE VALUE OF R IN THE RIGHT

,- AGUNENT. SUBPROGRAMS OF THIS FUNCTION INCLUDE: TDISTNe6.' •CALCULATE THE CONTINUITY CORRECTION

7 +6 -GNx 1+N*2a A TRANSFORM THE STATISTIC R INTO ONE THAT CAN BE USED WITH THE STUDENTill p x+ - x (N-2)* -iX-Yl*2i*0.511 L j DIST UNC O TO CALCULATE THE P-VALUE121 JPN-1- (N-2) TDISTN X

V SPEAR+SPEARP NC1;A;B-C-DEMN-DID2LzIM.RCDFI A THIS FUNCTION IS A SJkRGA FERAI' HSPY).T2 CALCULATES THE EXACT CUMULATIVE DIST. FOR R OR TE SAMPLE SIZE3 PASSED AS THE RIGHT ARGUMENT. BECA USE OF THE LARGE COMPUTER MEMORY

4 RE9UIREMENTSo N IS LIMITED TO SIX ON THE PC AND 7 ON THE MAINFRAME.5 a SUBPROGRAMS CALLED BY THIS FUNCTION INCLUDE: PERM6, A.B-07 a INITIALIZE VARIABLES, VECTORS, AND MATRICES.

*4 .- ks.a) 0 0

a .±S .0P ErERATSS AN Nxff ARRAY OF THE POSS:3LE 7ALUES OF DIFFERENCES12 A BETWEEN ANY TWO PAIRED RANKS BETWEEN SAMPLES.13 L1:B..+14 M B; +C-A15 A-A+16 *(B<N)/L117 a NOW CALCULATE THE SQUARES OF ALL POSSIBLE DIFFERENCES.I8 MeMj19 AJIL" PERM TO LIST ALL POSSIBLE PERMUTATIONS OF N NUMBERS20 DPRM N21 CALCULAE SIZE LIMIT OF FINAL VECTOR OF R STATS.22 :LIf-1 + (N_53 N *12)gOZrD231 IITII ZE AL B EFORE INDEXING OUT COMBINATIONS OF ALL POSSIBLE

103

JAIV

24 p SQUARED VALUES.25 AIQ26 8-1 N NO27 A TIS LOOP CALCULATES ALL POSSIBLE COMBINATIONS OF THE SQUARED VALUES.28 L2:A-A+29 1 ; IM rA;Dt;A330 2N) IL231 a DD DOWN ALL ROWS FOR EACH COLUMN TO SUM UP SQUARES COMBINATIONS.32 D2-+4E33 a ADD UP NUMBER OF DUPLICATED SUMS OF D-SQUARED VALUES AND COMPRESS3L4 a "ECTOR DOWN TO UNIQUE 7ALUES.35 L3:C1.C (L/D2)36 D1+D (D + /( (L/D2 )=D2))37 D2((/D2)4'2)/D238 *(0C1 )<LIM)1 L339 TRANSFORM SUM OF SQUARES VALUES TO SPEARMAN'S R STATISTIC.140 R-1-(6xCl).(Nx( N*2 -1))41 A CALCULATE CDF VALUES ASSOCIATED WITH THE R STATISTIC.1421 CDF-(+\D )+'Nt431 s FORM TWO ROW MATRIX FOR OUTPUT OF . STATS AND CDF 7ALUES.44 SPEAR (2,(pC1).)p(RCDF)V

V SPM+X SPMANP Z;C;D;DD;DI;D2;N;DENOMR;XX;YI;NS;NUMR;P;PV;PVAL:SU;SV;RHO;U;V;ARO;l UV1 X YiWW

A THIS FUNCTIOk S A SUBPROGRAM OF NONPAR LINEAR REGRESSION (NPLR)a AND SPEARMAN'S R SPMAN). !T COMPUTES THE SPEARMAN R STATST t

A AND ASSOCIATED P-VALUES. TEE LEFT ARGUMENT THAT IS PASSED ZS THE X4 SAMPLE: THE 21GHT ARGUMENT IS THE -' SAMPLE.

r5i z SUBPROGRAMS OF THIS FUNCTION !NCLJDE: TIES, TESX, SPEARP, SPAPROX,.I A ZNTEBP, AND THE VARIABLE PMATSP.7 ORDER I IN INCREASING ORDER OF X

10 a ORDER X IN INCREASING ORDERI1 U+-X [,$XJ

12 COMPUTE CURRENT RANKING OF I13 C+SV1&4 A NOW ORDER Y RANKS IN INCREASING ORDER15 D1.VCAVI16 IF TIES EXIST IN EITHER X OR I RANKED VECTOR USE MID-RANK METHOD17 DD.1 TIES DI18 XX-1 TIES U19 A FIND ORIGINAL RANKING OF I WITH TIES RESOLVED20 YI1DDiC221 A RECORD SIZE OF INPUT VECTOR22 N X23 a CALCULATE DIFFERENCES BETWEEN RANKS OF X AND I VECTORS24 D-XX-YI25 A DETERMINE THE SUN OF SQUARES OF THE DIFFERENCES26 D2+/(D*2)27 A OBTAIN THE NUMBER OF TIES IN EACH VECTOR USING THE TIESE FUNCTION28 UO1TIESK U29 V1.TIESl Dl30 SU ((+ /( U1*3 )) - (+/U1 )).123:1 SV ((+/ V1*3 )( V1 )232 NS+N ((N*2)-1)33 A CALCULAE T56 R S S TC INCLUDING THE CORRECTION FOR TIES34 NUM(N +( 6)xD2)+ ( 6 x(SU+SV))35 DENOMR ((NS-(12xSU )*.5 )x( (NS-(12xSV))*0.5)36 RBO.NQIMR+DENOMR37 AHHO IRHO38 (N>6)/Ll39 A CALL SPEARP TO CALCULATE THE RIGHT TAIL OF THE CDF OF R40 P SPEAAP N41 eL242 Ll:.(N>10 I)/L343 P+PMATSP (N-5) •J44 : CHANG SIZE bP P TO AN MxN MATRIX

46 CALL INTERP TO CALCULATE P-VALUE BI INTERPOLATIONL1 L2:?YAL.-4RHO rNTEAP P

*48 -. PVAL= !I/L7? L VAL-O. S

0 L7.51 CALCULATE P VALUE USING STUDENT T APPROX.52 L3:PVAL N SPAPROX ARHO

253 L7:SPM (RHO),PVALV

104I%4

,' .);/ : ;,'."j. .w TJ,%,': :',. : . .' .: ' ,.' ,%' %' % .';q__ .: ". .'.' ":l~lt'.''-' *4*-4

V P K fDISIN X;7A THIS PUNCTION IS A SUBPROGRAM OF THE CUMULATIVE PROBABILITY GENERATOR

2 FOR PEA MAN'S R (SPEARP). IT CALCULATES THE CDP AT X USING3 THE STUDENT'S T DIT WITH K DEGREES OF FREEDOM.

m THIS PNCTION CALLS ON THE 'F' DISTRIBUTION FUNCTION (FDISTN).

7 P.0.5k(2,K) FDISTN X*27 *g xtO~vtIs Pv/X *0.5.+V/P.9 -B X!A/VA C-V)/19X .. ..V/

V TIBB TIES B-CDIN;TY-ZK*-MLPP-NR-MMA THIS McfUN ok IS'ASJHRAk 4?b MEMALL'S B (KEN) SPEARMAN'S

. (SP.4A11P, 'USKAL'WA)L:S ±KRWL) MANN-WHITNEY (MANW)3 AND W.L'CXON fWISIG). rT CHEC..S HE RIGHT ARC. 7ECR FOR ':ES 4ND

4 CHANGES THE TIED POSITIONS OF THE LEFT ARG. BZ THE MIDRANK METHOD.51

7 A' F NO VECTOR OF RANKS IS PASSED; GENERATE ONE

1 L.NpO12 T.1.2 :( CHECKING FOR TIES 3Y :NCREMENTING THRU THE VECTOR

L,. .( COUNT NUMBER OF TIES; IF .VO TIES GO TO L2

16~ D--/C17 -7 *(D0) /L2i af RECORD WHERE TIES STARTED AND HOW MANY RANKS INVOLVED19 II.T. (D+1)20] a INCREMENT NEXT T BY THE NUMBER OF TIES ENCOUNTERED PLUS 121 L2 :T T (D-6122 al? T LESS THAN SIZE OF ORIGINAL VECTOR GO TO L3 AND START AGAIN AT NEW T3 -(T<N)IL3

25 a ASSIGN THE RANKS OF THE LEFT ARG. TO TI26 TI.BB27 a IF NO TIES FOUND QUIT28 *(Y=0)/029 Z-030 A LOCAE THE INDEXED POSITIONS OP TIED RANKS31 L5:PP.( (I1 ZJ)r 1)+((IL1+ZJ+I[2+ZJ 1)32 A FIND THE MIDRANK VALUE OF THESE RANKS33 NR. (+/TICPP) ) 1I2+Z J341 a ST UP VEC OR WITH ZEROS AND ONES; ONES WHERE TIE RANKS INVOLVED35 1.0

37 L :K K+C38 m-(TI=TICPPENK))39 J~.M)40 (K<IE2+Z1)/L441 ASET A VECTOR WITH ZEROS WHERE TIED RANKS OCCUR42 L.,MM43 a TRANSFORM ONES OF MM VECTOR TO MIDRANK VALUE44 MM-NRxMM45 a TRANSFORM ONES OF L VECTOR TO REMAINING UNCHANGED RANKING VECTOR46 L.TIXL47 A FILL IN MIDRANK VALUES48 TI-L+MM49 Z-Z+250 A DO THE SAME FOR ANY OTHER TIES INVOLVED BUT WITH NEWLY COMPUTED TI51 -(Y>Z)IL5

V TIE*TIESK AAAAkBLC D;g'Ta R ,SMA?) AND tSPVA1V1) .-;D iRUSXAL-WALL:S ,5 U AM ., • i -PE'ARMAN'

": A .IE AND ZEE :OTAL NUMBER JF 11-3 i; :HE 7ECTOR.

20 Tr p10

105

1 : V cx1NG FOR TIES BY INCREMEN':NG TBRU THE VECTOR

513, OUNT NUMBER OF TIES. IF NO TIES GO TO L2,11 D.I,/C

1 N SN NEXT T B7 TEE NUMBER OF TIES ENCOUNTERED PLUS I28 2: T-T (D+%)10 T.T LSED~HAN SIZE OF ORIGINAL VECTOR GO TO L3 AND START AGAIN AT NEW T

VARV RM B'CD'E:D1;E1a T5S f*ckC±iN IS A SUBPROGRAM OF THE MANN-WHITNEY TEST (MANy).

if) IT N GEEAESSTER SCHEME USED IN CALCULATING THER DIFFERECS IN SCALE (I ASSIGNED SMALLESTL2 ASSIGNED LARGEST 3

L& Zs:, D3~ f~NEXT RESr 4 SECOND SMZLZZST,-:.J T-vs ::LL ~?sE5 SAmPLE SIZE :S 1EACE.r). SAMPLE S:ZE :3 PASSED :N !BE R:,;T ,.C

D-0 .C. FIND FLOOR OF MIDPOINT OF 7ECTOR AND ASSIGN TO C

I SWS+S GENERATE RANKING VALUES LEFT HALF FIRSTE EL2, Dl

13 (X )=C)IL3

I.1' -( .iZ)<C)I"L2a: NOW CENERATE RIGHT HALF

26 19 .: D 120' rl+E lO "

21: -( (E):C)/L5!22 L6 :DIDl l2; &4 .=' o; i)=C)/L5 ?26 E+E DI "27' + (EI)<C)/L628 1", 4 oFZ) VECTOR IS-ODD VALUE CONCATENATE MIDDLE RANK IN BETWEEN HALFS29' L5:-( B B)O)/L730' a IF SIZE I$ EVEN CONCATENATE LEFT HALF WITs THE REVERSE OF THE RIGHT31' VAR E9132, -0 S

331 L7:VAR*E,B.(OE1)v

V WIL.WILP NN-N;A ;PT;NN;W-PPNMI a THIS FUNCTION IS A SUBPROGRAM OF TEX WILCOXON SIGNED RANK TEST2, a (WISIG). IT GENERATES THE CUMULATIVE DIST. FOR THE TEST STATISTIC3 a TEX QENERATQR USES A RECURSIVE FORMULA. NN=SAMPLE SIZE.5: NM.(L((+/INN)+2 +1*l"5 N-26, a SET P EQUAL TO PEROB. DIST. WHEN N EQUALS 2.7, P.1401I. L3:N N+".9,. A1. a SET T VECTOR TO PROPER SIZE OF ZEROS..- T+(+/(%N))POIF12 a IF AIN USE TRUNCATED RELATION TO COMPUTE OCCURRENCES.13, Lu:.(AIN)/LI.- a IF A>N USE FULL FORMULA TO COMPUTE OCCURRENCES.15. -(A>N)/L216. a WHILE (A-N) IS NEGATIVE TRUNCATE FORMULA TO AVOID A NEGATIVE INDEX.17, L1:T[A2.EAI18, A.A+I19. OL'4'. * IF A IS LARGER THAN THE LENGTH OF P GO TO L6.

"2 a INCE 'A-,V 3ECOMES POSITIVE: THE .ECURSZ7E FORMULA :AN 3E "YSE:.

25 a ONCE A IS LARGER THAN THE LENGTH OF P TRUNCATE FUlZTION AGAIN.26 L6:TEA.PC (A-N)327 L7:A-A+l29, A IF A AS AN INDEX HAS NOT EXCEEDED N(N+1)/2 GO AGAIN.29 i(A<I t)))/LL&'

.30 z CONVERT T INTO P AND CONCATENATE 1 FOR USE IN NEXT ITERATION OR OUTPUT.31 L:P.((+/(%N) pI32 'PNN+P33 WIL- +\PP) .(2*NN)34 a CHECK IF LENGTH OF INPUT VECTOR EXCEEDS NUMBER OF N'S GENERATED.35 - (NN>N) /L3

1.1 0 6 5..5

S d~5*? -~-~* .. '*~~~td*~ .. .. . - -, -

UnUW~lW~i Ryr M. . ~ nn nrzrr .- I ---- ' n

APPENDIX G

LISTINGS OF PROGRAMS USED TO GENERATE C.D.F. COMPARISONT A BLES

V EjNTEST:.N-ILPRA.-B:AjC;TAU;P;NUM;DEN; Z:ZZ ;ERRZZ ;M: ZZC;D;AA :PP;NUM4C;I;H;F

.:: 5T:5 ?RhP.m ETS EE F ... :M Rs.' ?OR .XEN.LS 3a SET SAM-OLE S5Zf AND ALPhA 7VALVES.

6AP_ S.-''

t121 PS-gC

'ii'~ - COMPUT? CUMUL:vr7 Z:ST AND ASSOCIATED STATS.:1'P-KENrALP NC. THIS LOOP :ALCVLATS ALPYHA 7ALUES AND APP:RXZMAIONS.

26 AA 6G ,N.27 PP-PP.P:2;A'28 A CQMPUTE NORMAL APPROXIMATION.29 NUM-3x(TAU~x(UNx N-1)DwC.5)30 DEN. (2K (2xN)+5))*O.531 Z NUM4 DEN32 IZZZZ(I-NORMCDF Z)33 aco*kPUjE NORMAL APPROXIMATION WITH CONTINUITY CORRECTION.

* 34 NUMC.3'CTALJ-AA x((NX (N-1))*0.5)*35 ZC-NUMC*DEN*36 Z ZC.Z (1-NORMCDF ZC)

137 (B<71)/L238 a OMPUTE ERROR DIFFERENCES.39 ErRRZZ*PP-Z Z40 ERRZZC.PP-ZZC41 aPRINT OUT TABLE OF VALUES.42 1-0.14 3 .- 9544 $-'TEST STAT. VALUE453 )#.M.l PROBED Z 13; FOR SAMPLE SIZE EQUAL TO 1,(2 0 NN),'

UG 46 L3[1.7 L5:J-294~48J L3:M-M5513 DAVE (180197) .J,610 (8p197),J3

N~4 -~j*(D:4 43Lp50 H. 1 I8 pS~511 -(Dxz 6*52 F-'E4G I 9.9999 >' OFMT(l 7 oKX)*53 F- 1 62 p?*54 j -H.55) -L7~56 L6:FS-9057j H-17o5

61 ? l Ll(7 5aPP LCI)62 iSF5

64 L7:M-M,EIj FS65 D-D+l66 1-167 -(NI 2N L3D68 N1:s.f!CC.F69 .1-198701 -L371 N2:S.'ERROR; NORMAL72 PP-S-RRZZ73 -L3

1 07

% ~ ~ .~~9*f * ** * * .. ~P.~. P. j-

S74] N3:S.'ERROR: NORM. w/CC75 PP*ZRRZZC76] -L377] L4:M

7 KWTsS m .A* PNL AN;;;RPV UEVA~ '~YE; ERPE Rl1CDP;6C ;kk ;l£; J; F F; S;:

S2 *a THIS PROGRAM GENERATES TABLES OF C.D.F. COMPARISONS FOR THE ZRUSL4L-3 A WALLIS TEST.4.

5 A-4 44 46. CC-27 ALPHA- 0.01 0.02 0.03 0.05 0.08 0.13 0.18

.8: -L2

10ALL UWAL TO GENERATE EXACT DISTRIBUTION FOR SELECTED SAMPLE s::Z.f 2 8-0L13 N+/A2 4 CC*CC+1t16 M-~001171 PP00

.20~ ?F-oC21 PFI-oO

1 22

C F P ;

24 LOWCATE POSITION OF EXACT CDP VALUE f. ALPHA.25 D-(+/ (CDFf.ALPRAEBl))26 -(Dzo)1L327 D-128 DETERMINE CORRESPONDING TEST STATISTIC 7ALUEf.29 z3:s-p::;D]30 KK*.KK 331 A RECOAD EXACT VALUE OF CDP.32 PVALUE-P[2-D]33 PP+.PP PVALIJE34 A COMP.TEk CORRSPONDING P-VALUE USING CHISQ APPROX.35 PVAL.(X I1 CHZSQ B36 PC-PC (N-VAL)37 A COPUtzE CORRESPONDING P-VAL51 USING F APPROXIMATION

39 P filp (pN-K)) FOISTY P40 p K F41 A COMPtITE COR.ESPONDINO P-VALUE USING F APPROX W/I1 LESS D.F. IN DENON42 PVF1. (K-1 3(X1-)-I)) FDISTN F43 PPlL(1-'V144 *(BC7)/L745 D-046 ERRS-.RP-PC47 ERRF._PP P48 ERRF1.PP-PFl49 a PRINT OUT TABLE OF VALUES.50 1-..151 '-52 S.'TEST STAT. VALUE53 M+M, P ROB[H Z G AVC463,:]; FOR A GROUP OF 3 SAMPLES CONSISTTNC OF ~

4, AND '(*cc)., obs.54 -,Ls:55: L5:9--*56 LS:M-M EI(17p'-' ),J,61p(Sp'-' ),J.57: .(D=55 1L9*58* F-pQ

,60~ P-.61 C-70

66 * c7)/L1167 *L1268 L10:.C-+169 P~e [(7 5 *KKEC])70 F.f'71 ill~ 3 Z1072 L12:NMM i F73 D-D+174 1-175 -(NI N2 N3,N4 L5)[D]76 Nl:S-rALPEA VALUE77 KK-PP781 .'+'

108

*o 1.Pia; CTISQOARE'ill;i 3iON; F C~sf

N' -UpjJoit P N/ DF'

V pq~~fjM~A~~j~;P_'!N -PW ;;;;,7NV~.F.E.....RR.RF

A-! P~?M ~ ~ :C U.: ~ i

1-1 074C : 1R M C C . ? : R L N S h:E s M L~

' SUN sMlEs RAK 4FZC SX-

.- M .-. 23i27 it181.1: : i~L:s

* SM EE APS FO LACEf OAMPLE.S

26 E*a':~

360 C C C1C'1 81 L..ff~a. VAsFRM LE TORSO~lv SMALLESTALVE

PIa LJMINEf CM;PNI VALUE CESPOCIN C APEAVALUE

a COI' OR*ESPONC:NC P-VALUE VSS.NG F APPROX W LESS r.F. :N ZENON

'51~ ? iPF: I FIU47T4W~52p @R-LPEA-P

a PR:N. OUT TABLE CF VALUES.

~2,-

p~ st, . -

.6 C:6:,, Li '(

IC' iS .F.FPJ

1 09

L I FNij.d RR /

2~ Mi:S-A PE

-8j j C

[so .f:jVOh1 V/IDl

:13 U ?0 00

'NUM :NNo16TTT 7:NOTLPA r

231 ME5.0 TST22 S PTE UMLATSIE ADIS ANDH 7ALSITDSAS

4 a THIS LOOP CCULATS APA VLE S ADA5O MTZ.S

2 -013 2 :6 U' 029 PA k

15 C-0. a2*0.

137 NUCa(..)( oM)2

24 a COTE LOODPN CACLTE APPIATIONSAD PORIATO

.3S0 a CMPTEERORDIFEENES31 ERRZZNOMA.APRXIATON32 gRR~.P- M +2)C

R3CZC.PP-OR ZZ. 4

37 PJmNTO ARE OF_ VALUES.#

110 ]cUC+% 3'% %fW CNOR~f iSUDIN5) APPROXIMATIONJ.'~ .?'~*.f~ ~ ~ ~ **~ p%~j~p~**

60S-$.TST STAT.' VALUE61I NM'I PROStU sq3: FOR SAMPLE SIZES N EQUAL TO ',(2 0 ONN),' AND

62 k*QUAL TO '(2 0 *MM)*

63 fS:J-19464 L3 ~M55)QV(817.,1(p9)66 8-1 Is OS67 -(DUO) 1668 P. 1 30 ZZ9 O FMT(1 7 OEK)169 F. 1 6 OF70 Ff..71 !f78'73

7" ~S.FS.Hf75 C.076 Lg-c'.+,77 F;.1 ( lPC379 , sPPiC)79 r..(C<7 I' L9Fso L7:M.M Li P81. D.D+1.83 -N V2 .N3 ,N4 N5 N6 .N7 LUS)914 N1:S.- 1EXACT .hp85 J-19886 -L387 N2:S'ErRR~oR; NvORMAL

89 -L390 N3:S.&ERROR; NORM. VICC,

g 2 -L3S93 NW4:S.'ERROR; T DIST914 PP.ERRTT95 .L396 NS:S.'ERROR; T V/CC97 P RTa8 -L3

991 N6:S-'ERROR; AVE T/Z10i6 PP.ERRAVEL101) -L3C102~ N7:S.'ErRROR; AVE TC/ZCC103 .iPP.EAATCZCF1014] -L3105~ L4:M

10 1. (NN 9 )/IL

v sIGNTEsT;N;ALPHA;3:A;C:E;P;Z;ZZ:ERHZZ:N:ZZC;D:PP;I:;F:?FS;S;J;ERRZZC;ZC

2A THIS PROGRAM GENERATES TABLES OF C.*D.P . COMPARISONS FOR THE SIGN TEST.3.14 a SET SAMPLE SIZE AND ALPHA VALUES.5. N.-236 ALPHA. 0.01 0.03 0.06 0.12 0.22 0.35 0.57 THIS LOOP INCREMENTS SAMPLE SIZE.8. L1:PP*PO

11 ZZ~pQ012: ZZC-pOO*13. N-N1+114 D,.015. B.016. A COMPUTE CUMULATIVE DIST AND ASSOCIATED STATS.

17 P-BINON Nisa THIS LOOP CALCULATES ALPHA VALUES AND APPORXIXATIONS.19L2:2*8.1

0 -LPHAERI

23 KK-KK,.T214 PPePP,PCA325 a COMPUTE NORMAL.AP3ROXIMATION.26 Z.(-( Q.N)05x (N*0 527 ZZ.ZZ, (NORMCF Z)28 A C3MPUTE N3N L APPROXIMAT~ION WITH CONTINUITI CORRECTION.30 Z14- C NORMCDFZ531 *<'71/t233 aRZ.PZ SMPUTE ERROR DIFFERENCES..34 EPRRZZC*PP-ZZC35 a PRINT OUT TABLE OF VALUES.36 X.0.1

Ile

371 Jr-% 95S39 )J.'TfSZ STAT. VALUE393 ., PRODC EL g J; FOR SAME SIZE EQUAL TO ',(2 0 *I),''40) ;L3413 S:J.194842] L:M.m 0l AVE(l8p197).J.61p(8o187),J3843 .D=845/L4I44 s I IS8LRS456 845 .(D: zz P DMT(1 7 pICK)U.7 P 1 62 pP48 FS.H.F'4 9 -0L750 L6: FS-spO51 8*17PS52 FS-FS.H53 C-054 L9:C.C.551 F' +1 (7 5 OPPEC3)56 ps.FS p'57 -(C'c73/L958 L7:M.M,E1J FS59 ! .D+lE60 1-1E61 -(NI N2 N3 L5) ED)~62 N 'SEXACT'C.D F.

65 N2ii4.'RROR; NORMAL 166 RP-'RZZ

L67 -3- 'ERROR. NORM. 'C'69 PP+ERRZZCS70 -L371L4:M)

7 SPMT'ST-N-ALPVA;B'A-C;AA-P Z-ZZ:fRRZZ:MZZC-D-PP;I;H;;S;S;;?;;T~C;

2: A THIS PROGRAM GENERATES TABLES OF C.D.F. COMPARISONS FOR SPEARMAN'S R.3.84 A SET SAMPLE SIZE AND ALPEA VALUES.5, N-e6, ALPHA* 0.01 0.02 0.03 0.05 0.08 0.13 0.187. a THIS LOOP INCREMENTS SAMPLE SIZE.8, Ll:PPv.pO

10 X0011 FS-po12 ZZ-pO13 ZZC4p 014 72-

16 N1.1+117 D-0

19 AA-i6+NX1I+N*220 a ACOMPUTE CUMULATIVE DIST AND ASSOCIATED STATS.21 P:P.PMATSP N-5-3)22 PC.(P E2:J0/ ;123 a T HIS ploc1 CALCULATES ALPHA VALUES AND APPORXIMATIONS.24 L2:8.8+125 C.APAB

27 RH P1A]28 KK*.r]C.RIO29 PP-PP, F 2e)30 a OMPUTE NORMAL APPROXIMATION.31 Z*RHOX C(N-i )*o.S)32 ZZ.ZZ, (1-NORMCDF Z)33 q :70MPUJT NORMAL APPROX:MATZrON WIZE CONTIJUIT7 :ORRECTION.~3L4 ZC-(RHO-AA4)x ((N-t )*0.352g Zzc-zZC,(-NO'RMCDF zc)36 aCOMPUTE STUDENVT 2' APPROXIMATION37 TT.TT N SPAPROXI RHO38 a CAMPUTE STUDENT T APPROXIMATION mITE CONTIuIT! CORRECTION.39 T'C.T2.C N SPAPROX RHO40 ILB7 /241 aCOMPUTE ERROR DIFFERENCES.42 ERRZ Z.PP-ZZ843 ERRZ ZC.PP-ZZCL4 iRRTC:-P-:Tk

46 a-. PRINT OUT TABLE OF VALUES.848 .7+195849 5.'TB5T STAT. VALUE I,Sol m.M. ,PROBER k fi): FOR SAMPLE SIZE EQUAL TO '.(2 0 *N),,'

53 13:N5M ,CIAVC (ISPI97 ),JSIp (0p197).4)54 .*D 6 RL

59 FS.H.7

61. Li:PS-oQ62 H17.S63 PSFR64 C-065 LO:.e66 F.' L' (7 5 vPPCC3)67 FS-FS )

69 L7:.L1 FS701 D-D+171* 1-!72 -(N N2N3 Nit N5,LS)ED4

75: -L376 N:S.' RROR; NORAL I/C77 PNaEPJZZC78 -L382 Nl&:S' ERROR: OM I ,903 PP.ERRzTC

2 5 N5:S-'ERROR; : D/CT863 PP-9RRTC

87) *L398 L4:Mi89 -ZN410)/Ll

VLiEST NiALPHAjB-A C T-P-N WDEPNZJZiR Z DP;?NrH 1 F Smtirc;Tc;N; icNtil; bEmok; DIJONkC:ERRA VEg; EARTT:; EATC E ER±Z;lC.!ER

mTHIS PROGRAM GENERATES TABLES OF C.D.F. COMPARISONS FOR THE WILCOXONmSZGNED-RANE TEST.

a SET' SAMPLE SIZE AND ALPHA VALUES.

7) ALPHA- 0.0180.02 0.03 0.05 0.08 0.1 10jill:PP 0 IS LOOP INCREMENTS SAMLESE

11 X..p12 F4$.po13 70 0~p14 1ZCso15 7.16 TTC-1017 NN+121 D.019 B*020 a COMPUTE CUMULATIVE DIST AND ASSOCIATED STATS.21 P.VILP N22 a THIS LOOP CALCULATES ALPHA VALUES AND APPORximATIONs. p23 L2:B.Iei24 C.AjLPjA EN)

e25 A. +/ PsC)26 A27 WixEKT28 PP.PP.PCA)29 A OMPU5E N3RMAL APPROXIMATION. 4

*0 NUM.T-((Nx(N.1 )e&u

:.-Num+DrN

1Z zZ,:PoRMCmP HAL WITH CONTINUITY CORRECTION.34COMPUTE NOR APRXMIN35 NUMC.(T+0.5)-((Nx(. *436 ZC+-NUMC.DEN v37 ZZC.ZZC,NORM5CDF Z39 OF COPTE SCDTNT T APPROXIMATION39 DErNOM.(t(N (DEN*2 +) (N-1 )- ((NUH*2).(N-1)))*.40 T-NUM4 DENOM41 T.TTz jN-) vIsTNTr42 A 6PUE STUDENT T APPROXIMATION WI1TH CONTINUITY CORRECTION. 4

43 NUMT.(NUJ044 DENQ.HC.(I (NX(DEN*2)).(N-1))-((((NUN)+0.5)*2).(N-1)))*0.545 NC.UM; *D Nom DSC C46 TTC4#-TTc( N-1)TSTTC

113

0-7 (Ic7)/L2 COMPUTE ERROR DIFFERENCES.

AR14 mI'Ci.PP ZZC C42

55a PRINT OUT TABLE OF VALUES.

57 J-19558 S-'TEST STLT. VALUE I

521 )F~jlPROBCVF 1 1(3; FOR SAMPLE SIZE EQUAL TO ',(2 0NW.,

2 3 (:855i GAVU1S8P197),J,61p(8p197),J364 a- 118 0s6~5 -(DbO)/L666 F*'B0<1 ZZ9 >1 GFMTC1 7 oKK)

~67 F 16 pP

F69 -L770LI:FS4.SO

[731 Cq-0

0754 1 q C. (7l aP 9C)

(C ?'''7

078 L7:M-M. P: s

80I1J *1 N2 N3,N4 NS.,6N7 L5)[,D]92~ N1i:NS' I EMAT ,C. b.L83 J-19844 *L385 N72S.''ROR NORMAL I

87 -L388 NjSIRO;NORM. V/CC'99 P-5RZ90 -L391 114:S-'ERROR; r DIST92 PP.-iERRTT

93 *L394 #5 :S.'ERROR; 7 Jv/CC I95 P-ERRTC96 -Z397 N9 :S.'EPRRQR; AVE T/Z Igo PP.ERRA V999 -L3

10 !N7S.'RROR; AVE TC/ZCI101 PPRRTCZC1102 -L3103 JL4:N

114~

INITIAL DISTRIBUTION LIST

No. Copies

1. Defense Technical Information Center 2Cameron StationAlexandria, Virginia 22304-6145

2. Library, Code 0142 2Naval Postgraduate SchoolMonterey, California 93943-5002

3. Superintendent, Code 53Jy 1Attn: Prof. T. JayachandranNaval Postgraduate SchoolMonterey, California 93943

4. Superintendent, Code 55Bm 1Attn: Prof. D. BarrNaval Postgraduate SchoolMonterey, California 93943

5. Superintendent, Code 55LaAttn: Prof. H. LarsonNaval Postgraduate SchoolMonterey, California 93943

6. Superintendent, Code 55Re 1Attn: Prof. R. ReadNaval Postgraduate SchoolMonterey, California 93943

6. Superintendent, Code 55Rh 1Attn: Prof. R. RichardsNaval Postgraduate SchoolMonterey, California 93943

7. Superintendent, Code 55Wd 1Attn: Prof. K. WoodNaval Postgraduate SchooLMonterey, California 93943

7. Commanding Officer 2Naval Special Warfare Unit - FourBox 3400FPO Miami, Florida 34051

115 'J

~ III

(PDF) O-A194 RN AkP WdORKSPACE FOR CONDUCTING … · considered: Sign Test, Wilcoxon Signed-rank Test, Mann-Whitney Test, Kruskal-Wallis Test, Kendall's B, Spearman's R, and Nonparametric - DOKUMEN.TIPS (2024)

FAQs

What is the difference between Kruskal-Wallis test and Wilcoxon signed-rank test? ›

The Wilcoxon test is intended to compared two groups. The Kruskal-Wallis test is intended compares multiple groups after the omnibus test rejects the null hypothesis that all groups have similar averages. In your setup, you are comparing exactly two dev. stage s within each genus.

Show Me More ›

What is the Mann Whitney U and Wilcoxon signed-rank test? ›

The Mann–Whitney U test / Wilcoxon rank-sum test is not the same as the Wilcoxon signed-rank test, although both are nonparametric and involve summation of ranks. The Mann–Whitney U test is applied to independent samples. The Wilcoxon signed-rank test is applied to matched or dependent samples.

When should you do a Mann-Whitney U test? ›

The Mann-Whitney U test is used to compare differences between two independent groups when the dependent variable is either ordinal or continuous, but not normally distributed.

Discover More ›

What are the disadvantages of Wilcoxon signed rank test? ›

Disadvantage: It is less powerful than the paired t-test - While the Wilcoxon test is useful when our data is not normally distributed, it is less powerful than the paired t-test when the data is normally distributed.

Tell Me More ›

What is the best graph for the Wilcoxon signed rank test? ›

You should graph the data for a Wilcoxon signed rank test the same way you would graph the data for a paired t–test, a bar graph with either the values side-by-side for each pair, or the differences at each pair.

Get More Info Here ›

How do you know when to use Wilcoxon signed rank test? ›

As the Wilcoxon signed-rank test does not assume normality in the data, it can be used when this assumption has been violated and the use of the dependent t-test is inappropriate. It is used to compare two sets of scores that come from the same participants.

Discover More Details ›

When should the Wilcoxon signed rank test be used? ›

Wilcoxon rank-sum test is used to compare two independent samples, while Wilcoxon signed-rank test is used to compare two related samples, matched samples, or to conduct a paired difference test of repeated measurements on a single sample to assess whether their population mean ranks differ.

Get More Info ›

Why use Wilcoxon signed rank test instead of t-test? ›

Flexibility of the Wilcoxon Signed Rank Test

This non-parametric nature renders it especially suitable for analyzing ordinal data or data that do not adhere to the assumptions of multivariate normality, often required by parametric tests like the t-test and F-test.

(PDF) O-A194 RN AkP WdORKSPACE FOR CONDUCTING … · considered: Sign Test, Wilcoxon Signed-rank Test, Mann-Whitney Test, Kruskal-Wallis Test, Kendall&#039;s B, Spearman&#039;s R, and Nonparametric - DOKUMEN.TIPS (2024)

FAQs

What is the difference between Kruskal-Wallis test and Wilcoxon signed-rank test? ›

When should you do a Mann-Whitney U test? ›

(PDF) O-A194 RN AkP WdORKSPACE FOR CONDUCTING … · considered: Sign Test, Wilcoxon Signed-rank Test, Mann-Whitney Test, Kruskal-Wallis Test, Kendall's B, Spearman's R, and Nonparametric - DOKUMEN.TIPS (2024)