GuideSelectingStatisticalTechniques OCR PDF

=
A Guide for
Selecting Statistical
Techniques for Analyzing
Social Science Data
Second Edition
Frank M. Andrews
LauraKlem
Terrence N. Davidson
Patrick M. O'Malley
Willard L. Rodgers
Survey Research Center
Institute for Social Research
The University of Michigan
1981
2
uewery 01 ContNH c...... In PuWlcatlon Dati

Mlln entry under till.:
A Ouldl for "'.cIlng e'ltlsUcllleehnlqUlI for

analyzing IOCII' science dill.
"l·He!. [)oM "058, ISR code no. 358r - T.p. Yel'SO.

Blbtlogrephy: p.
1. $oclallCteoou-StaUaUcal method.. I. Andrews,
Fr.'*- M. II. Unlv.aity 0' Mk:tttgM Survey Reeelrch
Cent....
HAZI.G8I7 1981 300'.28 81·7082
ISBN ().I7IM4-274-3 (pi*.) MCR2
Pubilihed by the Institute 'or Social Research,

The Unlverally of Michigan, Ann Arbor, Michigan 018108
Second Edlt5o" 1.1
5 7 8
Copyrto/ll 18111 ~ T1Io UnI_yof IoIk:1!toon. All RIghi. _
Mlnuflltltured In the United Stl'H of America
Cover Design by Carol lawrence

- z '.00 ,t • , G.'iJ¥ IJ!1I1lRl!" " ' 5 . C£ ••
- -
4 • ••
CONTENTS
Preface to the Second Edition vi
Instructions and Comments on the Use 1

ofthls Guide
The Decision Tree: Questions and Answers 3

Leading to Appropriate Statistics or
Statistical Techniques
Appendix A: Sources of Further Information 31

about Statistics Appearing In this Guide
Appendix B: Programs that Compute Statistics 43

Listed In the Guide
Appendix C: Some New or Rarely Used 59

Statistical Techniques
Glossary 63
References 67
PREFACE TO THE SECOND EDITION
This Guide Is Intended to help social sclenllsts select and functionally distinguish - those statistics and statls·
from the vast array at statistical techniques a particular tical techniques that are In common use In the social
stallsllc or technique that can be appropriately applied In a SCiences, that receive significant allentlon In social science
given analysis. The Guide Is addressed to pracllclng social statistics texts, or that seem to have high potential useful·
sclenllsts, data analysts, and graduate students who ness. About 150 statistics or stallstlcal techniques are In·
already have some knowledge of social science stallsllcs cluded In this Guide.
and who want a systematic but highly condensed overview The core of the Guide Is the 28 pages of sequential
of many of the statlsllcal techniques In current use and of questlons-and·answers that lead the user to an appropriate
the purposes for which each Is Intended. technique. This Is the "deCision tree." Preceding the "tree"
The popularity of the first edition of the Guide leads us to section Is a short set of Instructions about how to use the
hope that this substantially expanded and updated second tree and some comments suggesting alternative strategies
edition will also prove useful. The original version of this and certain cautions that should be kepi In mind. Three ap·
Guide became available In 1971, was revised and formally pendlces and a glossary follow the Iree. Appendix A cites
published by the Inslllute for Social Research In 1974, and specific pages In a major reference where each slatlstlc
has subsequently been through four Engllsh·language print· presented In the Guide Is discussed and Its means of com-
Ings. In addition, ISR has granted permission for editions In putation Is given. Appendix B Identifies the programs In five
French (Laval UniverSity, Quebec) and Hebrew (University of major software syslems and several special-purpose pro·
Haifa). This second edition contains nearly all of lhe material grams that compute given slatlstlcs. Appendix C covers
thaI appeared In the first edilion plus significant some additional statistical techniques Ihat were Judged to
expansions: the number of statistical techniques Included be too new or too rarely used to merit inclUSion In the
In the decision tree has been Increased by almost 50 decision· tree portion of the Guide but that seemed poten·
percent, with major additions being made to the coverage of lIally useful for social science data analysis. The Guide con·
muilivariate analysis; a glossary that defines technical cludes with a bibliography presenting Ihe full reference for
terms has been added; and Appendix B, which Indicates each cited book and arllcle.
where each statistic can be found In the output from com- For assistance In the preparation of Ihls Guide we are
puter software, now Includes detailed Information on grateful to Chrlsllne Zupanovlch and her colleagues In the
sources In the OSIRIS, MIDAS, SPSS, SAS, and BMOP soft· ISR Word ProceSSing Group, to Linda Stafford and her col·
ware systems. There has been a general updating through· leagues In the ISR Publishing Division, and to Eugene Lep·
out the Guide to Incorporate many of the stallsllcal and an- panen and his colleagues In the University of Michigan
alytical developments of the past decade. Technical lIIuslration Unit. Preparation of Ihe Guide has
No guide could Include all the stallstlcs ever proposed as been partially supported by the Computer Support Group of
useful for social science data analysis and this Guide ISR's Survey Research Cenler.
makes no claim to do so. Rather, It attempts to Include-
vi
INSTRUCTIONS AND COMMENTS ON THE USE OF THIS GUIDE
This Guide Is Intended to help a data an.lyst .elect sequence through the decision tree or to consult another
statistics or statl.tlcal techniques appropriate for the pur- source of Information.
poses and conditions of a particular analysis. In many analysis situations It Is possible to make a11erna·
To use this Guide, .t.rt with the question on page 3, tlve decisions about the nature of the variables, relation-
choose one of the answers pres.nted ther., and then con- ships, andlor goals, and these may result In the selection of
tinue along the "branches" of the decision tree as In- alternative final boxes. It Is always possible to use tech-
structed. Eventually you will arrive at a box that names a sta- niques that require less stringent assumptions than the
tistical technique andlor a statistical measure and/or a ones originally considered. For example. measures or tests
statistical test appropriate to your situation -If one was may be used that are appropriate for a weak.r scale of
known to the authors. Many of the technical terms used In measurement, or techniques appropriate for non-additive
the Guide are defined In the Glossary that begins on page situations may be used even though the variables actually
63. form an additive system. Note also that non-additive
The typical box contains one statistical measure (In the systems can sometimes be handled using an additive t.ch-
portion outlined by solid line.) and on. statl.tlcal te.t (In the nlque If an appropriate combination of variable. (e.g.,
dotted portion). In a few cases, several dlffer.nt m.asures, pattern variable, product variable) has been formed. Recall
or s.veral different tests, .re presanted In the same box. also thaI two-polnl nominal variables and ranks meel the
Thes•• re •••• ntl.lly .qulvalent from a functional point of deflnilion of Inlervally scaled variable •.
view, and comments to help you choose among them may
appear In an accompanying footnote. Sometime. a mea.ure
appear. without an accompanying test If none seemed par- Cautionary Comment.
ticularly appropriate, and sometimes a test Is listed without
any measure. 1. Welghled dala, missing data, small sample sizes, com-
Some branches of the tree terminate In boxes that are plex sample deSigns, and capitalization on ch.nce In fitting
empty. These Indicate situations for which the authors knew a slatistlcal model are sources of potential problems In data
of no approprlat. technique-Indeed, further statistical de- analYSis. The Guide does not deal with these complications.
velopment may be needed. If an analysis Is to be performed If one of these slluallons exists, Ihe Guide should be used
In such a case, It will be necessary to find an alternative with caution . (See note 91n Appendix C for a brief discussion
of sampling errors from complex samples.) example, It Is often possible to transform scores so that the
2. The statistical measures In the terminal boxes are de· transformed scores correspond to a normal distribution,
scrlptlve of the particular sample being examined. For some constitute an Interval scale. or relate linearly to another
statistical measures, the value obtained will also be a good variable.) Occasionally, It may be wise to eliminate cases
estimate of the value In the population as a whole, whereas with extreme values. For guidance on selecting appropriate
other statistics may underestimate (or overestimate) the transformations, see Kruskal (1978).
population value. In general, the amount of biBS Is relatively 5. Common assumptions for Inferences based on tech-
small and sometimes adjustments can be made for It. These niques using one or more Intervalty scaled variables (par-
adjustments are discussed In some statistics texts (but not ticularly when the Intervally scaled variable Is a dependent
In this Guide). If a statistic Is a biased estimator of the popu· variable) Include the following: first, that the observations
latlon value, It Is marked In this Guide with an asterisk. are Independent, I.e., the selection of one case for Inclusion
3. In principle, a confidence Interval may be placed In the sample does not affect the chances of any other case
around any statistic. It Is also possible to test the slgnlfl· being Included, and the value of a variable for one case In no
cance of the difference between values of a statistic cal· way affects the value of the variable for any other case;
culated for two non-overlapplng groups. These procedures second, that the observations are drawn from a population
are not Indicated In the Guide but are discussed In standard normally distributed on the Intervally scaled varlable(s); and
textbooks. third, If more than one variable Is Involved, that the Intervally
4. The Guide does not explicitly consider possible trans· scaled varlable(s) have equal variances within categories
formations of the data such as bracketing, using logarithms, of the other varlable(s), I.e., there Is homogeneity of
ranking, etc. Transformations may be used to simplify variance. Bivariate or multivariate normality Is also some·
analysis or to bring data Into line wllh assumptions. (For times assumed.
2
I
•
THE DECISION TREE:

QUESTIONS AND ANSWERS LEADING TO APPROPRIATE STATISTICS OR STATISTICAL TECHNIQUES
STARTING POINT
How many variables does the problem Involve?

r -__________________________ ~A~ ____________________________ ~
( 'I
On. Variable Two Vartabl.. More Than Two
Vlrllbll.
How do you want to tre,t the "arJabl.s with respect to sca/. of measurement?
,
_____________________ __________________
Bolh Both
-
Bolh
~
A~
One Intlrvl', One Intena•• One Onlnll,

Inllnll OnIlnel Nomlnol 0 .. Ordlnll OnoNomlnl1 0 .. Nomlnll
=========-=============-------=====~==========:·I
;::r
ONE VARIABLE
How do you want to 'r6al the variable with respect to scale of measurement?
r~----------------------~A~-------------------------
Nominal Ordinal Interval
\
I
, What do you want to know about the distri-
bution of the v8r/able? butlon of the '1Brlable?
I
What do you want to know about the dlstrl·
__------~A~________~ ~ ______~A~______~
( Conlral Dllparalon Frequencle. \ ( Cenlrol DI.panlon Froquonclo. \
Tendency Tendency
Inler-quartlle Relative
Relative Relative frequencies, e.g.,
frequency of frequencies, e.g., deviation
percentages
modal value percentages
or class Absolute
Absolute frequencies
frequencle.
N-Ules
4
•
(continued from page 4)
• One Interval variable
What do you want to know about the distribution of the variable?

r---------------------------__ ~A ____________________________~~
'I
I C.ntr.1
Tendency I
DI lpe,.. Ion
I
Symm.try Pe.ked nell F
requenc I
.1 Norm. lit y
Standard de. Skewness - Kurtosis·

vlatlon- To test departure. Kolmogoroy-
Ralatlve
To test departures from normality: for frequencies, e.g., Smlrnov one
Coefficient of N or••ter than 1000, sample test
from normality: for percentages
variation· refer the crltlcsl
N greater than 150,
rafer the critical ratio of the kUrtosis Absolute Lllilafors
Range- measure to a tabte extension of
Do you want to ratio of the skew- frequencies
neaa measure to a of the unit normal the KolmogorOy-
treat outlying curv.; for N between Smlrnov test
table of the unit N-Illes
cases differently normal curve; for 200 and 1000, refer
from others? N between 25 and the kurtosis measure Chi-square
.---~A~ _____________.. 150, refar the to a table for teating goodness-of-
kurtosis; lor N lesa fit test (xl)
( ~ skewness measure
than 200, use Geary's
Ve. No to a table for
testlno skewness. criterion. See also specltlc
I
Whet Is the form
testa fOf skewness
and kurtosis
or 'he
Wlnsorlzed mean
dlstrlbullon?
r -_ _ _ _.....
A ___--,
Sk._ ~
Trimmed mean
( Symm.trlc
Hampel esUmate
of location
SIwe10ht maan
-±-
~ M~lan
Mean
- Slased astlmator
TWO INTERVAL VARIABLES
Is 8 distinction made between 8 dependent and an Independent variable?
v..
rr--------~--------------~A------------~~----------~
No '\ -
I
Do you want to 'reat the relationsh ip 8$ linear?
I
Do you want 1o test whether the mBans on
the two vsr/ables are equal?
r_ _-::_ _ _ _-'A'-_ _ _ _ _ _ , ~
V.. No (r---~-----------------~~----~--------------
V.. No \
I I
Do you want to trBat the relationship as IInBar?
r -___________ A____________
0
Regression
coelflclent
Coefficients from
curvilinear l paired
t test for r VH ~ ,
(b or beta. tI)' regression I
I
l (F_
F testt
equals til_ ..JI
(b or betl, /'f) I.'
I F teat I
IL observatlons'·-·
______ .JI
What do you w.L to m.a.u,.?
r---~~~~_A~--~~-
'- _ _ ____
: (F equals I' tor
I'- ________
each coefticlent) J I ( Agreement COYl,lollon"\
I
Should there b. a panalty II th.
go 10
variables do not have the same
pagll 7
distributions?
• Biased astlmator.
t The assumptions In nola 5 on page 2 may apply. (r-------~~------~

YH ~ ,
I Beta Is a standardized version of b. See "standardized coefficient"
In Glossary.
Robinson's A Krlppendorff's
I The type of curvilinear regression relerred to here Is also known coefficient of
as polynomial regression. See note ~ In Appendix C fOf' further Intraclaas cor· agreement (T)
discussion. relation coeffl·
clent (r.)"
--The t test for paired observations Is appropriate for parallel meas-
ureS from matched calis as well as for repeated measures on a
single set of cases. See "matched samples" In Glossary.
6
7
(conflnued from page 6)
4 Two Interval variables • No dlstlnctlon Is made between a
dependent and an independent variable. The relationship is to

be trealed 8S linear. Covarlation Is to be measured
How many of the varIables are dIchotomous?
r~--~~----------------~~----~P--------------------~~--------~
Non. Ono Both "\
Is the dichotomous var/able a collapsing of a con-

tinuous vsr/able and do you went to estimate what
I
Are the varIables collapsfngs of continuous vari-
ables and do you want to est/male whst the cor-
the correlation would bell It were continuous? re/slJon would be If they were continuous?
~~ ____~A____~__~ ~~ ____~A_____~__~
( Yo. No \ ( Yo. No \
Pearson's product BIS8flal r' Pearson'. product Tetrachorlc rt Pearson's product

moment r' moment r (equals moment r (equals
I point biserial r) · .t phi) ··'
Refer critical I Refe, critical I
Do Fisher's r to Z I
trsnsformatlon and I ratio fOI biserial I Refer critical I ,atlo for tetra· : I
r to a table of the I I chortc I to • I Refer crltlca' I
refer critical r.tlo 1 ratfo for point
table 01 theI
: unit normal I blaerl.1 r to a I ,aUo for phi to I
10fZloatableo' I I unit normal I a table 01 the I
I curve. I I table of the I
I the unit norm.1 I L ________ J I unit normal : : curve. : I unit normal I
LC~~' _____ J L ________ ...l I curve.
'-- ________ J
I CUrvtl. L _______ ..JI
Biased "tlmator.
Both the talracharlc r and the biserial r depend on • atrlcl .lIump- • Pearson', r In thl, case I. mathematically equivalent to a point
tlon of the normality of the contlnuoua variable, that have been biserial r; the teata are almoat equIvalent.
dichotomized. Furthermore, the sampling error for both coeffiCient,
Is large when dichotomies are extreme. Nunnally (1978, pages I Pearson's r In this case Is mathematically equivalent to phi (see
135-137) advises against the use of Iheae coefficients. page 9); the tests are almost equivalent.
: . . £12
TWO ORDINAL VARIABLES
Is B distinction made between a dependent and 8n Independent var/able?
r~-------.~------------------~A~--~------------------------~
V.. No '\
I ~====
What do you
_______
w~nt 10 measure?
A~~~ _______
Somera' d
( Agreement Court.llon '\
dJ
I For N greater than la, refer the erillee' ratio I
: of S to 8 table of the unit normal curve; for I
1 N less than or equal to 10. refer d to 8 table I
L of____________________
critical values 01 S. JI Do you want to treat the fanks 01
the ordered c8legor/es 88 Interval scales?
rr---~----------~A~----------~--~
V.. No 'I
I I
Spearman's rho (rJ"
Kendall', lau ., tau b, or tau c
(1'., feu "el '
: When N Is 10 or larger, refer the crillca' valUe of
Goodman and Kruskal's gamma
I r. to a table of the t distribution; for N les8 than
I 10,_
.... _ reler
_ _r._to
_alable
_ _ _ of
__ critical
___ values
___ 01_r l ·_ __ JI hl*
Kim's d l
For N graatar than 10 reler the critical ratio :
of S to a table 01 the unit normal curve; lor
I N 'aas than or equal to 10, refer these statistics :
IL to a table 01 critical values of S.
___________________ ~ I
Biased estimator. statistics, tau a will be the smallest, and tau b, tau c, and Kim's d
will be Intermediate. This ordering Is because gamma Ignores all
t The data may be transformed to ranks and '. or Krlppendor"'s f ties (when present In the data-as Is usually the case), whereas the
used. See page e. other four statistics penalize for Ues In the sensa of reducing the
absolute value 01 tha statistic obtalnad. Unlike tau b and Kim's d,
I These stallstics differ with respect to how they treat pairs 0' cases tau c can attain ::k 1 even If the two variables do not have the same
that fall In the same category on one Of both of the varlabla • . number of categories. If there are no Ues on either variable the five
Excapt In axtrama casas (I.a., whara any 01 tha statl.tlcs equals 0 maasuras are Identical. Sea Goodman and Kruskal (1954), Kendall
or 1) tha absolute value 01 gamma will be the highest of the five (1970), Kendall and Stuart (1981), Stuart (1953), and Kim (1971).
8
9
TWO NOMINAL VARIABLES
Ars the variables both two-point scales?
r~--------v=----------------~A_----------~~------------~
V.. No \
What do you want to measure? Is a distinction made between a dependent and an Independent var/able?
r~----------~A~--~--~---
Symmetry Court.tlon " r~------------~--~A~----------------~
V.. No ....,
Yule's ot Do you want a statistic based on the number of cases In

Phi l.)' each category or on the number of cases In the mod.1 cat6-
: McNemar's tesl gar/as? A
I of symmetry"··
L ________ ...JI
Fisher's exact test I
I r~~=-------~------------~
StaUatlc BIHd on Number St.tlltlc a••ed on Number
\
I
Ref.r critical ratio ) 01 C.... In E.ch e.'IIIOfY 01 C.... In Mod.1 e••IIIO'Io.
of phi to a table I
at the unit normal
curve.
I
I
I
I
Goodman and Kruskal's
"symmetric lambda (>'", .>'.)
I Peareon 1 tau b Cl'.J
~~~u~(x~_J 1 Refer crilical raUo of tau b I

I
I Ret.r critical raUo of lambda
I to. table of the unit normal I I to. table of the unit normal I
L
I _____________
curve. .lI I'- curve
__ :... _________ ...1I
In this case, McNemar's test of symmetry I, equivalent to

Cochran's O.
I Pearaon chl·squares can be corrected for continuity (Yale's
• In this case, Yul,'s Q Is equivalent to Goodman and Kruskal's correction) but this Is controversial. See camilli and Hopkins (1978).
gamma and phi Is equivalent to Pearson's product moment r. tn
general, Q will be higher In absolute value than phi because 0 •• McNemar's lest ot symmetry la appropriate for parallel measures
IgnorlS pairs ot cases which talt In ,he same category on one Of from matched cases as well as fOf repeated meaaures on a single
both of the variables. set of cases. See "matched samples" In Glossary.
(contlnufHI from peg. 9)
• Two nominal variables. At least one of the variables Is not a
two-point scale. No distinction Is made between a dependent
and an Independent variable
What do you want to measure?
rr----~~==~------~~~---A~----------~~~--------~
Agreement Symmetry Cowerletlon "\
I
Do you want 8 statistic based on
the number of cases /n each cale-
gory or on the number of cases In
: McNemar'S lest I the modal categories?
Should there be B penalty If the symmetry·· ....I
L of______
variables do not have the S8me dis- ~ ________ ~A~~~~~~~
tributions? r
St.tI.tlc B••~ on Number St.tI.tlc B.ud on Number \
r-~----_A~------~ of C •••• In e.ch e.tegory at C.... In Mod.' e.tegorl••
r V.. No '\
I
Do you want a statlsllc whose up-
per limit varies with the number of
Symmetric lambda
().,u)
Scott's coeffi- Cohen's agree- categories and whose upper limit
cient of agree· ment coellicients, I
ment, pl(...) kappas IJr'S) may be less than one? I Refer critical ratio
I of symmetric lambda I
Refar critical ratios : rr-~~--A~--~~_

V.. No '\
I to a table of the
IL unit
I
normal curve. ...JI
_________
for COhen's Jr's to a
table of the unit nor- I
L mal curve.
I _________ JI
COntingency
coefficient (C)
Cramllr's V
I I Pearson I
I Pearaon I
L~~sqU~~j(t)' J I ohl·square (x1)1
L.. _ _ _ _ _ _ ..J
~
I Pearaon chl·squares can be corrected for continuity but this Is .. McNemar's tast of symmetry Is appropriate for parallel measures
conlroyer.lal. See Bradley at al. (19~. from matched case. as well as for tepealed maasures on a sIngle
set of cases. See "matched samples" In Glossary.
10
"
l ~,
• .ZS
11
TWO VARIABLES: ONE INTERVAL, ONE ORDINAL
Is the ordinal variable a two·polnt variable?

r~--YY=··~------------------~~--------------~N~O~--~~
Do you want to treat the ordlna' var'able as If It were b.stld on an

underlyIng normally distributed InteIVal variable?
r~------~Y·=·~---------------~~--------------~N~O------'
Jaapen's coefficient of
I
Do you want to treat the ordinal vaflable as If It wefe II monotonic
multlse,lal correlaUon · ,f transformation of 8n underlying Interval var/able?
Do Flaher'a r to Z trana. I
I 'ormation and ,ef.r critical I r~--~~------A~----~~
Y.. No
__~~
I raUo of Z to a table of I
L t~~nlt ~m!I.:~_ J
Mayer and
Roblnaon's M...
r 00 Flaher's r to Z trana. I
: fOrmation and reflN' critical ~
I ,atlo 01 Z to a tabfe of I
C~ ~~~~~curYe, __ J
Biased altlmator,
f Jaapen'a coefficient la the product moment correlation between the

Interva' variable and a transformation of the ordlna' variable, The
magnitude of this Ilaliltie Is sensitive to the asaumptlon or
normality.
• Any twc>polnt variable meets the criteria for an Intlrvally scaled

variable.
TWO VARIABLES: ONE INTERVAL, ONE NOMINAL
Is the Interval vBrlable dependent?
r~------------------~~~----------------------~
fu ~ ,
I
Do you want 8 measure of the strength of relationship between
the vllrlables O( 8 test of the slal/stlcal significance of d/flerences
between groups?
~~~~~~_ _~A~_ _ _ _ _ _ _ _~~~
r Me.lure of T•• , of "'\

Strength Significance
I
Do you want to descrlbB the relB' Is the nomInal vsrlable a two-poInt varIable?
tlonshlp In your da,a O( 10 estimate
It In the population which you have r---~------~A~----------_
V.. No '\
.ampled?
~~__~A~______~
( Dalcrlbl Estimate ,
~
I I
Omega' (.!)'"
I F test M I Intraclass correlallor}
L ____ _ J coefficient (tl)- ,I
Kelley'. epsllon l (fl)' ,I

I I
IL. _ _ _
F_test U
_ _ _ _ _ JI
• Biased estimator.
want to make Inlerencea about the total set 01 potential categories.
I The assumptions In nota 5 on page 2 may apply. (See Hays, 1973, page 525; Hays denotes the totraclass correlation
88 p. rather than ' 1') In most slluatlons II Is more appropriate to use
t If the nominal variable Is a two-point scale, the I teat Is an alter· the fixed effects model, I.• .• omega'. K.II.y's epsilon' Is used lor
nallve (because In such case F equals It). . xactly the same purpose as Hays' omega' but differs very slightly
In computation. Hays' omega' wa. apparently developed
IOmega' applies to the fi xed effects model, and the tnUaelass Independently 01 Kelley's earlier statistic. Kellay's epsilon' Is
correlation coefficient applies to the random effects model. Thus precisely equivalent to ela', after eta' Is adjusted for degrees of
omega' should be used If you want to make In'arence. only about fr..dom. Se. QI ••• and Haketlan (1989), Kelley (1935), and Hays
the specific calaga,'.' of the nominal variable which appear In the (1973; page 485).
data, whereas the !nlraclass earretal lon eoetflelentshould be used
If you view the particular categories that appear In the data as a • Any two-point variable meets the criteria for an IntetvaUy scaled
random sample from a larger set of potenllal categories and you variable.
12
- - -,--'-
13
(con tin"; from page 12) • There are two variables, one Interval and the other
nominal • The Interval variable Is dependent • Stallstlcal
stgnlflcance of dllferences between groups Is to be tasted
Are you willing to assume that the Intervally scaled verlable Is normally dis-
tributed In the population?
~ ____ ~__________ ________________
~A ~~
r Vo. No \
Do you went to test the equellty

of means or of variances of the Do you want to test the equality of
dependent vari.ble for different var'ances of the dependent v.r/-
categorIes 01 fhe/ndependent vari- abl. for different c81egorles of the
able? Independent vallable?
(__--------------~A~------~----~
__ ~ __________ ~A~ __ ~ ________ ~
M••ns Vlnlnc.. "\ ( ~ ~"\
I
Do you want to assume
homo,cedastlclty aCI088 Analysis 01
Is the nominal var/eble a two-point scale?
Analys's 01
levels 01 the Independent variance variance __-------JA~----~--~
variable? ( V.. No "\
I I
r V..
_ _- - - - - -_ _A~_ _~~--~
No"\
I
Bartlatt', 'Ulf ....JI
IL _______ I levene's
L ______ W t JI
Analys's of Analysis of
variance variance
I tt i I Welch statlstlet I
IL.. _F_test
_ _ _ ..JI I I
I I
I Brown-Forsythe I
I statlsllet :
I I
L lestt..
I t ______ .JI
Y..
/ No
/
I
Are Ihe cases (e.g., people) In one
I
Are the CBses (e.g., people) In one
category of the nominal variable
category of the nominal ver/able matched to the caSBS In each of
matched 10 the CBBes In the olher Ihe olher categories of that VBr/·
category of that var/able?·· able? ••
~ __~~____~A________~__~
r YII No , r.------:-:-----A'----::--~,
v.. No'
Within eech c.tegory olth. noml·

na/ variable, Is the distribution of
the Interva' vIr/able symmetrIc?
IIRandomization tesl I
I I
I Randomization
I I I
I Randomlzallon I
r~~~--~A----~--~
Y.. No ,
I for two Independent 1
IL aampl""
_________ .JI
1 teat lor matched :
I•_________
samplea tt J
II teat for Independent 1
L samples
________ tt
J
I I
I
I
I Walsh test I : Randomization test tor:
•l- _ _____ .JI Imatched palr. tt I
'-----------'
t The asaumptlons In not. 5 on page 2 may apply.
I If the nominal varla~e Is a tw~nl acale, the t le,t Is an ., See "matched sample," In Gloseary.
alternative (becaua. In such ca .. F equall t').
If In practice, randomization t •• t. ar. usually only applied when the
• It the nominal varlabl. la • two-point sca'., • special form of the number of ca.es Is very amall. With larger N's the Inlerval variable
t test may be u ••d. (See Hays, 1973, IIp. 4()ot and 410.) Is generally tr••ted .s an ordinal variable.
14
15
TWO VARIABLES: ONE ORDINAL, ONE NOMINAL
Is 8 distinction made between a dependent and an Independent var/able?
r~--~--------~A~--------~--~,
Y.. No '
I I
Is the nominal variable two-pofnt?
Is the ordlnalv8r'able dependent?
___,'
r VII
Is the nominal
r v..
.~/ab/a two-point?
..
No 'I
0 No' r~--~~--~A~----~
V.. No
I
A,e the cases (e.g., people) In one category
o( the nomina' variable matched to the cases I
In tha other category of that .arlab/e?""
r '"
V••
*
I
II
No
I
Somer,' d
" ArB the casas (e.g., people) In one category 01 the
nominal variable matched to the cases In each of
I For .Ionlllc.nce of Somers' d with N o,••ter I the other calegorles o( tha, var/able?""
: Sign test :
I then 10. r.ler the critical ratio of S to • table I __________ __
I Wilcoxon r
~~A~ ~~ ~
.Igned- : I 01 the unit normal curve; tor N less than or No 'I

L ------ .J
rank t ••
t I
:
equal to 10, refer d to a table of critlca'
valuea of S.
Yr' I
Freeman's coefficient
I Median teat *- of differentiation (Il" ·'
I
,I Mann·Whltney Utesl
, Friedman test I t(ruska'·Wallls test
I Kolmogorov-Smlrnov two sample test
I
ILRuns
_______________
test ~
I I
:'-- _____ JI
: Mod'.n , ..t (I..
I_________
mora than 2 groups) J
i
.. BI ••ed .allmetor.
I Meaaures 01 strength of relationship that .r. appropriate lor
unmatched data can alao be used descriptively here.
I The nominal variable may be treated as ordinal (In which case go to
, Thla coefficient Implicitly orders the nominal categories. Given n page 8) or a, Interval (In which case go to page 11).
nominal categories, thera are nl values lor Somera' d. Freeman's
theta I, equal to the highest 01 these d's. .. See " matched samples" In Glossary.
. 1,;; 'f"
~ .. . ,'
MORE THAN TWO VARIABLES
( V•• No
I
Do you want to trea, the relationships in-
I
Do you want to treat the relationships among
volving the covar'ale(s) as additive?' the var/ables 8S additive?'
r~~--------~A~----~~~
V.. No ~ r~--v~.~.--------~A~---------'N~O--~
I
Do you want to 'reat the dependent
I
variable and the covaria/e(s) as
Interval and the Independent var/·
able(s) 8S nominal?
( v..
~ __ ~ ______ ____
~A ____
~
No '\
I
I _ _ _
L Fiest'
__ __ _ _ JI
1 The assumptions In nota 5 on page 2 may apply.
, Nonadditivity can be represented within addlllv8 technique. by

using. pauern variable or a product variable. Another pOl8lblllly
Is to analyze subgroups alparately. See Glo... ry.
• Some analysis 01 covariance technique. assume statlstlc a,

Independence between all pairs of Independent variables.
16
17
(continued trom ".gff 18J
• More than two variables • No distinction Is made between

dependent and Independent variables
Do you want to measure a9rffement?
r~~~-----------------------JA~----------~~
Y.. No ,
I
How do you want to treat the varl-
I
Do you want to test whether the
abies with respect 10 scale of means (or proportions) on all var/·
measurement? abIes are equal?
rr~~~--A---~----~I~'
All Nomln.1 Alllnto,...1
(~~V~O~.-------------A---------~N-o----~'
All Ordln.1 Oth... I

Are a/l the variables
I
Do you want to lreal the relation·
two-poInt? ships among the variables 8S addl·
Light's agreement IntraclaSS correia· oU~ve~?~'______-'A_____~~~
coefficient (km) tlon coefficient r~--v~o~.---JA~'N"-O~--~\ r Y.. No "\
--~ (r,) •
: Reter critical raUo :
I at km to a table
I of the unit normal I
I Robinson's A I
Do you want 10 trsat a/l of the vBr/-
__________ J I
I curve.
L I
_ _ _t_ _ _ . J
:I.. _ _ _ _Ftest abies as
nominal?'
rr--~~~A~------~--~
Yo. No "\
Kendall's coetrl·
clent 01 concord·
,
IL Cochran's
________ au
-'I
I
I 1
ance (W)
For N greater Ihan , Multidimensional
contingency
7, uae x! teel for I Analysllof
W; lor N less than
or equal 10 7, re· I
I
, variance with
repeated measures
table analysis
fer. to • lable of : I Chl.aquare

I
L. critical
____ valu
__ ••
_01
__•. JI I
L. _ _ _ _teslt
_____F .JI IL. _ _tesl.'
_ _ _ _ _ _ ...JI
I SM note 3 In Appendix C.
• Biased estimator.
, There are varloua chl·square teat statistics Including Pearson,
t The assumptions In note 5 on page 2 may apply. maximum IIkeUhood, and Neyman.
parallel me.sures from matched
_'JN~o~n~'~d~d~,,~,y~,,~y~c~'ini~ij~i~wi"hln
_ _ _... using a pattern variable. Another
additive possibility
techniques by Is measures on a single set of cases.
• to I!.II!II!!III
- . - • .. _._ ............... - ~.IOIlIly.e
. .. . - ... .. _.- .... _ _ ,,_ ",.•• ur•• from matched
10 _n.ry:u eubgroup8 lopara.ely. See Oloe.lry. c.. . . . . well •• for repeated mea.ures on • single sel of cases
a 3
Se. "matched samples" In Glossary. .
,
(cOII,'nutd from page 17J • More than two variables. No distinction Is made between de-
pendent and Independent variables. Relationships are to be
treated as additive
Do you want to analyze patterns existing among "arlab/es or among Individual cases
(e.g., persons/?
cr"
~ ______________________ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\
A~
r v·~rl.'
Do you have two or more sets of Do you want to treat the Vlrlables
var/ables and do you want to as measured on Interval scales
measure the strength of the and relationships among them 8S
___ ~~
8ssoc/atlon between
____________ those sets?
A ______________ ~ ____ linear?
(- Y.. No ' r,,"-~y~
••~---A'--~N'-o----\
I
Do you want to tr8at ,he var/abla,
Does the analysis Involve (a) one Clustering techniques
as measured on Interval scales auch .s alngte linkage,
and relationships among them 8S group o/Indlvldual case. or (b/lwo
complete linkage, eyeraoe
linear? or more groups?' linkage, K·msans
r-----____ ____ __
(""--o::-no--=O-rou-p-~:wo'--or-M-ore--:O:-rou-P-'-\
A~ ~
r J~
: Wilks' lambda t
6 Q.type faclor
analyals
I Roy'S greatesl
I rool crnerlon' I
I ,
Plllal·Bartiett vt_J
IL ________
t The assumptions In note 5 on page 2 may apply,
t "Two or more groups" may mean dlstinci aeta 01 Indlvlduala, a sel

of Individuals observed on two or more occasions, etc.
18
19
(continued trom page f8)

dependent and Independent variables • Relationships are to be
trealed as additive • Patterns among variables are to be ana-
lyzed • One group of Indlvlduale
Do you want to explore covarlat/on among the variables (e.g., to examine their
relationships to underlying dimensions) or do you Want to find clusters of
IIBrlables tha, are more strongly related fa one another than 10 the fsmslnlng
variables?
~------~________~A~__________-=~~~
( Explon Co•• rlillon Find Cluo',,, ,
I
Do you want to treat the va,/abltls
as measured on Interval scales Clustering techniques
and the relationships among them such I' slng'e IInkege,
as IInear?1 complete linkage, lver.
age linkage, K·means
r~--------~Y~o.-----------A~--------------=NO~--'
I
Do you wIn' to explore the rel.-
I
Do you want 10 locate each of the
tlonshlps among the set of varl· variables In multidimensional
abIes or do you want to compate space?
the pattern of the relationships ~~ ____ ~A~ ______ ~ __
with 8 prespecilled pattern?
I Yn No "
"" V.I
"'" No
Non·metrlc multi·
dimensional
acallng techniques
rr-----------------A~--------------~
Explore Compo.. " Do you want to 'reat
R.IIUonlhlpo PIU.ml all 'IIr/ables as nominal?
I
Do you want to presBrve the metric
I
Do you want to preserve the metric
r__-----;:;:-:---~A~---::--,
V.I No '
units In which the variables were units In which the variables were
measured or to standardize them measured or to standardize them
by the observed var'ance o( BICh? by the observed variance of each?
~ ________ ~A __ ~~ _____ ~ ________ ~A~ ________ ~
( StljardlZ. OrlUljl Metrlc 'I ( Stondardlze Orlulnol Motrlc 'I Multldlmen·

slon.1
I
Confirmatory
I
Confirmatory
contingency
lable
analysiS
Faclor analysis
ot corralaUon
Factor analysis
ot varlance- tactor analysis factor analysis
ot varlanee-
,
matrix covariance of a standardized I Chl·square I
matrix
variance-covariance
matrix'
covariance
matrix
,I'- test.'
_____ J
I
: Maximum likelihood 1 I Maximum likelihood :

I chl·square (x.)t
1L chl·square
_________ (xl)t JI L _________ J1
, The assumpllons in note 5 on page 2 may apply.
t The variables should be standardized using the combined groups

(I.e., the observed group and the pre specified pattern) •• a I See nole 3 In Appendix C.
ret.rence. (Depending on the problem, this may or may not be
equivalent to using the correlaUon matrix tor the observed group.) • Ther. are various chl·.quare lest statistics Including Pearson,
See "standardized variable" In Glossary. maximum likelihood, and Neyman.
20
21
(contlnu.r1 from page f8)

dependent and Independent variables • Relationships are to be
treated as additive • Patterns among variables are to be ana·
Iyzed • Two or more groups of Indlvlduals l
Do you want to 8Xp/Ortllhe relationships among 8 set of variables In two or more

groups simultaneously or do you want to compare the similarity 01 the patterns
of the relationships among 8 set of var/ables either (a) across two or more
groups orlb) with a presp.clll.d pattern?
r~--~-----------------A~-----------------------
Explore Comp .... '\
Relltlon.hlpl Plttem.
Do you want to treal the var'ables

I
Do you want to preserve the metric
8S measured on IntelYa' scales units In which the variables wer.
and the relationships among them measured or 10 standardize them
as IInsar? by the observed variance of each?
rr---------~A~--------
Vo. No
__' rr----------~A~-----------
Stendonll.. O~glnol Mot~c '\
Three-way non·metrlc
Three-mode multidimensional scaling
factor analysis
techniques ConfirmatOf)' Confirmatory factor
factor analya'a analysis of varlanee-
of standardized covariance matrices
variance-covariance
t The auumpllons In note 5 on page 2 may apply. matrices' I
Maximum IIkallhood 1
t "Two or more groups" may mean distinct aeta of Indlvlduala, a aet I I chl·square (xl)t 1
Maximum likelihOod I L _______ .-l
of Indlvlduale obaerved on two or more occasions, etc. I chl·square b:1 )t I
L _______ ...l
I The variables should be standardized using the combined groups as
a ref.renee group. (This Is not the same as using the correlation
matrlce. for the separate groups.) See "standardized variable" In
Glossary.
• More than two variables • A distinction Is made between de·
pendent and Independent variables • There Is more than one de·
pendent variable
Is there more than one Independent variable?

r~------~=---------------_A~------------------------~--~
Y.. No "
I
Do you want to 'r8at the relation-
~ __ J~I~_
pendent var/able as nominally
ships among the variables as addl· scaled and a/l 01 the dependent
tlve?' variables 8S intervally scaled?
(--------------~~~------------~
v.. No '\ (
~~
V..
_ _ _ _ _ _ _ _A~_ _ _ _ _ _~~
No '\
I
Do you went to 'reat al/ the Do you want to test only whether
dependent varIables as Interval? the vectors of meens 8re equal for
( V.._______
~ A~~~~
No '\
aI/ categories of the Independent
variables?
I r~--:v:-.-. ___--A~_---:-:N-O-~\
Multivariate Profile
analysis of analysis I
t The assumptions In note 5 on page 2 may apply. varlance l
I Nonadditivity can be represented within additive techniques by I Wilks' lambda',f I Wilks' lambda',f
using a pattern variable or a product variable, Another possibility Is I I
10 analyze subgroups soparately, See Glossary, I Roy'S greatest I Roy's greatest
I root criterion I I I root crlterlon f
I Soma multivariate analyals of variance technlquee assume
stalistlcal Independence between al\ paint ot Independent variables. LP'~a~B~rtle~~ J I
IL Plllal·earUett
_______ Vt .JI
, If the Independent variable Is a two·polnt acale, Hotelllng's T' Is an

aitematlve (because In such cases the T' test Is equivalent to the
A·teSI), Mahalanobl,' 0' la another aUernatlve " n such a Clle,
22
• ~ .- .----
23
(contlnu&d from plge 22)
• A dlatlnctlon Is made between dependent and Independent
variables • There Is more than one dependent variable and more
than one Independent variable • Relationships among the varl·
abies are to be treated as additive
all the dependent and Independent vsrlsbles as Intsrval?

Do you went to tr.st
r~~v.=.~-----------J-'"~----------~N~O~\
~~
Do_you
__ want _trea'
_ _to _ _ all the_
relationships
~AL _ _ _8S _IInsar?
~~
r V.. - No \
DOBS the an"ys/s Include at 1.8st on. Intervening variable?'

~~ _ _ _ _--JA~________~
( V.. No '\
I
Does your analysis Include.' least
one 'a tent (I.e.• unmeasured) vari-
Canonical correlation
I Wilks' lambda t
abl.? I
I
I Roy's greatesl
I rool criterion' I
I I
L_':I11.~~~I~t~:' J
Structural Path
models with analysl.
'atent variable.
f The assumptions In note 5 on page 2 may apply.

, See Glossary.
.;
- .. -- .. -
(continued from pilgil 221
• A distinction Is made between dependent and Independent

variables • There Is more than one dependent variable and more
than one Independent variable • Relationships among the va,l·
able. are nolle be lrealed as addilive • Allihe dependenl varl·
abies are Interval
Do you want to treat 81/ the Independent var/abl., a' nomInal and t•• t a .8' 01 prespeclfled relationships?
rr--~~--------------A~
V.. __
------~~, No 1
Do you want to trea, a/l th.'nd.pendent Vlrl·

MulUvarl.'e ablas IS nominal or ordlna' and do you want
analyals 01
variance' to do an emplrlc.' search for 'trong ",/atlon·
ships?
Wilks' lambda t I
I rr--~-----A~----~--~,
V.I No
Roy', greatest :
root criterion'
I I
L!.1~.:B!~~t~~ j Multivariate
binary
segmentation
techntque.
I
The ...umpllon, In nole 5 on page 2 may apply.
Some multivariate analyall of variance technlqun asume

statlsllcal Independence between all pairs of Independent variables,
I
I
24 I
25
(continUed from plge 16,
• More than two variables • A distinction Is made between

dependent and Independent variables • There Is one dependent
variable • No coyarlate Is used 10 remove linear effects • Rela-
tionships among the variables are nol to be trealed as addilive
Do you want to do an amplrlcal search for strong relationships or to test 8 se' of prespecllled relationships?
rr--------~~~--------
S.arch
__ ~A~__________________~~--
r... "\
How do you want to treat the lIa,.
lables with respect to sClle of
measurement? Do you wlnt 10 'r881 the dependent
~---- __________ ~A~ ___________ vsrlabl. IS ordinal?
~~ ______A~____~~~
( Dopenclenl: Nomln.1 or In...".1 01110, '\
( Y.. No 'I
Inclepond.nt Nomln.1 0' Ordl..1
Do you want to 'rB8t 811 the Inde-

Binary segmentation pendent variables IS nominal?
technique. rr~y.-.------J~'----~N~O--'I
Multidimensional
contingency table
analysis baaed on
the cumulative
I This technique depends on a strict assumption 0' the normality 01 logistic dis-
the continuous varlab'e which " represented by the ordinal tribution'
dependent variable.
Chi-square :
IL _ _
tests'
_ _ _ _ _ -.JI
I There are various chl·square test statistics Including Pearson,
maximum likelihood, and Neyman.
Ii g
(conf/nued from peQ8 25) • More than two variables • A distinction Is made between de·
pendent and Independant variable. • There Is one dependent
variable • No covariate Is used to remove linear effects • Rela·
tlonshlps among the variables are not to be treated as additive
• A set of prespeclfled relationships Is to be tested • The de·
pendent variable Is not to be treated as ordinal
00 you
~
want to treat any ollheAIndependent
______________ verlables as ordinal?
....____________ ~~
r~ - ~,
I
Do you want to treat the dependent variable 8S Interval and all the Independent
vallables 8S nominal and do you want to assume homoscedasllclty?
••~~~~~~~~--A---~~~~~~~~N~O~'
r~~v~
I
Do you want to tr8ala1/ of the variables as nom/naI?l
~~~~~~~~~A....~~~~~~~
Analyals 0' ( V.. No '\
:L
variance'
F test' .J:
_______
I
Do you want to t,8a' the dependent
Do you want to do a hierarchical variable as Interval and all of Ihe
analysis? Independent variables 8S nominal?
r-~~~~__~A_____~~__~
rr~~---A....----~~
r V..
I
No
1
"I
Vi' c±JNO'\
Multidimensional Multidimensional Analystt of
contingency table contingency table variance using tt
analysis technique weighted least
analysis squares
allowing an unconstrained
I Chl.squa,e I design matrix
IL.. _ _______
testa' JI
I Chi· square I
I.... _ _ _tests'
________ J
t
t The assumptions In note 5 on page 2 may apply.
• There are various chl·square test statistics Including Pearson,
I Many analysis of variance techniques assume statistical maximum likelihood, and Neyman.
Independence between all pairs of Independent variables.
n Multidimensional contlngenc:y table analysis using weighted I.ast
t See note 3 In Appendix C. aquares may be appropriate.
?R
,
27
(continued ',om page 16)
• More than two variables • A distinction Is made between de-
pendent and Independent variables • There Is one dependent
variable • No covariate Is used to remove linear effects • Rela.
tlonshlps among the variables Bre to be treated 8S additive
r~--~~~-------------A~~~~-- ________
How do you want to treat the dependent vs".ble with respect to scale of measurement?
~~~~
Nomlnel
I
Ordlnel
Do you want to treat all the Inde·

pendent vaflables as Interval?
,.-_ _ _ _A..._ _ _ _ _ _~-
ctJ* Do you want to
Int.,..1
t~e8t
,
aI/ the Inde·

pendent variables 8S Interval?
,.-_ _ _ _ _A~_ _ _ _ __
( V.. No ,
r V.. No , I I
I
Do you want to trea' the relation-
Do you wsnt to tre., all the re18'
t/onshlps as IIns8r?
Dummy variable
regrellion or mul-
ships among ,halndopond.n, .arl· r'-V-.-.A ...._ _~_ _, tiple classification
abies .8 IIn.ar? Is the dependent ________
_____
vsrlable two-poInt? No 1 analysis
( V..
~
A
~
~
_
No I
Muiliple clIrvlllne.r
regresslon l
I
Wilks' lambda' I
I
Roy'. orealesl I
I Do you wlnt to treat aI/ 01 the Ind.
root criterion' I pendent variables as nominal?'
IL _ _ vt
_ _ _ _ _ _ _ .JI
Plllal-8arliett
I
__V..
I
---A~------__'"_~
No "\
Multidimensional contingency
table analysis
IL _ _ _Chi-square
_ _ _ _ _tests'
_____ J I
\v••
. \
Is there a very high proporllon In
one category of the dependent var-
Iable (e.g., 9O%)?
~~ _____________ ____________
A~ ~ __
( V.. No '\
Dummy variable regression

uSing weighted least squares Do you want to Bssume homoscedsstlclty?
or maximum likelihood, usually
on a transformed dependent r~~v~.-.----------A'---------N-O--~'
variable (e.g., on loglts)
Dummy variable Dummy variable regress/on

regression or mul· using weighted least squares
liple classification Of maximum likelihood, usually
analysis on a transformed dependent
variable (e.g., on loglts)
I The assumptions in nole 5 on page 2 may apply.
t See note 1 In Appendix C. See nota 3 In Appendix C.

~ The type 01 curvilinear regression referred to here is also known as
polynomial regression. See nole 4 in Appendix C for further • There are various chi·square lest statistics including Pearson,
discussion. maximum likelihood, and Neyman.
28
29
• More than two varlablss • A distinction Is made between

dependenl and Independent variables • There Is one dependent
variable • No covartate Is used to remove linear effects • Rela-
tionships among the variables are to be treated as addlUv8 and
linear • All the variables are Interval
Does the analysis (nclude B118.St one Intervening var/able?

~------------ ________A_________~~
rV.. - No '\
Does Ihe analysis Include at least Do you wanl 8 single measure of

one la'ent (I.e. , unmelJsured) varl- the relationship between the de-
BbIB? pendent variable and all the inde-
pendent var/ables taken together?
r-~~
V..
__~A'-__--.;=--
- No '\ rr.v~.~.--------_A~--------~N~O-'\
Structural
models
with
Path
analysis
Multiple correlation
(multiple regression)
I
Do you want a alst/Blle which as-
sIgns to .ach Independent variable
(A 1. , ••.• ,JI) some 01 the explainable variance
'atant
variable, In the dependent variable which
L-
I _ _Ft.stt
_____ ...JI that Independent varlabl. shares
with other Independent var/ables?
rr7.-------A~------~N~O~'
V.I 1
7"' _. - - - - - -- - --
v•• No
Regression Do you want a statistic that measures the

coeUlclent additional proportion of the total varIance In
(b or beta, ~)I the dependent varIable explaInable by each
I F lestt I Independent variable, over snd above what
IL.:(F______
equals 12) JI the other Independent variables can
expl.ln?'
~ __________ ~A~ ____________ ~
( Vo. No '\
I
Part correlaUon2
I
Do you wants stallsllc that meas-
Irflt4 ....•.,1* ures Ihe addlllons' ProportIon of
the lolal variance In the dependent
IL F_teet (F_ 2 varIable explainable by each Inde-
__ __ 1
equals _ )'..J:
pendent variable, over and above
what the other independent V8r'.
abies clln explain, expressed re/a·
tlve to the proportion of variance In
the dependent varIable unexplaln·
able by Ihe other Independent varl·
r-_~.=b/~e~S?~___A~__________~__~
( Vo. No '\
* Biased e,Umelor.
, The assumptions In nole 5 on page 2 may apply.
I Bela Is a standardized veralon 0' b. S .. ..standardized coefficient" Partial corralallon:

In Oloa .. ry.
(r'2.3, ... ,.)*
The additional proportion of the lotal variance explainable by a aet
0' Independent variables, over and above what the other
Independent varlablea can explain, can be measured by the
Do Fisher's r to Z trana.
formation and rafer crill· I
I
difference between the R2'S reaultlng 'rom two separate multiple cal ratio of Z to a tabla
correlation analysea. I of the unit normal curva,' :
I See Glossary. IL ________
F teat (F aquals 12)t __ ...lI
30
APPENDIX A
SOURCES OF FURTHER INFORMATION ABOUT STATISTICS
APPEARINO IN THIS OUIDE
A brief citation Is given below for each statistic and sta.

tlstlcal technique that appears In the Guide. A futi entry for
each cited work appears In the tist of references.
M... McNemar, 1969, p. 14

Distribution of relaUve frequencies Blalock, 1979, p. 31
Distribution of absolute frequencies McNemar, 1989, p. 5
Median McNemar, 1969, p. 14
Inter-quarUle deviation McNemar, 1969, p. 19
N·tlles McNemar, 1969, p. 19
Wlnsorlzed mean Dixon and Massey, 1989, p. 330

Trimmed me.n Andrews el al., 1972, p. 2B1
Hampel estimate of location Andrews at al., 1972, p. 2C3
Blwelght mean Mosteller and Tukey, 1917, p. 205
Mean McNemar, 1989, p. 18
Median McNemar, 1989, p. 14
Standard devlaUon Hays, 1973, p. 238
eoefflclent of vartatlon Blalock, 1979, p. 84
Range McNemar, 1989, p. 19
Skewness McNemar, 1969, p. 25
Critical ratio of skewness meaaura Snedecor and Cochran, 1967, p. 88
Table for testing skewness Snedecor and Cochran. 1967, p. 552
Kurtosis McNemar, 1989, p. 25
erUlcal ratio of kurtosis meaaure Snedecor and Cochran, 1967, p. 88
Table for telUng kurtosis Snedecor and Cochran, 1967, p. 552
Geary's criterion 'or kurlosls O'AgolUno, 1970
Distribution of relative frequencies Blalock, 1979, p. 31
Distribution of absolute frequencies McNemar, 1989, p. 5
N·llles McNamar, 1989, p. 19
Kolmogorov-Smlrnov one aample test Siegel. 1958. p. 07
Ulliefors teat Conover, 1971, p. 302
Chi-square goodneas-of-flt leal HaYtI, 1973, p. 725
Pagll •
Regrelslon coefficient Hays, 1973, pp. 623, 630

F test for regression coeUlcktnt Hays, 1973, p. 647
Coefficient from curvilinear regression Draper and Smith, 1986, p. 129; Heys, 1973, p. 875
F leal for coelilcleni from curvilinear regression Hays, 1973, p. 880
I lelt for paired observaUons Hays, 1973, p . • 2.
Robinson's A Robinson, 1957
Inlraclass correlation coltlclanl McNamar, 1989, p. 322
F test for Robinson's A (translate to Intraclass McNemar, 1989. p. 322
conelatlon coefflclant and lell as below)
F ta.t for Intraclasa correlation McNemar, 1989, p. 322
Krlppandorlf'. i Krlppendorff. 1970, p. 1<13
Pearson's product momant r Hays, 1973, p. 623
32
33
Flshe,'s , to z transformation and the critical ratio of Z Hay., 1973, p. 882

81serlal r McNemar, 1989, p. 215; Nunnally, 1978, p. 135
Critical ratio 'or blaerlal r McNemar, 1969, p. 217
Critical raUo 'or point biserial r McNemar, 1969, p. 219
Tetrachorlc r McNema" 1989, p. 221; Nunnally, 1978, p. 136
Critical ralio for tet,achoflc r McNemar, 1969, p. 223
Critical raUo for phi McNemar, 1969, p. 227
Pogo a
Somer,' d Somere, 1962

Critical ratio of S Kendall,1970,p.52
Standa,d error of 5, aSlumlng Ilea Kendall, 1910, p. 55
Table of critical values of S, asaumlng tie. Harshberger, 1971, p. 535
Spearman', rho Siegel, 1956, p. 202
Critical r.tlo for Spearman's rho Siege', 11~58, p. 212
Table of crlllc.1 values of rho Slogel, 1956, p. 284
Kendall'a tau a Kendall, 1970, p. 5
Standard error of S, assuming 00 ties Kendall, 1970, p. 51
Table of critical values of S, assuming no Ue. Kendall, 1970, p. 173
Kendall's tau b Kendall, 1970. p. 35
Kendall's tau c Kendall, 1970, p. 47
Goodman and Kruakat', gamma Hays, 1973, p. 800
Klm'ad Kim, 1871, p. 898
Pogo •
McNemar'. te.t of symmetry Siegel, 1956. p. 63 (when both variable. are two·polnt scales,
McNemar's test of symmetry and McNemar's tesl for the significance
of changes ate equivalent); Bowker, 1848
Yule'. Q Yule and Kendall, 1957, p. 30
Phi McNemar, 1989, p. 225
Critical raUo 01 phi McNemar, 1989, p. 227
Fisher's 1)llct tHt SIegel, 1856, p. 96
PearlOn chl·aquar. Hay., 1973, p. 735
Goodman and Kruskal'. tau b Blalock, 1979, p. 307
Crltlca' ratio 01 Goodman and Kruska"s tau b Goodman and Kruskal, 1972, p. 417
Aaymmetrlc lambda Hays, 1913, p. 747
CrtUel' ratio of lambda Goodman and Kruakat, 1983, p. 318
P.l0
ScoU', coefficient of agreement Krlppendorfl, 1970, p. 1~2

Cohen's agreement coefficients (kappas) Cohen, 1980; Cohen, 1988
Crltlca' ratio for Cohen', kappas Flels., Cohen. and Everitt. 1989
McNemar's t ••t of symmetry Bowker, 1SU8
Contingency coefficient Hays, 1973, p. 7~5
P••rlOn chl-aquI,e Hays, 1973,p. 730

C,am6r'a V Hays, 1973, p. 7~5 (Hsys call. It Cram.r'a statistic); Srlkant.n, 1970
Symmetric lambda Hayt, 1973, p, 7~9
Critical ratio of symmetric lambda Goodman and Kruakal, 1983, p. 321
'1841 11
J •• pen', coeltlclsnt of multls.rlal correlaUon Fr"mlln, 1985, p. 131

Flshe,'. , to Z tranaformatlon and the crltlca' ratio of Z Hays, 1973, p. 682; Harshbarger, 1971. p. 395
Mayer and Robinson'. Myu May... and Robinson, 19n
Fisher's r to Z tran,formaUon and the critical raUo of Z Mayer and Robinson, 1977; Hays, 1973, p, e&2
Page 12
Eta1 Have, 1973, p. e83

Omega· Hays. 1973, p. ~84
InUlell" corre'atlon coefficient Hays, 1973, p. 535
34
r ,
35
Kelley's apaUon' Kelley, 1935; Glass and HakaUan, 1969

F test for eta' , omega', Kelley's ep.llon', Hay., 1973, p. 471
and Intraclaas correlation coeffiCient
' .... 13-1 ..
Analysis of variance Hays, 1973. p. 457

F teat for analysl. of variance Hays, 1973, p. 471
Welch staUatic Brown and Foray the, 1974.
Brown·Forsythe ataU_tlc Brown and Forsythe, 1974.
t t.'1 Hays, 1973, pp. 404, 410
Bartl.tt', t8al Kirk, 1969, p. 61
Levene's W Brown and Forsythe, 1974b
Walsh t ••t Siegel, 1956, p. 83
Randomization leat for matched pelr.
Bradley, 1988, p. 78; Siegel, 1958, p. 88
Randomization test tor two Independent .ampl••
Bradley, 1968, p. 78; Siegel, 1956, p. 152
Rlndomlzallon test for matched aamplea
Bradley, 1968, p.80
Randomization test for Independent ,ampl ••
Btadley, 1988, p. eo
Sign t••t Siegel. 1956, p. 68

Wilcoxon signed-rank t.at SIOGol, 1956, p. 15
Somers' d Somers, 1982
Critical raUo of S Kendall, 1970, p. 52
Standard error of $, auumlng U•• Kendall, 1970, p. 55
Table of critical valU8S 01 S, a88umlng U•• Ha'lhbarger, 1971, p. 535
Median talt SIOGol, 1956, p. 111
Mann·Whltney U Siegel, 1958, p. 116
Kolmogorov·Smlmov two sampl. tes, Siegel, 1956, p. 121
Runa tall Siegel, 1958, p. 136
Friedman teat Haye, 1873, p. 785
~ - -- --- ------
Freeman's coefficient of differentiation Freeman, 1985, p_ 112

Kruskal-Wallls t8St Siegal, 1958, p. 1M
Median teat (for more than two groups) SI.gel, 1956, p. 179
Page l'
Covarlanca analysis Snedecor and Cochran, 1967, p. 419

F tesl lor covariance analysis Snedecor and Cochran, 1967, p. 424
Page 17
light's agreement coefflclenl Ught, 1971

Crltlc.1 ratio of Light's agreement coefficient lIgh., 1971
Kendall', coefficient 01 concordance (W) Sl.gel, 1958, p. 229
Chl-aquare test for W Siegel, 1958, p. 238
Table of crltlca' value. 01 a In the Kendall coefficient 01 concordance Sleg.I, 1958, p. 286
Intraclass eo".laUon coelflc~nt McNenuu, 1989,p.322
Robinson's A Robinson, 1957
F tesl for InlraclatlS correlallon coefflclenl McNemar, 1989, p. 322
F test for Robinson's A (translate to Intraclass conalaUon and tast .s Robinson, 1957, p. 23;
above) McNamar, 1969, p. 322
Cochran's a Siegel, 1958, p. 161
AnalY81s or variance with repeated me.aures McNemar, 1969, p. 338
F lest for analysla of variance with rapeated me.lures McNemer, 1989, p. 340
Muilldimenslon.l contlng.ncy ta~••naIYI's Siatillici Department, University of Chicago, 1973 (ECTA);
landis el aI., 1978 (GENCAl);
Flenberg, 1977 (O.n.... I)
Chl-squar. t.sts Flenberg, 1977, p. 38 (Pearlon and maximum likelihood)
Pogo"
Canonlca' correlation Cooley and lohnes, 1971, p. 188;

Harris, 1975, p. 132
36
" -
37
Wilks' lambda
Cooley and lohnes, 1971, p. 175;
Morrison, 1976, p. 222; HarriS, 1975, p. 143
Roy's greatest root criterion Morrison, 1976, p. 176;
Har,ls, 1975, pp. 103, 143
Plllal·Barllett V Morrison, 1978, p. 223
a ·type factor analysis Overall and Klett, 1972, p. 201; Gorsuch, 1974, p. 279
Clustering technqlues such as alnol8 IInkaoe, complete linkage,
Sneath and Sok.l, 1973
ayerage linkaoe, K.means
'a01l18-2O
Factor analysIs of correlallon maUh( Gorsuch, 1974

Factor analysis of yarlanc8-Covarlanc8 matrix GorsUCh, 1974, p. 271
Confirmatory taclor analysis of a standardized yarlanc.covarlance
Gorsuch, 1974, pp. 116, 168 (Ganeral);
matrix S6rbom and JOreskoo, 1976 (COFAMM)
Maximum likelihood chi· square GorSUCh, 1974. pp. 118, 139;
SOrbom and JOf••kog, 1978 (COFAMM)
Confirmatory factor analysis of v8rlance-covarlance matrix
GorsUCh, 1974, pp. 116, 168 (General);
SOrbom and JOfaskog, 1976 (COFAMM)
Maximum likelihood chl·aquare Gorsuch, 1974, pp. 118, 139;
SOrbom and JOr8skog, 1978 (COFAMM)
Non·metrlc multidimensional Icallng technique. Kruskal and Wish, 1978 (GenaraO;
Kruskal, 1984a, 1984b (MDSCAl);
Gultman, 1968; lingoes, Roskam, and Borg, 1979 (MINISSA);
Young and Torgerson, 1976 (TORSCA);
Takan., Young, and Deleeuw, 1977 (ALSCAl);
Krulkal, Young, and Seery, 1973 (KYST)
Multidimensional contingency table analyall
Statlatlcs Depa,tment, University
landis et at, 1978 (GENCAT);
0' Chicago, 1973 (ECTA):
Flenberg, 1977 (General)
Chl·aquare tesls Flenbero, 1977, p. 38 (Pearson and maximum likelihood)
Clustering techniques such as single linkage, complet. IInkaoa, Sneath and Sokal, 1973
avaraga linkage, K·means
'age 21
Three·mode factOf' analysis Gorsuch, 1974,p. 283

- -- -- - -- -_ ... .. -_.- .. __ .. _--
Three-way non·metrlc multidimensional Icallng techniques

Kruakal and Wish, 1978, p. 80 (Genera');
Cauoll and Chang, 1970 (INOSCAl);
Harshman, 1970 (PARAFAC);
Lingoes and BorO, 1978 (PINOIS);
Carroll, Pruzanaky, and Kruskal, 1980 (CANOELlNC);
Aamaa)" 1977 (MULTtSCAl);
Takane. Young, and Deleeuw, 1917 (ALSCAl);
Sands and Young, 1980 (ALSCOMP3)
Confirmatory factor analy.11 of standardized varlanca-covlrlance
Gorsuch, 1974, pp. 116, 251 (General);
matrices SOrbom and JOr.skag, 1976 (COFAMM)
Maximum likelihood chl.square
GorSUCh, 1974, pp. 118, 139;
SOrbom and JOraskog, 1978 (COFAMM)
Confirmatory laclor analyslll of var'anc.covarlance mattlee.
Gorsuch, 1974, pp. 118, 251 (General):
SOrbom and JOr•• kog, 1978 (COFAMM)
MaJCImum likelihood chl-aquare GorSUCh, 1974, pp. 118. 139;
"_22 SOrbom and JOreskog, 1978 (COFAMM)
Multivariate analyal8 of variance Cooley and Lohnes, 1971, p. 223;

Harris, 1975, p. 101;
Bock and Haggard, 1968
Wilks' lambda Cooley and lohnes, 1971, p. 175;
Monlaon, 1978, p. 222;
HarriS, 1975, p. 109; Ollon, 1978
Roy'. gr.8test rool criterion Morrison, 1978, p. 178;
Harris, 1975, pp. 103, 109; Ollon, 1976
Ptllal·Barllett V Morrison, 1978, p. 223; Olson, 1978
Profile Inalyals MorriSon, 1978, pp. 153.205
Wllka' lambda Morrtson, 1978, p. 222
Roy's greatest root crllerlon Morrison, 1978, p. 178
Plllal·Bartlett V Morrison, 1978. p. 223
"_23
Structura' model, with 'e'anl varlabl•• J6reskog and SOrbom, 1978

Path InalY81a KIHllnger and Pedhazur, 1973, p. 305
Canonlca' COffelatlon Cooley and Lohne., 1971, p. 188:
Ha,rls, 1975. p. 132
38
39
Wilks' lambda Cooley and Lohnes, 1971. p. 175;
Morrison, 1976, p. 222;
Hanl., 1975, p. 143
Roy'. gr.atest root criterion Morrison, 1976, p. 176;
HarriS, 1975, pp. 103, 143
PIIIIII·Bartlen V Morrison, 1976, p. 223
Multivariate analysis of variance Cooley lind lohnes, 1971, p. 223;

Ha"ls, 1975, p. 118;
Bock lind Haggard, 1968
Wilks' lambda Cooley and lohnes, 1971. p. 175;
Monlaon, 1976, p. 222;
Harris, 1975, p. 109; Olson, 1976
Roy's grea'est root criterion Morrison, 1976, p. 178;
Hanls, 1975, pp. 103, 109; Olson, 1978
Plllal·Bartlett V Morrison, 1978, p. 223; Olson, 1976
Multivariate binary segmentation ,echnlque. 01110, 1972 (MAID); 01110 and Shelley. 19U
Binary segmentation technique. Sonqolst, Oaker, and Morgan, 1974 (SEARCH, formerly known as AID)
Multidimensional contingency 'able analysis based on the cumulaUve Bock. 1975, p. 541 (Gene,al):
logistic distribution Bock and Vates, 1973 (MULTIQUAl)
Chl·square 'eata Bock, 1975, p. 518 (Pearson and maximum likelihOOd)
AnalysiS 01 variance McNemar, 1969, p. 325

F test for analyals at variance McNemar, 1969, p. 349
Multidimensional contlngency table analysis Statistics Department, University of Chicago, 1973 (ECTA);
Chl·square tests Flenberg, 1977, p. 38 (Pearson and maximum Ilkellhood)
Multidimensional contingency table analysis technique allOWing an landis at aI., 1978 (GENCAT)
unconstrained dHlgn matrix
---- ---
Chl·square tests Flenberg, 1977, p. 36 (Peaf80n and mBlClmum likelihOod)

Analysis of variance using weighted least squares Draper and Smith, 1968, p. 77;
Aao, 1965, p. 178
Page. 27- 28
MUllipie discriminant function Cooley and Lohnes, 1971, p. 243

Wilks' lambda Cooley and lohnes, 1971, p. 248
Roy's grealest root criterion Morrison, 1976, p. 178;
Harris, 1975, pp. 103, 109
Plllal·Barlletl V Morrison. 1976. p. 223
Dummy variable regression using weighted least squares or malelmum Draper and Smith, 1966, pp. 77, 134 (Weighted leas! squares-General);
likelihood DuMouchel , 197". 1978 (Malelmum IIkelihood- DREG);
landis et al., 1967 (GENCAT)
Dummy variable regreaslon or multiple elalllflcaUon analysis Draper and Smith, 1968. p. 13":
Andrews el al., 1973;
KerUng" and Pedhazur, 1973, p. 101
Multldlmen~onal contingency table analysis Andrews and Messenger, 1973 (MNA);
Statistics Department, University of Chicago, 1973 (ECTA);
landl. at al .. 1976 (GENCAT);
Chl·square tests Flenberg, 1977, p. 36 (Pearson and mllelmum likelihood)
Multiple curvilinear regression Neter and Waaserman, 197", p. 273
Paget 21-30
Structural models with latent variables Jtlreskog and SOrbom, 1978

Path analysis Kerllnger and Pedhazur, 1973. p. 305
Multiple correlation Hays, 1973, p. 707
F test for multiple correlaUon Hays, 1973, p. 709
Aegresslon coefficient Hays, 1973, pp. 7a.1, 708;
Kerllnger and POOhazur, 1973, pp. 58, 81
F teat tor (egression coefficient Kerllng.r and Pedhazuf, 1973, p. 66
Pari correlaUon MoNemar, 1969, p. 185
F test for part correlaUon McNemar, 1969, p. 321
40
41
ParUal correlation McNemar, 1989, p. 183

Flshe"s r to Z transformallon and the critical ratio 0' Z McNemar, 1989, p. 185
F lest for partIal correlation McNemar, 1969, p. 185
APPENDIX B
PROGRAMS THAT COMPUTE STATISTICS LISTED IN THE GUIDE
For many of the statistics and statistical techniques that explaining how to obtain It.
appear In the Guide, there exist one or more programs that In the following table, at least one program per column Is
calculate the statistic or use the technique. The entries In cited for each entry whenever possible. If multiple programs
this Appendix are Intended to guide the reader to an appro· could be cited, only the program or programs most fre·
prlate program or command. In some cases, the program or quently used for the particular purpose are listed. The appro·
command listed provides a functional approximation to the prlate program, command, or procedure was determined by
Indicated statistic (for example, many programs give prob· a review of the published documentation for each system; It
ability values rather than critical ratios). An asterisk follow· Is therefore possible that some errors, particularly of omls·
Ing a program name means thai the statistic, while not sian, may have been made. It Is Important to note the dates
printed, can be readily obtained or, In more complicated 01 the documentation (see References) as program pack·
cases, that there Is documentation In the User's Manual ages are constantly being Improved and augmented.
- - _.- - ---.-~ ....-- ..'
OSIRIS MIDAS SPSS SAS BMDP OTHER
Plge 4
Mod. TABLES HISTOGRAM FREQUENCIES UNIVARIATE P2D
ONEWAY
Distribution 01
relative frequencies TABLES HISTOGRAM FREQUENCIES UNIVARIATE P2D
ONEWAY CHART
Distribution of TABLES HISTOGRAM FREQUENCIES UNIVARIATE P2D

absolute frequencies ONEWAY CHART
Median TABLES DISTRIBUTION FREQUENCIES UNIVARIATE P2D
Inter-quartlle deviation TABLES' UNIVARIATE·· P2D
N·llies TABLES DISTRIBUTION UNIVARIATE
' .... 6
WlnlOrlzed mean P7D
Trimmed mean P2D
Hampel estimate of P2D

location
Blwelght mean P2D
Mean TABLES DESCRIBE CON DESCRIPTIVE UNIVARIATE PID

USTATS FREQUENCIES MEANS P2D
Median TABLES DISTRIBUTION FREQUENCIES UNIVARIATE P2D
Standard deviation TABLES DESCRIBE CONDESCRIPTIVE UNIVARIATE Pl0

USTATS FREQUENCIES MEANS P2D
Coefficient of variation UNIVARIATE PID

MEANS
•• SAS prints 0, - a,; our reference re'Ms to (a) - 0,)12 .
44
45
OSIRIS MICAS SPSS SAS B"'DP OTHER
Range TABLES DESCRIBE CQNOESCRIPTIVE UNIVARIATE PIC

USTATS fREQUENCIES P2D
Skewness TABLES DESCRIBE CONOESCRIPTIVE UNIVARIATE P2D

FREQUENCIES "'EANS
Criliesl ratio 01 P2D

skewness measure
Table lor teallng

skewness
Kurtosis TABLES DESCRIBE CON DESCRIPTIVE, UNIVARIATE P2D
FREQUENCieS MEANS
erillesl ratio 01 P2D

kurtosis measure
Table for testing

kurtosis
Geary's criterion
for kurtosis
Distribution 01 TABLES HISTOGRAM FREQUENCIES UNIVARIATE P2D

relative frequencies ONEWAY CHART
Distribution of TABLES HISTOGRAM FREQUENCIES UNIVARIATE P2D

absolute frequencies ONEWAY CHART
N·Ules TABLES DISTRIBUTION UNIVARIATE
KolmogoroY·5mlrno'i NPAR
one sample tesl
Lllliefors test
.. UNIVARIATE
Chl·square NPAA FREQ

goodness-ol·fit lest
•
Pogo.
Regression REGRESSN REGRESSION REGRESSION!
coefflclenl GLM P1R
REG P,F
F test for REGRESSN REGRESSION REGRESSION r
regr.8slon coefficient GLM PIR
REG
Coefficient from POLY REGRESSION", I
curvilinear regrasslon GLM PSR
ONEWAY
F lesl for POLY
coetllclent from REGRESSION' ,! GLM PSR
curvilinear regression ONEWAY
t test for paired PAIR T·TEST MEANSI P3D

observations
Robinson'. A
Inlraclass corralaUon ANOVA'

coefficient
F lest for Robinson's A

(Iranslale 10 Int,aclBSs
correlalion coefliclent
and lest as below)
F 18St for ANOVA

Intrlclass correlaUon
coefficient
Krlppendor'fs t
Pogo 1
Pearson's product MDC CORRELATE PEARSON CORR CORR
moment r PeD
MCORR CAOSSTABS P••
Fisher's r to Z MDC CORRELATE PEARSON CORR CORR
transformation and MCaRR CROSSTABS
the critical ratio of Z
8ls8rt., r
•• Requires a sequence of MIDAS commands. See SI.lIstle.' Research Laboratory, 1976, page 274,
t All cap.bllltI•• In SPSS REGRESSION are also available In NEW REGRESSION.
I Requires that the data analyzed be the dlff.rences between the paired observations.
46
47
Critical raUo tor

biserial r
Crltlca' raUo for

point blurlal ,
Tetrachorlc r P4F
Critical ratio for P4F

tetrachorlc r
Critical ratio TABLES' TWOWAY' CROSSTABS' FREO' P~F·

lor phi
p. . . .
Som8fs' d CROSSTABS FREO P4F
Crltlctll retlo 01 S TABLES CROSSTABS FREO P4F

NONPARCORR
Standard 8f'ror 01
5, assuming Uea
Table 0' critical values

of 5, allumlng tI.s
Spearman's rho ACORR NONPARCORR FREO P4F
Critical ratio for

Spearman's rho ACORR NONPARCORR FREO P4F
Table 0' critical

values for rno
Kendall's tau a TABLES NDNPARCORR
Standard enor 01
S, assuming no Ue.
·-
Table 0' critical values
of S, assuming no lies
Kendall's tau b TABLES RCORR CROSSTABS FREQ P4F

TWOWAV CORR
Kendall's lau c TABLES CROSSTABS FAEO .... P4F··
Goodman and TABLES RCORR CROSSTABS FREQ
Krusksl's gamma P4F
TWOWAV
Klm'ad
Pogo,
McNemar's le8t
0'symmetry
TWOWAV NPAR P4F
Yula's a
P4F
Phi TABLES! TWOWAV! CROSSTABS FREQt P4F
Critical fatlo 0' phi TABLES· TWOWAY· CAOSSTASS" FREQ" P4P
Fisher's exact lest TWOWAV CROSSTABS P4F
Pearson chi-square TABLES TWOWAV CROSSTABS FREQ P4F
Goodman and TWOWAY
Kruakal's tau b P4F
Critical rallo of
Goodman and Kruskal's
tau b
Asymmetric lambda TABLES TWOWAV CAOSSTABS FREQ P4F

Critical ratio of lambda TABLES FREQ P4F
Pogo 10
Scott's coefficIent
of agreement
•• SAS and BMOP reler to this 88 Stuart's lau c.

! For two dichotomous variables, Cram'r's V (In MIDAS, Cramltr's phI) 18 equivalent to phi,
48
49
Cohen's agreement TABLES

coeUlclenl1 (kappas)
Crltlca' raUo for TABLES

Cohen's kappas
McNema(s 18.1 of P4F

symmetry
Conllnoency coefficient TABLES TWOWAY CROSSTABS FREQ P4F

Pearson chl·aquare TABLES TWOWAY CROSSTABS FREQ P4F
C,am6r's V TABLES TWONAY CROSSTABS FREQ P4F
Symmetric lambda TABLES TWOWAY CROSSTABS FREQ P4F
Critical ratio 01 TABLES CROSSTABS FREQ P4F

symmetric lambda
P.lI
Jaspen', coefficient of
multl ••rlll correlation
Flaher'. r to Z
transformation and
the critical ,aUo of Z
Mayer and
RobInson'. MyU
Fisher's r to Z
transformation end
the critical raUo of Z
P... 12
Etaa ANOVA ANOVA BREAKDOWN GlM
MCA ANOVA ANOVA
Omega'
OLM
ANOVA
Intraclue correlallon ANOVA'
coefficient GLM
ANOVA
Kelley" epsilon!
F leat lor eta', omega', ANOVA ANOVA

Keney's epanonJ. and
BREAKDOWN GLM P1O'
ANOVA ANOVA
Intracla.. correlation
coefficient
Analyala of variance ANOVA ANOVA ANOVA GLM PIV

ONEWAY ANOVA PlO
BREAKDOWN
MANOVA
F test for analyal. ANOVA ANOVA ANOVA GLM P1V
0' variance
ONEWAY ANOVA P70
BREAKDOWN
MANOVA
Welch atallaUe P10
Brown-Forsythe PlO
ataUIUe
t lest T·TEST T·TEST P70

Bartleu's teal ANOVA ONEWAY OISCRIM P90
MANOVA
PlO
Walth 'eal
Aandomlzetlon teat
for matched pairs
Randomization t•• t for

two Independent aamplea
Randomlzallon tnt for
matched .ample•
•• In OSIRIS. Kelley'a ep.llon' I, labelled adJUlted etal .
50
51
OSIRIS MIOAS SPSS SAS BMOP OTHER
Randomization leat fOf

Independent aamplea
Pogo 15
Slon tMI TABLES· RPAIR NPAR MRANK P3S
WllcOKon slgned·rank TABLES· RPAIR NPAR UNIVARIATE P3S
teat
Somers' d CROSSTABS FREQ P4F
Crltlca' raUo of S TABlES CROSSTABS FREO P4F
Standard error of S,
.ssumlng lies
Tabte 01 crltlca' values

of S, assuming ties
Median test TWOSAMPLE NPAR NPAR1WAY

MRANK
Mann-Whitney U TABLES TWOSAMPLE NPAR NPAR1WAY P3S

MRANK
Kolmogorov-Smlrnov
two sample teat TWOSAMPLE NPAR
Runa test NPAR··
Friedman teat NPAR RANK· paS

RELIABILITY MRANK
Freeman'. coefficient
of differentiation
Kruskal·Waills lest TABLES KSAMPLE NPAR NPAR1WAY P3S

MRANK
Median teat (lor more KSAMPLE NPAR NPA.R1WA.Y

than 2 groups) MRANK
NK
P.,.
Covariance analysis MANOVA COVAR ANOVA GLM PIV

MANOVA F'2V
P4V
F t.s. for MANOVA COVAR ANOVA GLM PIV
covariance analya's MANOVA F'2V
P4V
P.17
LIght" agr••ment
eoefUclent
Crltlca' rallo of light'.

agl'Mment coefficient
kendall'. coefficient ACORR P3S
of concordonco (W)
Chl·squ... I••• for W RCORR P3S

Table of critical values
of • In the Kendall
coefficient of
concordonco
Inlraelaaa correlation ANOVA"
coefficient
Robtnson" A
F tnl for Intracl ... ANOVA
correlatton coefficient
F ,.,t lor Rob6nlOn'. A
.
(tranll•• to In.racl•••
cOffel_lion and t••t ••
_ )-
Cochran', a NPAR
RELIABILITY
AnilYIla of vatllnce RELIABILITY GLM F'2V

with repeated me••ur•• MANOVA ANOVA P4V
•• IN SPSS, thle leat 'S called Weld-Woltowill.
52
_. ~ - ........ .:.-~ -~ . ......>...-.......- - - - - - . . . - . . . " ->. . ..........' ''--'--- - ~ •• - .'-"---
53
OSIRIS MIDAS SPSS SAS BMOP OTHER
F teat for analysis of

variance with ,epealed RELIABILITY GLM P2V
measures ANOVA ANOVA P4V
Multldlmenslona'
contingency table FUNCAT P4F ECTA
analysis GENCAT
Chi-square testa
FUNCAT P4F ECTA
GENCAT
Page 18
Canonical correlation CANONICAL CANCORA CANCORA P6M

Wilks' lambda
CANCaRR CANCORR
Roys greatest root CANONICAL
criterion CANCORR
Plllal·BarUeli V
CANCORR
Q.type factor analysis FACTAN FACTOR FACTOR FACTOR P4M
Clustering techniques CLUSTER CLUSTER
8uch as single IInkag., CLUSTER P2M
complete linkage, FASTCLUS PKM
average linkage.
K-means
,..,..,1-20
Factor analysis of FACTAN FACTOR
cOffalatlon matrix
FACTOR FACTOR P4M
Factor analyal. of FACTOR
varlance-covartance FACTOR P4M
matrix
- - - - - - -- ---- - - ------------------- - - - ------------ - -- - -- -
Confirmatory faclor
analysla of • ROTATE
COFAMM
standardized variance-
COY.flance matrix
Maximum likelihood
chl-aquare COFAMM
Confirmatory factor
analysis 0' variance- ROTATE
COFAMM
covill_nee matrix
Maximum likelihOOd
chi-eqUalS COFAMM
Non-metric MINISSA
multidimensional AlSCAL MINISSA
ICailing techniques MeSCAL
TORse"
KYST
ALSCAL
FUNCAT P4F ECTA
GENCAT
FUNCAT P4F ECTA

GENCAT
Clustering technlquel CLUSTER CLUSTER
such II alng'e linkage, VARCLUS P1M
complete linkage,
aver-oe linkage,
K-meana
.'. -
AlSCAL INDSCAL
1oc:hnlquoo PARAFAC
PINOIS
CANDEUNC
MULTISCAl
"LSCAL
ALSCOMP3
55
Confirmatory lactor FACTOR COFAMM

analysis 01
standardized
variance-covariance
matrices
Maximum likelihood FACTOR COFAMM

chl·squa,e
Confirmatory faclor fACTOR COFAMM
analysis of variance·
covariance matrices
Maximum likelihood FACTOR COFAMM

chl·square
P.22
Multivariate analysis MANOVA MANOVA MANOVA GLM P'V
0' variance ANOVA
Wilks' lambda MANOVA MANOVA GLM P'V
ANOVA
Roy'S grealesl rool MANOYA MANOVA GLM p,y

criterion ANOVA
PIII.I·Bartiett V MANOVA GLM

ANOVA
Profile analysis PROFILE MANOVA GLM P'V

ANOVA
Wilks' lambda MANOVA GLM p,y

ANOVA
Roy's greatest rool PROFILE MANOVA GLM p,y

criterion ANOVA
Plllal·Barliett V MANOVA GLM

ANOVA
PIlIo23
======================---- -- ---- REGRESSION", f SYSAEG
Canonical correlation CANONICAL CANCORA CANCORR P6M
Wilks' lambda CANCORR CANCORR
Roy', gr••t.,t root CANONICAL CANCORR
criterion
Plllal·Bartlett V CANCORR
'_2A
Multivariate analysis MANOYA MANOYA GLM
of variance P4Y
ANOYA
Wilks' lambda MANOVA MANOVA OLM P4V
ANOVA
Roy'a gr•• t•• t root MANOVA OLM P4V
criterion ANOVA
PIII.I·S.rU." V MANOVA OLM
ANOVA
Multivariate tHnary MAID
segmentation
techniques
P_25
Binary segmentation SEARCH"
technique.
Multldlmen"ona' MULTIQUAL
contingency table
analysis based on the
cumulative logistic
dlstrlbuUon
Chl·square tests MULTIOUAL
P_25
Analys'l of variance ANOVA OLM P'V

MANOVA ANOVA
• • Formarly known as AID.

t All cap.bUIU•• 'n SPSS REGRESSION .f. also available 'n NEW REGRESSION.
56
57
F test for analysis ANOVA GlM PIV

01 variance MANOVA ANOVA
Multidimensional FUNCAl P4F ECTA

conllngency tabl.
analyals
ChH,quar. teata FUNCAT P4F ECTA
Multidimensional FUNCAT GENCAT

contingency table
ana'ysil lechnlque
allowing an
unconstrained
design matrix
Chl-aquare tests FUNCAl GENCAT
Analysis of Yarlance GlM P2V

using weighted le.st
square.
' . . . 27 - 28
Multiple discriminant DISCRIMINANT DISCRIMINANT DISCRIM P7M
function SEPARATE CANDISC
Wilks' lambda DISCRIMINANT CANDISC P7M
Roy's great.at root CANDISC

criterion
PIII.I·Bartlen V CANDISC
Dummy variable DREG FUNCAT P3R " GENCAT

regression using PAR"
weighted 1.lIt square.
or maximum likelihood
Dummy .... rl.ble REGRESSW REGRESSION" REGRESSION" , ,
regression or muiliple MCA GLM· P1R"
SELECT· ANOVA
classlflcallon analysis
Multidimensional MNA
conllngency table FUNCAT P4F ECTA
analysis GENCAT
Chl-squ.re tests
FUNCAT P4F ECTA
GENCIIT
Muiliple curvilinear
regre8l10n REGRESSION" , t GLM P1Ao
MANOVA
P• • 2I-30
Structural models with

lalenl variables LISREL
Path analysis
REGRESSION·, t SYSREG
Multiple cOfrelallon REGRESSN REGRESSION REGRESSION' GLM PIR
REG
F test 'or muiliple REGRESSN REGRESSION
correlallon REGRESSION t GLM PI R
REG
Regression coefficient REGRESSN REGRESSION REGRESSION t GLM PIR
REG
F tasl for regression REGAESSN REGRESSION
coefficient REGRESSION t GLM PIR
REG
Part COrrelation REGRESSN·· REGRESSION REGRESSION t
F tesl 'or pari REGRESSN REGRESSION
corralatlon REGRESSION", t
Partial correlation PARTIALS REGRESSION PARTIAL CORA GLM PeR

REGRESSN REGRESSION t REG
FI8her's r to Z
tr.naformatlon and the
critical raUo of Z
F lesl for patUal REGRESSN REGRESSION PARTIAL CORR GLM

correlaUon REGRESSION t REG
•• The square of tha part correlation Is printed; It Is labelled Marginal RSOD,

t All capabllliles In SPSS REGRESSION are also available In NEW REGRESSION_
APPENDIX C
SOME NEW OR RARELY USED STATtSTICAL TECHNIQUES
There are In the statistical literature many statistical performing multidimensional mappings simultaneously for
techniques that are not Included In this Guide for various separate groups so as to generate Information about how
reasons-they may be new and not yet well·known, or they the groups differ. An early algorithm for this type of analysis,
may be old and seldom used. Some of these techniques are INDSCAl (Carroll and Chang, 1970), has now been comple·
noted below. mented by several others that make fewer (or different)
assumptions and that are In other ways more powerful and
1. Multl.arlate analysis of ordinal dltl. general. These Include CANDELINC (Carroll, Puzansky, and
Developing methods of multivariate analysis appropriate Kruskal, 1980), PINDIS (Lingoes and Borg, 1976), MUlTlSCAl
to the uniquely ordinal properties of ordinal scales, Includ· (Ramsay, 1977), AlSCOMP3 (Sands and Young, 1980), and
Ing constructing coefficients that measure multiple and AlSCAl (Takane, Young, and Deleeuw, 1977). (In the de·
partial association among ordinal measures, has been ex- clslon tree, ' these are referred to as three-way non metric
tensively discussed In the methodological literature of the multidimensional scaling techniques.)
1970s but has proven to be a difficult problem. The Issues A second line of methodological Investigation has
are not yet resolved. Useful discussions of the problems, focused on the statistical significance of the obtained flts-
and references to other relevant literature, can be found In that Is, the probability that the correspondence between the
Blalock (1975), Kim (1975), and Mayer and Robinson (1977). multidimensional scaling solution and the observed dala
From a practical standpoint, most analysts who desire to could have been obtained purely by a random placement of
perform a multivariate analysle with ordinal measures disre- • specified number of pOints In a space of given dimension·
gard the uniquely ordinal aspects of their measures and allty; see Isaac and Poor (1974), langehelne (1980), MacCal·
treat them as either nominal scales or Interval scales. lum and Cornelius (1977), Spence and Graef (1974), and
Spence and Ogilvie (1973).
2. De.elopments In nonmetrle multidimensional seaUng. A third line of development has pursued "confirmatory"
Nonmetrlc multidimensional scaling has undergone con· multidimensional scaling - the attempt to fit data to an
slderable development and expansion In recent years existing structure; see Borg and Lingoes (1980), and Lingoes
through several distinct lines of methodological actl~lty. and Borg (1976).
One such line Is yielding a variety of different algorithms for
3. Developmenls In lechnlques lor muilldimensional believe these developments have not yet reached the point
conllngency table analysis. where most social science data analysts can routinely apply
Muilldimensional conllngency I~ble analysis has been them and expect to obtain better results than would be pro·
used mainly with nominal scales, but recent developments duced by more traditional approaches. Useful discussions
allow Its use with Interval scales that have a small number and reviews of biased estimation techniques (Including,
at categories. Because such applications are not yet com- particularly, "ridge regresslon'1 have been provided by the
mon, use 01 multidimensional contingency table analysis following authors: Darlington (1978), Dempster, Schatzoll,
with Interval scales Is not Included In the deciSion tree and Wermuth (1977), Fennessey and d'Amlco (1990), Raze·
portion 01 this Guide. For lurther Inlormatlon, see Flenberg boom (1979), and Smith and Campbell (1980).
(1977) and Landis et al. (1976).
8. Explorltory dltl In.lysls.
4. Polynomial regrellion and nonlinear regression. "Exploratory data analysis" Is a phrase associated with a
As used In this Guide, curvilinear regression relers to collection of techniques proposed by Tukey (1977) that are
polynomial regression, a type 01 regression that Is linear In Intended 10 let the analyst explore a set of data while
Its parameters but not In ItB variables (see Draper and Smith, making minimal assumptions. Although based on well
1966, page 129). This Is dillerent from a type of regression accepted slatlstlcal foundations, Tukey's terminology Is
that Is nonlinear In Its parameters, usually relerred to as nontraditional and his techniques are not yet widely used.
nonlinear regression (see Draper and Smith, 1966, p. 263). Summaries of some of his key Ideas can be found In Hartwig
(1979) and Lelnhardt and Wasserman (1978).
5. Reduced varlanc. regr.llion techniques.
When one Is attempting to predict a dependent variable 7. SurvivII analYlls.
using two or more predictor variables, the appropriate Techniques for survival analysis (I.e., the analysis of time
weights to be applied to those predictor variables can be Intervals between events) are not Included In the tree portion
expected to show substantial variation from one random of this Guide because, at least In the pasl, their appllcallon
sample to another If the correlations among the predictor In the social SCiences has largely been restricted to specifiC
variables are high. Sometimes this Is relerred to as "Insta· disciplines, such as demography. It Is poSSible, however,
blllty" 01 coelliclents that results from high multicollinearity that these techniques could profitably be applied to prob·
among the predictor variables. In recent years there has lems encountered In other contexts, such as studies of resl.
been considerable discussion In the statistical literature dentlal and occupational mobility, completion of education ,
about ways to achieve greater stability In regression coef· and retirement. Techniques to handle cases with Incomplete
ficlents by accepting certain biases. The underlying as· dala (censored data), data Involving competing risks, co·
sumptlon Is that It may be better to use coelliclents that varlales, and Interactions have been developed. Texts that
tend to be reasonably close to the Ideal (population) value describe such techniques Include Kalbfielsch and Prentice
bul that on average tend to come out sllghlly dillerent lrom (1980) and Gross and Clark (1975).
this value, rather than a coefficient that averages to Ihe
correct value over many samples but that In anyone sample 8. Informltlon theory Ind the Inllysls of contingency tabl.s.
may be very far all. Although theoretically Inlerestlng, we A measure of uncertainty. H, derived from Information
60
61
theory. can be used to measure the degree of association Irom unobserved (latent) Interval·scale variables with a
between two or more nominal variables. (The coefficient of blvarlate·normal distribution. Then Ihe "true" product·
association Is often called U.) More generally, Information moment correlation Is estimated by a measure called the
theory has been used to develop methods for analyzing polychorlc correlation coefficient (Olsson, 1979, 1980). The
multidimensional contingency tables. For detailS, see polychorlc coefficient Is a generalization to polychotomles
Gokhale and Kullback (1978). (scales with more than two points) 01 the tetrachorlc coef·
flclent, which Is a similar measure used In the case of two
9. Sempllng .rror. of atell.lle. lrom com pl•• de.lgns. dichotomous variables (see the cautionary footnote on
An assumption often required lor the use of Inferential page 7).
statistics Is that the observations are based on a simple
random sample from some population. This assumption Is 11. Tim••• rl•• enely.I • •
required because the estimates of sampling error a9sume Generally, time series analysis US9S regression tech·
that each observation Is Independent of all others. Often, nlques (often something other than ordinary least squares)
however, stratification or clustering Is used Instead 01 a to analyze or predict change. Economists have been the
simple random procedure, and this Introduces non· leaders among social scIentists In developing this area, but
Independence among the observations. Two programs are other social scientists Increasingly are finding time series
available In the OSIRIS IV software package that can be analysis to be relevant to their analytic problems. The Guide
used to estimate the sampling error of statistics from does not Include time series analysls-partiy because the
clustered or stratified samples: &PSALMS estimates the declslon·tree approach does not lend Itself well to the
sampling error of means, and &REPERR eSlimates the analysis of data of a special type (which Is the case with
sampling error 01 regression statistics. time series data), and partly because time series analysis
has not yet become widely used by social scientists (except
10, The polychorlc correletlon coelllcl.nt economists). However, because several of the major soft·
lor two ordlnel vartebl ••. ware packages now Include time series programs (BMDP,
It was pOinted out In the Instructions and Comments sec- MIDAS, SAS, SPSS), Increased use of these analytic tech·
tion of this Guide that ordlnally scaled variables may be nlques In the coming years seems likely. Introductions to
transformed to ranks, and the transformed data then treated time series analysis for social scientists can be found In
as Intervalty scaled. Another approach has been suggested Glass. Willson, and Gottman (1975), Hannan and Tuma
lor the case of two ordinal variables. This approach (1979), and McCleary et al. (1980).
assumes that the ordinal variables have been generated
L
GLOSSARY
ADDITIVE. A situation In which the best estimate of a dependent COMPLEX SAMPLE DESIGN. Any sample design that uses something
variable la obtained by simply adding together the appropriately com· olher Ihan simple random selecllon. Complex sample designs Include
puted effects 01 each of the Independent variables. Additivity Implies multl·stage selection, andlor slratltlcatlon, andlor clustering. For In·
the absence of Interactions. See ./50 INTERACTION. formation on the calculation of sampling errors of staUstlcs from
AGREEMENT. Agreement measures the extent to which two sels of complex designs, see nole 9 In Appendix C.
scores (e.g,. scores obtained from two ralers) ale Identical. Agreement COVARIATE. A variable thai Is used In an analysis to correcl, adjust, or
Involves a more stringent matching 0' two variables than does covarl· modify the scores on a dependenl variable befora Ihose scores are
atlon, which Implicitly allows one to change Ihe mean (by adding a relaled 10 one or more Independent variables. For example, In an
constant) andlor to change the variance (by multiplying by a constant) analysla of how demographic factors (age, sex, education, etc.) relata
for either or both variables before checking the match. to wage rates, monthly earnings might fhst be adjusted to lake
BIAS, The difference between the expected value of a statistic and the account of (I.e., remove effects attributable (0) number of hours
population value It Is Intended 10 estimate. See EXPECTED VALUE. worked, which In this example would be the covariate.
BIASED ESTIMATOR. A slatlstlc whose expected value Is not equal to COVARIATION. Covarlatlon measures the extent to which cases (e.g.,
the population value. See EXPECTED VALUE. persons) have the same relative positions on two variables. See also
BIVARIATE NORMAUTY. A particular form of distribution of two varla~es AGREEMENT.
that has the traditional "bell" shape (but not all bell· shaped dlstrlbu· DEPENDENT VARIABLE. A variable which the analyst Is trying to explain
tlons are normal), If ploUed In three-dImensional space, with the In terma of one or more Independent variables. The distinction
vertical axis showing the number of cases, the shape would be that of between dependent and Independent variables Is typically made on
a three-dImensional bell (It the varlancea on both variables were equal) theorellcal grounds-In terms of a particular causal model or to test a
or a ""reman's hat" (If the variances were unequal). When perfect .,.. particular hypothesis. Synonym: criterion variable,
variate normality obtaIns, the distribution of one variable Is normal for DESIGN MATRIX. A specification, expressed In matrix format, of the par·
each and every value of the other variable. S.. elso NORMAL Ilcular effects and comblnaUons of effects thai are to be considered In
DISTRIBUTION. an anaJysls,
BRACKETING. The operation of combining categories or ,anges of DICHOTOMOUS VARIABLE. A variable that has only two categories.
values of a variable so a8 to produce a small number of categories, Gender (malellemala) Is an example. S •• • 'so TWO-POINT SCALE,
Sometimes referred to as "collepslng" or "orouplng." DUMMY VARIABLE. A variable with Just two categories that reflects only
CAPITALIZATION ON CHANCE. When one Is searchlno for. maxImally part of the Informallon actually available In a more comprehensive
powerful prediction equation, chance fluctuations In a given semple variable. For example, the 'our-category variable Region (Northeast,
acl to Increase the predictive powe, obtained; since data from another South.ast, Central, West) could be the basIs for a two·category
sample from the same population will show different chance fluctu· dummy variable that would distinguish Northeast from all other
atlons, the equation derived for one sample Is likely to work less well regIons. Dummy variables often come In sets so as to reflect all of the
In any other sample. original Information. In our example, the four·category region variable
CAUSAL MODEL. An abstract quantitative representation of real·world defines four dummy variables: (1) Northeast vs. all other: (2) Southeast
dynamics (I.e., 0' the causil dependencies and other InterrelaUon· vs. all other; (3) Central va. all other; and (") West VI. all other. Alterna·
ships among observed or hypothetical variables). tlve coding procedures (which are equivalent In terms of explanatory
power but which may produce more easily Interpretable estimates) are MATCHED SAMPLES. Two (or more) samples selected In such a way that
effect coding and orthogonal coefficients. each case (e.g., person) In one semple Is matched -I.e., Identical
EXPECTED VALUE. A theoretical average value 0' a staUslic over an within specllled limits-on one or more preselected characteristics
Inllnlte number of samples from the same population. with a corresponding case In the other sample. One example of
HETEROSCEDASTICITY. The absence of hOlllOyenelty of variance. See matched ssmples Is having repeated measures on the same In·
HOMOGENEITY OF VARIANCE. dlvlduals. Another example Is linking husbands and wives. Matched
HIERARCHICAL ANALYSIS. As used on page 26 of the Guide, a hler· samples are different from Independent samples, where such case·by·
archlcal analysis Is one In which Inclusion 01 a higher order Inter· case matching on selected chsracterlstlcs has not been assured.
action term Implies the Inclusion of all lower order terms. For 8l18mple, MEASURE OF ASSOCIATION . A number (a statistic) whose magnitude
II the Interaction of two Independent variables Is Included In an ex- Indicates Ihe degree of correspondence-I.e., strength of relationship
planatory model, then the main affects 'or both of those variables are -belween two variables. An example Is the Pearson product·moment
also Included In the model. correletlon coefficient. Measures of association are different from sta·
HOMOGENEITY OF VARIANCE. A altuaUon In which the varlence on a tist/cal tests 01 association (e.g.• Pearson chl·square, F test) whose
dependent variable Is the ,ame (homogeneous) across aU levels of the primary purpose Is to assess the probability that the slrength of a rela·
Independent variables. In analysis of variance applications, several tlonshlp Is different from lOme preselected value (usually zero). See
staUslics .re avaUable for tesUng the homogeneity assumption (see .1,0 STATISTICAL MEASURE, STATISTICAL TEST.
Kirk, 1968, ps,ge 61); In regression applications, a lack of homogeneity MISSING DATA. Informallon that Is not available for a particular case
can be detected by examination of residuals (see Draper and Smith, (e.g.• person) for which at least some olher InformaUon Is available.
1966, page 86). In either case, a variance-stabilizing transformation This can occur for a variety of re8sons, Including a person's refusal or
may be helpful (see Kruskal, 1978, page 1052). Synonym: homosce- Inability to anSWer a qUestion, nonapplicability of a question, etc. For
dastlclty. Antonym: heterosceda.tlclty. useful discussions of how to overcome problems caused by missing
HOMOSCEDASTICITY. See HOMOGENEITY OF VARIANCE. data In surveys see Herle' (1976) and Kim and Curry (1971).
INDEPENDENT VARIABLE. A variable used 10 explain a dependent MULTIVARIATE NORMALITY. The form 01. dlstrlbullon Involving more
varlab4e. Synonyms: predictor variable, explanatory variable. See also than lwo variableS In which the distribution of one variable Is normal
DEPENDENT VARIABLE. for each and every combination of categories of all other variables.
INTERACTION. A situation In which the direction andlor magnitude of See Harris (1975, page 231) for s discussion 01 multivariate normality.
the relationship between two variables depends on (I.e., differs accord· See ."0 NORMAL DISTRIBUTION.
Ing to) the value of one or more other variables. When Intersctlon Is NOMINAL SCALE. A classification of cases which defines their equlva·
present, simple additive techniques are Inappropriate; hence, Inter- lence and non-equlvalencl, but Implies no quanlllaUve relationships
action is sometimes thought of as the absence of additivity. Syno· or ordering among them. Analyllc techniques appropriate for nomln·
nyms: nonadditivity, conditioning effect, moderating effect, contin- ally scaled variables are not affected by any one-to-one transformation
gency effect. S.. "SO PATIERN VARIABLE, PRODUCT VARIABLE. of the numbers assigned to the classes. See .'so SCALE OF
iNTERVAL SCALE. A scale consisting of equal·slzed units (dollars, MEASUREMENT.
years, etc.). On an Inlerval scale the distance between any two posl· NONADDITIVE. Nol additive. See ADDITIVE, INTERACTION.
tlons IS of known size. Results from snalytlc techniques appropriate NORMAL DISTRIBUTION. A particular form lor the distribution of a
for Interval scales will be affected by any non-linear tlansformatlon of varlabte which, when plotted, produces a "beli" shaped curVe-
the scale values. See a/.o SCALE OF MEASUREMENT. symmetrical, rl81ng smoothly from a small number of cases al both
INTERVENING VARIABLE. A variable which Is postulated to be a pre· extremes to a large number of cases In the middle. Not all symmetrical
dlctor of one or more dependent variables, and simultaneously pre· bell·,haped distributions meet the definition of normality. See Hays
dlcted by one Of more Independent 'Iarlables. Synonym: mediating (1973, page 296).
variable. NORMALITY. S.. NORMAL DISTRIBUTION.
KURTOSIS. Kurtosis Indicates the extenl to which 8 distribution Is more ORDINAL SCALE. A classification of caseS Into a set of ordered classes
peaked or flat·topped than a normal distribution. such that each case Is considered equal to, greater than, or less than
LINEAR. The form of a relationship amorlg variables such that when any every othet case. Analytic techniques appropriate for ordlnally scaled
two variables are plolted, 8 straight line results. A relationship Is variables are not affected by any monotonic transformation of the
linea, If the elleet on 8 dependent variable of a change of onl unit In numbers assigned to the classes. See also SCALE OF
an Independent variable Is the 8ame for all possible such changes. MEASUREMENT.
64
65
OUTLYING CASE (OUTLIER). A case (e.g., person) whose score on a vari· STANDARDIZED VARIABLE. A variable that h.s been transformed by
able deviates aubstantlally from the mean (or other measure of central multiplication of all scores by a constant andlor by the addition of a
tendency). Such ceaes can have disproportionately strong effects on constant to aU scores. Often thes. constants are selected so thai the
atatlatlcs. transform.d scor.s have a mean of z.ro and a variance (and slandard
PATIERN VARIABLE. A nominally sc aled variable whose categories deviation) 01 1.0.
Identify particular combinations (patterns) of SCOfes on two or more STATISTICAL INDEPENDENCE. A complele lack 01 cov.rlatlon between
other ...arlables. For e)(ample, a parly·by·gender pattern variable mlghl variables; a lack of association between variables. When used In an.'-
be developed by clasallylng people Inlo Ihe following .1)( categories: ysls of variance or covariance, statistical Independence between the
(1) Republican males, (2) Independent males, (3) Democratic males. (4) Independent ...arlables 's sometlm.s referred 10 as a balanced design.
Republican females, (5) Independ.nt females, (6) Democratic 'em ales. STATISTICAL MEASURE. A number (a slatlstlc) whose size Indleates the
A paUern variable can b. u.ed to Incorporate Interaction In multi· magnitude of some quantity of Interest-e.g., the strength of a rela-
...arlate analyals. lIonshlp, the amount of variation, the size of a difference, the level of
PRODUCT VARIABLE. An InteNally scaled variable whose scor.s are Income, etc. E)(amples Include me.ns, variances, correlation coeffi-
equal to the product obtained when the value. of two other ....rlables cients, and many others. Statistical measures are different from
are multiplied together. A product varl.ble can be used to Incorporate alatlstlcal tests. See ./so STATISTICAL TEST.
certain types 01 Interacllon In muilivariale analysiS. STATISTICAL TEST. A number (a statlsllc) that can be used to assess the
RANKS. The position of a particular case (e.g., person) relative to other probability that a slatlstlcal measure deviates from some preselected
cases on a defined scale-as In " 1st place," "2nd place," etc. Note that value (otten zero) by no more than would be e)(peeted due to the ope,a·
when the actual values of the numbers designating the relative posl· tlon of chance It the cases (e.g., persons) studied were randomly
tlons (the ranks) are used In analysis they are being treated as an Inler· selected from a I.rg.r population. E)(amples Include Pearson chi·
val scale, not .n ordinal sc.le. Se. ,1'0 INTERVAL SCALE, ORDINAL squa,e, F t.SI, t test, .nd many others. Statistical tests are different
SCALE. Irom statisUcal measures. See al80 STATISTICAL MEASURE.
SCALE OF MEASUREMENT. As used In this Guide, scale of measure· TRANSFORMATION. A change m.de to the scores of all cases (e.g., per·
ment re'ers to the nature of the assumptions one makes about the sana) on a varlabl. by the application of the same mathem.tlcal ope,·
properti.s of a v.rl.bl.; In particular, whether that variable meets the atlon(a} to each score. (Common op.ratlons Include addition of s
definition of nominal, ordinal, or Interval measurement. SH a/80 constant, mulllpllcaUon by • constant. taking logsflthmS, ranking,
NOMINAL SCALE, ORDINAL SCALE, INTERVAL SCALE. br.cketlng, .tc.) .
SKEWNESS. Skewness Is a m••sure 01 lack 01 symmetry of a distribu- TWO-POINT SCALE. It each case Is classllled Into one of two categories
tion. (e.g., yeslno, malelfemale, dead/.llve), the varl.ble Is a two-point scale.
STANDARDIZED COEFFICIENT. Wh.n an an.lyal" Is perlormed on For analytiC purposes, two-polnl scales can be treated as nominal
.... rlabl.s that have been standardized 10 that th.y have variances of scales, ordinal acales, or Interval scales.
1.0, the .stlma'es that reault.r. known •• standardized coefficients; WEIGHTED DATA. Weights are applied wh.n one wishes to adjust the
for e)(ampl., a regression run on original variables produces un- Impact of cas.s (e.g., persons) In the analYSis, e.g., to take accounl of
standardlz.d regr.sslon coefficients known as b'l. while a regr.sslon the number 01 population units that .ach cas. represents. In sample
run on st.nd.rdlz.d variables produces standardized regr.sslon coef· surveys weights ar. most likely to be used with data derived tram
flclents known .s betal. (In practice, both typ.s of coefficients can be sample deSigns h .... lng different s.leelion rates or with data having
estimated from the original ...arlables.) BI.lock (1987), Hargens (1978). m.rkedly dlffer.nt subgroup response rates.
and Kim and Mueller (1976) provide useful discussions on the use of
atandardlzed coefficient • .
REFERENCES
Andrews, D. F.; Bickel, P. J.j Hampel, F. A.; Huber, P. J.; Rogers, W. H.; 1290-1291.
and Tukey. J. W. Robust Eatlmar., of Location: Surv8Yllnd Advances. Bradley, J . V. D/strlbutlon·Free Statistical Tests. Englewood Cliffs, New
Princeton: Princeton UnIYsrslty Press, 1912. Jersey: Prentlce·Hell, 1968.
Andrews, F. M., and Messenger, A. C. Multivariate Nomina' SClIle Analy- Brown, M. B., and Forsythe, A. B. The small sample behavior 01 some sta·
,Is. Ann Arbor: Inlmute tor Social Research, The Unlyerslty of tlstlcs which test the equality of several means. Technometrlcs 16
Michigan, '973. (19748): 129- 132.
Andrews, F. M.; Morgan, J . N.; Sonqulst, J . A.; and Klem, L. Multiple Brown, M. B., and Forsythe, A. B. Robust (e,I8 for the equality of varl·
Classiflc.tlon Ana'y," . Second edition. Ann Arbor. Institute tor Social anceS. Journal 01 the American S,atlstlcsl Assoclat/on 69 (1974b):
Research, The University of Michigan, 1973. 364-367.
Blalock, H. M., Jr. Clus.llnt.rences, cloled populations, and measures Camilli, G., and Hopkins, K. D. Applicability 01chi-square to 2 x 2 contln·
of association. American Political Science R&v/ew 81 (1967): 130-136. gency tables with email expected cell frequencies. Psychological
Blalock, H. M.• Jr. Can we find a genuine ordinal slope analogue? In Bull&lIn 85 (1978): 163-167.
Sociological Melhodologyl976. edited by O. A. Heise. San Francisco: earrolt . J . ~.• and Chang, J. J. AnalYSis of Individual differences In multi·
Joney·Bass. 1975. dimensional scellng via an N·way generalization 01 "Eckart·Young"
Blalock. H. M., Jr. Soc/.1 St,tlstlcs. s.cond edition, revised. New York: decompoSition. Psyehomal,'ka 35 (1970): 283-319.
McGraw·HIII. 1979. Carroll, J. D.; Pruzansky, S.; and Kruskal, J. B. CANOELlNC: a genaral
IBMDP) Dixon, W. J., editor. BMDP SIaUsUc,1 Software 1981 Manusl. approach to multldlmenslonalanalys!s of many·way arrays with linear
Berkeley, California: Unlversl~y of California Press, 1981. constraints on parameters. Psychometrlks 45 (1980): 3- 24.
Bock, A. D. Mulllvariete Stetlsllcel Methods In Behavioral Research. New Cohen, J. A coefficient of agreement lor nominal scales. Educatlona'
York: McGraw·HIII, 1975. .nd Psychologic.' Me.surement 20 (1960): 37-48.
Bock, R. 0., and Haggard, E. A. The use of multivariate analysis of varl· Cohen, J . Weighted kappa: nominal ecale agreement with provision for
ance In behavioral research. In Handbook 0' Measurement and Icaled dleagreement or partial credit. Psychologlc.' Bulletin 70 (1968):
Ass.ssmenl In Behavioral Sciences, edited by D. K. Whltla. Reading, 213-220.
Massachusetts: Addlson·Wesley, 1968. Conover. W. J. Prectlc.1 Honper.metrlc Statistics. New York: John
Bock, R. 0 ., and Yates, G. MUlT/QUAl: log·lInear Analysis 01 Nomln,' Wiley, 1971 .
or Ordinal Qu,lItative Da,e by the M.thod of Maximum likelihood. Cooley, W. W., and lohnes. P. R. Multlv.rlate Data Analysis. New York:
User's Guide. Chicago: Nallonal Educational Resources, 1913. Wiley. 1971.
Borg, I., and LIngoes, J . C. A model and algorithm tor multidimensional O'Agostlno, R. B. Simple compact portable test of normality: Geary's tesl
scaling with external conal,alnts on Ihe distances. Psychometrlke .45 revisited. Psycho/ogIC., Bulletin 74 (1970): 138-1.40.
(1980): 25-38. Darlington. R. B. Aeduced variance regression. Psychological Bulletin 85
Bowker, A. H., A teat for symmetry In contingency tables. Journal of the (1918): 1238-1255.
Amerlc.n Stall'tlc.1 ASloc/.tlon 43 (1948): 512-574. Dempster, P.; Schatzofl, M.; and Wermuth, N. A simulation study of
Bradley, D. R.; Bradley, T. D.: McGrath. S. G.; and Cutcomb. S. O. Type I allernatlves to ordinary least squares. Journel of the American Ststls·
error rate of the chl·square test of Independence In R x C tablee that tle.1 Anoc/.tlon 72 (1977): 77-102.
have small expected freQuencies. Psychologic.' BullaUn 88 (1979):
Dixon, W. J., and Massey, F. J., Jr, Introductfon to St.tlst'c.' Ana/y,/!' Hannan, M. T., and Tuma, N. B. Methods for temporal analysis. In Annual
Third edition. New YOfk: McGraw·HIlI, 1969. R.ll/ew of Sociology: 1979, edited by A. Inkeles. Palo Alto: Annual
Draper, N, R., and Smith, H. Applied Regr6u1on AnalysIs. New York: Reviews, 1979.
Wiley, 1966, Hargens, l. A nole on standardized coelflclenls as structural paramo
DuMouchel. W. H. The regression ot a dichotomous variable. Unpub- eters. SOCiological Method. and Res."ch 5 (1978): 2.7- 258,
lished. Survey Research Center Computer Support Group, Institute for Harris, A. J . A Pr/mar of Mu/r/llarla'e St.tlsrlcs. New York: Academic
Social Research, University 01 Michigan, 197•. Press. 1975.
DuMouchel, W, H. On the analogy between linear and 10g·lInear regres· Harshbarger, T. R. Introductory Stetlsllcs: A Dacls/on Map. New York:
sian. Technlca' Aeport No. 87. Unpublished. Department of Statistics, Macmillan, 1971.
University 01 Michigan, March 1978.
Feinberg, S. E. Th. Ana/ysl, of Cross·CI.sslfled Date, Cambfldge,
Harshman, A. A. PAAAFAC: Foundations 0' the PAAAFAC procedure-
models and conditions for an 'explanatory' multl·model lactor analy·
MassachuseUs: The MIT Press, 1977.
Fennessey, J ., and d'Amico, A, Colllnearlty, ridge regression, and Investi·
sis. Working papers In phonetics 16. los Angeles: University
'ornla at Los Angeles, 1970.
0' Cal/·
gator Judgement. Soc/o'oglc.' M.thods and Reaearch 8 (1980): Hartwig. F. Exploratory D.t. An"ysls. Beverly Hills, California: Sage,
309-340. 1979.
Flelsa, J . L; Cohen, J.; and Everitt, B. S. Llrge sample standard errors of Hays, W. l. St.t/stlcs lor the Social SclMces. Second edliion. New York:
kappa and weighted kappa, Psychologlca' Bull."n 72 (1969): 323-327. Holt. Rinehart. and Winston, 1973.
Freeman, L C. Elamentary Appll~ St.tlstlc. for Sludants In Bahavlora' Hertel, 8 . R. Minimizing error varIance Introduced by missing data
Sclenca. New York: Wiley, 1965. routines In survey analysis. Sociological M.thods .nd R.seerch 4
Gllto. M. W. MAIO: A Honeywell 600 program for an automatlaed survey (1976): 459-474.
analysis. Sahali/oral Sc/.nce 17 (1972): 251-252. Isaac, P. O., and Poor, O. D. S. On the determination of appropriate
Glllo, M. W., and Shalley, M. W. Predictive modelling 01 multlvarlable and dimensionallly In data with error. Psychometrlk. 39 (1974): 91-109.
muilivariate data. Joum.' of th. Am.rlc.n SI.I/stlcal Assoc/ltlon 69 JOreakog, K. G., and SOrbom, O. LISRfL: Analy," of Linear Struclurel
(1974): 846-653. Relationships by the Method of Mu/mum Likelihood. Version IV.
GlaSl, G. V., and Haksllan, A. R. Measures of assoclallon In comparative User's GUide, Chlcego: National Educational Resources, 1978.
experiments: their development and Interpretation. Amerlc.n Educa- KalbflelKh, J . D., and Prentice, R. l. The St.tlstlcal An"ysls of Failure
tlon.1 Research Journ./8 (1969): 403-.1 •. Time Oat• . New York: Wiley. 1980.
Glaas, G. V.; Willson, V. L.; and Gottman, J . M. De"gn and An,'ys/. of Kelley. T. l. An unbiased cOffelallon ratio measure. Proceedings of the
Time S.r/es Experiments. Boulder, Colorado: Colorado Aasoclated Natlone' Ac.demy of Scl.nees 21 (1935): 55.-559.
UniverSity Preas, 1975. Kendall, M. G. Rank Correlation Methods. Fourth edition. London: Grlllln,
Gokhale, D. V., and Kullback, S. Theln/orm.tlon In Contingency Teb/es. 1970.
New York: Marcel Dekker, 1978. Kendall, M. G., and Stuart, A. Tha Advanced Theory 01 St.,lsllcs, Volume
Goodman, l. A., and Kruskal, W. H. Measures of association for cross 2. New York: Hainer, 1981.
claaslflcatlons. Journ.' of the Amer/c.n Stlltl.tlc,1 A.soclll"on .. 0 Kerllnger, F. N., and Pedhazur, E. J . Mull/pie Regr.5S/on In Behsvlor.1
(1954): 732-7"'. Research. New York: Holt, Alnehart .nd Winston, 1973.
Goodman, L A., and Kruskal, W. H. Measures of a18oclatlon for cross Kim, J . Predfctlve me•• ures ot ordinal association. Amerlc.n Journ.' of
classifications III: approximate sampling theory. Journ.' of the Am.r/· Socl%gy 76 (1971): 891-907.
can Sl8t1stlcal Assoc/.t/on 58 (1963): 310-384. Kim, J . Multivariate .nalysls of ordinal variables. Amerlc.n Journal of
Goodman, l. ..., .nd Kruskal, W. H. Measure of a..oclatlon for cross SociOlogy 81 (1975): 281-298.
classification IV: simplification 01 asymptotic vlrlances. Journ.1 of Kim. J ., and Curry, J . The treatment of mlaslngdata In multivariate analy-
the Amffrlcan S,.t/stlc.' Associlltion 67 (1972): .15-421 . sis. Socio/og/cill Methods and R.se"ch 6 (1977): 215-240.
Gorsuch, R. l. F'ctor An"ys/s. Philadelphia: W. B. Saunders, 197• . Kim, J ., and Mueller, C. W. Standardized and unstandardlzed coefficients
Gross, A. J ., and Clark, V. A. Survillal D/strlbutlons: Reliability Appllc" In causal analysis. Soc/olog/ca' Methods .nd ResetJ(ch • (1976):
tlon. In the Blamed/c.' SCiences, New York: Wiley, 1975. 423-438.
Guttman, L A general nonmetrlc technique for finding the smallest Kirk. A. E. ExperImental Design: Procedures for the Behavlora/ Sciences.
coordinate apace tor a configuration of points. Psychomatrlk. 33 Belmont, Caillornla: BrookstCole, 1968.
(1988): 469-506. Krlppendorlt, K. Bivariate agreement coefficients tor reliability of data.
68
69
In SociologIc.' Methodology; 1970, edited by E. F. Borgatt. and G. W. Mosteller, F., and Tukey, J, W. Oala Analysis end R~ression. Reading,
Bohrnstedt. San Francisco: Jo.sey·Ba•• , 1970. Mas.achusetts: Addlson.Wesley, 1977.
Kruskal , J. B. Multidimensional . callng by optimizing goodness of fit 10 a Neter, J., Ind Wasserman, W. Applied Line., Stet/stlcal Models, Home·
nonmetrlc hypothesis. Psychom.trlka 29 (1964a): 1-27. wood, illinois: Richard D, Irwin, 1974.
Kruskal, J . B. Nonmetrlc mullldimensional scaling: a numerical method. Nunnally, J. C. Psychom.trlc Theory. Second edition. New York:
Psychometrlk. 29 (1964b): 115-130. McGraw·HIII, 1978.
Kruskal, J . 8 . Transformations 01 data. In Infernatlona, Encyclopedia 01 Olson, C. L. On chOOSing a test atatiaUc In multivariate analysl. of
St.t/stlcs, Volume 2, edited by W. H. Kruskal and J . M. Tenur. New variance. Psychological Bulletin 83(1976): 579- 586.
York: Crowell Collier and Macmillan. Originally published 1968. Olsson, U. Maximum likelihood estimation 01 the polychorlc correlation
Copyright renewed In 1978 by The Free Press. coefficient. Psychomelrlk. 44 (1979): 443-460,
Kruskal, J . B., and Wish, M. Multldlm.nslon.1 Scaling. Beverly Hills, Cali· Olsson, U. Measuring correlation In ordered two·way conti ngenc y tables.
lornl.: Sage, 1978. Journal 01 Markellng Reselfch 17 (1980): 391-394.
Krusk." J . B.; Young, F. W.; and Seery, J. B. How to use KYST, I very [OSIRIS) Survey Research Center Computer Support Group. OSIRIS IV
fleKlble program to do multidimensional sc.llng and unfolding. User's M.nua/. Seventh edition. Ann Arbor: Inatltute for Social
Unpublished. Bell LaboratOfles, Murray Hills, New Jersey. 1973. Research, The University 01 Michigan. 1981.
Landis, J. A.; Stanish. W. M.; Freeman, J. L ; and Koch, G. G. A computer Overall, J. E" and Klett , C. J. Applied Mult/vlrl.te Analysis. New York:
program for the generalized chl·square analysis of c ategorlal data McGraw·HIII, 1972,
ualng weighted te.st squares (GENCAT). Computer Progflms In BIo- Ramsay, J. 0 , Maximum likelihood el Umalion In multidimensional
medicine 6 (1970): 196- 231 , scaling, Psychometrlkl 42 (1977): 241 - 288.
langehelne, R. Erwanete fitwerte fOr ZufaliakonliguraUonen In PINDIS, Rao, C. R. LIne" Statistical Inferance and Its Applications. New York:
Ze/tschrllt far Sodalpsychologle 11 (1980): 38-49. Wiley. 1965.
Lelnhardt, 5" and Wasserman, S, S. Exploratory data analysis: an Intra· Robinson, W. S. The statistical measurement of agreement. American
ductlon to selected methods. In Soclo'og/c.' Methodology 1979, SocIologIcal Revl.w 22 (1951): 17-25.
edited by K. F. Schuessler. San Francisco: Jossey·Bass, 1978. Rozeboom, W. W. Ridge regression: bonanza or beguilement? Psycho-
light, R. J. MelSures of response agreement for qualitative data: some log/c. ' Bulletin 86 (1979): 242-249,
generalizations and alternatives. Psychological Bu"e"n 78 (1971): Sands, R.• and Young, F. W, Component models lor three·way deta: an
385-377. alternating least squares algorithm with optimal scaling feltures.
lingoes, J . C., and Borg, I. Procrustean Individual difference scaling. Psychom.trlkl 45 (1980): 39- 68,
Journ.' of M"ketlng Re.elfch 13 (1978): 408-407. {SAS) SAS Institute, Inc, SAS Us.,'s Guide. 1979 Edition. Raleigh, North
lingoes, J . C.; Roskam, E, E,; and Borg, I. Oeom.trlc Representations 01 Carolina: SAS Institute, 1979.
Re/atlon.1 O.t• . Second edition. Ann Arbor: Mathesls Press, 1979. (SAS) SAS Institute, Inc, The SAS Supplementa' Library User'S Guide,
M.ccallum, R. C., and Cornelius, E. T. A Monte carlo Investigation at 1980 fdW,?n. Cary, North Carolina: SAS Institute, 1960.
recovery ot structure by ALSCAL P$ychome,rlke 42 (1977): 401 - 428. Siegel, S. Nonpar.metrlc Method. for the Behavioral Sciences. New
Mayer, L. S" and Robinson, J. A. Measures of essoclatlon for multiple York: McGraw·HIII , 1958.
regreSSion model. with ordinal predictor varl.ble• . In Soclologlca' Smith, G., and Campbell, F. A crllique 01 ridge regreSSion methods.
Methodology 1978, edited by K. F. Schuessler. San Francisco: Jossey· Journal 01 tha American Statistical Association 75 (1980): 74- 81 .
Baas. 1977, Sneath, P. H. A., and SotI:al, R. R. Hum_rica' Taxonomy, San Franc isco:
McCleary, R., and Hay, R. A., Jr., with Meidinger, E. E., and McDowall, D. W, H. Freeman, 1973.
Applied Time Serle. An.'y." lor the Social Sciences, Beverly Hills, Snedecor, G. W., and Cochran, W. G. Stetlstlcal Mftlhods. Sixth ed ition.
California: Sage, 1980. Ames, Iowa: The Iowa State University Press, 1967.
McNemar, O. Psychological Stat/stICB. Fourth edition. New York: Wiley, Somera, A. H. A new asymmetric measure ot assoclallon for ordinal
1!M19. varlablas, American Soclologlca' Review 27 (1982): 799-811.
(MIDAS) Fox, D, J ., and GUire, K. E. Documentarlon lor MIDAS. Third Sonqulst , J. A.; Baker, E. l.; and Morgan, J. H. Searching lor Structur• .
edition. Ann Arbor. StaUsUcal Research laboratory, The University of Revised edition, Ann Arbor: Institute for Social Research, The Univer-
Michigan, 1978, sity of Michigan, 1974.
Morrison, D. F. Mulf/varl,'a S/al/st/cal M.thods. Second edition. New SOrbom, D., and JDreskog, K. G. COFAMM: Conllrmatory Faclor Ana/y·
York: McGraw·HIII, 1976. sis with Model Mod.lllcerion. User's Guide, Chlc~go : National Educa·
tlonal Resources. 1978.
Spence, I., and Graef, J . The determination of Ihe underlying dimen- University of Michigan, 1976.
sionality 01 an empirically obtained matrix of proxlmltles. Multlva"a,. Statistics Department, University 01 Chlgaco. EClA program: descrip-
Behavioral R.s.arch 9 (1914): 331-342. tion lor users. Mimeographed paper, 1973.
Spence, l., and Ogilvie, J . C. A table of expected stre" values for random Stuart, A. The estimation and comparison of strengths 01 association
ranklngs In nonmetrlc multidimensional scaling. Multivariate In contlnency tables. BIometrika 40 (1953): 105-110.
Beha'llor.' Research 8 (1973): 511-517. T,kane, Y.; Young, F. W.; and Deleeuw, J . Nonmetrlc Individual dlf·
(SPSS) Nle, N. H.; Hull, C. H.; Jenkins, J. G.; Steinbrenner, K.; and 8enl, ferences multidimensional acallng: an alternaUng least squares
D. H. SPSS: Sf.tlstlcel P.ckage for the Social Sciences. Second method with optimal Bcallng t.atur.s. P.ychom.'rl~. 42 (t971): 7-67.
edition . New York: McGraW-Hili, 1975. Tukay, J. W. Exploratory Da'a Analysis. Reading, Massachusetts:
[SPSSI Hull, C. H., and Nle, N. H. SPSS UPDATE 7-9: New Procedures AddlsonWesley, 1977.
and Facilities for Re/e88as 7-9. New York: McGllw-HIII, 1961. Young, F. W., and Torgerson, W. S. TORSCA, a FORTRAN IV program for
Srlkantan, K. S. Canonlc.1 association between nominal measurements. Shepard·Kruska' multldlmenslonel scaling analysis. BehavIoral
Journ.' of the Amerlcen Statlatlcal Assoc/.tlon 65 (1970): 284-292. Scl.nce 12 (1976): 498.
Statistical Research labofalory. EI.m.ntary Stet/stlcs Using MIOAS. Yule, G. V" and Kendall, M. G. An Introduction to the Theory of S'at/sUcs.
Second edition. Ann Arbor: Slatlstical Research laboralory, The Fourteenth edition. london: Griffin, 1957.
70

GuideSelectingStatisticalTechniques OCR PDF

Uploaded by

Copyright:

Available Formats

GuideSelectingStatisticalTechniques OCR PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

GuideSelectingStatisticalTechniques OCR PDF

Uploaded by

Copyright:

Available Formats

=

uewery 01 ContNH c...... In PuWlcatlon Dati

A Ouldl for "'.cIlng e'ltlsUcllleehnlqUlI for

"l·He!. [)oM "058, ISR code no. 358r - T.p. Yel'SO.

Pubilihed by the Institute 'or Social Research,

Second Edlt5o" 1.1

Mlnuflltltured In the United Stl'H of America

Cover Design by Carol lawrence

Instructions and Comments on the Use 1

The Decision Tree: Questions and Answers 3

Appendix A: Sources of Further Information 31

Appendix B: Programs that Compute Statistics 43

Appendix C: Some New or Rarely Used 59

THE DECISION TREE:

How many variables does the problem Involve?

One Intlrvl', One Intena•• One Onlnll,

(continued from page 4)

• One Interval variable

What do you want to know about the distribution of the variable?

Standard de. Skewness - Kurtosis·

Is 8 distinction made between 8 dependent and an Independent variable?

t The assumptions In nola 5 on page 2 may apply. (r-------~~------~

dependent and an independent variable. The relationship is to

How many of the varIables are dIchotomous?

Is the dichotomous var/able a collapsing of a con-

Pearson's product BIS8flal r' Pearson'. product Tetrachorlc rt Pearson's product

Is B distinction made between a dependent and 8n Independent var/able?

TWO NOMINAL VARIABLES

Ars the variables both two-point scales?

Yule's ot Do you want a statistic based on the number of cases In

~~~u~(x~_J 1 Refer crilical raUo of tau b I

In this case, McNemar's test of symmetry I, equivalent to

What do you want to measure?

Refar critical ratios : rr-~~--A~--~~_

TWO VARIABLES: ONE INTERVAL, ONE ORDINAL

Is the ordinal variable a two·polnt variable?

Do you want to treat the ordlna' var'able as If It were b.stld on an

f Jaapen'a coefficient la the product moment correlation between the

• Any twc>polnt variable meets the criteria for an Intlrvally scaled

r Me.lure of T•• , of "'\

Kelley'. epsllon l (fl)' ,I

Do you went to test the equellty

M••ns Vlnlnc.. "\ ( ~ ~"\

Within eech c.tegory olth. noml·

t The asaumptlons In not. 5 on page 2 may apply.

.Igned- : I 01 the unit normal curve; tor N less than or No 'I

1 The assumptions In nota 5 on page 2 may apply.

, Nonadditivity can be represented within addlllv8 technique. by

• Some analysis 01 covariance technique. assume statlstlc a,

• More than two variables • No distinction Is made between

Do you want to measure a9rffement?

All Ordln.1 Oth... I

fer. to • lable of : I Chl.aquare

t The assumptions In note 5 on page 2 may apply,

t "Two or more groups" may mean dlstinci aeta 01 Indlvlduala, a sel

• More than two variables • No distinction Is made between

( Explon Co•• rlillon Find Cluo',,, ,

( StljardlZ. OrlUljl Metrlc 'I ( Stondardlze Orlulnol Motrlc 'I Multldlmen·

: Maximum likelihood 1 I Maximum likelihood :