Professional Documents
Culture Documents
Using Multivariate Statistics 7Th Edition Barbara G Tabachnick Full Chapter PDF
Using Multivariate Statistics 7Th Edition Barbara G Tabachnick Full Chapter PDF
Barbara G. Tabachnick
Visit to download the full and correct content document:
https://ebookmass.com/product/using-multivariate-statistics-7th-edition-barbara-g-tab
achnick/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
https://ebookmass.com/product/applied-statistics-from-bivariate-
through-multivariate-techniques-second/
https://ebookmass.com/product/applied-statistics-from-bivariate-
through-multivariate-techniques-second-edition-ebook-pdf-version/
https://ebookmass.com/product/etextbook-978-9351500827-
discovering-statistics-using-ibm-spss-statistics-4th-edition/
https://ebookmass.com/product/elementary-statistics-11th-edition-
allan-g-bluman/
Statistics for People Who (Think They) Hate Statistics:
Using Microsoft
https://ebookmass.com/product/statistics-for-people-who-think-
they-hate-statistics-using-microsoft/
https://ebookmass.com/product/discovering-statistics-using-ibm-
spss-statistics-north-american-edition-5th-edition-ebook-pdf/
https://ebookmass.com/product/introductory-statistics-using-
spss-2nd-edition-ebook-pdf/
https://ebookmass.com/product/statistics-informed-decisions-
using-data-4th-edition-ebook-pdf/
https://ebookmass.com/product/methods-of-multivariate-analysis-
wiley-series-in-probability-and-statistics-book-709-3rd-edition-
ebook-pdf-version/
SEVENTH EDITION
Using Multivariate
Statistics
Barbara G. Tabachnick
California State University, Northridge
Linda S. Fidell
California State University, Northridge
Acknowledgments of third party content appear on pages within the text, which constitutes an extension of this
copyright page.
Copyright © 2019, 2013, 2007 by Pearson Education, Inc. or its affiliates. All Rights Reserved. Printed in the
United States of America. This publication is protected by copyright, and permission should be obtained from
the publisher prior to any prohibited reproduction, storage in a retrieval system, or transmission in any form or
by any means, electronic, mechan ical, photocopying, recording, or otherwise. For infom1ation rega rding permis-
sions, request forms and the appropriate contacts within the Pearson Education Global R;ghts & Permissions
d epartmen t, please visit www.pearsoned.com/permissions/.
PEARSON and ALWAYS LEARNING are exclusive trad emarks owned by Pearson Education, Inc. or its
a ffiliates, in the U.S., and/ or other countries.
Unless otherwise indicated herein, any third-party trademarks that may appear in this work are the property of
their respective owners and any references to third-party trademarks, logos or other trade dress are for demon-
strative or descriptive purposes only. Sucll references are not intended to im ply any sponsorship, endorsement,
a uthorization, or promotion of Pearson's products by the owners of such mar ks, or any relationship between the
owner and Pearson Education, Inc. or its affiliates, a uthors, licensees or distributors.
Many of the d esigna tions by manufacturers and seller to distinguish their products are clain1ed as trademar ks.
Where those designations appea r in this book, and the publisher was aware of a trademark claim, the d esigna-
tions have been printed in initial caps or all caps.
1 18
5.7 Complete Examples of Regression Analysis 138 6.5.4.3 Specific Comparisons and
5.7.1 Evaluation of Assumptions 139 Trend Analysis ISS
5.7.1.1 Ratio of Cases to IVs 139 6.5.4.4 Effect Size 187
5.7.12 Normality, Linearity, 6.5.5 Alternatives to ANCOVA 187
Homoscedasticity, and 6.6 Complete Example of Analysis of Covariance 189
Independence of Residuals 139 6.6.1 Evaluation of Assumptions 189
5.7.1.3 Outliers 142 6.6.1.1 Unequal 11 and Missing Oat• 189
5.7.1.4 Multicollinearity and Singularity 144 6.6.1.2 Normality 189
5.7.2 Standard Multiple Regression 144 6.6.1.3 Linearity 191
5.7.3 Sequential Regression 150 6.6.1.4 Outliers 191
5.7.4 Example of Standard Multiple 6.6.1.5 Multicollinearity and Singularity 192
Regression with Missing Values 6.6.1.6 Homogeneity of Variance 192
Multiply Imputed 154 6.6.1.7 Homogeneity of Regression 193
5.8 Comparison of Programs 162 6.6.1.8 Reliability of Covariates 193
5.8.1 ffiM SPSS Package 163 6.6.2 Analysis of Covariance 193
5.8.2 SAS System 165 6.6.2.1 Main Analysis 193
5.8.3 SYSTATSystem 166 6.6.2.2 Evaluation of Covariates 196
6.6.2.3 Homogeneity of Regression Run 196
6 Analysis of Covariance 167 6.7 Comparison of Programs 200
6.1 General Purpose and Description 167 6.7.1 IBM SPSS Package 200
6.7.2 SAS System 200
6.2 Kinds of Research Questions 170
6.7.3 SYSTAT System 200
6.2.1 Main Effects of !Vs 170
6.2.2 Interactions Among IVs 170
6.2.3 Specific Comparisons and Trend
7 Multivariate Analysis of
Analysis 170 Variance and Covariance 203
6.2.4 Effects of Covariates 170 7.1 General Purpose and Description 203
6.2.5 Effect Size 171 7 2 Kinds of Research Questions 206
6.2.6 Parameter Estimates 171 7.2.1 Main Effects ofiVs 206
6.3 Limitations to Analysis of Covariance 171 7.2.2 Interactions Among IVs 207
6.3.1 Theoretical Issues 171 7.2.3 Importance of DVs 207
6.3.2 Practical Issues 172 7.2.4 Parameter Estimates 207
6.3.2.1 Unequal Sample Sizes, Missing 7.2.5 Specific Comparisons
Data, and Ratio of Cases to IVs 172 and Trend Analysis 207
6.3.2.2 Absence of Outliers 172 7.2.6 Effect Size 208
6.3.2.3 Absence of Multicollinearity
7.2.7 Effects of Covariates 208
and Singularity 172
6.3.2.4 Normality of Sampling Distributions 173 7.2.8 Re peated-Measures Analysis
of Variance 208
6.3.2.5 Homogeneity of Variance 173
6.32.6 Linearity 173 7.3 Limitations to Multivariate Analysis
6.32.7 Homogeneity of Regression 173 of Variance and Covariance 208
6.3.2.8 Reliability of Co,•ariatcs 174 7.3.1 Theoretical Issues 208
6.4 Fundamental Equations for Analysis 7.3.2 Practica I Issues 209
of Covariance 174 7.3.2.1 Unequal Sample Sizes,
Missing Data, and Power 209
6.4.1 Sums of Squares and Cross-Products 175
7.3.2.2 Multivariate Normality 210
6.4.2 Significance Test and Effect Size 177
7.3.2.3 Absence of Outliers 210
6.4.3 Computer Analyses of Small-Sample
7.3.2.4 Homogeneity of Variance-
Example 178 Covariance Matrices 210
6.5 Some Important Issues 179 7.3.2.5 Linearity 211
6.5.1 Choosing Covariates 179 7.3.2.6 Homogeneity of Regression 211
6.5.2 Evaluation of Covariales 180 7.3.2.7 Reliability of Covariates 211
6.5.3 Test for Homogeneity of Regression 180 7.3.2.8 Absence of Multicollinearity
and Singularity 211
6.5.4 Design Complexity 181
65.4.1 Wiiliin-Subjects and Mixed 7.4 Fundamental Equations for Multivariate
Wiiliin-Between Designs 181 Analysis of Variance and Covariance 212
6.5.42 Unequal Sample Sizes 182 7.4.1 Multivariate Analysis of Variance 212
vi Contents
14.4 Fundamen tal Equations for Slructural 14.7 Compa rison of Programs 607
Equations Modeling 535 14.7.1 EQS 607
14.4.1 Covariance Algebra 535 14.7.2 LlSREL 607
14.4.2 Model Hypotheses 537 14.7.3 AMOS 612
14.4.3 Model Specification 538 14.7.4 SAS System 612
14.4.4 Model Estimation 540
14.45 Model Evaluation 543 15 Multilevel Linear Modeling 613
14.4.6 Computer Analysis of
545 15.1 General Purpose and Description 613
Small-Sample Example
555 15.2 Kinds of Research Questions 616
14.5 Some Important Issues
14.5.1 Model Identification 555 15.2.1 Croup Differences in Means 616
557 15.2.2 Croup Differences in Slopes 616
14.5.2 Estimation Techniques
14.5.2.1 Estimation Methods 15.2.3 Cross-Level Interactions 616
and Sample Size 559 15.2.4 Meta-Analysis 616
14.5.2.2 Estimation Methods 15.2.5 Relative Strength of Predictors
and Nonnormality 559 at Various Levels 617
14.5.2.3 Estimation Metllods 15.2.6 Individual and Croup Structure 617
and Dependence 559
15.2.7 Effect Size 617
14.5.24 Some Recommendations
for Choice of Estim.1tion
15.2.8 Path Analysis at Individual
Metllod 560 and Croup Levels 617
14.5.3 Assessing the Fit of the Model 560 15.2.9 Analysis of Longitudinal Data 617
14.53.1 Comparative Fit Indices 560 15.2.10 Multilevel Logistic Regression 618
14.5.3.2 Absolute Fit Index 562 15.2.11 Multiple Response Analysis 618
Contents xi
15.3 Limitations to Multilevel Linear Modeling 618 15.7.1.1 Sample Sizes, Missing
15.3.1 Theoretical Issues 618 Data, and Distributions 656
618 15.7.1.2 Outliers 659
15.3.2 Practical Issues
15.3.2.1 Sample SUe, Unequal-11, 15.7.1.3 Multicollinearity
and Singularity 659
and l\1issing Data 619
15.7.1.4 Independence of Errors:
15.3.2.2 Independence of Errors 619
lntracLlss Correlations 659
15.3.2.3 Absence of Multicollinearity
and Singularity 620 15.7.2 Multilevel Modeling 661
15.4 Fundamental Equations 620 15.8 Comparison of Programs 668
15.4.1 Intercepts-Only Model 623 15.8.1 SAS System 668
15.4.1.1 The lnlercep~y Model: 15.8.2 IBM SPSS Package 670
Level-l Equation 623 15.8.3 HLM Program 671
15.4.1.2 The Intercepts-Only Model: 15.8.4 MlwiN Program 671
Level-2 Equation 623
15.8.5 SYSTATSystem 671
15.4.1.3 Computer Analyses
of Intercepts-Only Model 624
15.4.2 Model with a First-Level Predictor 627 16 Multiway Frequency Analysis 672
15.4.2.1 Level-l Equation fora 16.1 General Purpose and Description 672
Model with a Level-l
1'1\.>dictor 627 16.2 Kinds of Resea rch Questions 673
15.4.2.2 Level-2 Equations for a 16.2.1 Associations Among Variables 673
Model with a Level-l 16.2.2 Effect on a Dependent Variable 674
Pl\.>dictor 628 16.2.3 Parameter Estimates 674
15.4.2.3 Computer Analysis of a
Model with a Level-l 16.2.4 Importance of Effects 674
Predictor 630 16.2.5 Effect Size 674
15.4.3 Model with Predictors a t First 16.2.6 Specific Comparisons and
and Second Levels 633 Trend Analysis 674
15.4.3.1 Level-l Equation for 16.3 Limitations to Multiway Frequency Analysis 675
Model with Predictors at 16.3.1 Theoretical Issues 675
Both Levels 633
16.3.2 Practical Issues 675
15.4.3.2 Level-2 Equations for
Model with Predictors 16.3.2.1 Independence 675
at Both Levels 633 16.3.2.2 Ratio of Cases to Variables 675
15.4.3.3 Computer Analyses of 16.3.2.3 Adequacy of Expected
Model with Predictors at Frequencies 675
First and Second Le,·els 634 16.3.2.4 Absence of Outliers in the
15.5 Types of ML\11 638 Solution 676
15.5.1 Repeated Measures 638 16.4 Fundamental Equations for Multiway
15.5.2 Higher-Order ML\11 642 Frequency Analysis 676
15.5.3 Latent Variables 642 16.4.1 Screening for Effects 678
16.4.1.1 Total Effect 678
15.5.4 Nonnormal Outcome Variables 643
16.4.1.2 First-Order Effects 679
15.5.5 Multiple Response Models 644
16.4.1.3 Second-Order Effects 679
15.6 Some Important Issues 644
16.4.1.4 Third-Order Effect 683
15.6.1 lntraclass Correlation 644
16.4.2 Modeling 683
15.6.2 Centering Predictors and Changes 16.4.3 Eva luation and Interpretation 685
in Their In terpretations 646
16.4.3.1 Residuals 685
15.6.3 Interactions 648 16.4.3.2 J>aramctcr Estimates 686
15.6.4 Random and Fixed Intercepts 16.4.4 Compu ter Ana lyses of Small-Sa mple
and Slopes 648 Example 690
15.6.5 Statistical Inference 651
16.5 Some Important Issues 695
15.6.5.1 Assessing Models 651
16.5.1 Hierarchical and Nonhierarchical
15.6.5.2 Tests of Individual Effects 652
Models 695
15.6.6 Effect Size 653
16.5.2 Statistical Criteria 696
15.6.7 Estimation Techniques and 16.5.2.1 Tests of Models 696
Convergence Problems 653
16.5.2.2 Tests of Individual Effects 696
15.6.8 Exploratory Model Building 654
16.5.3 Strategies for Choosing a Model 696
15.7 Complete Example of MLM 655 16.5.3.1 IBM SPSS Hll.OGLINEAR
15.7.1 Evaluation of Assumptions 656 (Hierarchial) 697
xii Contents
A.6 Matrix "Division" (Inverses and B.7 Impact of Seat Belt Law 795
Determinants) 786 B.8 The Selene Online Educational Game 796
A.7 Eigenvalues and Eigenvectors:
Procedures for Consolidating Variance Appendix C
from a Matrix 788
Statistical Tables 797
C.l Normal Curve Areas 798
Appendix B C.2 Critical Values of the t Distribution
Research Designs for Complete for a = .05 and .01, Two-Tailed Test 799
C.3 Cri tical Values of the F Distribution 800
Examples 791 C.4 Critical Values of Chi Square (r) 804
B.1 Women's Health and Drug Study 791 c.s Critical Values for Squares Multiple
B.2 Sexual Attraction Study 793 Correlation (R~ in Forward Stepwise
B.3 Learning Disabilities Data Bank 794 Selection: a = .05 805
B.4 Reaction Ttme to Identify Figures 794 C.6 Critical Values for F~1AX (S2~1AX/S2~iiN)
B.S Field Studies of Noise-Induced Sleep Distribution for a = .05 and .01 807
Disturbance 795
B.6 Clinical Trial for Primary Biliary References 808
Cirrhosis 795 Index 815
Preface
ome good things seem to go on forever: friendship and updating this book. It is d iffi-
S cult to be lieve that the firs t ed ition manuscript was typewritten, with real cu tting and
pasting. The pub lisher required a paper manuscrip t w ith numbered pages-that was
almost our downfa ll. We cou ld write a book on multivariate statistics, bu t we couldn' t get the
same numbe r of pages (abou t 1200, doub le-spaced) twice in a row. SPSS was in release 9.0,
and the o ther p rogram we d emonstrated was BMDP. There were a mere 11 chapters, of which
6 of them were describing techniques. Multilevel and structural equation modeling were not
yet ready for prime time. Logistic regression and survival analysis were not yet popular.
Ma terial new to this edition includes a redo of all SAS examples, with a p retty new output
forma t and replacement of interactive analyses that are no longer available. We've also re-run
the IBM SPSS examples to show the new ou tput format. We've tried to update the references in
all chapters, including only classic citations if they d ate prior to 2000. New work on rela tive im-
portance has been incorpora ted in multiple regression, canonical correlation, and logistic regres-
s ion analysis-complete with d emonstrations. Multiple imputation procedu res for dealing with
missing data have been updated, and we've added a new time-series example, ta king ad vantage
of an IBM SPSS expert modeler that replaces p revious tea-leaf read ing aspects of the analysis.
Our goals in writing the book remain the same as in all previous ed itions-to p resent com-
plex s tatistical procedures in a way tha t is maximally useful and accessible to researchers who
are not necessarily statisticians. We strive to be short on theory but long on conceptual under-
s tanding. The statistical packages have become increasingly easy to use, making it all the more
critical to make sure that they a re applied w ith a good understanding of what they can and
cannot do. But above all else-what does it all mean?
We have not changed the basic format underlying all of the technique chapters, now 14 of
them. We start with an overview of the technique, followed by the types of research questions
the techniques are designed to answer. We then p rovide the cautionary tale-what you need to
worry about and how to deal with those worries. Then come the fundamenta l equa tions underly-
ing the technique, which some readers truly enjoy working through (we know because they help-
fully point out any errors and/ or inconsistencies they find); but other read ers discover they can
skim (or skip) the section without any loss to their ability to conduct meaningful ana lysis of their
research. The fundamental equations are in the context of a small, made-up, usually silly data set
for which compu ter analyses are p rovided- usually IBM SPSS and SAS. Next, we delve into is-
sues surrounding the technique (such as different types of the analysis, follow-up procedures to
the main analysis, and effect size, if it is not amply covered elsewhere). Finally, we provide one or
two full-bore analyses of an actual rea l-life data set together with a Results section appropria te for
a journal. Data sets for these examples are available at www.pearsontughered.com in IBM SPSS,
SAS, and ASCTI formats. We end each technique chapter with a comparison of features available
in IBM SPSS, SAS, SYSTAT and sometimes other specialized p rograms. SYSTAT is a statis tical
package that we reluctantly had to d rop a few editions ago for lack of space.
We apologize in advance for the heft of the book; it is not our intention to line the cof-
fers of cruropractors, p h ysical therapists, acupuncturists, and the like, but there's really just so
much to say. As to our friendship, it's still going strong despite living in d ifferent cities. Art has
taken the place of creating belly dance costumes for both of us, but we remain silly in outlook,
although serious in our analysis of research.
The lineup of people to thank grows with each ed ition, far too extensive to lis t: students,
reviewers, ed itors, and readers who send us corrections and point ou t areas of confusion. As
always, we ta ke full responsibility for remaining errors and lack of clarity.
Barbnrn G. Tabachnick
Linda S. Fidel/
xiv
Chapter 1
Introduction
Learning Objectives
1.1 Explain the importance of multivariate techniques in analyzing research
data
1.2 Describe the basic statistical concepts used in multivariate analysis
1.3 Explain how multivariate analysis is used to determine relationships
between variables
1.4 Summarize the factors to be considered for the selection of variables in
multivariate analysis
1.5 Summarize the importance of statistical power in research study design
1.6 Describe the types of data sets used in multivariate statistics
1.7 Outline the organization of the text
1
Chapter 17 attempts to foster such insights.
2 Chapter 1
In psychology, for example, we are less and less enamored of the simple, clean, laboratory
s tudy, in which pliant, first-yea r college s tudents each provide us with a single behavioral mea-
sure on cue.
In nonexperimental (correla tional or survey) research, the levels of the IV(s) are not ma-
nipulated by the researcher. The researcher can define the IV, bu t has no con trol over the
assignment of cases to levels of it. For example, groups of people may be categorized in to geo-
graphic area of residence (Northeast, Midwest, etc.), but only the definition of the variable is
under researcher control. Except for the military or p rison, place of residence is rarely s ubject
to manip ulation by a researcher. Nevertheless, a naturally occurring d ifference like this is often
considered an IV and is used to p red ict some other nonexperimental (dependen t) variable s uch
as income. In this type of research, the distinction between IVs and DVs is usually arbitrary and
many researchers prefer to call IVs predictors and DVs criterion variables.
In nonexperimental research, it is very difficult to attribu te causa lity to an IV. lf there is a
systematic d ifference in a DV associated with levels of an IV, the two va riables are said (with
some degree of confidence) to be related, but the cause of the relationship is unclear. Fo r exam-
p le, income as a DV might be rela ted to geographic area, bu t no causa l associa tion is implied.
Nonexperimenta l research takes many forms, bu t a common example is the survey.
Typically, many people are surveyed, and each respondent provides answers to many ques-
tions, producing a large number of variables. These variables are us ually interrelated in highly
complex ways, bu t univariate and bivaria te s tatistics are not sensitive to this complexity.
Bivariate correlations between all pairs of va riables, for example, cou ld not reveal tha t the 20 to
25 variables measured really represent only two or three "supervariables."
lf a research goal is to distinguish among s ubgroups in a sample (e.g., between Catholics
and Protestants) on the basis of a variety of attitudinal variables, we cou ld use severa l univari-
ate I tests (or analyses of va riance) to examine group d ifferences on each variable separately.
But if the variables are rela ted, which is highly likely, the resu lts of many t tests are misleading
and s tatistica lly suspect.
With the use of multiva riate s tatistical techniques, complex interrelationships among vari-
ables are revea led and assessed in s ta tistical inference. Further, it is possible to keep the overall
Type I error rate at, say, 5%, no matter how many variab les are tested.
Although most multivariate techniques were developed for use in nonexperimental re-
search, they are also useful in experimental research, in which there may be multiple IVs and
multiple DVs. With multiple IVs, the research is usually designed so that the IVs are indepen-
dent of each other and a straightforward correction for numerous statistica l tests is available
(see Chapter 3). With multiple DVs, a problem of inflated error ra te arises if each DV is tested
separately. Further, a t least some of the DVs are likely to be correlated w ith each o ther, so sepa-
rate tests of each DV reanalyze some of the same variance. Therefore, multivariate tests are used.
Experimenta l research designs with multiple DVs were unusual a t one time. Now, how-
ever, w ith attempts to make experimental designs more realistic, and with the availability of
computer programs, experiments often have several DVs. It is dangerous to run an experiment
with only one DV and ris k missing the impact of the IV because the most sensitive DV is not
measured. Multivaria te s tatistics help the experimenter design more efficien t and more realis tic
experiments by allowing measu rement of multip le DVs w ithout violation of acceptable levels
of Type I error.
One of the few cons idera tions not relevant to choice of statistical technique is whether the
da ta are experimen tal o r correla tional. The statis tical methods "work" whether the researcher
manipulated the levels of the IV or not. But attribution of causality to resu lts is crucially af-
fected by the experimental- nonexperimental d istinction.
If you have access to both packages, you are indeed fortuna te. Progra ms within the pack-
ages d o not completely overlap, and some p roblems a re better handled through one package
than the other. For examp le, doing severa l versions of the same basic ana lysis on the same set
of d ata is particularly easy with IBM SPSS, whereas SAS has the most extensive capabilities for
saving derived scores from d ata screening or from in termed iate ana lyses.
Chapters 5 through 17 (the chapte rs that cover the specialized multivaria te techniques)
offer explanations and illustra tions of a variety of p rograms2 within each package and a com-
parison of the features of the p rograms. We hope tha t once you understand the techniques, you
will be able to generalize to virtually any multivariate program .
Recen t versions of the programs are available in Windows, with menus tha t implemen t
most of the techniques illus trated in this book. All of the techniques may be imple mented
through syntax, and syntax itself is generated through menus. Then you may add o r change
syntax as d esired for your ana lysis. For example, you may "paste" menu choices into a
syntax window in IBM SPSS, ed it the resulting text, and then run the program . Also, syntax
genera ted by IBM SPSS menus is saved in the "journal" fi le (sta tistics.jnl), which may also
be accessed and copied into a syn tax window. Syn tax generated by SAS menus is recorded
in a "log" file. The con tents may then be copied to an inte ractive w indow, ed ited , and run .
Do not overlook the help fi les in these p rograms. Ind eed, SAS and IBM SPSS now p rovid e
the entire set of user manua ls online, often w ith more cu rrent information than is available
in printed man uals.
Ou r IBM SPSS demonstrations in this book are based on syntax generated through menus
whenever feasible. We would love to show you the sequence of menu choices, but space d oes
not permit. And, for the sake of pa rsimony, we have ed ited p rogram ou tpu t to illus trate the
mate rial that we feel is the most important for in terpretation.
With commercial computer packages, you need to know which version of the package you
are using. Programs are contin ua lly being changed, and not all changes are immed iately imple-
mented a t each facility. Therefore, man y versions of the various p rograms are sim ultaneously
in use a t different institutions; even at one ins titution, more than one vers ion of a package is
sometimes available.
Program upda tes are often corrections of errors discovered in earlier versions. Sometimes,
a new version will change the outpu t format but not its information. Occasionally, though,
there a re major revisions in one or more programs or a new program is added to the package.
Sometimes d efau lts change with upd ates, so tha t the outpu t looks different although syntax is
the same. Check to find ou t which version of each package you are using. Then, if you are us ing
a p rin ted manual, be su re that the manua l you a re using is consis tent with the vers ion in use at
you r facility. Also check updates for error correction in previous releases that may be relevant
to some of your p rev ious runs.
Except where noted, this book reviews Wind ows versions of IBM SPSS Version 24 and SAS
Version 9.4. Info rmation on availability and versions of software, macros, books, and the like
changes almost daily. We recommend the In te rnet as a sou rce of "keeping up."
2 We have retained descriptions of features of SYSTAT (Version 13) in these sections, despite the removal of
in the d ata that would be obvious if the data were p rocessed by hand a re less easy to spot
when processing is entirely by comp u ter. Bu t the compu ter packages have programs to grap h
and describe you r da ta in the simplest univariate terms and to display bivaria te relationships
a mong your variables. As discussed in Chap ter 4, these p rograms p rov ide p reliminary analy-
ses that are absolu tely necessa ry if the results of multivaria te programs are to be believed.
There a re also certain costs associated with the benefits of using multivariate p rocedu res.
Benefits of increased flexibility in research design, for ins tance, are sometimes paralleled by
increased ambiguity in interp retation of results. In add ition, multivariate results can be quite
sensitive to which ana lytic s trategy is chosen (cf. Section 1.2.4) and d o not always prov ide bet-
ter p rotection agains t statistica l errors than their univariate counterparts. Add to this the fact
that occasionally you still cannot get a firm statis tical answer to your research questions, and
you may wond er if the increase in complexity and difficulty is warranted.
Frankly, we think it is. Slip pery as some of the concepts and p rocedu res are, these statistics
p rovid e insigh ts into relationships among variables that may more closely resemble the com-
p lexity of the "real" world. And sometimes you get at least partia l answers to q uestions that
could not be asked at a ll in the univariate framework. Fo r a complete analysis, making sense of
your data usually requires a judicious mix of multivariate and univa riate statistics.
The ad dition of multivaria te statis tical methods to your repertoire makes data analysis a
lot more fun. If you liked univariate statistics, you will love multivariate statistics!3
form for analysis w ith many of the techniques4 because religions do not fall along a quantita-
tive con tinuum.
Discrete va riables composed of qualitatively differen t categories are sometimes ana lyzed
after being changed into a n umber of d ichotomous or two-level va riables (e.g., Ca tholic vs.
non-Catholic, Pro testant vs. non-Protestant, Jewish vs. non-Jewish, and so on until the degrees
of freedom are used). Reca tegorization of a d iscrete variable into a series of d ichotomous ones is
called dummy variable coding. The conversion of a discrete variable into a ser ies of dichotomous
ones is done to limit the relationship between the d ichotomous variables and o thers to linear
relationships. A discrete variable with more than two categories can have a rela tionship of any
shape w ith another variable, and the rela tionship is changed a rbitrarily if the assignment of
numbers to categories is changed. Dichotomous variables, however, with only two points, can
have only linear relationships with o ther variables; they are, therefore, app ropriately ana lyzed
by methods using correlation in which only linear relationships are ana lyzed.
The distinction between contin uous and discrete variables is not always clear. If you add
enough digits to the digital clock, for instance, it becomes for all p ractical pu rposes a con-
tin uous measu ring dev ice, whereas time as measu red by the analog dev ice can also be read in
discrete categories such as hou rs or half hours. In fact, any continuous measuremen t may be
rend ered discrete (or dichotomous) w ith some loss of informa tion, by specifying cu toffs on the
con tinuous scale.
The p roperty of variables that is crucial to the application of multivaria te p rocedures is
not the type of measuremen t so much as the shape of distribution, as d iscussed in Chapter 4
and in discussions of tests of assumptions in Chapters 5 through 17. Non-norma lly d is tribu ted
con tinuous va riables and dichotomous va riables with very uneven splits between the catego-
ries p resent problems to several of the multivariate analyses. This issue and its resolution a re
discussed at some length in Chap ter 4.
Another type of measurement that is used sometimes produces a rank o rder scale. This
scale assigns a number to each case to ind icate the case's position vis-a-vis o ther cases along
some d im ension. For ins tance, ranks are assigned to con testan ts (firs t place, second place,
third place, e tc.) to p rovid e an ind ication of who is the best-bu t not b y how much. A p roblem
with rank order measures is that their d is tribu tions are rectangula r (one frequency per num-
ber) instead of norma l, unless tied ranks are permitted and they pile up in the mid dle of the
distribution.
In p ractice, we often trea t variables as if they a re continuous when the underlying scale
is thought to be continuous, but the measured scale actua lly is rank o rder, the number of ca t-
egories is large-say, seven o r more, and the da ta meet o ther assumptions of the analysis. For
ins tance, the number of correct items on an objective test is technica lly not continuous because
fractiona l va lues a re not possib le, bu t it is thought to measure some underlying continuous
variable such as cou rse maste ry. Another example of a va riable with ambiguous measure-
men t is one measured on a Like rt-type scale, in which consumers ra te their a ttitudes towa rd a
p roduct as "strong ly like," "modera te ly like," "mild ly like," "neither like nor dislike," "mild ly
dislike," "moderately dislike," o r "strongly d islike." As mentioned previous ly, even d ichoto-
mous va riab les may be treated as if contin uous under some conditions. Thus, we often use the
term continuous, throughou t the remaind er of this book, whether the measured scale itself is
con tin uous o r the va riab le is to be treated as if con tin uous. We use the te rm discrete for vari-
ables with a few categories, whether the categories differ in type o r quantity.
4
Some multivariate techniques (e.g., logistic regression, SEM) are appropriate for all types of variables.
Introduction 7
Sampling has somewhat differen t connotations in nonexperim enta l and experimen tal
research. In nonexperimental research, you investigate relationships am ong variables in some
p redefined popu lation. Typically, you take elabora te p recautions to ensure that you have
achieved a representative sample of tha t popula tion; you d efine your population, and then do
your best to randomly sample from it.5
In experimental resea rch, you attemp t to create different popula tions by treating subgroups
from an originally homogeneous grou p differently. The sampling objective here is to ensure
that a ll cases come from the same population before you trea t them differently. Rand om sam -
p ling consists of randomly assigning cases to treatmen t groups (levels of the IV) to ensu re that,
before differen tial treatmen t, all s ubsamples come from the same popula tion. Statistical tests
p rovid e evid ence as to whether, after treatment, all samples still come from the same pop ula-
tion. Generalizations abou t treatment effectiveness are made to the type of indiv iduals who
participa ted in the experimen t.
5
Strategies for random sam pling are discussed in many sources, includ ing Levy and Lem enshow (2009), Rea
and Par ker (1997), and de Vaus (2002).
8 Chapter 1
Figure 1.1 Venn diagram for Y (income), X, (education), a nd X2 (occupational p restig e).
vertical s tripes represents the part p redictable from occupational prestige (the second IV, X2);
the circle for education overlaps the circle for income 35% and the circle for occupationa l pres-
tige overlaps 45%. Together, they account for 80% of the variability in income because educa-
tion and occupationa l p restige are orthogonal and do not themselves overlap. There are similar
ad vantages if a set of DVs is orthogonal. The overa ll effect of an IV can be pa rtitioned in to ef-
fects on each DV in an add itive fashion.
Usually, however, the variables are correlated with each other (nonorthogonal). IVs in nonex-
perimental d esigns are often correlated naturally; in experimental designs, IVs become correlated
when unequal numbers of cases are measured in different cells of the design. DVs are usually corre-
lated because individual differences among participants tend to be consistent over many attributes.
When variables are correlated, they have sha red or overlapping variance. In the example
of Figure 1.2, education and occupational p restige correlate with each o ther. Although the inde-
pendent con tribution made by education is s till35% and that by occupational prestige is 45%,
their join t contribu tion to p rediction of income is not 80%, but rather something smalle r d ue to
the overlapping a rea shown by the arrow in Figure 1.2(a). A major decision for the mu ltivaria te
analyst is how to handle the va riance tha t is predictable from more than one va riable. Many
multivariate techrUques have at least two s trategies for handling it, bu t some have more.
In s tanda rd analysis, the overlapping variance contribu tes to the size of su mmary statistics
of the overall rela tionship bu t is not assigned to either variable. Overlapping variance is dis-
regarded in assessing the contribu tion of each variable to the solution. Figu re 1.2(a) is a Venn
d iagram of a standard ana lysis in which overlapping variance is shown as overlapping areas
in circles; the unique contribu tions of X 1 and X 2 to p red iction of Yare shown as horizon tal and
ver tical areas, respectively, and the total relationship between Y and the combination of X1 and
X2 is those two areas plus the a rea w ith the arrow. If X1 is education and X2 is occupationa l
p restige, then in s tanda rd analysis, X1 is "credited w ith" the area marked by the horizonta l
lines and X2 b y the a rea marked by vertica l lines. Neither of the IVs is assigned the area d esig-
nated with the arrow. When X 1 and X2 substantially overlap each other, very little horizontal or
vertical area may be left for either of them, despite the fact that they are both related to Y. They
have essentially knocked each other ou t of the solu tion.
Figure 1.2 Standa rd (a) a nd sequentia l (b) a nalyses of the relationship between Y. x, , and
X2 . Horizonta l s had ing de picts varia nce assigned to X1 . Vertical sha ding depicts vari ance
assigned to X2 .
Introduction 9
Sequential ana lyses differ, in that the researcher assigns priority for entry of variables into
equations, and the first one to enter is assigned both unique variance and any overlapping vari-
ance it has with other variables. Lower-priority variables are then assigned on entry their unique
and any remaining overlapping variance. Figure 1.2(b) shows a sequential analysis for the same
case as Figure 1.2(a), where X1 (ed ucation) is given priority over X2 (occupational p restige). The
total variance explained is the same as in Figure 1.2(a), bu t the relative contributions of X1 and
X2 have changed; education now shows a s tronger relationship with income than in the standard
analysis, whereas the relation between occupa tional prestige and income remains the same.
The choice of strategy for dealing with overlapping variance is not trivial. If variables are
correlated, the overall relationship remains the same, but the apparent importance of variables
to the solution changes depending on whether a s tandard or a sequential strategy is used. If the
multivariate p rocedu res have a reputation for unreliability, it is because solutions change, some-
times d ramatically, when different strategies for entry of variables a re chosen. However, the
s tra tegies also ask different questions of the da ta, and it is incumbent on the researcher to deter-
mine exactly which question to ask. We try to make the choices clear in the chapters that follow.
(1.1)
If, for example, Y' is p redicted income, X1 is educa tion, and X2 is occupational prestige,
the best p red iction of income is obtained by weighting education (X 1) by W1 and occupational
prestige (Xv by W2 before summing. No o ther values of W1 and W2 produce as good a p red ic-
tion of income.
Notice that Equation 1.1 includes neither X 1 nor X2 raised to powers (exponents) nor a
product of X 1 and X2 . This seems to severely restrict multiva riate solu tions un til one realizes
that X1 could itself be a p roduct of two different variab les or a single variable raised to a power.
For example, X1 migh t be education squa red. A mu ltivaria te solution does not produce expo-
nents or cross-products of IVs to improve a solution, but the researcher can include Xs that
are cross-products of IVs or are IVs ra ised to powers. Inclusion of variables raised to powers
o r cross-products of variab les has both theoretical and p ractical implications for the solution.
Berry (1993) p rov ides a useful discussion of many of the issues.
The size of the W values (or some function of them) often reveals a great deal abou t the
relationship between DVs and IVs. If, for instance, the W value for some IV is zero, the IV is not
needed in the best DV- IV relationship. Or if some IV has a large W va lue, then the IV tends to
be important to the relationship. Although complications (to be explained later) prevent inter-
pretation of the multivariate solu tion from the sizes of the W values alone, they are nonetheless
important in most multivariate p rocedures.
The combination of variables can be considered a supervariab le, not d irectly measured
but worthy of interpreta tion. The supervariable may represent an underlying d imens ion that
predicts something or optimizes some relationship. Therefore, the attempt to understand the
meaning of the combination of IVs is worthwhile in many multivariate ana lyses.
In the search for the best weights to apply in combining variab les, computers do not try out
all possible sets of weights. Various a lgorithms have been developed to compute the weights.
Most algorithms involve manipulation of a correlation matrix, a va riance-covariance matrix,
o r a s um-of-squares and cross-products ma trix. Section 1.6 describes these ma trices in very
10 Chapter 1
simple terms and shows their development from a very small data set. Appendix A describes
some terms and manip ulations appropriate to matrices. In the fourth sections of Chap ters 5
through 17, a small h ypo thetical sa mple of d ata is analyzed by hand to show how the weights
are d erived for each analysis. Though this informa tion is useful for a basic understanding of
multiva riate statistics, it is not necessary for applying multiva riate techniques fru itfully to your
resea rch questions and may, sad ly, be skipped by those who are ma th averse.
6
The exceptions are analysis of s tructure, such as factor analysis, in which numerous correlated variables are
measured .
Introduction 11
sample size given a desired level of power (e.g., an 80% probability of acrueving a significant
result if an effect exists) and expected sizes of rela tionships. One of these p rograms that esti-
mates power for severa l techniques is NCSS PASS (Hintze, 2017). Many o ther p rograms are
rev iewed (and sometimes available as shareware) on the Internet. Issues of power relevan t to
each of the statistical techniques are d iscussed in Chapters 5 through 17.
Student x. x2 X3 x.
500 3.20
2 1 420 2.50 2
3 2 650 3.90
4 2 550 3.50 2
5 3 480 3.30 1
6 3 600 3.25 2
x2 X3 x•
x2 1.00 .85 - .13
x2 X3 x.
x2 7026.66 32.80 - 6.00
= 1.50
The off-diagonal elements of the sum-of-squares and cross-products matrix are the cross-
products- the sum of products (SP)-of the variables. For each pair of variables, represented
by row and colu mn labels in Tab le 1.4, the entry is the su m of the p roduct of the deviation
of one variab le around its mean times the deviation of the other variab le a round its mean.
N
SP{X1 Xk) = 2, ( X1i - Xi)(X1k - Xk) (1.3)
J=l
where j iden tifies the firs t variab le, k iden tifies the second variable, and all other
terms a re as defined in Equation 1.1. (Note tha t if j = k, Equation 1.3 becomes
identica l to Equation 1.2.)
!. = - 1- s (1.4)
N- 1
x.
35133.33 164.00 - 30.00
164.00 1.05 - 0.58
- 30.00 - 0.58 1.50
14 Chapter 1
The variance-covariance matrix is prod uced by d ividing every element in the su m-of-
squ ares and cross-products matr ix by N - 1, where N is the n umber of cases.
The correla tion matrix is derived from an S matrix by d ividing each sum-of-squ ares by it-
self (to produce the 1s in the main d iagonal of R) and each cross-p roduct of the S matrix by the
square root of the product of the sum-of-squ ared dev iations around the mean for each of the
variables in the p air. Tha t is, each cross-produ ct is d ivided by
For some multivariate operations, it is not necessary to feed the da ta ma trix to a computer
p rogra m. Instead , an S o r an R matrix is entered, w ith each row (representing a variable) s ta rt-
ing a new line. Often, consid erab le comp u ting time and expense are saved by entering one or
the other of these ma trices rather than raw d ata.
1.6.5 Residuals
Often a goal of analysis or test of its efficiency is its ability to reproduce the values of a DV o r the
correlation matrix of a set of variables. For example, we migh t want to p red ict scores on the GRE
(X2 ) of Table 1.1 from know ledge of GPA (X3 ) and gen der (X4). After app lying the p roper statis-
tical operations- a multip le regression in this case- a p red icted GRE score for each student is
computed by applying the p roper weights for GPA and gender to the GPA, and gender scores
for each stud ent. Bu t because we already ob tained GRE scores for the samp le of stud ents, we
are able to compare the p red icted score with the obtained GRE score. The d ifference between the
p red icted and obtained values is known as the residual and is a measure of error of p red iction.
In most analyses, the residuals for the entire sample sum to zero. Tha t is, sometimes the
p red iction is too large and some times it is too small, bu t the average of a ll the errors is zero.
The squared valu e of the residu als, however, p rovid es a measu re of how good the p rediction
is. When the p redictions a re close to the ob tained va lues, the squared erro rs are small. The way
that the residu als a re d is tribu ted is of fur ther interest in evaluating the degree to which the d ata
meet the assump tions of multivariate ana lyses, as d iscussed in Ch apter 4 and elsewhere.
1
You may find it helpful to read Chapter 18 now instead of waiting for the end.
2
If the effects of some IVs are assessed after the effects of other IVs are statis tically removed, the latter are called
covoriates.
15
16 Chapter2
2.1.1.1 BIVARIATE R Biva riate correlation and regression, as reviewed in Chapter 3, assess the
degree of relationship between two continuous variables, such as belly d ancing s kill and years
of musical training. Biva riate correlation measu res the association between two variables with
no d istinction necessary between IV and DV (depend en t variab le). Bivariate regression, on the
o ther hand, p redicts a score on one variable from knowledge of the score on another va riable
(e.g., predicts ski ll in belly dancing as measured by a single index, such as know led ge of steps,
from a single p red ictor, such as years of musical training).
The p redicted variable is considered the DV, whereas the predicto r is cons idered the IV.
Bivariate correlation and regression are not multivariate techniques, but they are integrated
in to the gen era l linear model in Chap ter 18.
2.1.1.2 MULTIPLE R Multiple correlation assesses the d egree to which one con tinuous vari-
ab le (the DV) is related to a set of o ther (usually) continuous variables (the IVs) that have been
combined to create a new, composite va riable. Multiple correlation is a bivariate correlation be-
tween the origina l DV and the composite va riable created from the IVs. For example, how large
is the association between the belly dancing skill and the set of IVs, such as years of musica l
training, body flexibility, and age?
Multiple regression is used to predict the score on the DV from scores on several IVs. In the
p receding examp le, belly dancing skill measured by knowledge of steps is the DV (as it is for
bivariate regression), and we have add ed body flexibility and age to years of musical training
as IVs. O ther examples are prediction of success in an educa tional p rogram from scores on a
number of aptitude tests, prediction of the sizes of earthquakes from a variety of geological and
electromagnetic variables, or stock market beh av ior from a variety of po litica l and economic
variables.
As for bivariate correla tion and regression, multiple correla tion emphasizes the d egree of
relationship between the DV and the IVs, whereas multip le regression emphasizes the p red ic-
tion of the DV from the IVs. In multiple correlation and regression, the IVs may or may not be
correlated with each other. With some ambiguity, the techniques also allow assessment of the
relative contribution of each of the IVs toward p redicting the DV, as discussed in Chapter 5.
2.1.1.4 CANONICAL R In canonical correlation, there are several continuous DVs as well
as several continuous IVs, and the goal is to assess the rela tionship between the two sets of
variables. For example, we might stud y the relationship between a number of indices of belly
dancing skill (the DVs, such as know ledge of s teps, ability to p lay finger cymbals, and respon-
s iveness to the music) and the IVs (such as flexibility, musical training, and age). Thus, canoni-
cal correla tion ad ds DVs (e.g., further ind ices of belly dancing s kill) to the sing le index of s kill
used in biva riate and multip le correlations, so tha t there are multiple DVs as well as multiple
IVs in canonical correlation.
A Guide to Statistical Techniques 17
2.1.1.6 MULTILEVEL MODELING In many research applications, cases are nested in (nor-
mally occurring) groups, wruch may, in tum, be nested in other groups. The quintessential
example is studen ts nested in classrooms, wruch are, in turn, nested in schools. (Another com-
mon example in volves repeated measures where, e.g., scores are nested within students who
are, in turn, nested in classrooms, which, in turn, are nested in schools.) However, students
in the same classroom are likely to have scores that correlate more rughly than those of s tu-
dents in genera l. This crea tes p roblems with an analysis that pools all students in to one very
la rge group, ignoring classroom and school designations. Multilevel modeling (Chapter IS) is a
somewhat complicated but increasing ly popular strategy for ana lyzing da ta in these situations.
2.1.2.1 ONE-WAY AN OVA AN D t TEST The two statistics, reviewed in Chapter 3, one-way
analysis of va riance (ANOVA) and t test, are strictly univariate in natu re and are adequately
covered in most standard statistica l texts.
possible to rand omly assign people to religious affi liation. In this situa tion, there could easily
be other systematic differences among groups, s uch as level of education, that a re also related
to attitude toward abortion. Apparent differences among religious grou ps might well be d u e to
differences in education rather than d ifferences in religious affilia tion. To get a "pu rer" measure
of the relationship between attitude and religious affiliation, attitude scores are first adjusted
for educational differences, that is, education is used as a covariate. Chapte r 6 also discusses
this somewha t p roblematical use of ANCOVA.
When there are more than two groups, planned o r post hoc comparisons are available in
ANCOVA just as in ANOVA. With ANCOVA, selected and/or pooled group means are ad-
justed for differences on covariates before differences in means on the DV are assessed .
2.1.2.3 FACTORIAL AN OVA Factorial ANOVA, reviewed in Chapter 3, is the s ubject of numer-
ous statistics texts (e.g., Brown, Michels, and Winer, 1991; Keppel and Wickens, 2004; Myers and
Well, 2002; Tabachnick and Fidell, 2007) and is introduced in most e lementary texts. Although
there is only one DV in factorial ANOVA, its place within the genera l linear mod el is d iscussed
in Chap ter 18.
2.1.2.4 FACTORIAL ANCOVA Factorial ANCOVA differs from one-way ANCOVA only in
that there is more than on e IV. The d esirability and the use of covariates are the same. For in-
s tance, in the educational therapy example of Section 2.1.2.2, another interesting IV migh t be
gender of the child. The effects of gender, the type of educational therapy, and their interaction
on the ou tcome are assessed after adjusting for age and prior degree of reading d isability. The
in teraction of gender w ith type of therap y asks if boys and girls differ as to which type of edu -
cational therapy is more effective after adjustmen t for cova riates.
2.1.2.5 HOTELLING'S T2 Hotelling's T 2 is used when the IV has only two groups and there are
sever al DVs. For example, there migh t be two DVs, such as score on an academic achievement
test and attention span in the classroom, and two levels of type of educational therapy, emphasis
on perceptual tra ining vers us emphasis on acad emic training. It is not legitimate to use sepa ra te
t tests for each DV to look for differences between grou ps because that infla tes Type I error due
to unnecessary multiple significance tests with (likely) correlated DVs. Instead, Hotelling's T 2
is used to see if groups d iffer on the two DVs combined . The researcher asks if there are non-
chance differences in the cen troids (average on the combined DVs) for the two groups.
Hotelling's T 2 is a special case of multivariate analysis of va riance, just as the t test is a
special case of univa riate ana lysis of va riance, when the IV has only two groups. Mu ltivaria te
analysis of variance is d iscussed in Chapter 7.
2.1.2.6 ONE-WAY MANOVA Multiva riate analysis of variance evaluates d ifferences among
centroids (composite means) for a set of DVs when there are two o r more levels of an IV
(groups). MANOVA is useful for the educational therap y examp le in the p reced ing section
with two groups and also when there are more than two groups (e.g., if a nontreatment con trol
group is added).
With more than two groups, planned and post hoc comparisons are ava ilable. For example,
if a main effect of treatment is found in MANOVA, it migh t be interesting to ask post hoc if
there are differences in the centroids of the two groups given differen t types of educationa l
therapies, ignoring the control group, and, possibly, if the centroid of the control group differs
from the centroid of the two educational therapy groups combined.
Any number of DVs may be used; the p rocedure deals with correlations a mong them, and
the entire analysis is accomplished within the preset level for Type I error. Once statis tically
s ignificant d ifferences are found, techniques are available to assess which DVs a re influenced
by which IV. For example, assignment to treatment group migh t affect the academic DV bu t not
a tten tion span.
MAN OVA is also ava ilab le when there are w ithin-subject IVs. For exam p l e, child ren might
be measured on both DVs three times: 3, 6, and 9 months after therapy begins. MANOVA is
discussed in Chapter 7 and a special case of it (profile analysis, in which the within-subjects
IV is treated multivariately) in Chapter 8. Profile analysis is an alternative to one-way
A Guide to Statistical Techniques 19
between-subjects MANOVA when the DVs a re all measured on the same scale. Discriminant
analysis is an alternative to one-way between-subjects d esigns, as described in Section 2.1.3.1
and Chapter 9.
2.1.2.7 ONE-WAY MANCOVA In add ition to dealing with mu ltiple DVs, multivariate ana ly·
sis of va riance can be app lied to p roblems when there are one or more covaria tes. In this case,
MANOVA becomes multivariate ana lysis of covariance-MAN CO VA. In the educationa l ther-
apy example of Section 2.1.2.6, it might be worthwhile to adjust the DV scores for pretreatment
differences in academic achievement and attention span. Here the covariates are pretests of the
DVs, a classic use of cova riance analysis. After adjustmen t for p retreatment scores, differences
in posttest scores (DVs) can be more clearly attributed to treatmen t (the two types of educa-
tiona l therapies plus con trol group tha t ma ke up the IV).
In the one-way ANCOVA example of religious grou ps in Section 2.1.2.2, it might be in-
teresting to test political liberalism versus conservatism and attitude toward ecology, as well
as attitude towa rd abortion, to create three DVs. Here again, differences in attitudes might be
associa ted with both d ifferences in religion and differences in educa tion (which, in tum, varies
with religious affiliation ). In the context of MANCOVA, education is the covariate, religious
affiliation the IV, and attitudes the DVs. Differences in a ttitudes among groups with d ifferent
religious affiliations are assessed after adjus tment for differences in education.
If the IV has more than two levels, planned and post hoc comparisons are useful, w ith
adjus tment for covariates. MANCOVA (Chapter 7) is available for both the main ana lysis and
the comparisons.
2.1.2.8 FACTORIAL MAN OVA Factorial MANOVA is the extens ion of MANOVA to designs
with more than one IV and multiple DVs. For example, gender (a between-subjects IV) might
be added to the type of educational therapy (another between-subjects IV) with both academic
achievement and a ttention span used as DVs. In this case, the analysis is a two-way between-
su bjects factoria l MANOVA that prov ides tests of the main effects of gend er and type of educa-
tiona l therapy and their in te raction on the centroids of the DVs.
Duration of therapy (3, 6, and 9 months) migh t be add ed to the design as a within·
su bjects IV with type of educational therapy a between-subjects IV to examine the effects of
d u ration, the type of educationa l therapy, and their in teraction on the DVs. In this case, the
analysis is a facto rial MAN OVA with one between- and one within-subjects IV.
Comparisons can be made among margins or cells in the design, and the influence of vari·
ous effects on combined or ind ividua l DVs can be assessed. For instance, the researcher might
p lan (or decide post hoc) to look for linea r trends in scores associated with duration of therapy
for each type of therapy separately (the cells) or across a ll types of therapies (the ma rgins). The
search for linear trend could be conducted among the combined DVs or sepa ra tely for each DV
with app rop riate adjustments for Type I error rate.
Virtually any complex ANOVA design (cf. Chap ter 3) w ith mu ltip le DVs can be analyzed
through MANOVA, given access to appropriate comp u ter p rograms. Factoria l MANOVA is
covered in Chapter 7.
2.1.2.9 FACTORIAL MANCO VA It is sometimes d esirable to incorporate one or more covari·
ates into a factorial MANOVA design to produce factoria l MANCOVA. For example, pretest
scores on academic achievement and attention span could serve as covariates for the two-way
between-subjects d esign with gender and type of educational therapy serving as IVs and post-
test scores on academic achievement and attention span serving as DVs. The two-way between-
s ubjects MANCOVA provides tests of gender, type of educational therap y, and their in teraction
on adjusted, combined cen troids for the DVs.
Here again, p rocedures a re ava ilable for comparisons among groups or cells and for evalu-
ating the influences of IVs and their interactions on the va rious DVs. Factoria l MANCOVA is
discussed in Chapte r 7.
2.1.2.10 PROFILE ANALYSIS OF REPEATED MEASURES A special form of MANOVA is
available when all of the DVs are measured on the same scale (or on scales with the same
20 Chapter2
psychometric p roperties) and you want to know if groups differ on the scales. For example,
you might use the subscales of the Profile of Mood States as DVs to assess whether mood p ro-
files d iffer between a group of belly dancers and a group of ballet dancers.
There are two ways to conceptualize this design. The firs t is as a one-way between-subjects
design in which the IV is the type of dancer and the DVs are the Mood States subscales; one-
way MANOVA p rov ides a test of the main effect of type of dancer on the combined DVs. The
second way is as a profi le study with one grouping variable (type of dancer) and the several
subscales; profile analysis p rovides tests of the main effects of type of dancer and of subscales
as well as their interaction (frequently the effect of greatest interest to the researcher).
If there is a grouping variable and a repeated measure such as trials in which the same
DV is measured several times, there are three ways to conceptualize the design. The first is
as a one-way between-subjects design w ith several DVs (the score on each trial); MANOVA
p rovides a test of the main effect of the grou ping va riable. The second is as a two-way between-
and within-subjects design; ANOVA p rovides tests of grou ps, tria ls, and their interaction,
but with some very restrictive assumptions that are likely to be violated. Third is as a p rofile
s tudy, in which p rofile analysis prov ides tests of the main effects of groups and trials and their
interaction, bu t withou t the restrictive assump tions. This is sometimes called t11e multivariate
approach to repeated-measures A NOVA.
Finally, you might have a between- and within-subjects design (groups and trials), in
which severa l DVs a re measured on each trial. For example, you might assess groups of belly
and ba llet dancers on the Mood States su bscales a t various points in their training. This ap pli-
cation of profile ana lysis is frequently referred to as doubly multivariate. Chapter 8 deals with all
these forms of p rofile analysis.
2.1.3.1 ONE-WAY DISCRI MINANT ANALYSIS In one-way discriminant analysis, the goal is
to predict membership in groups (the DV) from a set of IVs. For example, the researcher might
want to p red ict category of religious affiliation from attitude toward abortion, liberalism versus
conservatism, and attitude toward ecological issues. The analysis tells us if group membership
is predicted at a ra te that is significantly better than chance. Or the researcher might try to dis-
criminate belly dancers from ballet dancers from scores on Mood Sta tes subscales.
These are the same questions as those add ressed by MANOVA, but turned a round. Group
membership serves as the IV in MANOVA and the DV in discriminant analysis. If groups dif-
fer significantly on a set of va riables in MANOVA, the set of variables significantly p red icts
group membership in discriminant analysis. One-way between-subjects designs can be fruit-
fully ana lyzed through either p rocedure and are often best analyzed with a combination of
both p rocedures.
As in MANOVA, there are techniques for assessing the contribution of various IVs to the
p red iction of grou p membership. For example, the major sou rce of discrimination among reli-
gious groups might be abortion attitude, with little p redictability con tribu ted by political and
ecological attitudes.
In add ition, discriminant analysis offers classification p rocedures to evaluate how well in-
d ividual cases are classified in to their appropriate grou ps on the basis of their scores on the IVs.
One-way discriminant analysis is covered in Chapter 9.
variables are predicto rs of religious affilia tion, variables might be p rioritized according to their
expected con tribution to p rediction, with abortion attitude given the highest p riority, politi-
cal liberalism versus conservatism second p riority, and ecological attitude the lowest priority.
Sequential d iscriminant ana lysis first assesses the degree to which religious affiliation is p re-
d icted from abortion attitude at a better-than-chance rate. Gain in prediction is then assessed
with the addition of political attitude, and then w ith the add ition of ecological attitude.
Sequential analysis p rovides two types of usefu l information. First, it is helpful in eliminat-
ing p redictors that do not contribute more than p red ictors already in the analysis. For example,
if political and ecological attitudes do not add app reciably to abortion attitude in p redicting
religious affiliation, they can be dropped from further analysis. Second, sequential d iscrimi-
nant analysis is a covariance analysis. At each step of the hie ra rchy, higher-priority predicto rs
are covaria tes for lower-priority p red ictors. Thus, the analysis permits you to assess the contri-
bution of a predictor with the influence of other p redictors removed.
Sequential d iscriminant ana lysis is also useful for evaluating sets of predicto rs. For ex-
ample, if a set of con tinuous demographic variables is given higher p riority than an attitudinal
set in prediction of group membership, one can see if attitudes significantly add to prediction
after adjustment for demographic differences. Sequen tial discriminant ana lysis is discussed in
Chapter 9. However, it is usually more efficient to answer such questions through sequential
logistic regression, particularly when some of the predicto r variables are con tinuous and o thers
discrete (see Section 2.1.3.5).
2.1.3.3 MULTIWAY FREQUENCY ANALYSIS (LOGIT) The logit form of multi way frequency
analysis may be used to predict group membership when all of the predictors a re discrete. For
example, you migh t want to p red ict whether someone is a belly dancer or not (the DV) from
knowledge of gender, occupational category, and p referred type of reading material (science
fiction, romance, history, o r s tatistics).
This technique a llows eva luation of the odds tha t a case is in one group (e.g., belly dancer)
based on membership in va rious categories of p redictors (e.g., female professors who read sci-
ence fiction). This form of mu lti way frequency ana lysis is discussed in Chapter 16.
2.1.3.4 LOGISTIC REGRESSION Logistic regression allows prediction of group membership
when p red ictors are continuous, discrete, or a combination of the two. Thus, it is an alternative
to both d iscriminant analysis and logit analysis. Fo r example, p rediction of whether someone is
a belly dancer may be based on gender, occupational category, preferred type of reading mate-
rial, and age.
Logistic regression allows one to evalua te the odds (or p robability) of membership in one
of the grou ps (e.g., belly dancer) based on the combination of va lues of the predictor variables
(e.g., 35-year-old female professors who read science fiction). Chapter 10 covers logistic regres-
s ion analysis.
2.1.3.5 SEQUENTIAL LOG ISTIC REGRESSION As in sequential discriminant analysis,
sometimes p redictors are assigned priorities and then assessed in terms of their contribu tion
to p rediction of group membership given their p riority. For example, one can assess how well
the p referred type of reading material predicts whether someone is a belly dancer after adjust-
ing for d ifferences associated w ith age, gender, and occupational ca tegory. Sequential logistic
regression is also covered in Chapter 10.
2.1.3.6 FACTORIAL DISCRIMINANT ANALYSIS 1f grou ps are formed on the basis of more
than one attribute, prediction of group membership from a set oflVs can be performed th rough
factorial discriminan t analysis. For example, respondents might be classified on the basis of
both gender and religious affiliation. One cou ld use attitudes towa rd abortion, politics, and
ecology to p red ict gender (ignoring religion) or religion (ignoring gender), or both gender and
religion. Bu t this is the same problem as addressed by factoria l MANOVA. For a nu mber of rea-
sons, p rograms designed for d iscriminant analysis do not readily extend to factoria l arrange-
ments of groups. Unless some special cond itions are met (d. Chapter 9), it is usually better to
rephrase the research question so that factorial MANOVA can be used.
22 Chapter2
2.1.4 Structure
Another set of questions is concerned with the latent structure underlying a set of variables.
Depending on whether the search for structure is empirical or theoretical, the choice is principal com-
ponents, factor analysis, or structural equation modeling. Principal components is an empirical ap-
proach, whereas factor analysis and structural equation modeling tend to be theoretical approaches.
2.1.4.2 FACTOR ANALYSIS When there is a theory about underlying structure or when the re-
searcher wants to understand underlying structure, factor analysis is often used. In this case, the
researcher believes that responses to many different questions are driven by just a few underlying
structures called factors. In the example of mechanisms for coping with stress, one might hypoth-
esize ahead of time that there are two major factors: general approach to problems (escape vs. direct
confrontation) and the use of social supports (withdrawing from people vs. seeking them out).
Factor analysis is useful in developing and assessing theories. What is the structu re of per-
sonality? Are there some basic dimensions of personal ity on which people differ? By collecting
scores from many people on numerous variables that may reflect different aspects of personality,
researchers address questions abou t underlying structure through factor analysis, as d iscussed
in Chapter 13.
Logistic regression analyzes a discrete DV too, bu t the IVs are often a mix of con tinuous
and d iscrete variables. For tha t reason, the goa l is to predict the probability that a case will fall
into various levels of the DV rather than group membership per se. In this way, the analysis
closely resembles the familiar chi-square analysis. In logistic regression, as in all mu ltivaria te
techniques, the IVs are combined, bu t in an exponent rather than d irectly. That makes the anal-
yses conceptually more difficu lt, bu t well worth the effort, especially in the medical/biologica l
sciences where the risk ratios, a p roduct of logistic regression, are routinely discussed.
There are several procedures for examining structure (tha t become increasing ly "specu-
lative"). Two very closely aligned techniques are principal components and factor analyses.
These techniques are interesting because there is no DV (or, for that matter, IVs). Instead, there
is jus t a bunch of variables, w ith the goal of analysis to discover which of them "go" together.
The idea is that some latent, underlying structure (e.g., several different factors representing
components of personality) is driving similar responses to correlated sets of questions. The
trick for the researcher is to divine the "meaning" of the factors tha t are developed during
analysis. The technique of p rincipa l components p rovides an empirical solution while factor
analysis provides a more theoretical solu tion.
Structural equation modeling combines multiple regression with factor analysis. There are
one or more DVs in this technique, and the DVs and IVs can be both discrete and continuous, both
latent and observed. That is, the researcher tries to predict the values on the DVs (continuous or
d iscrete) using both observed IVs (con tinuous and discrete) and the latent ones (factors derived
from many observed variables during the analysis). Structural equation modeling con tinues to
rapid development, with expansion to MANOVA-like analyses, longitudinal analysis, sophisti·
cated procedures for handling missing data, poorly d istributed variables, and the like.
Multilevel modeling assesses the significance of va riables where the cases are nested into
different levels (e.g., s tudents nested in classes nested in schools; patients nested in wards
nested in hospitals). There is a DV at the lowest (student) level, but some IVs pertain to stu·
dents, some to classes, and some to schools. The analysis ta kes into accoun t the (likely) higher
correlations among scores of students nested in the same class and of classes nested in the same
school. Relationships (regressions) developed at one level (e.g., p red icting student scores on
the SAT from parental educationa l level) become the DVs for the next level, and so on.
Finally, we p resen t two techniques for analyzing the time course of events, survival ana ly-
sis and time-series analysis. One underlying IV for both of these is time; there may be other IVs
as well. In survival analysis, the goal is often to determine whether a treated group survives
longer than an untrea ted group, given the current standard of care. (In manufacturing, it is
called failure analyses, and the goal, for instance, is to see if a part manufactured from a new
alloy fails later than the part manufactured from the current alloy.) One advantage of this tech·
nique, a t least in medicine, is its ability to analyze da ta for cases that have disappeared for one
reason o r another (moved away, gone to another clinic for treatment, or died of another cause)
before the end of the study; these are called censored cases.
Time-series analysis tracks the pattern of the DV over multiple measurements (a t least 50)
and may o r may not have an IV. If there is an IV, the goal is to determine if the pattern seen in
the DV over time is the same for the group in one level of the IV as for the group in the other
level. The IV can be natura lly occurring or manipulated.
Generally, statistics are like tools- you pick the wrench you need to do the job.
frequ ently, a judicious mix of univa riate and multivaria te procedures to fully ans wer their re-
search questions. We recommend a flexible approach to data analysis in which both univaria te
and multivariate p rocedures are used to clarify the results.
<
One
One
(continuous) (continuous) Bivariate r
~"·~~ ....
Multiple (discrete)
- - - -- - Multiway frequency
analysis
<
One
(continuous)
<
None - - Factorial ANOVA
Multiple (discrete)
Some-- Factorial ANCOVA
Multiple
(continuous)
< One (d iscrete)
<
None __ One-way MANOVA
or Hotelling's
Significance
of group
differences
Multiple (discrete)
< None - - Factorial MANOVA
(Con tinued)
26 Chapter2
None _ Factorial
Multiple ____ Multiple discriminant function
(discrete) (continuous)< Sequential factorial
Some-- . . . .
d1scnmmant function
Multiple
(continuous - - Multiple (latent) - - - - - Factor analysis_ _ _-!
(theoretical)
observed)
Multiple Multiple (continuous Principal components
(latent) - - - - observed) - - - - - (empirical)
Table 2 .1 (Continued)
Survival a nalysis
None - - - None
(life tables)
Time
course of
events
One (time)
Results section when writing for an audience that is expected to be unfamilia r with the tech-
nique. When more than one major type of technique is available, there a re additional complete
examples using rea l data. Finally, a detailed compa rison of features available in the IBM SPSS,
SAS, and SYSTAT p rograms is mad e.
In working w ith these technique chap ters, it is suggested tha t the stud ent/resea rcher apply
the various analyses to some interesting large data set. Many data banks a re readily accessible
online, for example US census data a t https://www.census.gov /data.html.
Further, although we recommend methods of reporting multivariate results, it may be
inapp ropria te to repo rt them fu lly in all publications. Certainly, one would at least want to
mention that univaria te resu lts were supported and guided by multivaria te inference. Bu t the
details associated w ith a full disclosure of multivariate results at a colloquium, for ins tance,
might require more attention than one could reasonably expect from an aud ience. Likewise, a
full multiva riate analysis may be more than some journals are willing to p rint.
Learning Objectives
3.1 Explain hypothesis testing using statistical decision theory
3.2 Perform analysis of variance in different research design types
3.3 Summarize the factors that affect parameter estimation
3.4 Calculate the degree of association between variables using effect size
3.5 Use correlation and regression to determine the relationships between two
continuous variables
3.6 Use the chi-square test to assess the association between two discrete
variables
This chapter p rovides a brief review of univariate and biva riate statistics. Although it is p rob-
ably too "dense" to be a good source from which to learn, it is hoped that it will serve as a use-
ful reminder of ma terial a lread y mastered and will help in establishing a common vocabulary.
Section 3.1 goes over the logic of the s tatistical hypothesis test, and Sections 3.2 through 3.4
s kim many topics in analysis of variance and a re the background for Chapters 6 through 9.
Section 3.5 summarizes correlation and regression, which are the background for Chapters 5,
12, 14, 15, and 17, and Section 3.6 su mmarizes chi-squa re (y), which is the background for
Chapters 10, 14, and 16.
29
30 Chapter3
have smaller stand ard devia tions than d is tributions of scores, and the d ecrease is related toN,
the sample size. For the sample, then,
15
(T_ = -- = 3
y \125
The question being asked, then, is, "Does ou r mean, taken from a sample of size 25, com e from
a sampling distribution with f.Ly = 100 and ur = 3 o r does it come from a samp ling distribu-
tion with f.Ly = 108 and ur = 3?" Figure 3.1(a) shows the first sampling d istribution, defined as
the nu ll h ypothesis, HQ, that is, the sampling distribution of means calcula ted from a ll possible
samples of size 25 taken from a pop ulation where J.L = 100 and u = 15.
The sampling distribu tion for the null h ypothesis has a special, fond place in statistica l
decision theory because it a lone is used to define "best guess." A d ecision axis for retaining or
rejecting H 0 cu ts through the distribu tion so that the p robability of rejecting H 0 by mista ke is
sma ll. "Sma ll" is defined p robabilis tically as a. An er ror in rejecting the n ull hypo thesis is re-
ferred to as an a or Ty pe I error. There is little choice in picking a. Tradition and jou rnal editors
decree that it is .05 or smaller, meaning that the null h ypothesis is rejected no more than 5% of
the time when it is true.
With a table of areas under the standard normal d istribution (the tab le of z scores or s tan-
dard normal devia tes), the d ecision axis is placed so that the probability of obtaining a sample
mean above tha t poin t is 5% or less. Looking up 5% in Table C.1, the z corresponding to a
5% cutoff is 1.645 (between 1.64 and 1.65). Notice that the z scale is one of two abscissas in
Figu re 3.1(a). If the decis ion axis is placed where z = 1.645, one can translate from the z scale to
theY scale to p roperly position the decision axis. The transformation equation is
(3.2)
Equation 3.2 is a rearrangement of terms from the z test for a sing le sample: 1
z= - -
r- ,_. (3.3)
u_
y
1
The more usual proced ure for testing a hypothesis about a single mean is to solve for z on the bas is of the
sample mean and standard deviation to see if the sample mean is s ufficiently far away from the mean of the
sampling dis tribution under the null hypothesis . If z is 1.645 or larger, the null hypothesis is rejected.
Another random document with
no related content on Scribd:
may result from forcible manipulation are filled in by new bone, but
there do not seem to be any observations to confirm this statement.
The amount of force which must be employed is a matter for the
finest discrimination. The method includes complete anesthesia,
traction upon the spine in each direction from the location of the
deformity, and direct pressure force applied to the protection itself,
as by a sling passed around the body and just beneath the
projection, which can be used as a fulcrum upon which the rest of
the spine can be applied as a double lever, with the application, at
first, of gentle force, and, finally, sufficient to either satisfy the
operator that he should go no farther or that the desired effect has
been obtained. Immediately after completion of the maneuver a
snugly fitting plaster jacket should be applied and the patient kept
absolutely at rest in bed.
Fig. 259 Fig. 260
SACRO-ILIAC DISEASE.
Under this name is included a tuberculous condition of the bony
tissues on either side of the sacro-iliac synchondrosis, or of the
cartilage itself, similar to that which produces the special caries
described above. It is an uncommon expression of tuberculous
disease occurring often in the young, identical in pathology with
other tuberculous bone lesions, and giving rise to peculiar
symptoms, mainly because of its location. Early in the course of the
disease these may consist of mild discomfort in the lower abdomen,
irritability of the bladder and bowels, disinclination for exercise, while,
as the disease becomes more pronounced, there will be actual pain,
intensified by standing, relieved by lying down, often severe at night,
usually referred along the course of the sciatics. A most significant
symptom is the tenderness and complaint produced by firm pressure
made upon both sides of the pelvis, thus forcing tender surfaces
against each other. In the later stages of the disease abscess may
develop and present either externally in the lumbar region or
internally, breaking into the pelvis and appearing perhaps in the groin
or close to the perineum. The disease is usually unilateral, and will
cause characteristic limping and aggravated pain upon standing on
the limb of the affected side. Naturally this limb will be spared in
every possible way. It is likely to be mistaken for sciatica or lumbago,
in neither of which diseases is there any tenderness at the sacro-iliac
joint such as can be evoked by pressure from the sides of the pelvis.
It also has to be distinguished from hip disease by the fact that
motions at the nip are not interfered with, and from Pott’s disease of
the lower spine, which usually causes prominence of the spinal
processes and local tenderness in a different region.
The surfaces and tissues involved are extensive and the disease
is always serious. It is one of the most chronic of all such affections,
and too often tends to suppuration, with its slow but inevitable
consequences, or to dissemination. Thus of 38 cases with abscess
reported by Van Hook only 3 recovered.
Treatment.—Treatment should consist of absolute rest, with
traction, so long as the symptoms are active, and
avoidance of all irritation when patients rise from bed. Abscess due
to sacro-iliac disease should be radically attacked, especially if this
can be done early. Intrapelvic pus collections may require trephining
of the pelvic walls or resection of some portion of the ilium, by which
complete evacuation may be made and drainage be amply provided.
When the joint itself is thoroughly broken down the case will have a
hopeless aspect.
NON-CARIOUS DEFORMITIES.
TORTICOLLIS; WRYNECK.
This term includes a peculiar postural deformity by which the head
is rotated and inclined abnormally to one side in a more or less fixed
position. As to the causes of the deformity two will be considered:
Congenital causes include:
1. Injury to the sternomastoid muscle at birth, which is perhaps the
commonest.
2. Abnormal intra-uterine position and pressure.
3. Arrest of muscular development.
4. Intra-uterine myositis, the muscles being sometimes found
actually altered in structure.
5. Defective development of the upper vertebrae or such distorted
growth as is often met along with other deformities, e. g., club-foot.
The acquired causes include:
1. Traumatisms, either direct, as by injury to the muscles, such as
may happen from gunshot wounds, etc., or follow operations by
which the spinal accessory has been injured, or by burns, and other
lesions which cause much cicatricial contraction.
2. Reflex activity in connection with disease of the lymph nodes,
deep cervical abscesses, parotid phlegmons or tumors, etc.
Whitman states that tuberculous disease of the cervical nodes
caused the condition in 50 per cent. of over 100 cases analyzed by
him.
3. Reflexes from the eyes, as Bradford and Lovett have described
from the orthopedist’s standpoint, and Gould from that of the oculist,
refractive errors causing the head to be held in unnatural positions in
order to improve vision.
4. Compensation in high degrees of rotary lateral curvature, the
effort being to keep the head facing to the front.
5. Myositis, usually rheumatic, but sometimes a sequel of the
infectious fevers, or even of gonorrhea.
6. Habitual deformity, the result of occupation or sheer bad habit.
7. Tonic or intermittent spasm leading to spastic contractures
whose causes are difficult to seek, but appear to inhere in the central
nervous system.
8. Paralyses of certain muscles, permitting lack of opposition and
consequent deformity.
Pathology.—According to circumstances significant pathological
changes may be found in the affected muscles. These
are usually the sternomastoid and the trapezius, although in long-
standing or complicated cases the deeper muscles of the neck may
also participate. A long contracted muscle may change almost into
mere fibrous tissue.
The secondary effects of contraction of the sternomastoid and the
trapezius are really far-reaching and noteworthy. The jaw may be
drawn down and to one side, so that teeth do not appose each other
as they should, or perhaps even do not meet. Compensatory
curvatures occur also in the spine and there is well-marked change
in gait and in most of the body habits. In the young and rapidly
growing cranial and facial asymmetry also become pronounced. The
later results and deformities of torticollis are not to be mistaken for
congenital elevation of the scapula, sometimes known as
“Sprengel’s deformity,” which consists not merely in elevation, but in
rotation of the shoulder-blade so that its lower angle is too near the
spine. There may be some limitation of motion of the scapula and of
the arm. Sprengel accounted for this abnormality by maintenance of
the intra-uterine position of the arm behind the back. The acute
forms of torticollis occur nearly always in acute phlegmons of one
side of the neck, and should subside with the other and causative
lesions. Nevertheless from such spasm may develop a chronic form
which may persist.
The position of the head varies with the muscles particularly
involved and the associated spasm. The sternomastoid muscle
alone will draw the mastoid down toward the sternum, with rotation
of the face to the other side. When the trapezius is involved the head
is drawn backward and the chin raised. The more the platysma,
scaleni, splenii, and deep rotators are involved the more complex
becomes the condition, to such an extent even that in serious cases
it is almost impossible to decide which muscles really are at fault.
When the superficial muscles are involved they can usually be
distinctly felt to be firm and contracted, while the sternomastoid will
stand out like a cord. Pain is a rare complaint, but a feeling of
tenderness or soreness is not unusual.
The spasmodic or intermittent form is less common, but more
difficult to account for and even to treat. It seems to be due to
choreiform spasm of those muscles which produce it, and here the
condition is reflex, the causes lying deeply in the nervous system. In
some instances, however, they are of ocular origin and can be
relieved by correcting refractive errors. Intermittent spasm is usually
absent during sleep and quiescent in the recumbent position; it is
usually confined to one side.
Diagnosis.—In the matter of diagnosis it is necessary mainly to
eliminate only spinal caries, while as between
involvement of the anterior and posterior groups of muscles the
determination is made by palpation and inspection.
Treatment.—There are few morbid conditions whose cause it is
more necessary to discover. Could this be done
operative treatment would be less often demanded. Treatment
should depend, therefore, on the exciting cause and the possibility of
its removal. The spasmodic or intermittent form may spontaneously
subside. Cases of essentially ocular origin need the services of the
oculist, and other acute cases usually subside with the successful
treatment or the subsidence of their causes. On the other hand,
chronic cases usually need either mechanical or operative treatment.
The most common operation for relief of torticollis is simple
tenotomy of the sternomastoid, taking care to divide the sheath and
everything which resists, and, at the same time, to avoid the external
jugular vein as well as the deeper structures. Mere tenotomy of one
or both of its lower tendons is an exceedingly simple measure, but in
serious cases an open division will permit of more thorough work.
Here an incision made one inch above the clavicle and parallel to it
will permit division of everything which resists and also any
recognition of that which should be spared. In any event the position
of the head should be immediately rectified, and kept so either by
plaster or starch bandage, or by a traction apparatus applied to the
head, the body being in the recumbent position, while later some
efficient and well-fitting brace should be worn for some time. The
posterior cases, i. e., those where the posterior muscles are
involved, afford greater operative difficulty, muscles involved lying
too deeply and being in too close relation with important vessels and
nerves to justify the ordinary wide-open division. Nevertheless in
extreme cases there need be no hesitation in extirpating completely
those muscles which are primarily and mainly at fault. The writer has
removed the sternomastoid and the trapezius, with sections of the
still deeper muscles, and has seen nothing but benefit follow the
procedure. It should be resorted to when repeated anesthesia with
forcible stretching and a suitable brace fail to give relief. These forms
of wryneck which are due to contraction of muscles infiltrated from
the presence of neighboring phlegmons, etc., will usually subside
with massage and semiforcible stretching under an anesthetic. They
need conservative rather than operative treatment. Attack upon the
spinal accessory and the deep cervical nerves will be described in
the chapter on Surgery of the Nerves. It, however, will rarely be
justified, since the primary causes inhere not so much in those nerve
trunks as in the nerve centres. Such operations are usually of
questionable benefit, and cases should be carefully watched before
being submitted to them.