Sei sulla pagina 1di 39

1

These websites contain information for mediation analysis as well as downloadable


software to conduct mediation analysis.
David Kenny
http://davidakenny.net/kenny.htm
This site has general information on mediation.
David MacKinnon
http://www.public.asu.edu/~davidpm/ripl/
This site has mediation information and computer programs.
Kristopher Preacher
http://www.psych.ku.edu/preacher/
This site has computer programs.
Comments and Short Answers to Even Numbered Questions for each Chapter
Chapter 1.
Comments: Chapter 1 is dense but should provide an introduction to mediation relations
and how they differ from other relations like moderation and confounding. A purpose of
the chapter is to provide a variety of examples of the mediation model so that the reader
appreciates mediation for more than just the role of mediation in their particular field. A
good student activity would be to take one of the examples described in this chapter and
learn more about it by reading some of the research cited in the chapter.
1.2. If the disease was transmitted by person to person contact, the mediators targeted
would be variables that would reduce contact among persons. Some possible mediators
are quarantining of infected individuals in hospitals, reducing all physical contact among
persons, and screening of persons for disease before they are contagious.
1.4. The four-way interaction is an example of a relation among four variables, whereby a
three-way interaction differs across levels of the fourth variable. A mediation relation
may change across levels of a fourth variable such as a moderator. A moderator effect
may be mediated such that the mediator explains the relation of the moderator to the
outcome.
1.6. There are thousands of articles with the words moderator or mediator in the title.
1.8. a. Simon (1954) describes a negative correlation between marital status and candy
consumption, a negative correlation between candy consumption and age, and a positive
correlation between martial status and age. When age is held constant, there is a nearly
zero correlation between marital status and candy consumption because the correlation
between marital status (independent variable) and candy consumption (dependent
variable) is spurious, because both variables are related to age (confounder). It does not
make sense that age could be a mediator, i.e., that marital status causes age.
b. Simon (1954) reports a positive correlation between percentage of employees who are
married and absenteeism. When housework is held constant, the correlation was nearly
zero. Amount of housework is a mediator such that marriage (independent variable) leads
to housework and more housework leads to more absenteeism (dependent variable).
Simon notes that common sense was used to conclude that the third variable was a
confounder in 1.8a and a mediator in part 1.8b, even though similar statistical information
was obtained.

2
1.10. More intelligent workers tend to get bored easily and produce less in the assembly
line so boredom may be considered a mediator. It is also likely that intelligent workers
will, in general, be more productive so there is both an increase in productivity and a
decrease in productivity through boredom--an inconsistent mediation model. Adding
boredom to the equation will increase the magnitude of the relation between intelligence
and productivity, a suppression relation.
1.12. Note that the test of statistical significance of the overall relation may be
underpowered so that there may actually be a significant relation between the
independent and the dependent variable in the population. Example 1.10 provides a
possible example where there is a zero relation between two variables that is explained
when a mediator is included in the analysis. There are many possible examples of
opposing mediation relations that have an overall nearly zero relation. For example, as a
sports team gets more successful, the success increases confidence which leads to better
performance but the increased performance leads opposing teams trying harder to win
which reduces subsequent performance. Incarceration of criminals may reduce
recidivism to crime through rehabilitative activities but may also increase recidivism by
putting persons in an environment where the norm is to commit crimes and inmates may
learn new criminal activities. Similarly, interventions that select high risk adolescents for
special programs may have beneficial effects but may also introduce adolescents to a less
healthy social norm. A program to promote safe sex behaviors may have beneficial
effects on risky sexual behavior but may also increase interest in sexual activity. Other
examples are described in Chapter 5.
Chapter 2.
Comments: A goal of chapter 2 is to provide a wide range of examples of mediation
relations. Another goal is to provide background for mediation in prevention and
treatment studies. A good student activity would be to summarize mediation relations in
the students area of interest. There is another interesting example of mediation in social
psychology that I almost used in this chapter. Isen and Levin (Isen, A. M. & Levin, P. F.
(1972). Effect of feeling good on helping: Cookies and kindness. Journal of Personality
and Social Psychology, 21, 384-388) describe two studies that suggest that an
intervention (X) that makes people feel good (M) leads to more helping behavior (Y).
The intervention used to make people feel good was giving them cookies or having
subjects find a dime in the coin return of the public telephone.
2.2. Action theory refers to the actions taken to change mediators in the X to M part of
the mediation model. Action theory is most relevant when X codes an intervention or
manipulation because typically actions are taken to change mediators. Action theory
when X is not an intervention (or does not represent assignment to a group) is not as clear
as when X is an intervention. In this case X does not represent manipulation or actions to
change M. In this case the action theory refers to how the X variable may change the
mediator. Conceptual theory refers to the theory for how the mediators are related to the
dependent variable. This is typically how theory is conceived in terms of how existing
variables are related to the dependent variable.

3
2.4. In drug prevention, interventions to change norms are delivered which are expected
to change norms to be less tolerant of drugs and the norm change leads to reduced drug
use.
2.6. In a program to reduce dating violence, Foshee et al. (1998) targeted changing
norms, decreasing gender stereotyping and improving conflict management.
2.8. As described in section 2.10.3, there are many ways that mediators for drug
prevention programs may be selected including a review of prior mediation analyses,
theories of drug use such as social learning theory, and empirical studies of variables that
correlate with drug use.
2.10a. Seedlings are most likely a surrogate because presumably more seedlings leads to
more potato yield because there are more plants. And it is likely that the number of
seedlings completely mediated fertilizer effects on potato yield. On the other hand, the
number of seedlings may be considered a measure of how much fertilizer was delivered
and more fertilizer would then lead to more potatoesso in this case seedlings may be a
mediator.
b. Norms is most likely a mediator because the relation of norms and assault is not as
direct as for most surrogates. It would be difficult to defend norms as a surrogate such
that the intervention effect on norms is the same as the intervention effect on assault. It is
also unlikely that norms completely mediate the relation of socioeconomic status on
assault. There are many other predictors of assault.
c. Carotene is likely a mediator as the amount of alfalfa eaten by cows increases carotene
consumed which increases Vitamin A in butter produced from the cows milk. On the
other hand, there may be a large relation between carotene and Vitamin A in butter so that
a measure of carotene consumed could be used in place of amount of Vitamin A in butter
made from the cows milk. More evidence for a surrogate is that the relation between
alfalfa consumed and Vitamin A in milk is likely to be entirely through carotene in cows
milk.
d. Fighting may be more accurately considered as a mediator rather than a surrogate. It is
a surrogate in that fighting is probably closely related to adult incarceration. On the other
hand, many more persons in fights in sixth grade are not incarcerated as an adult. It is
also likely that there are many influences of an intervention on adult incarceration besides
fighting in sixth grade. Persons incarcerated for non violent crimes as well as violent
crimes that may be related to fighting in the sixth grade. This example illustrates how
surrogates tend to be strongly related to the outcome variable and conceptually closer to
the dependent variable.
Chapter 3
Comments: A useful exercise is to prove that ab = c - c, either algebraically based on c =
c + ab or based on writing the formula for coefficients (i.e., c) as in MacKinnon, Warsi, &
Dwyer (1995, page 45-46) and shown below.
(1)
c = Cov(Y,X)/Var(X)
: from Equation 3.1
(2)
Cov(Y,X) = c' Var(X) + b Cov(M,X) : taking covariance of Y
(from Equation 3.2) and X
(3)
c = b Cov(M,X)/Var(X) + c'
: substitute (2) into (1)
(4)
c = ab + c'
: a = Cov (M,X)/Var(X)
from Equation 3.3

4
(5)

c - c' = ab

Plots of the mediated related relations are described in more detail in Fritz & MacKinnon
(2007), A graphical representation of the mediated effect. Behavior Research Methods.
Sample SAS, SPSS, and R programs to make the plots described in the article are given at
http://www.public.asu.edu/~davidpm/ripl/mediate.htm. Note that the plots are difficult for
many to understand especially Figure 3.5 because it includes so much information. An entire
class period may be needed to clearly describe the plot. Describing the plots in Figures 3.3
and 3.4 first provides a simpler introduction to the information in more complicated plots that
show the mediated effect.
Note that the water consumption example was created during a hot summer day in
Phoenix, Arizona.
There is a lot of information in this chapter. There will likely be a tendency to get bogged
down in the assumptions described in this chapter. Keep in mind that models and data to
address assumptions are covered in later chapters. For example, time ordered relations from
X to M to Y are a critical assumption of the mediation model and longitudinal data can
provide insight into these longitudinal relations as described in chapter 8. But understanding
material in chapter 8 requires understanding material in earlier chapters. The purpose of this
chapter is not to advocate cross-sectional studies but understanding this material is required
before applying more complicated models.
3.2.a. a b c c' .1696
b. Equation 3.6/3.7 = .02467; Equation 3.8 = .02490; Equation 3.9/3.10 = .02473;
Equation 3.11/= 3.12 = .02461
c. LCL = .12124; UCL = .21795
d. Test 1, Assess XY: ta 5.6292, p .01 , there is a significant program effect on the
dependent variable.
Test 2, Assess XM: t a 11.0625, p .01 , there is a significant program effect on the
mediating variable.
Test 3, Assess M1Y: t b 8.7744, p .01 , there is a significant relation of the
mediating variable to the dependent variable adjusted for the independent variable.
Test 4, Assess XY: t c ' 2.8910, p .05 , there is a significant relation of the
independent variable to the dependent variable adjusted for the mediating variable.
Because this path is statistically significant there is not complete mediation. This
significant direct effect coefficient would not satisfy the causal step criteria outlined in
Judd and Kenny (1981b) but would satisfy criteria in Baron and Kenny (1986) because
the total effect is larger than the direct effect. Note that James, Mulaik, & Brett (2006)
confirmatory approach described in section 3.13 would lead to the conclusion that this is
a partial mediation model because the direct effect is statistically significant.
e. The results are consistent with a mediation process by which the program changed
peers as an information source for nutrition behavior which in turn changed nutrition
behavior.
Here is SAS code used to compute quantities for question 3.2.

5
data a;
input c sec cp secp b seb a sea mse ssx N;
ta=a/sea;tb=b/seb;tc=c/sec;tcp=cp/secp;
varab=a*a*seb*seb+b*b*sea*sea;
se36=sqrt(varab);
se37=(a*b*(sqrt(ta*ta+ tb*tb)))/(ta*tb);
se38=sqrt(sec*sec+secp*secp-2*(mse/(N*ssx)));
se39=sqrt(varab+sea*sea*seb*seb);
se310=(a*b*(sqrt(ta*ta+tb*tb+1)))/(ta*tb);
se311=sqrt(varab-sea*sea*seb*seb);
se312=(a*b*(sqrt(ta*ta+tb*tb-1)))/(ta*tb);
ccp=c-cp;ab=a*b;
*Normal theory confidence limits with first order standard error;
ucl=ab+1.96*se36; lcl=ab-1.96*se36;
*Causal Steps are the same as the t-tests for regression estimates, c a, b, and cp;
cards;
.3552 .0631 .1856 .0642 .1711 .0195 .9912 .0896 1.1289 .2459 1227
;
proc print; var ccp ab;
proc print;var se36 se37 se38 se39 se310 se311 se312;
proc print;var ab se36 lcl ucl;
proc print; var tc ta tb tcp ccp;
run;
3.4. The water consumption data are simulated data so it is a unique situation where the
population model is known. In a real study, population values are not known.
Assumptions are briefly addressed here for this example. For mediation analysis of real
data, consideration of assumptions can be very complicated including investigation of
assumptions that can be tested as well as some assumptions that can only be addressed
theoretically or in a program of research. Statistical assumptions of the mediation
regression analysis include: (1) Correct functional form. Linear relations are assumed.
Figures 3.3, 3.4, and 3.5 suggest that linear relations are appropriate. The usual methods
of investigating nonlinear relations could be applied here including plots of residuals as a
function of independent variables. A test of the XM interaction was nonsignificant (as
addressed in Chapter 10) so it appears that this interaction is not substantial in the study.
If the interaction was statistically significant then the size of the mediated effect is likely
to depend on the value of X and contrast methods described in Chapter 10 would be
applicable. (2) No omitted influences. Additional variables could be incorporated in the
analysis such as an additional mediator or covariate in the analysis. It may also be
sensible to consider how large a correlation an omitted variable would have to be to
substantially reduce the mediated effect. (3) Accurate Measurement. The measures are
valid and reliable. Measures like temperature and water consumed are likely to be
accurately measured. Self-reported thirst may be more problematic but this could be
improved with a latent variable model described in Chapter 7. No measures of reliability
or validity are given for these examples but psychometric studies are important for a real
study to obtain test-retest correlations and other measures of validity and reliability. To

6
some extent, the measures must be reasonably accurate to obtain a significant mediated
effect because error in the mediator tends to reduce power to detect a mediated effect. (4)
Well-behaved residuals. Plots of residuals in the equations ought to resemble a normal
distribution as the residuals for the water consumption example.
Inferential assumptions are often more complicated than the assumptions of the
regression analysis because they require information from other studies and in some cases
there is not sufficient information to address the assumption. There is much overlap
between these assumptions and the statistical assumptions described above. (5) Temporal
Precedence. The random assignment of participants to temperature ensures that this
variable comes before self-reported thirst and water consumed. The timing of the
measures in the study was also such that temperature, thirst, and water consumed were
measured in a temporal sequence. (6) Micro Versus Macro Mediation Chain. It is possible
to provide more measures of the entire mediation chain including measures of blood
volume, actual work during the session, etc. that would provide a more detailed
micromediational chain. (7) Measurement timing. It is assumed that the actual timing of
measures was adequate to capture the mediated effect. It is important in other studies to
consider if the mediated effect would be present with other times of data collection. (8)
Normally distributed X, M, and Y. If X is binary then the methods described in this
chapter are accurate. Measures of skewness and kurtosis indicate a normal distribution.
Plots of residuals and actual distributions suggest that the variables are normally
distributed. In chapter 12, resampling methods that do not make as many statistical
assumptions led to the same research conclusions as described in this chapter. (9)
Normally distributed product of coefficients. Confidence limits based on the distribution
of the product and resampling methods led to the same conclusions as the analysis. The
confidence limits from the distribution of the product and resampling methods are
asymmetric consistent with the often nonnormal distribution of the product of regression
coefficients. (10) Omitted influences. This issue is the same statistical assumption as
discussed for the regression analysis but here the focus is on the extent to which inference
regarding the mediated effect is accurate and refers to substantive aspects of this
assumption. If other data were available it may be worth investigating whether the
mediated effect differs by subgroups, such as the fitness of participants in the study (fit
versus normal moderator for this example is described in Chapter 10). There may also be
other, more important, mediators of this relation. For the water consumption example,
perhaps actual blood volume is a better mediating measure. (11). Causal inference.
Because temperature is randomized, the relation of temperature to self-reported thirst and
the relation of temperature to water consumed may be considered a causal effect (with
some assumptions discussed in Chapter 13). The relation self-reported thirst to water
consumption is a reasonable assumption as a causal relation based on other research. The
size of the relation of temperature to water consumption adjusted for self-reported thirst
may be more difficult to defend as a causal effect. (12). Theoretical versus Empirical
Mediator. One benefit of the water consumption study is that future studies can be
designed based on these results to replicate and extend the mediated effect such that
temperature affects self-reported thirst and this affects water consumed. These future
studies help provide more evidence for a theoretical mediator.

7
3.6.
Figure 3.3 Relation of X and Y
5

Y=3.395

c=0.645

Y 3
Y=2.75

1
0.0

0.2

0.4

0.6
X

0.8

1.0

Figure 3.4 Relation of X and M


5

Y=3.21
M

a=0.627

Y=2.58

1
0.0

0.2

0.4

0.6
X

0.8

1.0

Y=i3+c'X+bM (Equation 3 for X=1)

c'
ab=c-c'

Chapter 4
Comments: There is a lot of information in this chapter. There are many unsolved issues
discussed in the chapter. For example, the ideal effect size measure for mediation models
is not yet determined. This chapter may be of most interest to more technically inclined
readers. The Monte Carlo simulation program can be used for a wide variety of
simulation studies. Note that the original version of this program was written around
1987 and parts of if have been used in many simulations studies both by me and my
collaborators but also by other researchers. The program can be adapted for longitudinal
data, categorical data analysis, and missing data, for example. A simulation program for
covariance structure models is described in MacKinnon, D. P. (1992). Statistical
simulations in CALIS. In Sugi 17. Proceedings of the Seventeenth Annual SAS Users
Group International Conference 1992, 1199-1203. Note that simulation studies can also
be conducted in SPSS and major covariance structure analysis programs.
4.2. Taking absolute values of the coefficients before computing the proportion mediated
measure changes the meaning of the proportion measure. Now all relations in the model
are positive so the total effect is the sum of the absolute value of all of the effects in the
model. The new measure is the proportion of the total relations between X and Y that is
attributable to the mediator. One of the reasons that absolute values are taken is to ensure
that the proportion will not be less than zero or greater than 1. The measure is subject to
sampling variability as the regular proportion measure so it requires large sample sizes or
effect sizes for it to be a stable measure.

10

4.4. The population correlation between X and Y for a equal to .5, b equal to .1 and error
variances equal to 1, rXY = Cov[X,Y]/(Var[X] Var[Y] ), is equal to .04969. The required
sample size to have .8 power to detect a correlation this size is approximately 3172
participants. The number of required participants is very large because the relation
between X and Y (first step in the Baron and Kenny causal steps test) is equal to the
product of the a and b relations which can be very small.
4.6. For each case, all possible pairs of elements in the formula for the variable are
written out, values that are 0 by definition are deleted, and terms are then combined and
simplified to give the following variances of X, M and Y.
Cov [X, X] = Var [X] = Cov (X, X ) = 2X
Cov [M, M]= Var [M]= Cov (aX + e3, aX + e3) = a 2 2X + 2e3
Cov [Y, Y] = Var [Y] = Cov(bM + c'X + e2, bM + c'X + e2)
= b 2 (a2 2X + 2e3) + 2bc'a 2X + c'2 2X + 2e2
4.8. The product of standardized regression coefficients for a and b for the analysis of
standardized data equal to (rXM(rMY-rXYrXM)/(1-rXM2) with partial derivatives of (rXM2rMY +
rMY -2 rXMrXY)/((1-rXM2)2), -rXM2/(1-rXM2), and rXM/(1-rXM2), for rXM, rXY, and rMY,
respectively. The covariance matrix among the correlations is pre and post multiplied by
the vector of partial derivatives to obtain the multivariate standard error of the product of
a and b for standardized data as described in Bobko and Reick (1980). The covariance
matrix among the correlations is given in Olkin & Siotani (1976) and a SAS program to
compute the standard errors of functions of correlation coefficients adapted from
Appendix B of MacKinnon et al., (2002) is shown below. For the water consumption data
the t-test for the product of standardized regression coefficients is equal to 2.246. The ttest for the raw minus partial correlation is also included in the program and is equal to
1.997 for the water consumption data.
data a; input rxm rxy rmy nobs;
/*revised program from Appendix B, MacKinnon et al., 2002 Psychological Methods, 7,
83-104;
*note r corresponds to correlation;
*x, m, and y represent the independent, mediating, and dependent variables, respectively;
*variance of the correlations from Olkin & Siotani 1976 and Olkin & Finn (1995);*/
vrxy=((1-rxy*rxy)*(1-rxy*rxy))/nobs;
vrmy=((1-rmy*rmy)*(1-rmy*rmy))/nobs;
vrxm=((1-rxm*rxm)*(1-rxm*rxm))/nobs;
*covariances among the correlations from Olkin & Siotani 1976 and Olkin & Finn
(1995)t;
crxyrmy=(.5*(2*rxm-rxy*rmy)*(1-rxm*rxm-rxy*rxy-rmy*rmy)+rxm*rxm*rxm)/nobs;
crxyrxm=(.5*(2*rmy-rxy*rxm)*(1-rxy*rxy-rmy*rmy-rxm*rxm)+rmy*rmy*rmy)/nobs;

11
crmyrxm=(.5*(2*rxy-rmy*rxm)*(1-rxm*rxm-rxy*rxy-rmy*rmy)+rxy*rxy*rxy)/nobs;
*olkin and finn (1995);
*partial correlation or correlation with mediating variable removed;
rxyi=(rxy-rmy*rxm)/sqrt((1-rmy*rmy)*(1-rxm*rxm));
*difference between simple and partial correlations;
diff=rxy-rxyi;
*partial derivatives of the difference between simple and partial correlations;
opd1=1-(1/(sqrt(1-rmy*rmy)*sqrt(1-rxm*rxm)));
opd2=(rxm-rxy*rmy)/((sqrt(1-rxm*rxm))*(1-rmy*rmy)**(1.5));
opd3=(rmy-rxm*rxy)/((sqrt(1-rmy*rmy))*(1-rxm*rxm)**(1.5));
ovar=opd1*opd1*vrxy+opd2*opd1*crxyrmy+opd3*opd1*crxyrxm+opd1*opd2*crxyrm
y
+opd2*opd2*vrmy+opd2*opd3*crmyrxm+opd1*opd3*crxyrxm+opd2*opd3*crmyrxm
+opd3*opd3*vrxm;
ose=sqrt(ovar);
zolkin=diff/ose;
polkin=1-probnorm(zolkin);
*bobko & rieck (1980) product of standardized a and b;
bobko=rxm*(rmy-rxy*rxm)/(1-rxm**2);
*partial derivatives for product of standardized a and b;
bpd1=((rxm*rxm*rmy+rmy-2*rxm*rxy)/(1-rxm*rxm)**2);
bpd2=(-(rxm*rxm)/(1-rxm*rxm));
bpd3=(rxm/(1-rxm*rxm));
bobkovar=((bpd1**2)*vrxm)+((bpd2**2)*vrxy)+((bpd3**2)*vrmy)+
(2*bpd1*bpd2*crxyrxm)+
(2*bpd2*bpd3*crxyrmy)+(2*bpd1*bpd3*crmyrxm);
bobkose=sqrt(bobkovar);
zbobko=bobko/bobkose;
pbobko=1-probnorm(zbobko);
cards;
.371 .361 .490 50
;
proc print;
var diff ose zolkin polkin bobko bobkose zbobko pbobko;
run;
4.10. The derivation of the asymptotic covariance among parameter estimates described
in Bollen (1989, p.107-109) is applied to the mediation model described in the text to

12
, the estimates of a and b parameters,
show that the covariance between a
and b
equals zero.
.
and b
Independence of a

Model: Y = c X + bM + 1
M = aX + 2
X, 2 and 3 are pairwise independent.
Let 2X = c1, 21 = c2, 22 = c3 and normal distributions.
The likelihood function for the estimation of the parameters a, b and c ( = a, b, c) is:
FML = log () + tr (S-1()) - logS - (p + q) where
()=
c1
ac1
abc1 + cc1

2
2
ac1
a c1 + c3
a bc1 + acc1 + bc3

2
2 2
abc1+cc1 a bc1+acc1+bc3 b (a c1 + c3)+ 2abcc1+c2c1+ c2
= c1c2c3 and
Var (X) COV (X,M) COV (X,Y)
S = COV (M,X) Var (M) COV (M,Y)
COV (Y,X) COV (Y,M)
Var(Y)
Is defined as the matrix of sample variances and covariances (Note that this is different
from the text where these quantities (e.g., COV(X,M)) are population values.
tr (S-1 ()) = Sum of elements on the main diagonal of S-1 ()
= (VarX) (a2c1c2 + c2c1c3 + c2c3)
+ [COV (X,M)] (bcc1c3 ac1c2) [COV (Y,X)] cc1c3
+ [COV (M,X)] (bdc1c3 ac1c2)
+ (Var M) (b2c1c3 ac1c2) + [COV (M,Y)] (-bc1c3)
+ [COV (Y,X)] (-cc1c3)
+ [COV (M,Y)] (-bc1c3) + (Var Y) (c1c3)
log S and (p + q) are constants.
Taking partial derivatives of FML and setting them equal to zero we get the estimates of a,
b and c(c1, c2, c3 are assumed known). The inverse of the matrix of the second
and c
, b
of
derivatives of FML gives the variances and covariances of estimates a
a, b and c respectively. The only elements of FML that depend on a, b and c are the
elements of tr(S-1 ()). By inspection we can see that in tr(S-1 ()) there are no cross
2FML
terms like ab or ac. So when we take second partial derivatives we get
= 0 and
ab

13

2FML
= 0. This implies that the estimate of a is uncorrelated with the estimates of b
ac '
and c.
Chapter 5
Comments: One important point to emphasize in this chapter is that with multiple
mediators it is often unclear if the mediators should be treated as if they are all in a
vertical line between X and Y as in the two mediator model in this chapter or if there are
relations among the multiple mediators such as X to M1 to M2 to Y. In this way, students
will start to think of the complexity of models discussed in later chapters. The true model
for the relations among multiple mediators is complicated and will likely require
experimental studies to manipulate different mediators in order to understand how the
mediators are related. The single mediator model often has X representing a randomized
intervention so that it is only the M to Y relation that may be reversed (i.e., Y to M) or
correlational. With multiple mediators there are many possible patterns among the
mediators and Y, even when X represents a randomized study. Theory and prior research
are required to specify complex mediation models because the number of possible
relations among variables is very large. This following paper describes useful description
of methods for multiple mediator models: Brown, R. L. (1997). Assessing specific
mediational effects in complex theoretical models. Structural Equation Modeling, 4(2),
142-156.
5.2. MacKinnon et al., (2001) investigated the mediation relations in a program to
prevent anabolic steroid use among high school football players. There were three
dependent variables, intent to use anabolic steroids, nutrition behaviors, and strength
training self-efficacy. Twelve mediators were evaluated: Knowledge of the effects of
anabolic steroids, perceived coach tolerance of steroids, team as an information source,
peers as an information source, ability to turn down drug offers, perceived peer tolerance
of drug use, normative beliefs about anabolic steroid use, perceived severity of anabolic
steroid use, perceived susceptibility to anabolic steroid use, belief in media
advertisements, reasons for using anabolic steroids, reasons for not using anabolic
steroids. The mediators were evaluated one mediator at a time and also together in a
multiple mediator model. The models were estimated using a covariance structure model
and the estimates of the mediated effects were tested for each mediator using methods
described in Chapter 5.
Here is a SAS program to compute the quantities in the text of Chapter 5 and exercise
5.1.
/* SAS program to compute quantities in Chapter 5 (text example and
exercise 5.1: MacKinnon, D. P. (2007) Introduction to Statistical
Mediation Analysis. Mahwah, NJ: Erlbaum.). Note that there are two sets of coefficients
for the
expectancy example; one set has four decimal places and the other set
of coefficients uses all of the decimal places in the SAS output. With

14
more decimal places c-c' is equal to ab to more than four decimal
places.
*/
Data a;
Input c sc cp scp b1 sb1 b2 sb2 #2 a1 sa1 a2 sa2 mse varx sb1b2 N;
*Compute t-values for coefficients;
ta1=a1/sa1;ta2=a2/sa2;tb1=b1/sb1;tb2=b2/sb2;tc=c/sc;tcp=cp/scp;
*Computes mediated effects and variances of coefficients;
a1b1=a1*b1; a2b2=a2*b2; ccp=c-cp;totab=a1b1+a2b2;
vara1=sa1*sa1;vara2=sa2*sa2;varb1=sb1*sb1; varb2=sb2*sb2;
*Equation 5.5;
vara1b1=vara1*b1*b1+varb1*a1*a1;vara2b2=vara2*b2*b2+varb2*a2*a2;
sa1b1=sqrt(vara1b1);sa2b2=sqrt(vara2b2);ta1b1=a1b1/sa1b1;ta2b2=a2b2/sa2b2;
*Equation 5.6;
sa1b1t=(a1b1*sqrt(ta1*ta1+tb1*tb1))/(ta1*tb1);
sa2b2t=(a2b2*sqrt(ta2*ta2+tb2*tb2))/(ta2*tb2);
*Equation 5.7;
Stot57=sqrt(vara1*b1*b1+varb1*a1*a1+vara2*b2*b2+varb2*a2*a2+2*a1*a2*sb1b2);
*Equation 5.8;
Stot58=sqrt(vara1b1+vara2b2+2*a1*a2*sb1b2);ttot=totab/stot58;
/*General formula for standard error of the total mediated that uses
covariances. Most of these covariances are zero with real data. For
the multiple regression analysis in chapter 5, only sb1b2 is
nonzero as in Equation 5.7 and 5.8;
Stotcov=sqrt(vara1b1+vara2b2+2*a1*a2*sb1b2+2*b1*b2*sa1a2+
2*b1*a1*sa1b1+2*b1*a2*sa1b2+2*a1*b2*sa2b1+2*a2*b2*sb2a2);
ttotcov=totab/stotcov;
*/
*Equation 5.9;
covccp=mse**2/(N*varx);rccp=covccp/(sc*scp);
Stot59c=sqrt(sc*sc+scp*scp-2*covccp);
Stot59r=sqrt(sc*sc+scp*scp-2*rccp*sc*scp); tdiff=(c-cp)/stot59c;
*Equation 5.10;
Stot510=sqrt(vara1*b1*b1+varb1*a1*a1+vara2*b2*b2+varb2*a2*a2-2*a1*a2*sb1b2);
*Equation 5.11;
Stot511=sqrt(vara1b1+vara2b2-2*a1*a2*sb1b2);tequal=(a1b1-a2b2)/stot511;
/*General formula for standard error of the difference between two
mediated effects. Most of these covariances are zero with real data.
For the multiple regression analysis in chapter 5, only sb1b2 is
nonzero as in Equations 5.10 and 5.11;
Stotcov=sqrt(vara1b1+vara2b2-2*a1*a2*sb1b2-2*b1*b2*sa1a2+
2*b1*a1*sa1b1-2*b1*a2*sa1b2-2*a1*b2*sa2b1+2*a2*b2*sb2a2);
ttotcov=totab/stotcov;
*/
*Confidence Limits;
lcla1b1=a1b1-1.96*sa1b1;ucla1b1=a1b1+1.96*sa1b1;

15
lcla2b2=a2b2-1.96*sa2b2;ucla2b2=a2b2+1.96*sa2b2;
*c sc cp scp b1 sb1 b2 sb2 a1 sa1 a2 sa2 #2 mse varx sb1b2 N;
Cards;
-.0014 .0603 .0044 .0635 -.4830 .0647 .3365 .0562
.3441 .0471 .0542 .0129 10.1913 125.5615 .004 300
.7078 .1734 .1122 .2073 .5690 .1568 .5297 .1696
.8401 .1580 .2219 .1460 8.3879 84.87 .0079 40
.707760 .17342970 .112152 .20731147 .569029 .15681205 .529720 .16963747
.840138 .15795758 .221903 .14601522 8.3879 84.87 .0079 40
;
Proc print;
Run;
Here is a SAS program to compute quantities for Exercise 5.3.
/* SAS program to compute quantities in Chapter 5 Exercise 5.3.,
MacKinnon, D. P. (2007) Introduction to Statistical
Mediation Analysis. Mahwah: NJ: Erlbaum.
*/
Data a;
Input c sc cp scp b1 sb1 b2 sb2
#2 b3 sb3 b4 sb4 #3 a1 sa1 a2 sa2 a3 sa3 a4 sa4
#4 mse sb1b2 N;
*Compute t-values for coefficients;
ta1=a1/sa1;ta2=a2/sa2;tb1=b1/sb1;tb2=b2/sb2;tc=c/sc;tcp=cp/scp;
ta3=a3/sa3;ta4=a4/sa4;tb3=b3/sb3;tb4=b4/sb4;
*Computes mediated effects and variances of coefficients;
a1b1=a1*b1; a2b2=a2*b2; a3b3=a3*b3; a4b4=a4*b4; ccp=ccp;totab=a1b1+a2b2+a3b3+a4b4;
vara1=sa1*sa1;vara2=sa2*sa2;vara3=sa3*sa3;vara4=sa4*sa4;
varb1=sb1*sb1; varb2=sb2*sb2;varb3=sb3*sb3; varb4=sb4*sb4;
*Equation 5.5;
vara1b1=vara1*b1*b1+varb1*a1*a1;vara2b2=vara2*b2*b2+varb2*a2*a2;
vara3b3=vara3*b3*b3+varb3*a3*a3;vara4b4=vara4*b4*b4+varb4*a4*a4;
sa1b1=sqrt(vara1b1);sa2b2=sqrt(vara2b2);ta1b1=a1b1/sa1b1;ta2b2=a2b2/sa2b2;
sa3b3=sqrt(vara3b3);sa4b4=sqrt(vara4b4);ta3b3=a3b3/sa3b3;ta4b4=a4b4/sa4b4;
*Equation 5.6;
sa1b1t=(a1b1*sqrt(ta1*ta1+tb1*tb1))/(ta1*tb1);
sa2b2t=(a2b2*sqrt(ta2*ta2+tb2*tb2))/(ta2*tb2);
sa3b3t=(a3b3*sqrt(ta3*ta3+tb3*tb3))/(ta3*tb3);
sa4b4t=(a4b4*sqrt(ta4*ta4+tb4*tb4))/(ta4*tb4);
*Equation 5.10;
da1b1a2b2=a1b1-a2b2;
Stot510=sqrt(vara1*b1*b1+varb1*a1*a1+vara2*b2*b2+varb2*a2*a2-2*a1*a2*sb1b2);
*Equation 5.11;

16
Stot511=sqrt(vara1b1+vara2b2-2*a1*a2*sb1b2);tequal=(a1b1-a2b2)/stot511;
*Confidence Limits;
lcla1b1=a1b1-1.96*sa1b1;ucla1b1=a1b1+1.96*sa1b1;
lcla2b2=a2b2-1.96*sa2b2;ucla2b2=a2b2+1.96*sa2b2;
lcla3b3=a3b3-1.96*sa3b3;ucla3b3=a3b3+1.96*sa3b3;
lcla4b4=a4b4-1.96*sa4b4;ucla4b4=a4b4+1.96*sa4b4;
* c sc cp scp b1 sb1 b2 sb2 #2
b3 sb3 b4 sb4 #3 a1 sa1 a2 sa2 a3 sa3 a4 sa4
#4 mse sb1b2 N;
Cards;
-.095112 .05284463 -.279107 .05449605 .281868 .05323348 .059703 .05495897
.335794 .05427165 .114041 .05350677
.237149 .04621743 .131117 .04478027 .242897 .04526981 .243409 .04589909
1.03824 .004 400
;
Proc print;
Run;
Chapter 6
Comments: The material in Chapter 6 is more complicated than the first five chapters.
Ideally the
reader of Chapter 6 has exposure to structural equation
models and
matrices. If the reader does not have this prior exposure, it
makes sense to
skip the sections of this chapter dealing with deriving
standard errors
and instead focus on the examples including the two factor
model and the
socioeconomic status. Note that describing these more
complicated
models is usually easier to understand if done in regression
form as the
examples in Mplus and EQS. The problem with the
regression
form of specifying these models is that the role of error
variances and
residuals can be difficult to describe. One benefit of this
chapter is that
often students have their own data or can use published
data sets like
the one in exercise 6.4. In this way, the chapter can be
made more
compelling by having the reader use their own data.
The following
to investigate
Taylor, A. B.,
three-path
6.2a. The

6.2b. The
LISREL

1 12 13 14x1
y1 0 0y1 1
21 2 23 24x2
y 2 21 0 y 2 2
31 32 3 34x3
y3 31 320y3 3
x 4

paper describes a simulation study of different approaches


three path mediation effects including Equation 6.23.
MacKinnon, D. P., & Tien, J.-Y. (in press). Tests of the
mediated effect. Organizational Research Methods.
equations in matrix form:

parameter estimates from LISREL are below:


Estimates (Maximum Likelihood)

17

BETA
GRADES EDUCEXPE OCCUASPI
-------- -------- -------GRADES - --EDUCEXPE 0.405150
(0.032273)
12.554022

--

--

OCCUASPI 0.157912 0.549593


(0.036799) (0.037601)
4.291224 14.616388

--

GAMMA
INTELLIG SIBLINGS FATHEDUC FATHOCCU
-------- -------- -------- -------GRADES 0.525902 -0.029942 0.118966 0.040603
(0.030646) (0.029630) (0.037600) (0.037135)
17.160610 -1.010527 3.163956 1.093375
EDUCEXPE 0.160270 -0.111779 0.172719 0.151852
(0.032147) (0.026414) (0.033716) (0.033108)
4.985583 -4.231814 5.122768 4.586508
OCCUASPI -0.039405 -0.018825 -0.041333 0.099577
(0.033907) (0.027737) (0.035593) (0.034836)
-1.162145 -0.678684 -1.161246 2.858434

PHI
INTELLIG SIBLINGS FATHEDUC FATHOCCU
-------- -------- -------- -------INTELLIG 1.000000
(0.051232)
19.519221
SIBLINGS -0.100000 1.000000
(0.036407) (0.051232)
-2.746735 19.519221

18
FATHEDUC 0.277000 -0.152000 1.000000
(0.037590) (0.036642) (0.051232)
7.368924 -4.148214 19.519221
FATHOCCU 0.250000 -0.108000 0.611000 1.000000
(0.037341) (0.036437) (0.042453) (0.051232)
6.695038 -2.964033 14.392382 19.519221
PSI
Note: This matrix is diagonal.
GRADES EDUCEXPE OCCUASPI
-------- -------- -------0.650995 0.516652 0.556617
(0.033351) (0.026469) (0.028516)
19.519221 19.519221 19.519221
6.2c. There are many possible indirect effects. Here are three of them:
Intelligence to grades to occupational aspirations 11 31 = .08297
2
2
2
2
2
Applying Equation 6.16: 2

= (.525402)2(.036799)2 +(.157912)2 (.0306462)+ 2 (.525402) (.157912) (.000000)


= .0199312
z = 4.16
Normal theory LCL=.0439, UCL = .1220 using Equations 3.4 and 3.5 and distribution of
the product LCL= .0446, UCL = .1228 using Equations 4.37 and 4.38 and PRODCLIN.

11 31

11

31

Intelligence to educational expectations to occupational aspirations 21 32 = .08808


Applying Equation 6.16:
2 = (.160270)2(.037601)2 +(.549593)2 (.0321472)+ 2 (.160270) (.549543) (.00000)
21 32
= .0186672
z = 4.72
Normal theory LCL=.0515, UCL = .1247 using Equations 3.4 and 3.5 and distribution of
the product LCL= .0525, UCL = .1257 using Equations 4.37 and 4.38 and PRODCLIN.
21

32

Applying Equation 6.16:


Number of siblings to educational expectations to occupational aspirations 12 31 =
-.061443

12 32

= (-.111779)2(.037601)2 +(.549593)2 (.0264142)+ 2 (-.111779) (.549593)

19
(.000000) = .0151132
z = -4.06
Normal theory LCL=-.0911, UCL = -.0318 using Equations 3.4 and 3.5 and distribution
of the product LCL= -.0918, UCL =-.0325 using Equations 4.37 and 4.38 and
PRODCLIN.
12

32

All three indirect effects tested were statistically significant. The indirect effects are
consistent with intelligence leading to improved grades which leads to greater
occupational aspirations and intelligence leads to greater educational expectations which
leads to greater occupational aspirations. As in the achievement example described in the
chapter, number of siblings is negatively associated with educational aspirations which
are, in turn, positively related to occupational aspirations.
The two indirect effects from intelligence to occupational aspirations, 11 31 and 21 32 ,
were tested for equality using Equation 6.18. Only the correlation (covariance) between
31 and 32 was nonzero. All other correlations involved in the formula were zero as
shown below in a section of the covariance and correlation matrix among parameter
estimates. Note, for example, the covariance, -.000573, between BE 3,1 and BE 3,2 is
used in the calculation of the standard error comparing two mediated effects.
Part of the Covariance Matrix of Parameter Estimates
BE 2,1 BE 3,1 BE 3,2 GA 1,1 GA 1,2 GA 1,3
-------- -------- -------- -------- -------- -------BE 2,1 0.001042
BE 3,1 0.000000 0.001354
BE 3,2 0.000000 -0.000573 0.001414
GA 1,1 0.000000 0.000000 0.000000 0.000939
GA 1,2 0.000000 0.000000 0.000000 0.000054 0.000878
GA 1,3 0.000000 0.000000 0.000000 -0.000179 0.000110 0.001414
GA 1,4 0.000000 0.000000 0.000000 -0.000120 0.000014 -0.000807
GA 2,1 -0.000548 0.000000 0.000000 0.000000 0.000000 0.000000
GA 2,2 0.000031 0.000000 0.000000 0.000000 0.000000 0.000000
Part of the Correlation Matrix of Parameter Estimates
BE 2,1 BE 3,1 BE 3,2 GA 1,1 GA 1,2 GA 1,3
-------- -------- -------- -------- -------- -------BE 2,1 1.000000
BE 3,1 0.000000 1.000000
BE 3,2 0.000000 -0.413983 1.000000
GA 1,1 0.000000 0.000000 0.000000 1.000000
GA 1,2 0.000000 0.000000 0.000000 0.059255 1.000000
GA 1,3 0.000000 0.000000 0.000000 -0.155187 0.098581 1.000000
GA 1,4 0.000000 0.000000 0.000000 -0.105199 0.012960 -0.578140

20
GA 2,1 -0.527960 0.000000 0.000000 0.000000 0.000000 0.000000
GA 2,2 0.036583 0.000000 0.000000 0.000000 0.000000 0.000000

11 31 21 32

= .018667 + .015113 - (-.11178)(.16027) (-.00057282) = .025482

The difference between the two mediated effects equals -.005116 with a standard error of
.02548 so it does not appear that the equality of the two indirect effects can be rejected
because the difference between the two mediated effects is about one fifth of the size of
its standard error. Of course there are many other potential tests of equal mediated effects.
Below are total indirect effects from the LISREL output. Note that many of the total
indirect effects are statistically significant. In general there are substantial indirect effects
of intelligence and fathers occupation on both educational experience and occupational
aspirations. Number of siblings and fathers occupation had significant indirect effects
on occupational aspirations but not educational expectations. And grades had a significant
indirect effect on occupational aspirations. In general background variables had
substantial relations with occupational aspirations. Kerckhoff (1974) and subsequent
researchers present interpretations of these results including some discussion of the
results that may or may not be consistent in other generations.
AMBITION AND ATTAINMENT KERCKHOFF
Indirect Effects of X on Y
INTELLIG SIBLINGS FATHEDUC FATHOCCU
-------- -------- -------- -------GRADES - ---EDUCEXPE 0.213069 -0.012131 0.048199 0.016450
(0.021029) (0.012043) (0.015710) (0.015102)
10.132186 -1.007269 3.068019 1.089252
OCCUASPI 0.288231 -0.072828 0.140201 0.098909
(0.026605) (0.018876) (0.024530) (0.023765)
10.833758 -3.858306 5.715393 4.161984
Indirect Effects of Y on Y
GRADES EDUCEXPE OCCUASPI
-------- -------- -------GRADES - --EDUCEXPE - -

--

--

21
OCCUASPI 0.222668
(0.023381)
9.523449

--

--

6.4. There are three indirect effects of number of siblings on respondents income:
through respondent education, 3-> 1 -> 3, 13 31 = -.0456 (se=.0090, t = -5.05),
respondent occupation, 3 -> 2 -> 3, 23 32 =-.0326 (se=.0085, t= -3.65), and through
respondent education to respondent occupation, 3 -> 1 -> 2 -> 3, 13 21 32 = -.0703
(se=.0073, t=-9.62). Equation 6.23 was used to compute the standard error of the threepath mediated effect and Equation 6.16 was used to compute the standard error of the
two-path mediated effects. All the indirect effects are much larger than their standard
errors with relatively small confidence intervals, at least in part because of the large
sample size. The results are consistent with a model whereby more siblings are associated
with less respondent income because of less education and a lower SES occupation and
also through the association that less education leads to a lower SES occupation. It is
important to keep in mind that these relations are descriptive and may not reflect a causal
relation among variables. The results do provide the basis for a future study, perhaps a
randomized experiment that may address the descriptive relations observed among these
variables.
The order of parameters in the columns of the V and V matrices in Sobel (1987, p. 169)
was 11, 12, 13, 21, 21, 22, 23, 31, 32, 31, 32, 33. The columns of the V and V
matrices in Sobel (1986, p. 173) was 11, 12, 13, 21, 22, 23, 31, 32, 33, 21, 31, 32. The
order of the parameters differs so the elements in the columns differ, making the V and
V matrices different in the two articles. Note also that the order of the elements in the
covariance matrix among the parameters must correspond to the different ordering of
parameters. The rows of the V and V matrices correspond to all the elements in the
and matrices, respectively. The matrices are different because the order of parameters is
different.
6.6. Sobel (1982; 1986) and Stone and Sobel (1990) investigated the total indirect
effects rather than specific indirect effects. Some specific indirect effects were
investigated because there was only one mediator between the independent and
dependent variable so the total indirect effect is the same as the specific indirect effect.
Bollen (1987) investigated specific indirect effects but not for the achievement model
discussed in this chapter. Stone and Sobel (1990) suggest sample sizes of 200 or more for
recursive models and 400 or more for nonrecursive models based on the two models
simulated in that paper. The achievement model discussed in this chapter was the
recursive model studied in Stone and Sobel (1990). One of the total indirect effects
studied in the Stone and Sobel article was the indirect effect of fathers occupation to
respondents education to respondents occupational status, 1221. In this case the total
indirect effect is the same as the specific indirect effect. Stone and Sobel conducted a
simulation study where the population values for the achievement model discussed in this
chapter were the estimates in earlier analyses of the model. A total of 500 replications at
sample sizes of 50, 100, 200, 400, and 800 were used to ascertain how close observed

22
estimates in the 500 replications for each sample size were to the population values. Table
1 in Stone and Sobel presents results on the relative bias of the indirect effect estimate
and standard error as a function of sample size. Relative bias is the difference between
the estimate and the population value divided by the population value. Relative bias was
one measure used to evaluate the statistical performance of the total indirect effect and
standard error estimators. For the indirect effect estimate of fathers occupation to
respondents education to respondents occupational status, the relative bias was always
less than .05 and was less than .01 at a sample size of 400 or greater. Similarly, the
standard error for this total indirect effect had less than or equal to .05 relative bias at
sample sizes of 100 or greater and relative bias less than .01 at a sample size of 200 or
greater.
Chapter 7
Comments: The three factor mediation model example described in the text had a
nonzero covariance between the GA(1,1) and BE(2,1) coefficients in the mediated effect.
That is, the a coefficient, GA(1,1),was correlated with the b coefficient, BE(2,1). As a
result this correlation (actually the covariance) was included in the standard error of the
mediated effect. The EQS program output in the CD accompanying this book has the
correlation among the estimates. For the LISREL program with this book, add pc to get
the covariance and correlation matrix among the parameter estimates and nd=6 to get
numbers to 6 decimal places in the output. The LISREL output line should be OU MI
RS EF MR SS SC pc nd=6. The section of the LISREL output with the covariance and
correlation among the parameter estimates matrices is shown below. In the output,
BE(2,1) corresponds to the b coefficient, GA(1,1) corresponds to the a coefficient and
' coefficient. The correlation between BE(2,1) and GA(2,1)
GA(2,1) corresponds to the c
is equal to .051518 and the covariance is equal .000231. Equation 6.16 in the text is given
below with the covariance term added to incorporate the covariance between the
coefficients where a corresponds to in the equation and b corresponds to in the
equation.
2 2 2 2 2 2

= (-.41458)2(.048115)2 +(.26629)2 (.0929932)+ 2 (-.41458) (.26629) (.051518) =


0009602
so the standard error equals the square root of .0009602 which is .030988. The normal
theory confidence limits are LCL = -.17114 and UCL = -.049665. Using the PRODCLIN
program and the following input, -.41458 .092993 .26629 .048115 .051518 .05 for the
following quantities:
a coefficient, standard error of the a coefficient, b coefficient, standard error of the b
coefficient, correlation between a and b , and .05 for 95% confidence limits. The
distribution of the product confidence limits are LCL = -.17114 and UCL = -.049665,
which differ slightly from the normal theory confidence limits. Note that in most
situations the a and b coefficients from multiple regression are uncorrelated.

11 21

Selected LISREL Output for covariance and correlations among parameter estimates

23
Covariance Matrix of Parameter Estimates
LY 2,1 LY 3,1 LY 5,2 LY 6,2 LX 2,1 LX 3,1
-------- -------- -------- -------- -------- -------LY 2,1 0.005905
LY 3,1 0.003709 0.006753
LY 5,2 0.000000 0.000000 0.004714
LY 6,2 0.000000 0.000000 0.003361 0.005108
LX 2,1 0.000000 0.000000 0.000000 0.000000 0.063004
LX 3,1 0.000000 0.000000 0.000000 0.000000 0.037509 0.045651
BE 2,1 0.000818 0.000895 -0.000699 -0.000624 0.000000 0.000000
GA 1,1 0.001265 0.001586 0.000000 0.000000 -0.011984 -0.010174
GA 2,1 0.000002 -0.000051 -0.000004 -0.000003 0.000040 0.000034
PH 1,1 0.000000 0.000000 0.000000 0.000000 -0.027605 -0.023540
PS 1,1 -0.005462 -0.006410 0.000000 0.000000 0.000032 0.000045
PS 2,2 -0.000001 0.000029 -0.003639 -0.003462 0.000000 0.000000
TE 1,1 0.001223 0.001920 0.000000 0.000000 0.000000 0.000000
TE 2,2 -0.001574 0.001230 0.000000 0.000000 0.000000 0.000000
TE 3,3 0.000434 -0.002896 0.000000 0.000000 0.000000 0.000000
TE 4,4 0.000000 0.000000 0.000595 0.000313 0.000000 0.000000
TE 5,5 0.000000 0.000000 -0.001252 0.000269 0.000000 0.000000
TE 6,6 0.000000 0.000000 0.000921 -0.000673 0.000000 0.000000
TD 1,1 0.000000 0.000000 0.000000 0.000000 0.006766 0.005853
TD 2,2 0.000000 0.000000 0.000000 0.000000 -0.014395 0.009638
TD 3,3 0.000000 0.000000 0.000000 0.000000 0.007895 -0.009042
Covariance Matrix of Parameter Estimates
BE 2,1 GA 1,1 GA 2,1 PH 1,1 PS 1,1 PS 2,2
-------- -------- -------- -------- -------- -------BE 2,1 0.002315
GA 1,1 0.000231 0.008648
GA 2,1 0.001054 -0.000264 0.004574
PH 1,1 0.000000 0.006509 -0.000022 0.015707
PS 1,1 -0.001267 -0.001211 -0.000093 0.000002 0.011585
PS 2,2 0.000420 0.000005 -0.000066 0.000000 -0.000002 0.005528
TE 1,1 0.000284 0.000468 -0.000007 0.000000 -0.001959 0.000004
TE 2,2 0.000003 0.000060 -0.000015 0.000000 -0.000139 0.000008
TE 3,3 -0.000026 -0.000474 0.000115 0.000000 0.001105 -0.000066
TE 4,4 -0.000083 0.000000 0.000000 0.000000 0.000000 -0.000402
TE 5,5 0.000132 0.000000 0.000001 0.000000 0.000000 0.000370
TE 6,6 -0.000075 0.000000 0.000000 0.000000 0.000000 -0.000210
TD 1,1 0.000000 -0.001561 0.000005 -0.003562 -0.000002 0.000000
TD 2,2 0.000000 0.000006 0.000000 0.000348 -0.000057 0.000000
TD 3,3 0.000000 0.000007 0.000000 0.000406 -0.000067 0.000000

24
Covariance Matrix of Parameter Estimates
TE 1,1 TE 2,2 TE 3,3 TE 4,4 TE 5,5 TE 6,6
-------- -------- -------- -------- -------- -------TE 1,1 0.006577
TE 2,2 0.000162 0.006538
TE 3,3 -0.001292 -0.002777 0.006101
TE 4,4 0.000000 0.000000 0.000000 0.001910
TE 5,5 0.000000 0.000000 0.000000 -0.000438 0.002181
TE 6,6 0.000000 0.000000 0.000000 0.000249 -0.001630
TD 1,1 0.000000 0.000000 0.000000 0.000000 0.000000
TD 2,2 0.000000 0.000000 0.000000 0.000000 0.000000
TD 3,3 0.000000 0.000000 0.000000 0.000000 0.000000

0.003327
0.000000
0.000000
0.000000

Covariance Matrix of Parameter Estimates


TD 1,1 TD 2,2 TD 3,3
-------- -------- -------TD 1,1 0.042451
TD 2,2 -0.000348 0.025794
TD 3,3 -0.000406 -0.014313 0.013286
Chapter 7 Three Factor Latent Variable Model
Correlation Matrix of Parameter Estimates
LY 2,1 LY 3,1 LY 5,2 LY 6,2 LX 2,1 LX 3,1
-------- -------- -------- -------- -------- -------LY 2,1 1.000000
LY 3,1 0.587342 1.000000
LY 5,2 0.000000 0.000000 1.000000
LY 6,2 0.000000 0.000000 0.684983 1.000000
LX 2,1 0.000000 0.000000 0.000000 0.000000 1.000000
LX 3,1 0.000000 0.000000 0.000000 0.000000 0.699387 1.000000
BE 2,1 0.221285 0.226413 -0.211465 -0.181495 -0.000010 -0.000016
GA 1,1 0.177061 0.207521 0.000000 0.000000 -0.513427 -0.512048
GA 2,1 0.000436 -0.009204 -0.000784 -0.000673 0.002361 0.002352
PH 1,1 0.000000 0.000000 0.000000 0.000000 -0.877494 -0.879082
PS 1,1 -0.660405 -0.724697 0.000000 0.000000 0.001171 0.001958
PS 2,2 -0.000227 0.004796 -0.712822 -0.651540 0.000000 0.000000
TE 1,1 0.196257 0.288075 0.000000 0.000000 0.000000 0.000000
TE 2,2 -0.253329 0.185110 0.000000 0.000000 0.000000 0.000000
TE 3,3 0.072273 -0.451195 0.000000 0.000000 0.000000 0.000000
TE 4,4 0.000000 0.000000 0.198372 0.100306 0.000000 0.000000
TE 5,5 0.000000 0.000000 -0.390617 0.080542 0.000000 0.000000
TE 6,6 0.000000 0.000000 0.232584 -0.163248 0.000000 0.000000

25
TD 1,1 0.000000 0.000000 0.000000 0.000000 0.130834 0.132963
TD 2,2 0.000000 0.000000 0.000000 0.000000 -0.357080 0.280850
TD 3,3 0.000000 0.000000 0.000000 0.000000 0.272865 -0.367152
Correlation Matrix of Parameter Estimates
BE 2,1 GA 1,1 GA 2,1 PH 1,1 PS 1,1 PS 2,2
-------- -------- -------- -------- -------- -------BE 2,1 1.000000
GA 1,1 0.051518 1.000000
GA 2,1 0.323864 -0.041930 1.000000
PH 1,1 -0.000001 0.558490 -0.002571 1.000000
PS 1,1 -0.244623 -0.121008 -0.012805 0.000121 1.000000
PS 2,2 0.117320 0.000749 -0.013093 0.000000 -0.000229
TE 1,1 0.072871 0.062061 -0.001231 0.000000 -0.224452
TE 2,2 0.000831 0.007917 -0.002654 0.000000 -0.015959
TE 3,3 -0.006843 -0.065222 0.021863 0.000000 0.131471
TE 4,4 -0.039493 0.000000 -0.000146 0.000000 0.000000
TE 5,5 0.058569 0.000000 0.000217 0.000000 0.000000
TE 6,6 -0.026914 0.000000 -0.000100 0.000000 0.000000
TD 1,1 0.000001 -0.081466 0.000375 -0.137937 -0.000073
TD 2,2 0.000028 0.000427 0.000006 0.017309 -0.003319
TD 3,3 0.000045 0.000694 0.000010 0.028108 -0.005390

1.000000
0.000641
0.001383
-0.011393
-0.123846
0.106634
-0.049001
0.000000
0.000000
0.000000

Correlation Matrix of Parameter Estimates


TE 1,1 TE 2,2 TE 3,3 TE 4,4 TE 5,5 TE 6,6
-------- -------- -------- -------- -------- -------TE 1,1 1.000000
TE 2,2 0.024751 1.000000
TE 3,3 -0.203898 -0.439695 1.000000
TE 4,4 0.000000 0.000000 0.000000 1.000000
TE 5,5 0.000000 0.000000 0.000000 -0.214756 1.000000
TE 6,6 0.000000 0.000000 0.000000 0.098686 -0.605027
TD 1,1 0.000000 0.000000 0.000000 0.000000 0.000000
TD 2,2 0.000000 0.000000 0.000000 0.000000 0.000000
TD 3,3 0.000000 0.000000 0.000000 0.000000 0.000000
Correlation Matrix of Parameter Estimates
TD 1,1 TD 2,2 TD 3,3
-------- -------- -------TD 1,1 1.000000
TD 2,2 -0.010529 1.000000
TD 3,3 -0.017098 -0.773138 1.000000

1.000000
0.000000
0.000000
0.000000

26
Here is a short SAS program to compute the standard error.
data a;
input a sea b seb rab;
ab=a*b;
sab=sqrt(a*a*seb*seb+b*b*sea*sea);
sabr=sqrt(a*a*seb*seb+b*b*sea*sea+2*a*b*(rab*sea*seb));
tsab=ab/sab;tsabr=ab/sabr;
covab=rab*sea*seb;sab=sqrt(a*a*seb*seb+b*b+sea*sea+2*a*b*(rab*sea*seb));+
lcl=ab-1.96*sabr;
ucl=ab+1.96*sabr;
cards;
-.414584 .092993 .266294 .048115 .0515
;
proc print;
run;
7.2 Matrices for the two mediator model with three indicators of X, M1, M2, and Y. x,
, and are the same as on page 179.
0
0
y11

0
0
y 21
y 31
0
0

0
0 y 42
0
y 0 y 52

0
0 y 62
0
0
y 73

0
y 83
0
0
0
y 93

11
0

0
0

0
0

0
0

0
22
0
0
0
0
0
0
0

0
0
0
0
0
0
33 0
0
0 44 0
0
0 55
0
0
0
0

0
0
0
0

0
0
0
0

0
0
0
0
0

0
0
0
0
0

0
0
0
0
0

0
0
0
0
0

66 0
0
0
0 77 0
0

0
0 88 0
0
0
0 99

27
0
0
31

0
0
32

0
0
0

0
0
11

21 22
0
0
0 33

2
3
7.4. More measures of the same latent variable will improve reliability of the measure, all
else being equal. The Spearman-Brown formula can be used to calculate reliability of a
scale if items are added or removed (see Allen & Yen, 1979; Crocker & Algina, 1986;
Lord & Novick, 1968).
7.6. Binary X, four indicators of the mediator and four indicators of the dependent
variable. There is the indirect effect of to 1 to 2. If X is a binary variable coding
exposure to a randomized intervention, then the mediated effect refers to the relation of X
to the latent dependent variable through the latent mediating variable. There is also the
indirect effect of 1 to 2 to Y, which would reflect the relation of the latent mediating
variable to the latent dependent variable to each of the indicators (Y variables) of the
latent dependent variable. The same matrices are needed for this situation as in the
answer to exercise 7.1 but with new x and matrices shown below to reflect that there
is one X variable. Typically in this model x11 will be fixed to 1 and 11 would be fixed to
0 to reflect that there is one measure of with perfect reliability. As described in section
7.12, an adjustment for measurement error may be obtained by fixing 11to 1 minus
reliability times the variance of X.
x x11
11
Chapter 8
Comment: It is difficult to give general guidelines for the estimation of longitudinal
mediation models because of different hypotheses about growth over time and the
specifics of research areas. The models in this chapter provide a starting point for models
relevant for research. Note also that the longitudinal mediation models are first described
in this chapter or have only recently been proposed. Future simulation studies and
application of the models to real data should clarify the best set of models and model
comparisons for your research. Note that the data sets used for illustration contain a
subset of the participants in the original study. An additional exercise for this chapter

28
would be to further analyze these data to find a good model. The data for these models
are real so the data have the usual idiosyncrasies and challenges of applying models to
real data. One of these idiosyncrasies is in the development of individual models for
media and nutrition for the LGC model, the second slope for nutrition behavior is very
small and may not be justified as a predictor. I kept the second slope in the model as it
codes the delivery of the program. Readers may want to consider alternative longitudinal
models for these data.
Note that a program to estimate a latent difference score model is included on the CD
accompanying this book (Chap8_LDS.mplus.inp).
A good additional reading for this chapter is Maxwell, S. E. & Cole, D. A. (2007). Bias in
cross-sectional analyses of longitudinal mediation. Psychological Methods, 12, 23-44.
This interesting article demonstrates the bias possible in cross-sectional mediation
analysis. The authors do this by comparing the cross sectional indirect effect to the
longitudinal indirect effect. One interesting task would be to apply the calculations in this
article to the examples in this chapter and your own longitudinal data. It is also helpful to
consider whether the longitudinal mediation relation described in the article is really the
mediation relation that researchers wish to investigate. It is also useful to consider the
points of this article for the case where X represents randomization to conditions. It
would also be interesting to think of situations where cross-sectional mediation may be
more important than longitudinal mediation.
8.4. The autoregressive Model II gives the same chi-square as the autoregressive model
described in the book because the contemporaneous mediated effects are equivalent to
allowing correlated residuals for the mediator and dependent variable at each wave for
this model. Adding the three additional parameters (paths from intentions to severity at
each measurement after baseline) in Model III for these data significantly improves the fit
of the model (103.257 - 83.694 = 19.563 with three degrees of freedom). However, the
RMSEA is .117. When allowing errors in adjacent measures of intent and severity after
the first wave, contemporaneous mediation, and correlated residuals at each wave, the
model fit improves dramatically 8.75 (p = .1139) with a difference of 74.944 on seven
degrees of freedom which is highly significant. The RMSEA is also acceptable, .042.
However, there are several ambiguities in the model including that some relations
between perceived severity and intention to use steroids becomes positive. Perhaps a
latent growth curve model or a latent difference score model would be more appropriate
for these data. It may also be helpful to construct a measurement model for intentions and
perceived severity that would include the four measures of each construct as indicators of
a latent construct.
8.6. The equations for the latent growth curve models would be the same as 8.21 to 8.29
with t or time equal to 5. The autoregressive equations would just include lag 1 relations
among variables as in the autoregressive model equations described in the chapter.
Chapter 9

29
Comments: This chapter only touches on the possibilities for multilevel mediation
models. These models have many applications. Multilevel mediation models have only
recently been applied so their usefulness is to be determined. The ability to investigate
mediational processes at different levels is an exciting aspect of these models. There is
plenty of research on methodology possible in this area as well. Effect size measures,
ways to choose multilevel versus other longitudinal models, and models at three or more
levels are just a few of the possibilities for future development. Note that Keenan Pituch
and colleagues have conducted simulation studies of mediated effect standard errors in
multilevel mediation models (Pituch, K. A., Stapleton, L. M., & Kang, J.-Y. (2006). A
comparison of single sample and bootstrap methods to assess mediation in cluster
randomized trials. Multivariate Behavioral Research, 41, 367-400).
9.2. Equations for random slopes are described in Equations 9.15, 9.16, 9.17, 9.18, and
9.19. Include two more Level 2 predictors, one for the intercept in the Level 1 Equation
9.15 for M and one for the Level 1 Equation 9.16 for Y to model random intercepts.
9.4. The Sampson, Raudenbush, and Earls (1997) article modeled for data from 343
neighborhoods in Chicago, Illinois. There are many possible answers to this question just
based on the description in Section 9.5. In the article, the first level of analysis was within
person and represented the responses of each person to several questions, the second level
of analysis was a model for each persons average response to the questions within
neighborhoods, and the third level had a model where the mean of each neighborhood
was predicted. The article examined the relation of social composition of the
neighborhood (concentrated disadvantage, immigrant concentration, and residential
instability) to the mediator of collective efficacy to violence (perceived violence, violent
victimization, and homicide). A difference in coefficients method was used to assess
mediation such that collective efficacy was added to the model that had concentrated
disadvantage, immigrant concentration, and residential instability as predictors of
violence. When collective efficacy was added to the model the coefficients for the three
social composition variables were reduced consistent with a mediation effect. It is not
exactly clear how the significance test was conducted although the Clogg, Petkova, and
Haritou (1995) article was cited and this method consists of testing whether the
coefficient relating the mediator to the dependent variable is statistically significant. It
would be helpful to test whether the product of the coefficient relating social composition
to collective efficacy and the coefficient relating collective efficacy to violence was
statistically significant using product of coefficient methods described in this book. The
test could be conducted at the neighborhood or level 3 of the multilevel model. More
cities could be added to form an additional level of analysis; for example if the same data
were obtained from 20 cities in the United States. Another level could be longitudinal
data from the neighborhoods so that there were four waves of data for example within
each neighborhood. There are many other levels that could be investigated. See the
following article for a fascinating review of neighborhood effects: Sampson, R. J.,
Morenoff, J. D., & Gannon-Rowley, T. (2002). Assessing Neighborhood effects: Social
processes and new directions in research. Annual Review of Sociology, 28, 443-478. See
page 467-8 for the importance of indirect effects in this research.

30
Chapter 10
Comments: Models with mediation and moderation are an active area of methodological
development and application. These models are important because these ideas most
clearly match many hypotheses in psychology and other fields. A recent paper by Edward
also describes moderation and mediation models (Edwards, J. R. & Lambert, L. S.
(2007). Methods for Integrating Moderation and Mediation: A General Analytical
Framework Using Moderated Path Analysis. Psychological Methods, 12, 1-22.) describes
a method similar to the one in this chapter.
10.2. An important part of investigating mediation and moderation is to clearly specify
effects to be tested prior to statistical analysis either based on theory or prior empirical
work. Tests of moderation are also conducted to investigate whether a mediation process
does not differ across groups to help provide evidence for generalizability of the results
across groups. It is often helpful to test each moderator and mediator relation separately
in order to understand the patterns in the data. Models with moderation and mediation are
complex and often require detailed thinking to understand them. Plots of relations are
also helpful. After these first tests, then models with both moderation and mediation are
estimated including the models in this chapter. Note that sometimes large sample sizes
are needed to detect effects with these models.
10.4. There are several options for the analysis of these data. Equations 10.15 and 10.16
were expanded to model two moderators and their interaction effects. Note that there are
only 25 participants in each group defined by fit and recorded groups making power to
detect effects low. There is an interaction effect on both the mediator and dependent
variable where the fitness group effect depends on whether the participant was recorded
or not as shown in the means below. For fit persons, more water was consumed if
recorded but less water was consumed if not recorded. Similarly, more self-reported thirst
was reported by fit persons if recorded but less thirst was reported by fit persons if not
recorded.
Means for M and Y water consumption study
Y
M
Not Recorded Normal
2.92 2.92
Not Recorded Fit
3.56 3.72
Recorded Normal
3.56 3.20
Recorded Fit
2.48 2.68
The results are consistent with more fit persons being more sensitive to their own thirst
and drinking more water in response. The recording factor did not interact with the
mediator or the fit moderator suggesting that being recorded did not alter the mediation
relations. However sample size was very small. Below are the values of the mediation
relations for each of the four groups defined by whether the experiment was recorded or
not and whether the participants were in the fit or normal groups. For the most part the
coefficients are comparable across groups with the exception of the a coefficient which
appears to differ from the other groups. One very tentative interpretation of these results
is that when fit persons are recorded some of them want to prove how fit they are by not

31
consuming as much water thereby reducing the relation between self-reported thirst and
water consumed.

' se
a se
c
c
se
se
b
Not Recorded Normal
.3527 .1910 .2387 .1945 .4456 .2563 .2558 .1490
Not Recorded Fit
.3507 .1910 .0905 .2250 .3534 .1824 .7361 .2063
Recorded Normal
.2808 .1934 .1041 .1886 .4397 .1772 .4
.2057
Recorded Fit
-.1066 .2783 -.0986 .3130 -.0144 .2374 .5492 .2499
Chapter 11
Comments: Many readers from psychology may not be familiar with methods for
categorical outcomes such as logistic or probit regression. A brief description of the odds
ratio and predicted logits and odds are described to provide some preliminary information
for readers unfamiliar with logistic regression. Survival analysis is another important
methodology for categorical variables. It is likely that the same issues described in this
chapter are also present in survival analysis. Here are some other recent papers on
mediation with categorical dependent variables. Huang, B., Sivaganesan, S., Succop, P.,
Goodman, E. (2004). Statistical assessment of mediational effects for logistic mediational
models Statistics in Medicine, 23, 2713-2728. and Li, Y., Schneider, J. A., & Bennett, D.
A. (2007). Estimation of the mediation effect with a binary mediator. Statistics In
Medicine, 26, 3398-3414.
11.2. c = -3.985 (se=.755), odds ratio =2.7360, LCL =1.8787 UCL = 3.9846
For each day of eating grilled meat per week, the person was 2.7360 times more likely to
get pancreatic cancer.
' = 1.050 (se = .229), odds ratio =2.8574, LCL = 1.8251,UCL = 4.4734
c
For each day of eating grilled meat per week, the person was 3.8574 times more likely to
get pancreatic cancer adjusted for the relation of blood fats to pancreatic cancer.
b = 1.737 (se = .276), odds ratio = 5.6820, LCL = 3.3112,UCL = 9.7501
For each one unit increase in blood fats a person is 5.6820 times more likely to get
pancreatic cancer, adjusted for days of eating grilled meat.
a = .2149 (se = .075), ,LCL = .0684, UCL = .3613
For each day of eating grilled meat per week, persons had an increase of .2149 units of
fat in their blood.
11.4. The mediated effect from logistic regression analysis of the data in 11.3 was equal
to .1487 with a standard error of .04977, z = 2.9877 (LCL =.0512 UCL = .2462). With
Mplus the mediated effect estimate was .088 with a standard error of .030 (z = 2.935) and
LCL = .030 and UCL = .146. The z values are very close but the estimate and confidence
limits differ because the Mplus coefficients are standardized. It is the standardization that
makes Mplus the program of choice for mediation analysis for models with categorical
and continuous measures. The asymmetric confidence intervals from the distribution of
the product were LCL = .0609 and UCL = .2554. Using Mplus to obtain the percentile
bootstrap asymmetric confidence limits discussed in Chapter 12 yields LCL = .036 and
UCL = .155 . Note that the model indirect statement is used to obtain the estimate of the
mediated effect and its standard error and the cinterval(symmetric) command generates
confidence limits. The !bootstrap = 1000 (the ! indicates a comment in Mplus so this

32
command was ignored in the analysis below) and cinterval(bootstrap) program
commands are necessary for the bootstrap analysis.
title: Exercise 11.3
Data:
File is "c:\Bookfinal\Chap11_ex11.3.txt";
Variable:
names are id x m y;
usevariables are x m y;
categorical y;
analysis:
type is general;
!bootstrap=1000;
estimator is wlsmv;
iterations=1000;
convergence=.00005;
model:
Y on x m;
m on x;
model indirect:
y ind x;
output: samp mod standardized cinterval(symmetric)
Selected Mplus output
MODEL RESULTS
Estimates
Y
X
M

ON

M
X

ON

S.E. Est./S.E.

Std

StdYX

0.085 0.072
0.261 0.066

1.169 0.085 0.099


3.926 0.261 0.300

0.337

0.073

4.618

0.337

0.344

Residual Variances
M
1.218

0.203

6.004

1.218

0.882

TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS


Estimates
Effects from X to Y

S.E. Est./S.E.

Std

StdYX

33
Total
0.173 0.069
2.488 0.173 0.202
Total indirect
0.088 0.030
2.969 0.088 0.103
Specific indirect
Y
M
X
Direct
Y
X

0.088

0.030

2.969

0.088

0.103

0.085

0.072

1.169

0.085

0.099

CONFIDENCE INTERVALS OF MODEL RESULTS


Lower .5% Lower 2.5% Estimates Upper 2.5% Upper .5%
Y
X
M

ON

M
X

ON

-0.102
0.090

-0.057
0.131

0.085
0.261

0.227
0.391

0.271
0.432

0.149

0.194

0.337

0.481

0.526

Residual Variances
M
0.695

0.820

1.218

1.615

1.740

TOTAL, TOTAL INDIRECT, SPECIFIC INDIRECT, AND DIRECT EFFECTS


Lower .5% Lower 2.5% Estimates Upper 2.5% Upper .5%
Effects from X to Y
Total
Total indirect

-0.006
0.012

0.037
0.030

0.173
0.088

0.309
0.146

0.352
0.164

Specific indirect
Y
M
X

0.012

0.030

0.088

0.146

0.164

34
Direct
Y
X

-0.102

-0.057

0.085

0.227

0.271

Chapter 12
Comments: Over the last decade or so, resampling methods have been incorporated in
many structural equation modeling programs including EQS and Mplus. In many of
these programs, conducting a resampling analysis is as simple as including a keyword
and the number of resamples. Resampling methods are a good general purpose approach
to confidence limit estimation for mediation effects in simple and complex mediation
models. It is important to note that the resampling methods yield more accurate
confidence limits but the interpretational difficulties of mediation relations remain
including temporal precedence, omitted variables, and measurement error. Note that a
few simple keywords will allow for bootstrap estimation of all the data in this book that
include individual data including the models with moderation and mediation, categorical
dependent variable models, and structural equation models. The effects of non-normality
on estimation of mediation is described in Finch, J., West, S., and MacKinnon, D. P.
(1997) Effects on non-normality on mediated effect estimates. Structural Equation
Modeling, 4, 87-107. Note also that there can be problems with the bootstrap if the
process generating the data is not the same as the way the bootstrap samples are taken.
The following article describes this issue for measures of fit in structural equation
modeling: See Bollen, K. A. and Stine, R. A. (1992). Bootstrapping goodness-of-fit
measures in structural equation models. Sociological Methods & Research, 21, 205-229.
12.2. Approximate tests are used rather than exact randomization tests because of the
sheer number of data sets that would need to be analyzed with an exact randomization
test. Instead, an approximate randomization test takes a random sample of all the possible
data sets.
12.4. It is not possible to conduct a typical resampling analysis for the achievement
model described in Chapter 6 because the raw data are not available. If the raw data were
available, a resampling method could be used by adding the bootstrap= 1000 command in
the Analysis section and Cinterval(Bootstrap) Cinterval(bcbootstrap) in the Output
section to get bootstrap and bias corrected bootstrap confidence intervals.
12.6. Fishers quote is important at several different levels. It is interesting that Fisher
used the permutation test as a check on his t-test suggesting that Fisher himself thought
the test the most accurate. In some respects, he is correct that a resampling test shows
ignorance about the underlying distribution and other detailed aspects of a variable and
area of research. However, as described in this chapter, there are instances where a
resampling method is the method of choice if an analytical solution for a statistic is not
known or the distribution of a statistic is known to be nonnormal, for example. Fishers
other fascinating comment is about trusting persons who are experts in a substantive field

35
rather than statisticians. I think this is probably true if the statistician is not acquainted
with a field of substantive research. The ideal situation occurs when the substantive
researcher is knowledgeable about statistical methods and the statistician is well
acquainted with a substantive area of research. If this is not possible, the statistician and
the substantive research must keep an open mind and be aware that statistical models are
only abstractions and science requires rigorous tests of substantive ideas.
Chapter 13
Comment: Causal inference in mediation models is an important area for future research,
especially to identify ways to test the assumptions of the models and sensitivity to
violation of assumptions. Small group activities may be very helpful in the classroom.
Often the persons most adamant about causal inference approaches and very critical of
mediation analysis are technical researchers who have not had much exposure to having
to make decisions based on actual data. This is sensible given the difficult mathematical
aspects of determining a cause. However, making a decision based on data entails taking
into account the practical strengths and limitations of a research study in a way to make
the best possible decision. Other researchers want to claim causal mediation relations
without considering the sometimes insurmountable assumptions as described in this
book. These two views suggest a series of small group activities that may help convey
these issues. A group activity in a course of mediation analysis could require small
groups to discuss two alternative approaches to causal inference. One approach argues
that studies without randomization should not be conducted and that mediation analysis
should not be attempted because of the difficulties associated with identifying true
mediators. Resources may be better spent on other research projects where randomization
is possible. Another point of argument could be that humans will never know true causes
so it is pointless to make the establishment of causal inference as the most important
aspect of research. Is causal inference needed to make sound decisions about research?
This latter approach appears to be the one favored by eminent statistician, Karl Pearson,
who considered it best to focus on the values of statistics and not focus on underlying
causes as humans were not especially capable of identifying causes. Here is a quote from
Pearsons famous book on the scientific method, Science in no case can demonstrate any
inherent necessity in a sequence, nor prove with absolute certainty that it must be
repeated (Pearson p. 113, Grammar of Science, 1957; originally published in 1892).
There are several outstanding articles that I was not able to include in the book. Here are
some them. They could be assigned as additional reading for this chapter.
Kaufman, J. S., MacLehose, R. F., & Kaufman, S. (2004). A further critique of the
analytical strategy of adjusting for covariates to identify biologic mediation.
Epidemiologic Perspectives & Innovations, 1, 4. Available on-line at http://www.epiperspectives.com/content/1/1/4
Judea Pearls papers are available at http://bayes.cs.ucla.edu/jp_home.html

36
Pearl, J. (June, 2001). Direct and Indirect Effects. UCLA Cognitive Systems Laboratory,
Technical Report (R-273), In Proceedings of the Seventeenth Conference on Uncertainty
in Artificial Intelligence, San Francisco, CA: Morgan Kaufmann, 411-420, 2001.
The articles by James Robins are available at the website,
http://biosun1.harvard.edu/~robins/
Robins J. M. (1989). The control of confounding by intermediate variables. Statistics in
Medicine, 8:679-701.
Robins, J. M., Hernn M., & Brumback B. (2000). Marginal structural models and causal
inference in epidemiology. Epidemiology, 11(5):550-560.
Robins, J.M. (2003). Semantics of causal DAG models and the identification of direct
and indirect effects. In: Highly Structured Stochastic Systems, Eds by P. Green, N. Hjort
and S. Richardson, Oxford University Press.
Steyer, R. (2005). Analyzing individual and average causal effects via structural equation
models. Methodology, 1, 39-54.
Ten Have, T., Elliott, M., Zanutto, E., & Datto, C. (2004). Causal models for randomized
encouragement trials in treating primary care depression. Journal of the American
Statistical Association, 99, 8-16.
Ten Have, T., Joffe, M., Lynch, K., Maisto, S., Brown, G., & Beck, A. (2007), Causal
mediation analyses with rank preserving models, Biometrics, 63, 926934.
The Kaufman et al. (2004) article is critical of the investigation of direct and indirect
effects based on causal inference for mediation models. It is one of the most accessible
articles describing the limitations of studies of direct and indirect effects. In general the
approach in this book is based on a program of research and distillation of information
from many sources while the focus of this criticism is the difficulty of making causal
inferences from a single study. Information on mediation relations is a goal of many
research projects so the idea is to obtain as much clear information as possible. In many
cases, avoiding a decision about data is not possible and a decision must be made.
Second, mediation relations are indeed unknown and potentially very complicated. They
force deeper statistical methods and deeper thinking. That is why the investigation of
mediation relations is so interesting.
13.2. Holland applied the Rubin causal model to the case of mediation using an
encouragement design. He made several important observations regarding the extent to
which the usual approaches to investigating mediation are limited. His first contribution
was a new notation for the causal structure relating the three variables in the single
mediator model based on counterfactual conditions. He showed how this notation was
different from the structural equation notation and he used this to demonstrate several
challenging aspects of causal inference from the single mediator model. In particular,
with X representing random assignment, the c and a coefficients reflect causal relations

37
but the b and c coefficients do not because c and a are relations estimated between a
variable randomized to participants and a dependent variable. With some assumptions,
including the exclusion restriction that all of the effect of X on Y is through M, Holland
demonstrated that b may be treated as a causal effect as well.
13.4. Satisfying the assumptions of instrumental variable methods is one of the
challenges of the analysis. It is ideal to obtain some information from the data set
regarding these assumptions and to conduct programs of research so that the
reasonableness of the assumptions can be more accurately addressed.
a. X is exogenous is ensured if X reflects random assignment. There are certainly
research situations where instrumental variables would be very useful but X is not a
measure of random assignment. In this case some information must be used to satisfy
this assumption.
b. Effect of the instrument is large. This may be difficult to obtain in practice but makes
sense given that the relation between the instrument and mediator must be substantial for
prediction to be accurate.
c. SUTVA is probably realistic in many situations whereby the potential outcomes for a
participant do not depend on their assignment.
d. Monotonicity is generally reasonable as the relation of X to M is likely to be the same
for all participants (but see answer to Exercise 13.6 below).
e. Exclusion restriction is unrealistic in behavioral research but may be sensible in some
situations and it may be possible to design a research study so that the exclusion
restriction holds.
The extent to which violation of these assumptions affects research conclusions and
estimates of effects is a critical area of future research. It is likely that these assumptions
are violated in actual research studies. The important question is how much the
assumption must be violated to invalidate a research conclusion.
13.6. Defiers define a unique group of persons who do the opposite of what they are
assigned, i.e. get the treatment if in the control group and get the control if in the
treatment group. It is probably unlikely that there are many persons like this in a real
research project. One example of this situation may occur where there is reactance
among subjects. For example, adolescents who feel that experimenters are reducing their
freedom attempt to regain their freedom by doing the opposite of what the experimenter
wishes. Another possible example may occur in situations where persons in the control
group resent that treatment group participants got the treatment and find ways to get the
treatment themselves (called resentful demoralization by Cook and Campbell, 1979).
Alternatively, persons in the treatment group may prefer the control condition because it
is less time demanding or taxing in some way.
Chapter 14
Comments: Designs to investigate mediation processes is probably the most important
part of this chapter. A useful exercise is to come up with examples of each type of
experimental design (i.e., enhancement, blockage, purification) for the readers research

38
area. It is also useful to consider that any mediation study should have a qualitative
component both to validate hypothesized mediators but also to get some information on
mediators that are not included in the quantitative mediation analysis. Issues regarding
mediation meta-analysis are also interesting to discuss including when it makes sense to
include information from separate studies of the X to M and the M to Y relation.
14.2. Randomization of units to conditions ensures that any omitted variables will be
balanced between the groups so that any difference between groups is attributable to the
manipulation and not another variable. Randomization assumes that the assignment is
indeed random and a large number of units are assigned. Because of randomization the
estimate of the effect of randomization can be considered a causal effect because the
FACE equals the ACE. If X represents random assignment then the a and c coefficients
represent causal effects.
14.4. There are many answers to this question but the most straightforward answer is that
a program of research would be conducted consisting of randomized experimental studies
of X to M and M toY relations, design of future studies based on the results of mediation
analysis, enhancement and blockage designs, and qualitative approaches. The goal is to
provide a pattern of significant results demonstrating a mediation process. In drug
prevention research, a first group of studies by several researchers investigates a
prevention program targeting several mediators. As part of mediation analysis of these
programs, several critical mediating variables are identified and these are further
investigated by randomly assigning components to change the mediators to different
groups of participants. Alternative explanations of mediation relations are investigated
with comparison mediators and dependent variables to further specify the critical
mediators. Qualitative data from participants in these studies ideally provide
corroborating evidence that the mediators identified in the statistical analysis are the
important ones. Evidence for the mediator is obtained in laboratory and real world
settings and alternative detailed experimental studies.
14.6. If only one study can be conducted, then it makes sense to apply the most detailed
methods to extract the maximum amount of information. Before the study is conducted
care must be taken to base an intervention on theory. Ideally the study results will provide
evidence for competing theories. Measures are developed so that they have excellent
validity and reliability. Detailed causal inference methods as well as specific
confirmatory hypotheses should be specified. After the data are collected, a set of specific
tests of hypotheses would be conducted and then after the planned tests, detailed post-hoc
analysis of the data would be conducted to extract the maximum amount of information
form the data. After these data analysis is fully understood, then a second study is
planned to further test theory and confirm important mediators. These types of studies are
more common in applied research and public health research where it is very expensive
to conduct research. If it is easy to conduct experiments it seems that the best strategy is
to conduct a program of randomized experimental studies where less attention is paid to
extracting the most information from a study and more attention is paid to the
specification and testing of detailed hypotheses that can be easily replicated by other
researchers with other samples. Investigation of mediators with college populations is an

39
example of this research context. Of course, ideally results from studies that are difficult
to conduct are further investigated in other contexts such as the research laboratory where
it is easier to conduct the research.
Chapter 15
Comment: One exercise for this chapter is for students to make up their own questions. It
might also be useful for students to indicate how the quote at the beginning of each
chapter is related to the chapter content. It is also interesting for students to determine
why complete names are given for some of the quotes. Another exercise would be for
students to describe a technical or substantive future topic for investigation.

Potrebbero piacerti anche