You are on page 1of 10

Psychological Methods © 2011 American Psychological Association

2011, Vol. 16, No. 1, 34 – 43 1082-989X/11/$12.00 DOI: 10.1037/a0021763

Estimating, Testing, and Comparing Specific Effects in Structural Equation


Models: The Phantom Model Approach
Siegfried Macho Thomas Ledermann
University of Fribourg University of Connecticut

The phantom model approach for estimating, testing, and comparing specific effects within structural
equation models (SEMs) is presented. The rationale underlying this novel method consists in representing
the specific effect to be assessed as a total effect within a separate latent variable model, the phantom
model that is added to the main model. The following favorable features characterize the method: (a) It
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

enables the estimation, testing, and comparison of arbitrary specific effects for recursive and nonrecursive
This document is copyrighted by the American Psychological Association or one of its allied publishers.

models with latent and manifest variables; (b) it enables the bootstrapping of confidence intervals; and
(c) it can be applied with all standard SEM programs permitting latent variables, the specification of
equality constraints, and the bootstrapping of total effects. These features along with the fact that no
manipulation of matrices and formulas is required make the approach particularly suitable for applied
researchers. The method is illustrated by means of 3 examples with real data sets.

Keywords: phantom models, specific effects, bootstrapping, SEM

In many practical applications, researchers using structural The SEM literature provides two main approaches for address-
equation modeling (SEM) face the problem of assessing specific ing the estimation, testing, and comparison of specific effects. The
effects that cannot be classified as direct, indirect, or total effects, first one employs matrix methods (Bollen, 1987, 1989). Specific
as estimated by most SEM programs. In structural models, the effects are estimated by means of (differences of) matrix products,
direct effect of variable X on Y is the effect that is not mediated by and the associated standard errors of estimates are computed by
any other variable in the model. The indirect effect from X to Y is means of the delta method in matrix format (Sobel, 1986). The
the sum of all mediated effects between the source variable X and matrix method might not be well suited for all applications and
the final outcome variable Y. The total effect is the sum of the users, for the following reasons: First, the use of complex matrix
direct and indirect effect. For example, a scientist might be inter- formulas (involving vector operators and Kronecker products)
ested in the specific effect of a source variable X on a final outcome might have a deterrent effect on users interested primarily in
variable Y that is mediated by a subclass of the mediators (inter- applied research. Second, with complex specific effects, involving
vening variables) involved in the effect of X on Y only. The many paths, the computation of standard errors of the specific
evaluation of such specific effects calls for special techniques that, effect becomes quite complex and error prone. Finally, in general
unfortunately, are not implemented in all SEM programs and the method does not permit the user to bootstrap confidence
require some expertise in matrix algebra to execute. intervals. At present, the bootstrapping of confidence intervals of
Moreover, not only the estimation and testing of single specific arbitrary specific effects and comparison of effects via Bollen’s
effects but also a comparison of different specific effects (or of (1987, 1989) method is only possible with the programs Mplus,
direct, indirect, and total effects) might be of substantial interest.
Mx, and OpenMx. However, bootstrapped confidence intervals are
For example, a researcher might want to compare the effect of a
usually superior to those computed by means of the delta method,
source on a final outcome variable that is mediated by two groups
especially with small sample sizes (e.g., Bollen & Stine, 1990;
of intervening variables. Again, most commonly used SEM pro-
MacKinnon, Lockwood, & Williams, 2004).
grams provide no direct solution to this problem.
A second approach to the estimation and comparison of specific
effects was proposed by Cheung (2007). According to this method,
a latent variable, a so-called phantom variable (Rindskopf, 1984),
This article was published Online First February 7, 2011. is added to the model as an extra variable with a direct effect from
Siegfried Macho, Department of Psychology, University of Fribourg, this latent variable to one of the main variables in the model. The
Fribourg, Switzerland; Thomas Ledermann, Department of Psychology, variance of this latent variable is fixed to zero, and the structural
University of Connecticut. coefficient assigned to the path from the phantom variable to the
We would like to thank Michael Munz for his comments on an earlier main variable is restricted by means of a formula representing the
draft of the article. Additional material comprising raw data, AMOS specific effect or a contrast between two or more effects. This
graphics files with models, and R source files implementing OpenMx
strategy forces the program to provide estimates and standard
programs of the examples is available online (http://www.unifr.ch/psycho/
site/units/allpsy/team/Macho/research/sem). errors for the structural coefficient represented by the formula. The
Correspondence concerning this article should be addressed to Siegfried insertion of the additional variable with fixed variance and the
Macho, Department of Psychology, University of Fribourg, Rue Faucigny direct effect constrained as a function of the other model param-
2, CH 1700 Fribourg, Switzerland. E-mail: siegfried.macho@unifr.ch eters does not influence the estimation of parameters for the main
34
THE PHANTOM MODEL APPROACH 35

model. In addition, the method permits the bootstrapping of con- Third, phantom models are of functional significance only; that
fidence intervals of the coefficient representing the specific effect. is, they have no substantive meaning, as they are employed to
The approach of Cheung requires a SEM program, like LISREL, attain a specific objective with respect to estimating and testing
Mplus, or Mx, that enables the specification of nonlinear con- aspects of the main model only. In our applications, phantom
straints on parameters. However, most SEM results reported in models are constructed in order to force the SEM program to
articles were obtained by means of SEM programs that do not provide estimates and tests of specific effects and of their contrasts.
permit the specification of such nonlinear constraints (cf. Jackson, However, we do not exclude the possibility that phantom models
Gillaspy, & Purc-Stephenson, 2009, p. 14). In addition, with may be used for other purposes, too.
complex effects and contrasts, the method reaches its limits, as the For the present applications, phantom models are entirely dis-
elaboration of the relevant algebraic expressions for testing or connected from the main model. This feature is not regarded as
comparing specific effects becomes difficult and error prone (cf. mandatory, however. There might be applications of phantom
the discussion below). models that require a connection between the phantom and the
In sum, these two methods may not be appropriate in every main model.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

application and for all researchers who would like to estimate, test, The characterization of phantom models can be summarized as
This document is copyrighted by the American Psychological Association or one of its allied publishers.

and compare specific effects. By contrast, the phantom model follows: Phantom models consist entirely of latent variables, and
approach is well suited with complex models and for users em- all parameters are constrained. They are added to the main model
ploying SEM software with a graphical interface enabling the to serve specific purposes. In our applications, phantom models are
representation of a model by means of a causal diagram. Conse- employed to obtain estimates and tests of specific effects and
quently, it may serve as a viable complement to existing methods. comparisons of effects. The name phantom model was inspired by
The utility of the phantom model approach derives from the Rindskopf’s (1984) phantom variables that are also of functional
following favorable characteristics: First, the method does not significance only (in this case, to specify complex constraints on
require the elaboration of complex formulas that may be not only parameters). After this characterization of phantom models, we
daunting for many users but also error prone. Second, the method next present the phantom model approach for estimating and
is not limited to SEM programs allowing for the specification of comparing specific effects.
arbitrary functional restrictions on parameters. Third, the method
enables the bootstrapping of confidence intervals for specific ef-
fects. Finally, the approach is enormously flexible by allowing the The Phantom Model Approach
estimation, testing, and comparison of arbitrary specific effects in
recursive and nonrecursive models with manifest and latent vari- In the following section, the construction of phantom models for
ables. estimating specific effects or contrasts is explicated. A specific
The present article is structured as follows: First is an abstract effect is regarded as an effect of a source variable X on an outcome
specification of phantom models as well as the application of the variable Y that does not conform to the standard classification of
method to the estimation and comparison of specific effects. Sec- direct, indirect, and total effects. Typically, a specific effect makes
ond is an illustration of the application of phantom models to up a portion of the indirect effect between X and Y that encom-
estimating, testing, and comparing specific effects, using examples passes a subset of the paths and/or variables constituting the
from the literature. Finally, in the Discussion section, the phantom indirect effect. Specific effects are of focal interest in mediation
model approach is compared with the existing approaches. research (e.g., Ledermann & Macho, 2009; MacKinnon, 2000).
For example, assume that the effect from X on Y is transmitted by
two parallel mediating paths. In this case, it is usually of interest
Phantom Models for Estimating Specific Effects
which of the two mediated effects is significant.
Concerning the comparison of effects, the phantom model ap-
In the following section, we first present a characterization of
proach may be used for contrasting two specific effects. Consider
phantom models. This is followed by an explication of how to use
once again a model with two parallel mediating paths transmitting
phantom models for assessing specific effects and contrasts be-
the effect of X on Y. A comparison of the specific effects can foster
tween effects.
a better understanding of the significance of the intervening vari-
ables (e.g., Ledermann & Macho, 2009). The comparison of ef-
Phantom Models fects is not restricted to specific effects, however. For instance,
below, in Example 2, we apply the method for comparing the
Phantom models are characterized by the following features: indirect with the direct effect.
First, phantom models are made up entirely of latent variables. The phantom model approach for estimating specific effects and
Second, their parameters are completely constrained; that is, they contrast is based on the following idea: If the specific effect or
are fixed either to specific values or to functions of the other contrast of interest is represented as a total effect within a phantom
parameters. By consequence, the addition of one or more phantom model, then SEM programs enabling the estimation of total effects
models to the main model does not influence the estimation or provide point estimates of the specific effect or the contrast be-
testing of the latter. In the present case, the parameters are either tween effects. In addition, confidence intervals may be obtained by
fixed or equated to one of the parameters of the main model. means of bootstrapping.
Due to the fact that all available SEM packages enable the Modeling specific effects by means of phantom models differs
specification of these two types of restriction, the method is from the way they are represented in the other two approaches: In
generally applicable. Cheung’s (1987) and Bollen’s (1987, 1989) approach specific
36 MACHO AND LEDERMANN

effects are represented algebraically, by means of formulas (poly- 7. Insert direct effects from the final outcome variables of
nomial equations) and matrix formulas, respectively. the two phantom models to variable t_ and fix the path
Like the other methods, the phantom model approach assures coefficient of one effect to 1 and that of the other to ⫺1.
the correct estimation of specific effects or contrasts as long as the
specific effect or contrast is modeled accurately as a total effect This guidance for building a phantom model representing a
within the phantom model, provided that total effect can be esti- specific effect or the contrast between two effects as a total effect
mated properly for the given model. Thus, concerning the correct can be generalized in a straightforward way for testing simultane-
estimation of the specific effect or contrast, it does not matter ously multiple specific effects or for comparing more than two
whether the effect is represented by means of an algebraic expres- effects (cf. the example presented in the Discussion section).
sion or by a phantom model. The given instruction lists the single steps in an abstract manner
only. Concerning the concrete implementation of a phantom
model, it is important to note that the phantom model approach
Guidance to Build Phantom Models Representing requires the reconstruction of a portion of the main model as a new
Specific Effects and Contrasts
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

latent variable model. By consequence, with most SEM programs,


This document is copyrighted by the American Psychological Association or one of its allied publishers.

the construction of the phantom model consists essentially in the


The construction of a phantom model that represents the specific reproduction and adjustment of an already existing model struc-
effect of interest consists in the following steps: ture. To be more specific, by means of a graphical interface, like
AMOS graphics, the construction of the latent model consists
1. Identify the variables and direct effects in the main basically in copying existing portions of the main model and then
model making up the specific effect to be tested. changing the variable names as well as specifying all variables as
latent. Similarly, by means of program syntax, for many programs,
2. Add a new latent variable for each variable identified in
like AMOS BASIC, EQS, RAMONA, and SAS PROC CALIS, the
the first step, and add direct effects between these new
model is specified by means of a set of equations. In this case, the
latent variables so that the structure of the resulting latent
variable model is an exact replicate of that portion of the equations defining the relevant portion of the main model may
main model that makes up the specific effect of interest be copied and the names of variables modified. Furthermore,
(no residual terms are required). with programs like AMOS BASIC, SAS PROC CALIS, and
RAMONA, where equality constraints are specified by means of
3. Specify equality constraints on path coefficients of the identical names, the copying of relevant equations implies that the
newly constructed latent variable model: Each path co- required equality constraints are set correctly.
efficient is restricted to the value of the respective coef- With programs, like LISREL and Mplus, enabling the represen-
ficient in the main model. tation of specific effects by means of nonlinear constraints on
parameters, the representation of a specific effect by a polynomial
4. Specify a restriction on the variance of the exogenous equation may be more favorable than the construction of a phan-
variable within the phantom model (e.g., set the variance tom model for representing the specific effect, at least in those
parameter to 1). Instead of fixing the variance parameter, cases where the specific effect or contrast is not too complex (cf.
it is also possible to specify equality constraints between our discussion of this issue in the Discussion section).
the variance of the exogenous phantom variable and its The application of the phantom model approach is subject to
counterparts in the main model (cf. Example 1, below). two limitations. First, the method cannot generally be applied for
testing and comparing standardized effects. Second, with missing
The latent model resulting from Step 2 to Step 4 is a phantom data most SEM programs do not permit the bootstrapping of
model representing the specific effect of interest as a total effect. confidence intervals. In this case, confidence intervals for specific
In order to contrast two specific effects, the variables and direct
effects or contrasts may be obtained by means of parametric
effects for each of the two effects to be compared have to be
bootstrapping. More specifically, the following two-stage proce-
duplicated as separate structures. That is, for each of the contrasted
dure may be applied: In the first step, a consistent estimate of the
effects, a separate phantom model representing the respective
covariance matrix has to be computed. This may be achieved by
effect is constructed. In addition, the last step of the previous
fitting the saturated model (or another model representing the
specification has to be replaced by the following steps:
relations in the population adequately) to the observed data using
5. Insert two additional latent variables into the phantom the full information likelihood information available for most SEM
model, an exogenous variable s_ and a target variable t_, packages (cf. Arbuckle, 1996). Alternatively, specialized pro-
and fix the variance parameter of s_ to 1 or any other grams, like NORM (Schafer, 1999) and SPSS Missing Values
admissible value. (Throughout this article we adopt the (SPSS, 2007), may be used for estimating the covariance matrix. In
notational convention of adding an underline character to the second step, confidence intervals are estimated by means of
names of variables making up the phantom model.) parametric bootstrapping via the covariance matrix estimated in
the first step.
6. Insert direct effects from s_ to the exogenous vari- Having specified the construction of phantom models to repre-
ables of the two phantom models representing the sent a specific effect or a contrast between effects as total effects,
single effects to be compared and fix the path coeffi- we now illustrate the application of the method using data and
cients to 1. models from applied research.
THE PHANTOM MODEL APPROACH 37

Illustrative Examples items, whereas the self-evaluation of participants comprised a


single measure only. Answers to questions were provided on a
In the following section, we illustrate the method by means of 4-point rating scale. The data were collected at different points of
three examples using real data sets. The first case demonstrates the time: The variables CHS, TP, SRGA, and SR were measured 2
testing of a specific effect within a recursive model. In the second weeks after the variables PM, HOC, and SE.
example the direct effect of the exogenous variable on the outcome Our model differs from that of Tsai et al. (2007) in three
variable is compared with the indirect effect by means of the same respects. First, no control variables were included in our model.
recursive model as in Example 1. The final example illustrates the Tsai et al. incorporated two control variables into their model in
comparison of two indirect effects within a nonrecursive model. addition to the main variables shown in Figure 1 (the inclusion of
The data of Examples 1 and 2 result from Study 1 of Tsai, Chen, these variables resulted in a worse fit). Second, no correction of
and Liu (2007, Table 2). The data of the third example are from a attenuation was performed. Except for the self-rating of task per-
study of Kurdek (1998, Table 4). formance, the variables represented by the rectangles of Figure 1
represent composite measures. Tsai et al. used Cronbach’s alpha to
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

Example 1: Testing a Specific Effect correct for measurement errors. However, due to the high reliabili-
This document is copyrighted by the American Psychological Association or one of its allied publishers.

ties (␣ ⬎ .80 for each of the measures), the effect of the correction
Tsai et al. (2007) investigated variables mediating the effect of is not substantial (this is evidenced by the fact that our path
positive moods on task performance of insurance sale agents. The coefficients are nearly identical to theirs). Third, two direct effects
model in Figure 1 is a modified version of their Model 3 (cf. Tsai (PM 3 TPF and SE 3 HOC) as well as one covariance arc
et al., 2007, Figure 2). The variable positive moods (PM) repre- between ␨TP and εSRGA were added to the model. The modification
sents a composite of answers to 14 items measuring the retrospec- of the original model was inspired by the fact that the former
tive evaluation of positive moods. The variable helping other provided no acceptable fit to the data: ␹2(9) ⫽ 62.55, root-mean-
coworkers (HOC) is a composite of four items measuring partic- square error of approximation (RMSEA) ⫽ .14. In addition, the
ipants’ helping behavior with respect to coworkers. The variable modification resulted in a substantial improvement: ⌬␹2(3) ⫽
coworker helping/support (CHS) represents four items measuring 50.80. Note also that the added effects are not implausible (e.g., a
the estimation of participants’ support by their coworkers. Self- higher estimated self-efficacy might lead participants to spend
efficacy (SE) comprises 10 items measuring participants’ estimates more time in helping other coworkers).
of their efficacy on the job. The variable task persistence (TP) is a The model of Figure 1 provides an acceptable fit: ␹2(6, N ⫽
composite of two items measuring participants’ ability to over- 306) ⫽ 11.75, p ⫽ .068, RMSEA ⫽ .056, 90% CI [.000, .103].
come obstacles. Participants’ task performance (TPF) was evalu- When degrees of freedom are low, as in this case, a sample size of
ated by participants themselves (self-rated goal attainment 306 will not yield high precision in the RMSEA estimate. As a
[SRGA]) and by their supervisors (supervisor rating [SR]). The consequence, the 90% confidence interval spans a wide range of
evaluation of the supervisors was measured by means of eight RMSEA values, from perfect to unacceptable.

HOC CHS SR

1 1 1

Helping Other CHS . HOC Coworker Supervisor


Coworkers Helping / Support Rating
(HOC) (CHS) (SR)
HOC. PM
2
TPF.CHS 1
PM
CHS. PM
TP.CHS

Positive Moods Task


TPF. PM
(PM) Performance
TP. PM
HOC. SE (TPF)
SE. PM
1

TPF . SE SRGA.TPF
TPF .TP
TPF

Self-Rated Goal
Self-Efficacy Task Persistence
Attainment
(SE) (TP)
TP. SE (SRGA)

1 1 1

SE TP SRGA

Figure 1. A model explaining the effect of positive moods on task performance. Adapted from “Test of a
Model Linking Employee Positive Moods and Task Performance,” by W.-C. Tsai, C.-C. Chen, and H.-L. Liu,
2007, Journal of Applied Psychology, 92, p. 1577. Copyright 2007 by the American Psychological Association.
38 MACHO AND LEDERMANN

For substantive reasons, a researcher might be interested in the inclusion of this additional variable does not influence the estima-
effect of PM on TPF that is mediated by the variable SE. Thus, the tion and testing of the specific effect of interest [cf. the AMOS
specific effect of interest consists in that component of the indirect graphics file on our website at http://www.unifr.ch/psycho/site/
effect of PM on TPF that is transmitted by the variable SE. It can units/allpsy/team/Macho/research/sem].)
be conceived of as the product of the direct effect of PM on SE and Table 1 contains the estimated effects as well as bootstrapped
the total effect of SE on TPF. Stated differently, the specific effect standard errors and 95% confidence intervals for the each of the
of interest is the total effect of PM on TPF minus the effect of PM three examples. The bootstrapping results (for each of the three
on TPF that is not transmitted by the variable SE. examples) are based on 5,000 bootstrap samples. Due to the fact
A model representing the focal specific effect as a total effect that only the covariance matrices were available, we performed
results from the main model in Figure 1 by eliminating all direct parametric bootstrapping: Instead of drawing bootstrap samples
effects of PM on the other variables except for the direct effect of from the actual sample, we drew bootstrap samples from a multi-
PM on SE. The phantom model in Figure 2 represents this specific variate normal distribution using the sample covariance matrix.
effect as total effect. It differs from the main model in two The results of two methods for computing bootstrapped confidence
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

respects: The variables are all latent, and the direct effects of PM intervals, the percentile method and the bias-corrected percentile
This document is copyrighted by the American Psychological Association or one of its allied publishers.

on the variables other than on SE have been omitted. All other method, are presented (for a basic introduction to these methods,
direct effects as well as all variables of the main model are see Mooney & Duval, 1993; Stine, 1989). The results indicate a
replicated within the phantom model, as they are part of the significant specific effect, as zero is not included in the confidence
specific effect. Thus, the phantom model comprises exactly those interval. The bootstrapped confidence intervals are practically
phantom variables and direct effects that are required for repre- identical for both methods of computing confidence intervals.
senting the focal specific effect within the main model as total
effect. The phantom model in Figure 2 does not contain variables Example 2: Comparing Two Effects
representing residual terms because they are not required for
estimating the total effect of PM_ on TPF_ within the phantom The second example concerns the contrast of two effects. Using
model (phantom variables corresponding to the variables in the again the model shown in Figure 1, we compared the indirect
main model are symbolized by the name of the variable plus an effect of positive moods on task performance with the respective
underline). direct effect. Figure 3 depicts two possible phantom models that
The symbols denoting path coefficients within the phantom enable the estimation of the difference between these two effects.
model are identical to the corresponding coefficients within the The phantom model in Figure 3A is constructed according to the
main model, thus indicating the equality of coefficients. Similarly, instruction given above. It comprises two partial models. The
the same symbol ␾PM2 was used to denote the variance of the upper submodel, consisting of the variables PMi_, SE_, HOC_,
exogenous variables, thus indicating that both variances were CHS_, TP_, and TPFi_, represents the indirect effect of PM on
restricted to be equal. With the phantom model of Figure 2 having TPF. The lower partial model, comprising variables PMd_ and
been added to the main model of Figure 1, the SEM program TPFd_, represents the direct effect of PM on TPF. The two partial
provides the specific effect as a total (or indirect) effect in the models are connected to a source variable s0_ and a target variable t0_.
output. (Unfortunately, AMOS [graphics] does not provide an By fixing the coefficient of the direct effect TPFi_ 3 t0_ to 1 and
output of the total [or indirect] effect of PM_ on TPF_ for the that of TPFd_ 3 t0_ to ⫺1, the indirect effect represented by the
model shown in Figure 2. This is probably due to the fact that the upper partial model is multiplied by 1, whereas the direct effect
program interprets variable PM_ as an error term associated with represented by the lower partial model is multiplied by ⫺1. As a
variable SE_. To prevent this misinterpretation of the variable result, the total (or indirect) effect of s0_ on t0_ represents the
PM_, and to force the program to provide the required output of difference between the indirect and the direct effect of PM on TPF.
the total effect, one can insert a further latent variable, say, The model in Figure 3B depicts a phantom model that achieves
dummy_, with an arrow: PM_ 3 dummy_. The associated coeffi- the same result as the model of Figure 3A. This model is a perfect
cient may be fixed to an arbitrary value, say, 1. Clearly, the copy of the main model, with one exception: The direct effect from

Helping Other CHS . HOC Coworker


Coworkers Helping /Support
(HOC_) (CHS_)
2 TPF .CHS
PM

Positive Task
Moods HOC. SE TP.CHS Performance
(PM_) (TPF_)
SE. PM
TPF . SE TPF .TP

Self- Task
Efficacy Persistence
(SE_) TP.SE (TP_)

Figure 2. A phantom model representing the specific effect of positive moods on task performance that is
mediated by the variable self-efficacy.
THE PHANTOM MODEL APPROACH 39

Table 1
Estimated Specific Effects or Contrast and Bootstrapped Standard Errors and 95% Confidence Intervals for Example 1–3
95% CI
Percentile method Bias corrected
Effect Contrast being tested Valuea SE LL UL LL UL
1 Specific effect (phantom model of Figure 2) 0.081 0.026 0.034 0.134 0.036 0.139
2 Contrast between direct and indirect effect (phantom model of Figure 3) ⫺0.173 0.090 ⫺0.348 0.006 ⫺0.338 0.016
3 Contrast between two specific effects (phantom model of Figure 5) 0.123 0.158 ⫺0.195 0.425 ⫺0.198 0.425
Note. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
a
Point estimate of the specific effect or the difference of two effects being compared.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

PM_ to TPF_ is split into two parts with the phantom variable p0_ 1 indicate that this hypothesis cannot be rejected because the
This document is copyrighted by the American Psychological Association or one of its allied publishers.

in between. The path coefficient of the direct effect PM_ 3 p0_ confidence intervals of both methods include zero.
was set equal to the coefficient of the direct effect of PM on TPF
within the main model (as indicated by the identical symbols),
Example 3: Comparing Effects of a
whereas the coefficient of the direct effect p0_ 3 TPF_ was fixed
to the value ⫺1. As for the model in Figure 3A, the current Nonrecursive Model
phantom model represents the difference between the indirect and
the direct effect of PM on TPF as a total (or indirect) effect of PM_ The final example is intended to demonstrate that the phantom
on TPF_. The construction of the phantom model in Figure 3A is model approach works equally well with nonrecursive models. A
a bit more demanding than that of Figure 3B. However, the common nonrecursive model is Kenny’s (1996) mutual influence
principle for constructing the former (but not the latter) is gener- model, which was designed to examine reciprocal influences in
ally applicable, as illustrated by the subsequent example. dyadic data (see also Woody & Sadler, 2005).
With one of these phantom models having been constructed, The model depicted in Figure 4 was used by Kurdek (1998) to
standard errors of the difference as well as confidence intervals test the mutual influence of marital quality and symptoms of
may be bootstrapped, thus enabling a test of the hypothesis that the depression in married couples. Specifically, according to the re-
indirect effect is identical to the direct effect. The results in Table ciprocal change hypothesis of Kurdek, the deterioration of marital
quality is accompanied by an increase of depressive symptoms. On
the other hand, depressive symptoms are conceived as risk factors
for marital distress. Kurdek employed the Dyadic Adjustment
CHS. HOC
HOC_ CHS_ Scale to measure marital quality (MQ). The reliability of the index
TPF .CHS
HOC. PM score (32 items), quantified by means of Cronbach’s alpha, was
CHS . PM

PMi_ HOC. SE
TP.CHS
TPFi_ quite high: ␣ ⱖ .89. The variable depressive symptoms (DS)
TP. PM
TPF . SE represents the sum of 20 items from the Symptom Checklist-90 –
1 SE. PM
TPF .TP
1 Revised. Again, the reliability was high: ␣ ⱖ .86. As with the
1 SE_ TP_
s0_ TP. SE t0_
previous examples, no correction for attenuation was performed.
Again, this may be justified by the high reliabilities, as measured
1 by Cronbach’s alpha. The model in Figure 4 provides a good fit to
1
the data: ␹2(12, N ⫽ 136) ⫽ 16.46, p ⫽ .171, RMSEA ⫽ .052,
TPF . PM
PMd_ TPFd_ 90% CI [.000, .109]. The low number of degrees of freedom and
A the small sample size result in the wide 90% confidence interval of
the RMSEA.
CHS . HOC One of the issues of Kurdek (1998) concerned the question
HOC_ CHS_
CHS . PM whether the decline in marital quality is similar for husbands and
TPF .CHS
HOC. PM wives. In a similar spirit, it might be interesting to note whether the
1 TP.CHS

PM_ TPF . PM
p0_
1
TPF_
variables measured during the 1st year of the relationship exert the
TP. PM TPF . SE
same influence on the marital quality in the 4th year, for both
SE. PM
HOC. SE TPF .TP partners. The phantom model in Figure 5 enables one to address
SE_ TP_
this issue by providing a comparison of the joint influence of
TP. SE 1st-year variables on the marital quality in the 4th year for hus-
bands and wives.
B
The phantom model is made up of two submodels. The upper
Figure 3. Two phantom models representing the comparison of the
one comprising the variables H1MQ_, H1DS_, H4MQ_, and
indirect effect of positive moods on task performance with the direct effect. H4DS_ represents the husbands, and the lower submodel com-
HOC ⫽ helping other workers; CHS ⫽ coworker helping/supporting; prises the corresponding variables for the wives: W1MQ_, W1DS_,
PM ⫽ positive moods; TPF ⫽ task performance; SE ⫽ self-efficacy; TP ⫽ W4MQ_, and W4DS_. The exogenous phantom variable s0_ whose
task persistence. variance was fixed to 1, is linked to each of the 1st-year variables
40 MACHO AND LEDERMANN

Husband’s γH .
MQ H MQ Husband’s
1
Marital Quality Marital Quality ζH MQ
Year 1 (H1MQ) Year 4 (H4MQ)

βH .
DS H MQ βH .
MQ H DS

Husband’s γH .
DS H DS Husband’s
Depressive Symptoms Depressive Symptoms
1 ζH DS
Year 1 (H1DS) Year 4 (H4DS)
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

Wife’s γW .
MQ W MQ Wife’s 1
Marital Quality Marital Quality ζW MQ
Year 1 (W1MQ) Year 4 (W4MQ)

βW .
DS W MQ βW .
MQ W DS

Wife’s γW .
DS W DS Wife’s
Depressive Symptoms 1 ζW
Depressive Symptoms DS
Year 1 (W1DS) Year 4 (W4DS)

Figure 4. A model representing the influence of marital quality and depression at two time points. Adapted
from “The Nature and Predictors of the Trajectory of Change in Marital Quality Over the First 4 Years of
Marriage for First-Married Husbands and Wives,” by L. A. Kurdek, 1998, Journal of Family Psychology, 12,
p. 506. Copyright 1998 by the American Psychological Association.

with path coefficients equal to 1 for each of the direct effects. The methods of Cheung and Bollen can be simulated by means of
phantom variable H4MQ_, representing the marital quality of phantom models, thus highlighting that the latter approach is as
husbands in the 4th year of marriage, is linked to the target variable powerful as the other methods. This is followed by a discussion of
t0_ with coefficient 1, whereas the corresponding variable W4MQ_ the conditions that render the phantom model approach more or
for the wives is linked to t0_ with coefficient ⫺1. As a result, the less favorable compared with the other methods.
total effect from s0_ to t0_ represents the difference between The method of Cheung (2007) may be implemented by using a
husbands and wives of the joint effect of the 1st-year variables on phantom model that consists of two latent variables s0_ and t0_
marital quality in the 4th year. with a direct effect from s0_ on t0_. The variance of s0_ has to be
As for the previous examples, we performed parametric boot- fixed to an arbitrary value (say, 1.0), and the path coefficient is
strapping, using 5,000 samples, for estimating standard errors and determined by the formula representing the specific effect or the
confidence intervals. The results did not show significant differ- comparison to be assessed.
ences between husband and wives, which is in general agreement Concerning the method of Bollen (1987, 1989), it can also be
with the other results of Kurdek (1998). demonstrated that the matrix manipulations for computing specific
The examples presented demonstrate the application of the effects, as proposed by Bollen, may be simulated by means of
phantom variable approach for estimating, testing, and comparing phantom models. By consequence, all specific effects that can be
specific effects. The bootstrapping of confidence intervals renders handled with Bollen’s method can be managed by means of
the method extremely convenient for testing specific hypotheses in phantom models, too. The following illustration uses one of Bol-
complex models. The application of this method to the testing of len’s examples. Figure 6 depicts the path model employed by
arbitrary effects or contrasts is straightforward. Bollen (1987, Figure 3; 1989, Figure 8.16). He considers those
fractions of the indirect effects of the exogenous variables X1, X2,
Discussion and X3 on the endogenous variables Y2 and Y3 that are mediated
by Y1.
In the introduction, we presented the methods of Cheung (2007) The approach comprises three steps. First, indirect effects of
and Bollen (1987, 1989), and we claimed that the phantom model exogenous on endogenous variables are estimated for the complete
approach may be regarded as a valuable complement to these model. Second, indirect effects are estimated by means of a model
methods. In the following section, we first demonstrate that the with all coefficients representing the direct effects on Y1 as well as
THE PHANTOM MODEL APPROACH 41

specific effect in a direct way and not, like the model in Figure 7A,
.
H 4 MQ H 1 MQ
H 1MQ _ H 4 MQ _ by means of subtracting effects of two models.
The preceding exposition demonstrated that phantom models
1
.
H 4DS H 4MQ
are as powerful as existing approaches. However, the phantom
1 .
H 4MQ H 4DS
model approach enables the testing and comparison of specific
t0_ effects that exceed the potential of existing matrix methods. Con-
.
H 4DS H 1DS
1 1 H 1DS _ H 4 DS _ sider once again our third illustrative example that compared
s0_ 1 husbands’ and wives’ joint effect of the two 1st-year variables on
marital quality in the 4th year. The method of Bollen cannot be
1 .
W 4 MQ W 1 MQ used directly to handle this comparison of joint effects. The phan-
W 1MQ _ W 4 MQ _
tom model in Figure 5 enables the estimation and testing of the
1 required comparison, however. (The matrix method can be ex-
.
W 4DS W 4MQ
tended by augmenting the structural matrices with phantom effects
.
W 4MQ W 4DS
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

[i.e., direct effects between phantom variables and the variables in


This document is copyrighted by the American Psychological Association or one of its allied publishers.

.
W 4DS W 1DS the main model]. This amounts to the use of an elementary type of
W 1DS _ W 4 DS _
phantom model.)
The previous discussion revealed various qualities of the phan-
Figure 5. A phantom model with the total effect of variables s0_ on t0_ tom model approach for testing and comparing specific effects.
representing the comparison between husbands and wives of the joint However, as mentioned above, under certain conditions, the other
effect of the two 1st-year variables on marital quality in the 4th year. methods may be more convenient than the phantom model ap-
H1MQ, H4MQ ⫽ husband’s marital quality Years 1 and 4; H1DS, proach. We now discuss the conditions favoring the phantom
H4DS ⫽ husband’s depressive symptoms Years 1 and 4; W1MQ, W4MQ ⫽ model approach. First, the phantom model approach may be re-
wife’s marital quality Years 1 and 4; W1MQ, W4MQ ⫽ wife’s depressive garded as the method of choice for those users employing the
symptoms Years 1 and 4. programs AMOS and EQS, which do not allow for the represen-
tation of specific effects by means of algebraic expressions (ac-
cording to Jackson et al., 2009, these users make up about 50% of
the direct effects from Y1 on other (endogenous) variables set to the SEM users).
zero. Third, the indirect effects of these two models are subtracted. Second, the method is useful in general with complex models
The resulting difference corresponds to the specific effect in ques- like those of Examples 2 and 3. In both cases it may be possible to
tion. work out the formulas representing the comparison of the effects
The phantom model in Figure 7A emulates these operations. and to apply Cheung’s (2007) method instead of building a phan-
The upper partial model, with variable names containing the su- tom model. However, the phantom model approach will still be of
perscript 1, represents the complete model. The lower submodel, great value for those users with little expertise or motivation to
with variable names involving the superscript 2, represents the work out complex formulas, as the phantom model approach
model with the coefficients of direct effects to and from variable requires only the basic capability of implementing latent variable
Y1_[2] (representing Y1) set to zero. In Figure 7A the respective models and specifying constraints on parameters on the platform
direct effects are represented by dashed arrows. The subtraction of employed. Moreover, with complex models, working out formulas
effects is performed by connecting the endogenous variables cor- is relatively error prone, as they may comprise many terms, and it
responding to Y2 and Y3 within the two submodels to the target is quite easy to miss one of them. For example, the formula in
variables t1_ and t2_, with coefficients 1 and ⫺1 assigned to the
respective arrows. In addition, the exogenous variables denoting
X1, X2, and X3 in the two partial models are connected to the source
variables s1_, s2_, and s3_, with coefficient 1 on all arrows. The Y1

total effects of s1_, s2_, and s3_ on t1_ and t2_ represent those 1
portions of the indirect effects of X1, X2, and X3 on Y2 and Y3 that
11
are mediated by variable Y1. X1 Y1
31
The phantom model in Figure 7A can be simplified: Variables 21
31 Y3

s1_, s2_, and s3_, as well as X1_[2], X2_[2], X3_[2], and Y1_[2] with 1
associated arrows, may be deleted, and the arrows that emanate 12
21
from X1_[2], X2_[2], and X3_[2] may be replaced by respective X2
32
Y3
arrows originating in X1_[1], X2_[1], and X3_[1]. 22

The more complex model of Figure 7A was used to demonstrate


in detail how Bollen’s (1987, 1989) procedure for estimating 13
33
32

specific effects can be implemented by means of phantom models. X3 Y2


A much simpler and intuitively more appealing solution that leads 23

to the same outcome is presented in Figure 7b. This phantom 1


model results from the full model by deleting all paths from Y2

exogenous to endogenous variables that do not include variable Y1.


This model is intuitively more appealing because it represents the Figure 6. An example path model taken from Bollen (1989, p. 383).
42 MACHO AND LEDERMANN

X 11_ 11
Y1 _1
31
21
31

12
21
1 X 21_ 32 Y3 1_
1 22
s1_
1
13 33
32 1
X 31_ Y21_
1 23 1
1 t2_
s2_ t1_
1 0 2
X 1 2_ Y 1_

31 0 1 1
1 21
1
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

s3_ 1 0 0
This document is copyrighted by the American Psychological Association or one of its allied publishers.

X 22_ 32
Y3 _2
22

0 33
32

X 32_ Y2 2_
23

A
1
X1_ 11

Y1_
31
12
1
X2_ 21 Y3_
13

1 Y2_ 32

X3_

Figure 7. Two phantom models for estimating and testing the effect of the exogenous variables X1, X2, and X3
on Y2 and Y3 mediated by Y1, for the model of Figure 6. The phantom model in Figure 7A simulates the approach
of Bollen (1987, 1989). The phantom model in Figure 7B represents the same specific effect.

Example 2 contrasting the direct and the indirect effect of positive Furthermore, for nonrecursive models with more than two vari-
moods on task performance takes on the following algebraic form: ables making up a nonrecursive cycle (see, e.g., Berry, 1984), the
algebraic method is no longer practicable, as formulas for specific
␥ SE.PM ⫻ 关␤ TP.SE ⫻ ␤ TPF.TP ⫹ ␤ TPF.SE ⫹ ␤ HOC.SE ⫻ ␤ CHS.HOC effects involving cycles including more than two variables repre-
sent entries of inverted matrices with more than two rows and
⫻ 共␤ TP.HOC ⫻ ␤ TPF.TP ⫹ ␤ TPF.CHS兲] ⫹ ␥ TP.PM ⫻ ␤ TPF.TP columns, respectively. Without the help of mathematical software,
these formulas are usually too complicated to be worked out
⫹ 共␥ HOC.PM ⫻ ␤ CHS.HOC ⫹ ␥ CHS.PM兲
analytically.
⫻ 共␤ TP.CHS ⫻ ␤ TPF.TP ⫹ ␤ TPF.CHS兲 ⫺ ␥ TPF.PM The phantom model approach might be less favorable for users
employing programs, like Mplus and LISREL, that enable the
The formula for estimating and testing the comparison in Ex- specification of nonlinear constraints, specifically in case of simple
ample 3 is even more complex. In addition, errors of omission are models. For example, the comparison of the direct with the indirect
difficult to detect. The incorporation of the phantom model, on the effect for the three-variables mediator model (see, e.g., Shrout &
other hand, should be less error prone for users with some expe- Bolger, 2002) is more easily accomplished with the method of
rience in implementing SEMs (this is certainly the case if the Cheung (2007). In this case, the phantom model approach is
program contains a graphical interface), as it requires only “more indicated only for users of SEM programs that do not allow for the
of the same,” namely, the capability to implement a latent variable specification of nonlinear constraints. The same is true for other
model and to specify fixed and equality constraints. Thus, with models of low to medium complexity. Finally, for users of Mx or
complex effects or contrasts, we would recommend the use of the OpenMx who are probably familiar with the matrix formulations
phantom model representation additionally to the algebraic repre- of SEM models, it may be convenient to use the phantom model in
sentation as a test for the correctness of the worked out formula. connection with the matrix formula for computing total effects (cf.
THE PHANTOM MODEL APPROACH 43

the R files on our website employing OpenMx to implement the Kurdek, L. A. (1998). The nature and predictors of the trajectory of change
examples). in marital quality over the first 4 years of marriage for first-married
The phantom model approach offers a convenient and powerful husbands and wives. Journal of Family Psychology, 12, 494 –510.
method for estimating, testing, and comparing arbitrary specific doi:10.1037/0893-3200.12.4.494
Ledermann, T., & Macho, S. (2009). Assessing mediation in simple and
effects within the SEM framework. The method demands from
complex models. Manuscript submitted for publication.
users the ability to implement a latent variable model representing
MacKinnon, D. P. (2000). Contrasts in multiple mediator models. In J.
the specific effect or contrast as a total effect and to use the Rose, L. Chassin, C. C. Presson, & S. J. Sherman (Eds.), Multivariate
bootstrapping capabilities of the SEM program. It does not require applications in substance use research: New methods for new questions
the manipulation of matrices or complex formulas, and its appli- (pp. 141–160). Mahwah, NJ: Erlbaum.
cation is with nonrecursive models as easily as with recursive ones, MacKinnon, D. P., Lockwood, C. M., & Williams, J. (2004). Confidence
including manifest as well as latent variables. In addition, the limits for the indirect effect: Distributions of the product of resampling
approach allows for the bootstrapping of confidence intervals. methods. Multivariate Behavioral Research, 39, 99 –128. doi:10.1207/
These features render the method particularly suitable for applied s15327906mbr3901_4
Mooney, C. Z., & Duval, R. D. (1993). Bootstrapping: A nonparametric
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.

researchers who would like to test and compare specific effects


approach to statistical inference. Newbury Park, CA: Sage.
This document is copyrighted by the American Psychological Association or one of its allied publishers.

that were hitherto— due to the complexity of methods required or


Rindskopf, D. (1984). Using phantom and imaginary latent variables to
due to restrictions of SEM programs— out of reach.
parameterize constraints in linear structural models. Psychometrika, 49,
37– 47. doi:10.1007/BF02294204
References Schafer, J. L. (1999). NORM: Multiple imputation of multivariate contin-
uous data under a normal model [Computer software]. Retrieved from
Arbuckle, J. L. (1996). Full information estimation in the presence of http://www.stat.psu.edu/⬃jls/misoftwa.html
incomplete data. In G. A. Marcoulides & R. A. Schumacker (Eds.), Shrout, P. E., & Bolger, N. (2002). Mediation in experimental and non-
Advanced structural equation modeling: Issues and techniques (pp. experimental studies: New procedures and recommendations. Psycho-
243–277). Mahwah, NJ: Erlbaum. logical Methods, 7, 422– 445. doi:10.1037/1082-989X.7.4.422
Berry, W. D. (1984). Nonrecursive causal models. Thousand Oaks, CA: Sobel, M. (1986). Some new results on indirect effects and their standard
Sage. errors in covariance structure models. In N. B. Tuma (Ed.), Sociological
Bollen, K. A. (1987). Total, direct, and indirect effects in structural methodology (Vol. 16, pp. 159 –186). Washington, DC: American So-
equation models. In C. C. Clogg (Ed.), Sociological methodology (Vol. ciological Association.
17, pp. 37– 69). Washington, DC: American Sociological Association. SPSS. (2007). SPSS Missing Values 17.0: Reference manual. Chicago, IL: Author.
doi:10.2307/271028 Retrieved from http://support.spss.com/ProductsExt/SPSS/ESD/17/Download/
Bollen, K. A. (1989). Structural equations with latent variables. New User%20Manuals/English/SPSS%20Missing%20Values%2017.0.pdf
York, NY: Wiley. Stine, R. (1989). An introduction to bootstrapping methods: Examples and
Bollen, K. A., & Stine, R. A. (1990). Direct and indirect effects: Classical ideas. Sociological Methods & Research, 18, 243–291. doi:10.1177/
and bootstrap estimates of variability. In C. C. Clogg (Ed.), Sociological 0049124189018002003
methodology (Vol. 20, pp. 115–140). Oxford, England: Blackwell. Tsai, W.-C., Chen, C.-C., & Liu, H.-L. (2007). Test of a model linking
Cheung, M. W.-L. (2007). Comparison of approaches to constructing employee positive moods and task performance. Journal of Applied
confidence intervals for mediating effects using structural equation Psychology, 92, 1570 –1583. doi:10.1037/0021-9010.92.6.1570
models. Structural Equation Modeling, 14, 227–246. Woody, E., & Sadler, P. (2005). Structural equation models for inter-
Jackson, D. L., Gillaspy, J. A., Jr., & Purc-Stephenson, R. (2009). Report- changeable dyads: Being the same makes a difference. Psychological
ing practices in confirmatory factor analysis: An overview and some Methods, 10, 139 –158. doi:10.1037/1082-989X.10.2.139
recommendations. Psychological Methods, 14, 6 –23. doi:10.1037/
a0014694
Kenny, D. A. (1996). Models of non-independence in dyadic research. Received June 26, 2009
Journal of Social and Personal Relationships, 13, 279 –294. doi: Revision received August 3, 2010
10.1177/0265407596132007 Accepted September 2, 2010 䡲

You might also like