Professional Documents
Culture Documents
2010 Moderation SEM Lin
2010 Moderation SEM Lin
2010 Moderation SEM Lin
Zhonglin Wen
Center for Studies of Psychological Application, South China Normal University, China
Herbert W. Marsh
Department of Education, Oxford University, UK
Huey-Shyan Lin
Department of Nursing Management, Fooyin University, Taiwan
Correspondence should be addressed to Zhonglin Wen, Center for Studies of Psychological Application, South
China Normal University, Shipai, Guangzhou, 510631, China. E-mail: wenzl@scnu.edu.cn
374
DOUBLE-MEAN-CENTERING STRATEGY 375
research by Kenny and Judd (1984; Jöreskog & Yang, 1996). Marsh, Wen, and Hau (2004,
2006) reviewed approaches leading to what they referred to as the mean-centered constrained
approach based on Algina and Moulder’s (2001) extension of earlier models by Jöreskog and
Yang (1996). Although related to earlier work by Jaccard and Wan (1995), Jöreskog and Yang
(1996) argued that the appropriate models for latent interaction effects typically require the
inclusion of a mean structure—a feature of their model that was not included by Jaccard and
Wan. The critical feature of the mean-centered constrained approach proposed by Algina and
Moulder was that each of the indicators of the first-order term was mean-centered, but in other
respects it was like the model proposed by Jöreskog and Yang (1996).
In contrast to these constrained approaches, Marsh et al. (2004, 2006) proposed the uncon-
strained approach that they demonstrated was much easier to specify. As with the constrained
approaches, they used a mean-centering strategy and, following Jöreskog and Yang (1996;
Jöreskog, 1998; Yang, 1998), they argued that their model—like the constrained model—also
required that researchers include a mean structure. However, their unconstrained approach
was fundamentally different in that their model did not contain any complicated nonlinear
constraints (that were highly dependent on normality assumptions) to define relations between
product indicators and the latent interaction factor. Marsh et al. (2004) argued that an important
advantage of their unconstrained approach was that it was much easier for applied researchers to
implement. Importantly, their simulation results demonstrated that their unconstrained approach
typically resulted in similar results to the constrained approach when the assumptions of
normality imposed by the constrained approach were met, but consistently performed better
than the constrained approach when normality assumptions were violated.
The mean-centered unconstrained approach had important advantages over the mean-centered
constrained approach, but still required analyses be based on a mean structure, making the
analyses somewhat cumbersome for applied researchers. However, subsequent developments
have called into question the need for including the mean structure. Noticing that the centering
strategy can reduce or eliminate correlations among first-order and interaction (product) factors
and associated multicollinearity problems, Little, Bovaird, and Widaman (2006) used the
orthogonalizing strategy based on earlier work by Lance (1988) in combination with the Marsh
et al. (2004) unconstrained approach. Their simulation results indicated that results based on the
orthogonalizing strategy were similar to the mean-centering strategy when the unconstrained
approach was used. Marsh et al. (2007) compared the mean-centering strategy and orthogonal-
izing strategy, pointing out the similarities and differences between the two strategies. One of
the important similarities is that the interaction effects are invariant across the two strategies
whereas the first-order effects might differ—sometimes substantially. An important difference
between the two emphasized by Marsh et al. (2007) is that the mean structure is a necessary
component of the interaction model when using the mean-centering strategy (Marsh et al.,
2004, 2006) but apparently not when using the orthogonalizing strategy (Little et al., 2006).
When using mean-centered data, Jaccard and Wan (1995) and Coenders, Batista-Foguet,
and Saris (2008) also ignored the mean structure of the interaction model but did not fully
explain the theoretical or mathematical rationale for doing so, limitations on when this was
appropriate, and consequences when assumptions that allowed the mean structure to be ignored
were violated. In this investigation, we propose and recommend a new strategy—double-mean-
centering—that provides theoretical support for ignoring the mean structure of the interaction
model. We also show its merit in relation to the mean-centering and orthogonalizing strategies.
376 LIN ET AL.
When the effects of two manifest variables (X1 and X2 ) and their interaction (X1 X2 ) on an
outcome Y is considered, the following multiple regression equation is widely recommended
to estimate interaction effects (see, e.g., Aiken & West, 1991; Cohen & Cohen, 1983) under
supposition that the independent variables X1 and X2 are reasonably continuous variables
Y D “0 C “1 X1 C “2 X2 C “3 X1 X2 C e (1)
where “1 and “2 represent the first-order effects, “3 represents the interaction effect, and e is
a random disturbance term with zero mean that is uncorrelated with X1 , X2 and X1 X2 . One
problem is the product term X1 X2 might be highly correlated with the independent variables
X1 and X2 (so-called multicollinearity), which leads to instability in estimation of coefficients.
By mean centering the independent variables X1 and X2 (see, e.g., Aiken & West, 1991;
Cohen, Cohen, West, & Aiken, 2003), the preceding multicollinearity problem is typically
reduced substantially. Especially when X1 and X2 are bivariate normally distributed, the
(centered) product term .X1 X 1 /.X2 X 2 / is uncorrelated with X1 X 1 , and uncorrelated with
X2 X 2 , because the third-order moment is zero under normal assumption. The corresponding
regression equation with interaction term is
That is,
Y D .“00 “01 X 1 “02 X 2 C “03 X 1 X 2 / C .“01 “03 X 2 /X1 C .“02 “03 X 1 /X2 C “03 X1 X2 C e 0 (3)
X1 X2 b0 b1 X1 b2 X2
which is uncorrelated with both X1 and X2 (i.e., they are orthogonal). The corresponding
regression equation with interaction term is
That is,
Y D .“000 “003 b0 / C .“001 “003 b1 /X1 C .“002 “003 b2 /X2 C “003 X1 X2 C e 00 (5)
Therefore, the interaction effects based on the mean-centering and orthogonalizing strategies
are invariant, although the corresponding first-order effects might differ substantially.
DOUBLE-MEAN-CENTERING STRATEGY 377
Double-Mean-Centering
Now we combine the mean-centering and orthogonalizing strategies by first mean-centering
each of the observed variables and then orthogonalizing them. The orthogonalization is to find
the residual by regressing product term on individual variables (Little et al., 2006). When X1 and
X2 are bivariate normal, the product term .X1 X 1 /.X2 X 2 / is uncorrelated with X1 X 1 ,
and uncorrelated with X2 X 2 . The residual term formed by regressing the product term
.X1 X 1 /.X2 X 2 / on X1 X 1 and X2 X 2 is: .X1 X 1 /.X2 X 2 / .X1 X 1 /.X2 X 2 /
(refer to Appendix A for more details). Therefore, orthogonalizing leads to double-mean-
centering: First center each of the observed variables, and then form the product term and
recenter it.
For the general case (without normality assumption), .X1 X 1 /.X2 X 2 / can be correlated
with its components. However, we still can construct a regression model with a double-mean-
centered interaction term:
Y D “000 000
0 C “1 .X1 X 1 / C “000
2 .X2 X 2/
(6)
C “000
3 Œ.X1 X 1 /.X2 X 2/ .X1 X 1 /.X2 X 2 / C e 000
That is,
Y D Œ“000
0 “000
1 X1 “000 000
2 X 2 C “3 .X 1 X 2 / “000
3 .X1 X 1 /.X2 X 2 /
(7)
C .“000
1 “000 000
3 X 2 /X1 C .“2 “000 000
3 X 1 /X2 C “3 X1 X2 C e
000
“000
1 “000 0
3 X 2 D “1 “03 X 2 ; “000
2 “000 0
3 X 1 D “2 “03 X 1 ; “000 0
3 D “3
Hence, the interaction effects are identical based on double-mean-centering and mean centering.
Of particular importance, the first-order effects are also identical for the two strategies.
The only difference between the mean-centering and double-mean-centering strategies is the
intercept terms. When Y is also mean-centered, the intercept term “000 0 in Equation 6 is zero,
whereas “00 in Equation 2 is generally not zero. This implies that when all observed variables
are mean-centered, the intercept term is not necessary with double-mean-centering, whereas
the intercept term is always necessary with single-mean-centering. Although this distinction
is typically of little importance in the interpretation of the results (as the intercept is usually
ignored), it makes a substantial difference in the ease with which the model is specified,
particularly for applied structural equation modeling (SEM) researchers who are not familiar
with the extension of structural equation models to include mean structures.
˜ D ”1 Ÿ1 C ”2 Ÿ2 C ”3 Ÿ1 Ÿ2 C — (8)
378 LIN ET AL.
Marsh et al. (2007) argued that need for a mean structure or not is a key difference between the
mean-centering and orthogonalizing strategies. That is, the mean-centering strategy must always
include a mean structure—except the special case in which the correlation between the two first-
order terms is exactly zero (see Marsh et al., 2004, 2006); with the orthogonalizing strategy,
however, the mean structure is unnecessary (Marsh et al., 2007). Next we demonstrate that
the mean structure is also unnecessary with the double-mean-centering strategy. For example,
suppose that x1 ; x2 ; x3 are the indicators of Ÿ1 , and x4 ; x5 ; x6 are the indicators of Ÿ2 . Denote
x C as the mean-centered variable of x; that is, x C D x E.x/. The measurement equations
are:
The product term x2C x5C (say) then is mean-centered again (i.e., double-mean-centering), and
is denoted as .x2C x5C /C . Under assumptions that Ÿ1 , Ÿ2 , — and all •s have zero means, and each
is uncorrelated with the other (except that Ÿ1 and Ÿ2 are allowed to correlate), we have
where
•25 D œ2 Ÿ1 •5 C œ5 Ÿ2 •2 C •2 •5
Ÿ1 Ÿ2 E.Ÿ1 Ÿ2 /
is the interaction construct and has a zero mean. Therefore, when we use
x1C ; x2C ; x3C I x4C ; x5C ; x6C I .x1C x4C /C ; .x2C x5C /C ; .x3C x6C /C
as the indicators (SPSS syntax for creating double-mean-centered indicators is given in Ap-
pendix B) to model a latent interaction framework, we actually analyze the following structural
equation
Hence, ˜ has zero mean. In the present double-mean-centering strategy, all Y-indicators were
mean-centered as well as the X-indicators. As a result, the mean structure is unnecessary for
DOUBLE-MEAN-CENTERING STRATEGY 379
either X-indicators or Y-indicators (i.e., £0x s, £0y s in LISREL specification) or for the latent
factors (i.e., ›0 s in LISREL specification).
We have shown that the mean structure is unnecessary with the double-mean-centering
strategy, even though the double-mean-centering strategy is not always a special case of the
orthogonalizing strategy (for which we already know obviates the need for the mean structure;
Marsh et al., 2007). As emphasized earlier, eliminating the need for the mean structure greatly
simplifies the latent interaction model, as shown in a comparison of the path diagrams with
double-mean-centering (Figure 1a) and single-mean-centering (Figure 1b).
It is easy to understand that single-mean-centering without mean structure as proposed by
Jaccard and Wan (1995) and by Coenders et al. (2008) leads to the same estimates of both
the interaction and the first-order effects as the double-mean-centering. Therefore, the path
diagram corresponding to single-mean-centering without mean structure is Figure 1a. That is,
the interaction structure is Ÿ1 Ÿ2 E.Ÿ1 Ÿ2 /, not Ÿ1 Ÿ2 . Otherwise, when Ÿ1 Ÿ2 is used as the
interaction structure, the path diagram is Figure 1b, and the mean structure is necessary.
Both the double-mean-centering and orthogonalizing strategies eliminate the need for the
inclusion of a mean structure. However, the orthogonalizing strategy involves a cumbersome
two-step procedure in which preliminary multiple regressions must be conducted before the
indicators are used in the latent interaction model. This can be particularly tedious when
there are many indicators of the first-order constructs. Because we have now shown that the
mean structure can be eliminated with our double-mean-centering strategy that is so much
easier to conduct, much of the appeal for the orthogonalizing strategy is lost. However, we
further illustrate a potential flaw in the orthogonalizing strategy in the next section, showing
that the strategy causes structural inconsistency. That is, the interaction structure reflecting the
orthogonalized product indicators is not always orthogonal with Ÿ1 and Ÿ2 , as assumed by the
orthogonalizing strategy.
Unlike mean-centering and double-mean-centering, orthogonalizing does not start with indi-
cators that have been mean-centered. Rather, orthogonalizing generally starts with original
indicators (Little et al., 2006). To show its structural inconsistency, however, we consider the
special case in which each indicator has a zero mean. For example, suppose that x1 ; x2 ; x3 are
the indicators of Ÿ1 , and x4 ; x5 ; x6 are the indicators of Ÿ2 . The measurement equations are:
x1 D Ÿ1 C •1 ; x2 D œ2 Ÿ1 C •2 ; x3 D œ3 Ÿ1 C •3 (13)
x4 D Ÿ2 C •4 ; x5 D œ5 Ÿ2 C •5 ; x6 D œ6 Ÿ2 C •6 (14)
As usual, we assume that Ÿ1 , Ÿ2 , — and all •s have zero means, and each is uncorrelated with
the other (except that Ÿ1 and Ÿ2 are allowed to correlate).
In the orthogonalizing strategy, each product term is regressed on all of the first-order
indicators x1 ; x2 ; x3 ; x4; x5 ; x6 to produce a residual term. Taking x2 x5 as an example, the
(a)
(b)
FIGURE 1 Path diagram of latent interaction model using the unconstrained approach based on the double-
mean-centering strategy (a) and the mean-centering strategy (b). Boxes represent observed measured variables,
ovals represent latent constructs, and triangles represent intercept terms.
380
DOUBLE-MEAN-CENTERING STRATEGY 381
residual is:
O_x2 x5 D x2 x5 .b0 C b1 x1 C b2 x2 C b3 x3 C b4 x4 C b5 x5 C b6 x6 /
D .œ2 œ5 Ÿ1 Ÿ2 C œ2 Ÿ1 •5 C œ5 Ÿ2 •2 C •2 •5 /
where
•25 D œ2 Ÿ1 •5 C œ5 Ÿ2 •2 C •2 •5 .b1 •1 C b2 •2 C C b6 •6 /
The residual O_x2 x5 is used as an indicator of the interaction construct (Marsh et al., 2007)
where
b0
bQ0 D (16)
œ2 œ5
b1 C œ 2 b2 C œ 3 b3
bQ1 D (17)
œ2 œ5
b4 C œ 5 b5 C œ 6 b6
bQ2 D (18)
œ2 œ5
It can be shown that O_x1 x4 , O_x2 x5 , O_x3 x6 all reflect the same interaction construct through
tedious but elementary computations.
On the other hand, using the orthogonalizing strategy, we directly obtain the following
interaction construct:
Ÿ1 Ÿ2 .“0 C “1 Ÿ1 C “2 Ÿ2 / (19)
structure Ÿ1 Ÿ2 .bQ0 C bQ1 Ÿ1 C bQ2 Ÿ2 / reflecting the orthogonalized product indicators is not
always orthogonal with Ÿ1 and Ÿ2 . Therefore, there exists a problem with the consistency of
the orthogonalizing strategy unless the third-order moments are zero. This problem, when it
occurs, results in a nonrandom bias for each indicator of the interaction construct.
However, when the third-order moments are zero, the orthogonalizing and double-mean-
centering strategies are equivalent in the sense that both of them have the same first-order
and interaction effects (see Appendix A and C). Hence, there is no reason to use the more
cumbersome orthogonalizing strategy instead of the double-mean-centering strategy that is
so much easier to implement. Importantly, the problem with the consistency of estimates
based on the orthogonalizing strategy is eliminated with the double-mean-centering strategy
in that the interaction construct is always Ÿ1 Ÿ2 E.Ÿ1 Ÿ2 /. That is, the double-mean-centered
indicators (e.g., .x1C x4C /C , .x2C x5C /C , .x3C x6C /C / always reflects the interaction construct Ÿ1 Ÿ2
E.Ÿ1 Ÿ2 /.
To illustrate the results caused by the structural inconsistency of the orthogonalizing strategy,
we use simulated data following the Marsh et al. (2004) and Little et al. (2006) designs. The
(simulated) population data were created from a model with
where Ÿ1 and Ÿ2 were standardized variables, the variance of ˜ was 1, and the correlation
coefficient between Ÿ1 and Ÿ2 was 0.3. The loadings relating each indicator to its first-order
factor were all 0.7. Ÿ1 and Ÿ2 are simulated from the following four distribution conditions:
N.0; 1/, ¦2 .16/, ¦2 .8/, and ¦2 .4/. Except for N.0; 1/, the third-order moments of .Ÿ1 ; Ÿ2 / are
not zero. Like the Little et al. (2006) design, 1,000 replications were generated with sample
size of 1,500 for each distribution condition. The results used the unconstrained approach
(Marsh et al., 2004) with the double-mean-centering and orthogonalizing strategies shown in
Table 1.
For the double-mean-centering strategy, all of the theoretical first-order effects were 0.4,
interaction effects were 0.2, and the correlations between Ÿ1 and Ÿ2 were 0.3 under the four
distribution conditions. All of the maximum likelihood estimations (MLE) in Table 1 showed
small and acceptable levels of bias (less than 2%). Although the standard errors were likely
underestimated, the biases were obviously reduced when Satorra and Bentler (1988) corrected
standard errors were used.
For the orthogonalizing strategy, all of the theoretical interaction effects were still 0.2, and
the correlations between Ÿ1 and Ÿ2 were still 0.3 under the four distribution conditions. However,
the theoretical first-order effects were rather different. When the orthogonalizing strategy was
applied, the structural equation ˜ D 0:4Ÿ1 C 0:4Ÿ2 C 0:2Ÿ1 Ÿ2 C — became
(continued )
383
384
TABLE 1
(Continued )
Mean Value
Note. See Figure 1 for the definition of various parameters and models fit to each of the four distribution conditions considered here. % bias D Value 100%,
where Value D theoretical parameter (for effects and correlation), or SD of parameter (for SE).
DOUBLE-MEAN-CENTERING STRATEGY 385
where
The preceding formula can be obtained by applying Cramer’s rule (Cheney & Kincaid, 2009)
to Equations 26 and 27 in Appendix C. Then the theoreticalpfirst-order effects under N.0;p1/,
3 3 3
¦2 .16/, ¦2 .8/, and ¦2 .4/ conditions were 0.4, 0:4 C 130 2, 0:4 C 65 , and 0:4 C 65 2,
respectively. The theoretical first-order effects were still 0.4 under the normal distribution
condition. It can be seen from Table 1 that all of the parameter estimates other than the
interaction effect showed small and acceptable levels of bias (less than 2%). For example,
2
under theh ¦ .16/ condition,
the percentage
p i
bias of the first-order effect is approximately equal to
p
3 3
100% 0:4315 0:4 C 130 2 = 0:4 C 130 2 . 0:2668%/ when the orthogonalizing
strategy is adopted. Even if under the ¦2 .16/ condition, which introduced the smallest violation
of the normality assumption, the bias of interaction effect estimation was more than 16%. The
more serious the violation of the normality assumption and thus the assumption that the third-
order moments of .Ÿ1 ; Ÿ2 / are zero, the larger the bias of interaction effect estimation.
SUMMARY
TABLE 2
Comparisons Among the Mean-Centering, Orthogonalizing and Double-Mean-Centering
Double-Mean-
Characteristic Mean-Centering Orthogonalizing Centering
severely biased estimates when the assumption of normality is violated. As the result, the
proposed double-mean-centering strategy is recommended because of its ease of integration
into SEM. The double-mean-centering strategy can be combined with the pair-match strategy
in the formation of product indicators with the unconstrained approach recommended by Marsh
et al. (2004) and be conducted in any popular commercial SEM software.
In summary, we have considered a number of approaches to estimating latent interaction
effects that are particularly relevant when data are not normally distributed. This is important,
because nonnormal data are the rule and normally distributed data are the exception (Micceri,
1989; also see Malgady, 2007; Pyzdek, 1995). Historically, approaches to latent interactions
(e.g., Kenny & Judd, 1984) introduced highly restrictive constraints based on the assumption of
normally distributed predictor variables that result in inaccurate estimates when this assump-
tion is violated. Subsequent research (e.g., Jöreskog & Yang, 1996) demonstrated that this
constrained approach must be based on a full mean structure analysis, further complicating the
analyses and not resolving the inherent dependency on normally distributed predictor variables.
Although mean-centering the indicators (e.g., Algina & Moulder, 2001) reduced the amount of
bias to some extent, the constraints were still fundamentally dependent on normally distributed
predictor variables. However, Marsh et al. (2004, 2006) proposed an unconstrained approach,
demonstrating that removing the constraints altogether had little effect on solutions when
predictors were normally distributed and substantially improved the accuracy of estimation
for interaction effects when they were not. Building on this unconstrained approach, Little
et al. (2006; also see Marsh et al., 2007) proposed a two-stage orthogonalizing strategy to
be used with the unconstrained approach as an alternative to the mean-centering strategy.
Here we demonstrate a new double-mean-centering strategy that is superior to the mean-
centering and orthogonalizing strategies used in combination with the unconstrained approach
(Marsh et al., 2004, 2006), as outlined in Table 2. Based on this investigation, we recommend
that the double-mean-centering strategy should be used whenever estimating latent interaction
effects.
DOUBLE-MEAN-CENTERING STRATEGY 387
ACKNOWLEDGMENT
This research was funded in part by a grant from the National Natural Science Foundation of
China (30870784) to Zhonglin Wen.
REFERENCES
Aiken, L. S., & West, S. G. (1991). Multiple regression: Testing and interpreting interactions. Newbury Park, CA:
Sage.
Algina, J., & Moulder, B. C. (2001). A note on estimating the Jöreskog–Yang model for latent variable interaction
using LISREL 8.3. Structural Equation Modeling, 8, 40–52.
Cheney, W., & Kincaid, D. (2009). Linear algebra: Theory and applications. Sudbury, MA: Jones & Bartlett.
Coenders, G., Batista-Foguet, J. M., & Saris, W. E. (2008). Simple, efficient and distribution-free approach to interaction
effects in complex structural equation models. Quality & Quantity, 42, 369–396.
Cohen, J., & Cohen, P. (1983). Applied multiple regression/correlation analysis for the behavior sciences. Hillsdale,
NJ: Lawrence Erlbaum Associates.
Cohen, J., Cohen, P., West, S. G., & Aiken, L. A. (2003). Applied multiple regression/correlation analysis for the
behavior sciences (3rd ed.). Mahwah, NJ: Lawrence Erlbaum Associates.
Jaccard, J., & Wan, C. K. (1995). Measurement error in the analysis of interaction effects between continuous predictors
using multiple regression: Multiple indicator and structural equation approaches. Psychological Bulletin, 117(2),
348–357.
Jöreskog, K. G. (1998). Interaction and nonlinear modeling: Issue and approaches. In R. E. Schumacker & G. A.
Marcoulides (Eds.), Interaction and nonlinear effects in structural equation modeling (pp. 239–250). Mahwah, NJ:
Lawrence Erlbaum Associates.
Jöreskog, K. G., & Yang, F. (1996). Nonlinear structural equation models: The Kenny–Judd model with interaction
effects. In G. A. Marcoulides & R. E. Schumacker (Eds.), Advances in structural equation modeling techniques (pp.
57–88). Hillsdale, NJ: Lawrence Erlbaum Associates.
Kenny, D., & Judd, C. M. (1984). Estimating the nonlinear and interactive effects of latent variables. Psychological
Bulletin, 96, 201–210.
Lance, C. E. (1988). Residual centering, exploratory and confirmatory moderator analysis, and decomposition of effects
in path models containing interactions. Applied Psychological Measurement, 12, 163–175.
Little, T. D., Bovaird, J. A., & Widaman, K. F. (2006). On the merits of orthogonalizing powered and product term:
Implications for modeling interactions among latent variables. Structural Equation Modeling, 13, 497–519.
Malgady, R. G. (2007). How skewed are psychological data? A standardized index of effect size. Journal of General
Psychology, 134, 355–359.
Marsh, H. W., Wen, Z., & Hau, K. T. (2004). Structural equation models of latent interactions: Evaluation of alternative
estimation strategies and indicator construction. Psychological Methods, 9, 275–300.
Marsh, H. W., Wen, Z., & Hau, K. T. (2006). Structural equation models of latent interaction and quadratic effects. In
G. Hancock & R. Mueller (Eds.), A second course in structural equation modeling (pp. 225–265). Greenwich, CT:
Information Age.
Marsh, H. W., Wen, Z., Hau, K. T., Little, T. D., Bovaird, J. A., & Widaman, K. F. (2007). Unconstrained structural
equation models of latent interactions: Contrasting residual- and mean-centered approaches. Structural Equation
Modeling, 14, 570–580.
Micceri, T. (1989). The unicorn, the normal curve, and other improbable creatures. Psychological Bulletin, 105, 156–
166.
Pyzdek, T. (1995). Why normal distributions aren’t [all that normal]. Quality Engineering, 7, 769–777.
Satorra, A., & Bentler, P. M. (1988). Scaling corrections for chi-square statistics in covariance structure analysis.
In American Statistical Association 1988 Proceedings of the Business and Economics Sections (pp. 308–313).
Alexandria, VA: American Statistical Association.
Yang, F. (1998). Modeling interaction and nonlinear effects: A step-by-step LISREL example. In R. E. Schumacker &
G. A. Marcoulides (Eds.), Interaction and nonlinear effects in structural equation modeling (pp. 17–42). Mahwah,
NJ: Lawrence Erlbaum Associates.
388 LIN ET AL.
APPENDIX A
ORTHOGONALIZING IS EQUIVALENT TO DOUBLE-MEAN-CENTERING
UNDER THE NORMAL DISTRIBUTION ASSUMPTION
X1 X2 .b0 C b1 X1 C b2 X2 /
This implies that the residual of regressing X1 X2 on X1 , X2 is the same as the residual of
regressing .X1 X 1 /.X2 X 2 / on X1 X 1 and X2 X 2 , where the slopes are b1 X 2 and
b2 X 1 respectively, and the intercept is b0 C b1 X 1 C b2 X 2 X 1 X 2 .
Because X1 X 1 and X2 X 2 have bivariate normal distribution with zero mean, all the
third-order moments are zero and then .X1 X 1 /.X2 X 2 / is uncorrelated with X1 X 1 and
X2 X 2 . From the standard definition of the slope in regression, we obtain zero slopes when
regressing .X1 X 1 /.X2 X 2 / on X1 X 1 and X2 X 2 , that is, b1 X 2 D 0 and b2 X 1 D 0.
Furthermore, the intercept is .X1 X 1 /.X2 X 2 /, that is, b0 C b1 X 1 C b2 X 2 X 1 X 2 D
.X1 X 1 /.X2 X 2 /. Hence, the residual is .X1 X 1 /.X2 X 2 / .X1 X 1 /.X2 X 2 /, that
is, double-mean-centering.
Importantly, the residual of regressing X1 X2 on X1 , X2 is also .X1 X 1 /.X2 X 2 /
.X1 X 1 /.X2 X 2 /. Thus, orthogonalizing always leads to double-mean-centering even if
the first-order terms are not mean-centered. On the other hand, the double-mean-centering is
obviously the orthogonalizing when .X1 ; X2 / is bivariate normally distributed. Hence these two
strategies are equivalent.
APPENDIX B
SPSS SYNTAX FOR CREATING DOUBLE-MEAN-CENTERED
INDICATORS
AGGREGATE
/OUTFILE=*
MODE=ADDVARIABLES
OVERWRITEVARS=YES
/PRESORTED
/BREAK=GroupNo
/mY1 = MEAN(Y1) /mY2 = MEAN(Y2) /mY3 = MEAN(Y3)
/mX1 = MEAN(X1) /mX2 = MEAN(X2) /mX3 = MEAN(X3)
/mX4 = MEAN(X4) /mX5 = MEAN(X5) /mX6 = MEAN(X6).
* GroupNo is set to 1 when there is only one group of samples.
DOUBLE-MEAN-CENTERING STRATEGY 389
APPENDIX C
LATENT INTERACTION CONSTRUCT IS CONDITIONAL CONSISTENCY
WITH THE ORTHOGONALIZING STRATEGY
The residual
cov.xi ; O_x2 x5 / D 0; i D 1; 2; ; 6:
Hence,
where ¥11 D var.Ÿ1 /, ¥12 D cov.Ÿ1 ; Ÿ2 /, ¥13 D cov.Ÿ1 ; Ÿ1 Ÿ2 /, ™11 D var.•1 /. Similarly, we
have
¨_Ÿ1 Ÿ2 D Ÿ1 Ÿ2 .“0 C “1 Ÿ1 C “2 Ÿ2 /
œ2 ¥11 b1 C .œ22 ¥11 C ™22 /b2 C œ2 œ3 ¥11 b3 C œ2 ¥12 b4 C œ2 œ5 ¥12 b5 C œ2 œ6 ¥12 b6 D 0 (31)
œ3 ¥11 b1 C œ3 œ2 ¥11 b2 C .œ23 ¥11 C ™33 /b3 C œ3 ¥12 b4 C œ3 œ5 ¥12 b5 C œ3 œ6 ¥12 b6 D 0 (32)
œ5 ¥21 b1 C œ5 œ2 ¥21 b2 C œ5 œ3 ¥21 b3 C œ5 ¥22 b4 C .œ25 ¥22 C ™55 /b5 C œ5 œ6 ¥22 b6 D 0 (34)
œ6 ¥21 b1 C œ6 œ2 ¥21 b2 C œ6 œ3 ¥21 b3 C œ6 ¥22 b4 C œ6 œ5 ¥22 b5 C .œ26 ¥22 C ™66 /b6 D 0 (35)
DOUBLE-MEAN-CENTERING STRATEGY 391