Moderated Mediation in Career Development Research

Mediation

Technically, mediation is a statement about causality (MacKinnon, Lockwood, Hoffman, West, & Sheets, 2002; Preacher & Hayes, 2004; Rose, Holmbeck, Coakley, & Franks, 2004). That is, a mediation model specifies that the causal effect of one variable on another variable is through the effect of a third variable. The term “indirect effect” is sometimes used interchangeably with the term mediation (MacKinnon et al., 2002). Mediation studies are often conducted with the aim of helping to increase understanding of the processes that underlie the relation between two (or more) variables.

Using variables that are prominent in career development research, Figure 1 shows the concept of mediation in a path diagram (Baron & Kenny, 1986; Frazier, Tix, & Barron, 2004). In Figure 1, perseveration, the mediator (M), is hypothesized to serve as the mechanism through which X and Y come to be related. That is, the mediation model specifies that the causal effect of perfectionism on career indecision is through the effect of the third variable, perseveration. In other words, perfectionism tendencies lead an individual to perseverate about career goals, which, in turn, causes an individual to become more indecisive regarding career choices. As shown in Figure 1, perfectionism has an effect on the mediator, perseveration. This effect is represented by the symbol a in the figure. Additionally, after accounting for the effect of perfectionism on career indecision, perseveration is hypothesized to have an effect on career indecision, which is represented by the symbol b in the figure. Furthermore, the symbol c′ represents the effect of perfectionism (the predictor) on career indecision (the outcome) after accounting for the effect of perseveration (the mediator).

OPEN IN VIEWER

Figure 1.

Mediation model for career decision-making.

There are three general frameworks that have been proposed for testing mediation. These frameworks are the causal steps approach, the product of coefficients approach, and the bootstrap approach. Judd, Kenny, and Baron (Baron & Kenny, 1986; Judd & Kenny, 1981) developed the causal steps approach. According to this approach, four conditions must be met to establish mediation. The first condition that must be met under the causal steps approach is that there must be a relation between X and Y. For example, there must be a relation between perfectionism and career indecision. The second condition is that there should be a relation between X and the mediator. For example, there should be a significant relation between perfectionism and perseveration. The third condition specifies that when both the mediator and the predictor variables are included in the model, the mediator is a significant predictor of the outcome variable. For example, when we include perfectionism and perseveration in the model, perseveration should be a significant predictor of career indecision. The fourth condition is that the coefficient relating the predictor variable to the outcome variable is reduced when the mediator is included. For example, when we compare the model that includes both perfectionism and perseveration in the prediction of the career indecision to the model that does not include perseveration, the coefficient representing the link between perfectionism and career indecision should be lower when perseveration is in the model than when it is not. If the coefficient is reduced so that it is no longer significantly different from zero, then this is referred to as complete mediation. If the coefficient is reduced but remains statistically significant, then this is referred to as partial mediation (James & Brett, 1984).

The causal steps approach has been used often in the counseling literature. However, simulation studies and advances in our understanding of mediation have indicated that there are several limitations with this approach. A primary weakness is that there is low power for detecting a mediation effect if it exists (Rucker, Preacher, Tormala, & Petty, 2011). Additional weaknesses are that there is not a joint test of the four conditions, the approach does not provide a direct estimate of the indirect effect, and the approach does not provide a test of the significance of the indirect effect (MacKinnon et al., 2002).

To address some of the limitations of the causal steps approach, a second framework for testing mediation, the product of coefficients approach, was proposed. Under this framework, the mediation effect is the product of a and b (MacKinnon, Warsi, & Dwyer, 1995). That is, the effect is the product of the coefficients for the prediction of the mediator from the predictor variable (the a path) and the prediction of the outcome from the mediator variable (the b path). In the example mediation model, the effect would be computed as the product of the link between perfection and perseveration and that between perseveration and career indecision. A statistic for the indirect effect, or the cross product a × b, is then obtained by dividing the effect by its standard error and comparing the statistic to a critical value to evaluate its significance (MacKinnon et al., 2002). The standard error developed by Sobel (1982) is often used. The Sobel test overcomes the weaknesses of the causal steps approach by providing a test of the indirect effect. However, a primary limitation is that the estimate for the indirect effect (the a × b product) is compared with the normal distribution, although distributions of products are often positively skewed and not normally distributed (MacKinnon, Lockwood, & Williams, 2004). Thus, the Sobel test violates the assumption of normality, leading to low power.

The third framework for testing mediation is the bootstrap approach. Bootstrapping is a resampling strategy in which data are resampled from the original sample, with replacement. The resampling process is repeated k times. Usually, k is 1,000 or greater. The bootstrapping technique produces a reference distribution that can be used for significance testing and confidence interval (CI) estimation. Bootstrapping allows researchers to make statistical inferences without an a priori assumption about the distribution of a given statistic, such as a normal distribution. Given that the a × b product representing the indirect effect does not follow a normal distribution, bootstrapping has been shown to be particularly useful for estimating and testing indirect effects (Bollen & Stine, 1990; MacKinnon et al., 2004; Mallinckrodt, Abraham, Wei, & Russell, 2006).

Currently, research indicates that the bootstrap approach to testing mediation is superior to both the causal steps approach and product of coefficients approach with regard to power (MacKinnon, 2008; MacKinnon et al., 2004). Additionally, the bootstrap approach can be used even when sample sizes are small to moderate, such as in the range of 20–80 (Efron & Tibshirani, 1993). Therefore, bootstrapping is currently the recommended approach for testing mediation hypotheses.

When the bootstrapping method is applied to testing mediation, an indirect effect is calculated for each of the k bootstrapped samples, and the average of the k indirect effects serves as the estimated indirect effect (MacKinnon et al., 2004). A sampling distribution of the indirect effect is then obtained. CIs can be derived based on this distribution, and there is evidence that the bias-corrected CI is the most robust (MacKinnon et al., 2004). Effect sizes can also be computed to provide an indication of the strength of the indirect effect (Hayes, 2013; Preacher & Kelley, 2011). The bootstrap approach for testing mediation can easily be tested using existing statistical software such as IBM SPSS (version 21) and Mplus (Muthén & Muthén, 1998–2012).

Moderation

Although mediation focuses on the processes underlying a relation between variables, moderation serves to explain the conditions under which one variable (X) influences another variable (Y; Baron & Kenny, 1986; Kraemer, Wilson, Fairburn & Agras, 2002). An example of a moderator model in the field of career development is that the effect of career indecision (X) on life satisfaction (Y) depends on the level of working alliance in career counseling (the moderator, W). It may be that the link between career indecision and satisfaction with life is apparent only among individuals who experience a positive working alliance in the process of career counseling. Specifically, for those who experience a positive working alliance, there may be a negative relation between indecision and satisfaction. However, among those with negative working alliances, there may be a nonsignificant relation.

Although moderation effects can be very important in a practical sense, these effects tend to be small in magnitude and typically account for a small amount of unique variance when entered last in a regression analysis. For this reason, it can be very hard to detect true moderating effects in research, including career development research. Researchers must work hard to increase the power of moderator tests that are conducted, such as conducting power analyses to detect a sufficient sample size and ensuring that reliability estimates for all measures are as high as possible.

Depending on the levels of measurement for the variables of interest, tests of moderation can easily be conducted using approaches such as analysis of variance and multiple regression. Significant moderation can be probed using procedures described by Aiken and West (1991) to calculate simple slopes. Various statistical programs have been written to assist researchers in testing moderation and calculating and plotting simple slopes, such as Process by Hayes (2013), ModGraph by Jose (2008), and an Excel worksheet developed by Dawson and Richter (2006).

Moderated Mediation

The concept of moderated mediation represents a combination of mediation and moderation effects among variables in one model. That is, a moderated mediation model specifies that the mediation effect is dependent on the level or value of a moderator (Hayes, 2013; James & Brett, 1984; Rose et al., 2004). In other words, a moderated mediation model proposes that the process underlying the relation between two variables is different in different contexts (e.g., school and work environment). Such a model is essentially mediational at its base, and the moderator contextualizes the mediation effect.

Preacher, Rucker, and Hayes (2007) describe at least five ways in which a moderated mediation effect can operate. One way that moderated mediation can operate is that there may be an interaction between the predictor variable and the mediator in predicting the outcome variable. For example, there may be an interaction between perfectionism and perseveration in predicting career indecision. A second way is that there may be an interaction between the predictor variable and the moderator in predicting the mediator. For example, there may be an interaction between perfectionism and gender in predicting perseveration. A third way is that there may be an interaction between the mediator and moderator in predicting the outcome variable. For example, perhaps there is an interaction between perseveration and gender in predicting career indecision. A fourth way is that both paths a (the link between the predictor and mediator) and b (the link between the mediator and outcome) might be moderated, but by different variables. For example, the link between perfectionism and perseveration might be moderated by gender whereas the link between perseveration and career indecision is moderated by race. Finally, paths a and b might be moderated by the same variable. For example, gender may moderate the link between perfectionism and perseveration, as well as the link between perseveration and career indecision. Many more moderated-mediation models have also been described (e.g., Edwards & Lambert, 2007; Hayes, 2013) that can be useful for career development researchers.

Regardless of the model being tested, moderated mediation models are associated with the following two equations: (a) prediction of the mediator variable from the predictor variable and (b) prediction of the outcome variable from the predictor and mediator variables. Each of these equations would include any relevant moderators along with relevant interaction terms (Preacher, Rucker, & Hayes, 2007). For example, to test a moderated mediation hypothesis in which paths a and b are moderated by the same variable, two equations would be computed. One, there would be an equation for the prediction of perseveration (the mediator). That equation would include an interaction between perfectionism and gender. Two, there would be an equation for the prediction of career indecision (the outcome variable), which would include an interaction between perseveration and gender, along with other relevant terms. The significance of the coefficient representing the interaction term provides a test of moderated mediation.

Significant moderated mediation can be probed in a manner that is similar to the approaches recommended by Aiken and West (1991). Specifically, conditional indirect effects can be derived by computing the indirect effect at various levels of the moderator (Hayes, 2013). For example, the equation for computing the conditional indirect effects derived from Figure 2 would be

I n d i r e c t e f f e c t = a b + e × W .

1

OPEN IN VIEWER

Figure 2.

Moderated-mediation model for predicting vocational behavior.

In this equation, a is the coefficient for the prediction of the mediation (e.g., perseveration) from the predictor (e.g., perfectionism), b is the coefficient for the prediction of the outcome variable (e.g., career indecision) from the mediator (e.g., perseveration), e represents the effect of the moderator variable (e.g., gender), and W represents a specific level of the moderator (e.g., 0). Results of the analysis may show that the mediation effect only holds for females, but not males, which could be important to counselors. As with simple mediation analyses, bootstrapping can be used to calculate CIs for the conditional indirect effects. Moderated-mediation hypotheses can be tested using various statistical programs such as SPSS and Mplus, and scripts have been created to assist researchers to use these statistical methods (e.g., Hayes, 2013).

Use of Moderated Mediation in Career Development Research

It is evident that there has been increasing interest in tests of mediation and moderation in recent years. For example, a quick search in PsycInfo for “mediation and moderation” revealed fewer than 10 published articles in the 1980s, over 40 in the 1990s, nearly 300 in 2000s, and a projection of over 500 in the 2010s, thus demonstrating a growing appeal among researchers. However, currently, there appears to be a preference for researchers in career counseling and associated fields of counseling to examine simple relations among variables and use more modest statistical methodologies rather than advanced methodologies (Tracey, 2010). In line with this observation, a literature search on PsycInfo revealed over 18,000 publications for the term “mediation,” but only 2% of those (about 330 articles) were published in counseling-related journals such as those sponsored by the American Counseling Association. A search for the term “moderation” revealed nearly 3,500 publications overall, with about 3% in counseling-related journals. These results suggest that researchers in career counseling and related fields have focused on examination of direct relations between variables, such as “what is the relation between dysfunctional career thoughts and vocational behavior?” The infrequent use of more advanced approaches may be a function of several factors, including the pressure to publish articles at a faster rate to meet tenure demands, limited awareness of advanced statistical approaches, and limited access to resources that explain the statistical approaches in practical ways (Sharpe, 2013).

Yet, clients seeking satisfying and meaningful careers are presenting to career counselors with complex career-related problems. Work has been identified as a means of survival and power, social connection, and self-determination (Bluestein, 2006; Niles & Harris-Bowlsbey, 2013). Moreover, Bluestein (2006) identifies that social barriers can impact career decisions when fully considering issues of race, gender, sexual orientation, disability status, and social class. These concerns set the stage for researchers to ask more sophisticated questions, such as “what is the process underlying the relation between dysfunctional career thoughts and vocational behavior?” Current questions represent progressively more complicated career problems that are evident in the modern world of work. Moreover, such questions can easily be addressed given advances in theoretical and statistical understanding.

Moderated mediation analyses advance the analysis of simple relations among variables and can be used to help address complex research questions. This method of inquiry addresses the “by what process” and “under what conditions” questions. The examination of the processes and conditions under which relations between variables occur will ultimately allow pinpointing of specific areas to target for effective interventions and policies. For example, suppose a researcher finds that perfectionism is positively associated with career indecision. That is, suppose that individuals who have higher levels of perfectionism tend to be more indecisive in their career decision making. Although an important finding, this information alone will not sufficiently help career practitioners to develop effective interventions to alleviate the condition of career indecision. Rather, understanding of the underlying process by which the relation between career indecision and perfectionism occurs will lead to more informed career counselors and practitioners who can then better develop meaningful interventions. Mediating variables help us understand causal processes and moderating variables contextualize the relations between variables. Tests of mediation and moderation are quickly becoming dominant and more accepted ways to test processes, but the use of these methods of inquiry remains rare in counseling-related research (Tracey, 2010), including the field of career development.

Moderated mediation can help advance career development research and practice by identifying the moderating and mediating variables that affect the emerging needs within the modern world of work. Specifically, these questions could include: By what process does use of social networking and/or access to technology affect the efficacy of career counseling activities? By what process do governmental incentives for hiring practices affect the employment outcomes of returning veterans? Or, does an examination of the moderating effect of gender, race, or ethnicity affect the application of emerging theories utilized by career counselors? Subsequently, we provide an illustrative example of testing a moderate mediation hypothesis in career development research, using the Mplus statistical software.

Illustrative Analysis

To illustrate how a moderated mediation question can be tested, we propose that the mediated relation between dysfunctional career thoughts and vocational behavior via parental relationship quality is moderated by gender, such that the indirect effect is stronger for females than for males. The illustrative analysis model in Figure 2 depicts this hypothesis in which path b is hypothesized to be moderated by gender. Using the relations depicted in Figure 2 as a basis, we simulated normally distributed data for a pre-specified correlation matrix. This correlation matrix is available from the first author upon request. We generated a sample size of 200, which is a minimum sample size that has been suggested as adequate for testing moderated sized structural equation models (MacCallum, Browne, & Sugawara, 1996; Ullman, 1996) and is also a sample size that one might encounter in the field of counseling research (Heppner, Wampold, & Kivlighan, 2008).

One statistical software program that is available for testing moderated mediation hypotheses is the Mplus program (Muthén & Muthén, 1998–2012). Mplus is a latent variable modeling program that can be used to perform a variety of analyses, allowing for increased versatility. The Mplus script for testing a model with path b moderated is shown in the Appendix. The exclamation points in the script indicate comments. As shown, there are two regression equations modeled simultaneously: one equation for the mediator variable (i.e., parental relationship quality, in this case) and the second for the outcome variable (i.e., vocational behavior). Because our hypothesized model is one in which the moderator affects path b, the relevant interaction term is included in the outcome variable regression equation.

The “Model Constraint” command in the script is used to compute the conditional indirect effects. Under this command, specific indirect effects can be given a label via the NEW option and an equation is written specifying how to calculate the indirect effect. For a model with path b moderated, the conditional indirect effect equation is as presented in Equation 1. In this equation, W represents a specific value of the moderator. For a dichotomous variable coded 0 and 1, the values 0 and 1 would be substituted in for W. Both equations are included in the example script. For a continuous moderator, one can specify values that represent ±1 SD from the mean, as is often done in basic moderation analyses, and include these equations in the script.

Finally, under the output command, bias corrected bootstrapped CIs can be requested. The desired number of bootstrap draws can be specified prior to the MODEL command. In the present example, 5,000 bootstrapped samples were requested.

To assess the fit of a hypothesized model to the data, various goodness of fit indices can be examined. In general, the χ² statistic is often presented in the literature. However, the root-mean square error of approximation (RMSEA; Browne & Cudeck, 1993; Steiger & Lind, 1980) and comparative fit index (CFI; Bentler, 1990) have been found to be more robust indices (e.g., less susceptible to sample size bias) than the χ² statistic. As a rule-of-thumb, RMSEA values smaller than 0.10 (Browne & Cudeck, 1993) and CFI values above 0.95 (Hu & Bentler, 1999) are considered favorable. However, CFI values above 0.90 are also tenable (Bentler, 1990) and thus 0.90 continues to be a widely used cutoff value (Van Lieshout, Cleverley, Jenkins, & Georgiades, 2011).

The goodness of fit information for the illustrative model indicated a good fit (χ² [4] = 7.88, RMSEA = 0.07, CFI = 0.96, standardized root mean square residual = 0.04), which was expected given that the data were simulated. Table 1 displays selected model results from the Mplus output. Within this output, the estimated regression coefficients and their associated significance values, along with the estimated indirect effects, are provided. As can be seen in the output, the interaction effect in the regression equation for vocational behavior is significant, as indicated by a significant two-tailed p value for the variable labeled Parental Relationship Quality × Gender (PARRXGEN). This result indicates that the moderated mediation effect is significant.

OPEN IN VIEWER

Table 1.

Selected Mplus Output.

	Estimate	SE	Estimate/SE	2-Tailed p Value	Lower 5% CI	Upper 5% CI
PARRSHPQ ON
DYSFCT	0.35	0.07	5.28	<.01	0.24	0.51
VOCBEH ON
PARRSHPQ	0.19	0.06	2.99	<.01	0.08	0.34
DYSFCT	−0.09	0.06	−1.47	.14	−0.21	0.09
GENDER	3.34	1.19	2.81	<.01	0.05	0.34
PARRXGEN	0.50	0.06	8.52	<.01	0.40	0.64
GENDER WITH DYSFCT	1.00	0.36	2.77	<.01
Means
DYSFCT	50.00	0.71	70.89	<0.01
Intercepts
PARRSHPQ	32.50	3.38	9.62	<0.01
VOCBEH	18.83	4.64	4.06	<0.01
New/additional parameters
IndLow	0.07	0.03	2.60	<0.01	0.03	0.15
IndHi	0.24	0.05	4.41	<0.01	0.15	0.39

Note. The table has been modified from the exact Mplus output to conform to APA standards. APA = American Psychological Association; PARRSHPQ = parental relationship quality; DYSFCT = dysfunctional career thoughts; VOCBEH = vocational behavior; PARRXGEN = parental relationship quality × gender; CI = confidence interval.

Also in Table 1 are the indirect effects at specific levels of the moderator, under the new/additional parameters heading. In this table, IndLow represents the indirect effect when the moderator is zero and IndHi represents the indirect effect when the moderator is one, which are common values for dichotomous variables in research. Examination of the conditional indirect effects in this output indicates that while both effects are significant, the effect is stronger for the group indicated by a value of one on the moderator (e.g., females) versus the group indicated by a value of zero value on the moderator (e.g., males). CIs are also provided in the output, including the CIs for the conditional indirect effects, as shown in Table 1. By default, 90% (upper and lower 5%), 95% (upper and lower 2.5%), and 99% (upper and lower 0.5%) CIs are provided. CIs that do not contain zero suggest significant effects.

Further Considerations

Mediation, moderation, and moderated mediation are methods that can be used to address complex career development and vocational-related questions. However, various considerations should be taken prior to conducting career counseling outcome research using the aforementioned techniques. Foremost, theory and prior research must be used to determine which models to test. As previously mentioned in the section describing moderated mediation, there are at least five ways in which a moderated mediation effect can hold. Research conclusions regarding the associations between variables could differ depending on the model tested. To guard against reaching incorrect conclusions, career development educators and vocational researchers should develop hypotheses that are grounded in a well-established body of literature, such as self-efficacy beliefs theory, P-E congruence theory, and goal setting theory. Competing models can be tested to rule out alternative hypotheses. Tests of moderated mediation hypotheses can also serve to inform theory development in the field.

An additional consideration is that to establish causality, the predictor variable, X (e.g., perfectionism), must be related to the outcome variable, Y (e.g., career indecision), and also precede Y in time (Rosenthal & Rubin, 1982). Short of conducting counseling research using experimental designs, researchers are limited in the causal statements that they can make about the association between variables when using cross-sectional and/or correlational designs to test mediation hypotheses. Therefore, researchers must exercise caution in drawing causal conclusions when mediation analyses are conducted, particularly in cases where data are cross sectional and/or correlational. A final consideration is that the models described here can be extended to include multiple mediators, multiple predictor variables, multiple outcome variables, multilevel models, longitudinal model, and combinations of continuous and categorical variables. These more complex models can be tested using structural equation modeling in Mplus or any of the other existing software for conducting structural equation modeling analyses.

Implications for the Field of Career Counseling and Development

The development of moderated-mediation research will continue to be highly influential far into the future, judging by the number of meaningful publications sprouting forth in recent years in related fields of education and psychology (e.g., Edwards & Lambert, 2007; Frazier et al., 2004; Muller, Judd, & Yzerbyt, 2005). Moreover, current CACREP standards emphasize the importance of training in research methods and data-analytic techniques in the preparation of graduates from its accredited programs (CACREP, 2009). Contemporary career development and vocational-related difficulties clearly require the use of research methodology that matches emerging complex research questions. This article focused on introducing to some, and refreshing for others, the concept of moderated mediation and how it can be applied in the career development field. We encourage interested readers to also consult comprehensive literature to gain a deeper understanding of this methodology, such as Hayes (2013) and MacKinnon (2008).

Research methods in career development and other counseling-related fields must continue to move beyond examination of direct effects to help explain the processes underlying relations and the contexts under which relations hold. Some of the most esteemed concepts in the field of career development, such as those embedded in Holland’s (1973) theory of personality characteristics and vocational interest, the social cognitive theory of career development, career construction, and cognitive information processing theory of career development, to name a few, continue to prove useful in the development of career interventions. The field can benefit by expanding on these theories through methodological rigor. With the increased complexity of career-related issues in the 21st century world of work, career development research, policy, and practice will profit by widening its methodological lens to include approaches such as those presented in this article. Counseling researchers, particularly those interested in career development research, are called to explore moderated mediation in future research and support future researchers in their training on this essential methodology.