All Things in
Tests of Mediation
Jeffrey B. Vancouver
Charlie M. Thompson
Covariance tests of mediation are
questioned regarding their internal
validity. Specifically, eight alternative
explanations for a finding of mediation
are described. An alternative method
using a moderation design is proposed.
Implications regarding the potential use,
and advantages and disadvantages of this
approach are described.
Figure 1. Mediator and moderator models. 3
• Few concepts confuse researchers more readily than the
distinction between mediation and moderation.
– Mediation refers to a process or mechanism through which
one variable (i.e., exogenous) causes variation in another
variable (i.e., endogenous) (see Figure 1a).
– Moderation refers to the influence of a process or mechanism
on the degree or kind of co-variation between the exogenous
and endogenous variables (see Figure 1b).
• Conceptually and analytically important
– Mediation is analyzed via differences in partial and whole
– Moderation is analyzed via interactions.
• We argue that studies designed to test for moderation
may provide stronger tests of mediation than the partial
and whole covariance approaches typically used (e.g.,
Baron & Kenny, 1986; Bing, Davison, LeBreton, &
LeBreton, 2002; James & Brett, 1984).
– The fundamental issue is the weak internal validity of passive
observational methods (Runkel & McGrath, 1972; Shadish,
Cook, & Campbell, 2002).
– Specifically, the issue is that the covariance methods for
testing mediation are susceptible to numerous alternative
explanations (Shadish, et al., 2002).
• Figure 2 shows alternative models when the exogenous/independent
variable is manipulated.
• Figure 3 shows alternative models when all variables are passively
• Incidence of problem
– Use of the covariance methods are widespread, particularly in
the applied psychological literature (Bing, et al., 2002;
MacKinnon et al., 2002),
– whereas use of the moderation method, either experimentally
or passively, to assess mediation is rare
– Figure 4 shows the results of a review of the Journal of
Applied Psychology from 1999 through 2002.
• Potential solution
– Use moderation (especially manipulated) to assess hypothetical
mediators (see Figure 5).
– Consider “automotive science example.”
Alternatives Explanations for
Positive Test of Mediation
A. "Endogenous" variable causes "mediator" 7
B. "Exogenous" variable causes "mediator" and endogenous variable independently.
C. "Exogenous" variable causes "mediator" more quickly then it causes the endogenous
Figure 2. Alternative models when the exogenous/independent variable is
A. "Endogenous" variable causes the "exogenous" variable, which causes the "mediator".
B. The “mediator” mediates effect of the "endogenous" variable on the "exogenous" variable.
C. The “mediator” causes the "endogenous" variable, which causes the "exogenous“ variable.
D. The “mediator” causes "exogenous" variable, which causes the "endogenous" variable.
E. Some fourth variable causes two or more of the others in some order other than the one
Figure 3. Additional alternative models when no variable is manipulated.
76 (27%) 99 (35%)
tested for 19 tested for
mentioned the moderator
implied a mediator
Figure 4. Incidence of tests of mediation and moderation in the Journal
of Applied Psychology (volumes 84-86). 11
The Automotive Science Example
Research Question: Why does pressing the gas pedal increase speed of car?
Hypothesis 1: Drive train mediates relationship between pedal and speed.
Hypothesis 2: Alternator mediates relationship between pedal and speed.
Results of standard covariance approach to tests of mediation:
Both hypotheses are supported.
Moderation research designs:
Hypothesis 3: Wrench in drive train will affect
relationship between pedal and speed.
Hypothesis 4: Disconnecting alternator from car will affect relationship.
Results of moderation tests:
Hypothesis 3 was supported but Hypothesis 4 was not!
Conclusion: Drive train is a mediator but alternator is not.
Figure 5. Test of mediation using moderation. 13
• Tests of mediation through manipulated moderation are
– Confounds in manipulation can lead to invalid interpretations.
– No manipulation possible or possible manipulation weak.
• Tests of mediation through passive observation are
likely to be . . .
– heavily confounded or
– analytically suspect (power, analysis issues).
– manipulation studies are generally more internally valid,
– require clever designs (e.g., task load paradigm to moderate
attentional resources available), and
– represent many of the issues I/O Psychologists find important
• E.g., why do context/job dimensions matter?
• Or, when are they likely to matter given the underlying mechanisms? 14
• We are not suggesting that covariance techniques for
assessing mediators should be abandoned. There will
always be cases where it is the only viable choice or other
considerations (e.g., external validity) suggest favoring
the approach. Rather, our objectives are threefold:
– First, we wanted to add some humility to the discussion
sections of covariance mediator studies.
– Second, we wanted to motivate researchers to develop
manipulations or measures of moderators that might allow for
cleaner inferences of mediation.
– Finally, we wanted to increase the likelihood that researchers
will consider the understandings their findings of moderation
might provide regarding mediating mechanisms.
Baron, R. M. & Kenny, D. A. (1986). The moderator-mediator variable distinction in
social psychological research: Conceptual, strategic, and statistical considerations.
Journal of Personality and Social Psychology, 51, 1173-1182.
Bing, M. N., Davison, H. K., LeBreton, D. L., & LeBreton, J. M. (2002, April). Issues
and improvements in tests of mediation. Poster presented at the annual meeting of the
Society for Industrial and Organizational Psychology, Toronto, Canada.
James, L. R. & Brett, J. M. (1984). Mediators, moderators and tests of mediation.
Journal of Applied Psychology, 69, 307-321.
MacKinnon, D. P.., Lockwood, C. M., Hoffman, J. M., West, S. G., & Sheets, V. (2002).
A comparison of methods to test mediation and other intervening variable effects.
Psychological Methods, 7, 83-104.
Runkel, P. J. & McGrath, J. E. (1972). Research on human behavior: A systematic guide
to method. New York: Holt, Rinehart, and Winston.
Shadish, W, R, Cook, T. D. & Campbell, D. T. (2002). Experimental and Quasi-
Experimentation Designs for generalized causal inference. Boston: Houghton