Mediation Analysis Examples

1
Mediation Analysis Examples

• David P. MacKinnon, Arizona State University
• Health Maintenance Consortium, Washington, DC,
  March 19, 2007

Application of mediation analysis and special
topics Firefighter study (PHLAME) Health
promotion for adolescent girls (ATHENA) Mediterran
ean diet study (MLP) Special topics from
questions Categorical outcomes, path analysis,
moderators and mediators.
2
Research Design Requirements

• Longitudinal study containing a baseline and at
  least one follow-up measurement. More follow-ups
  allow for more comprehensive modeling of
  intervention effects.
• Delivery of a program or programs to at least one
  goup of units. Effects in the program group are
  compared to a comparison or control group.
• Randomization may be at the unit of individual,
  high school, fire station shift etc.
• Measurement of at least one major outcome
  variable.
• Measurement of at least one hypothesized
  mediating variable.

3
Mediation model in Prevention
Knowledge
Beliefs
Program
Drug Use
Norms
Skills
4
Reasons for Mediation Analysis in Prevention
Research

• Mediation is important for prevention science.
  Practical implications include reduced cost and
  more effective interventions.
• Mediation analysis is based on theory for the
  processes underlying programs. Action theory
  corresponds to how the program will affect
  mediatorsthe X to M relation. Conceptual Theory
  focuses on how the mediators are related to the
  dependent variablesthe M to Y relation (Chen,
  1990, Lipsey, 1993).

5
Two-wave Longitudinal Model
b1
POST-TEST OUTCOME
BASELINE OUTCOME
b5
BASELINE MEDIATOR
b2
POST-TEST MEDIATOR
c3
a4
Mediated effecta4b5 Direct effect c3
PROGRAM
6
Autoregressive Model with Time-Ordered and
Contemporaneous Mediation
Note All residuals are correlated
7
Latent Growth Curve Model
8
Latent Growth Model (LGM)

• LGM model change over time by estimating an
  intercept and slope for change in variables.
  These models can be used to investigate mediation
  by estimating change over time for the mediator
  and change over time for the outcome. The
  relation between the change in the mediator and
  change in the outcome represents the b path
  (Cheong et al. 2003).
• The causal direction of correlated change is
  ambiguous. Another LGM estimates change in the
  mediator at earlier time points and relates to
  change in the outcome at later time points
  providing more evidence for temporal precedence
  of the mediator.

9
PHLAME Mediation

• Fairchild, A.J., Elliot, D., MacKinnon, D. P.,
  Moe, E. L., et al.
• Diane Elliot, Principal Investigator
• Funded by National Cancer Institute
• 5 a day project, Office of Behavioral and Social
  Sciences Research (OBSSR)

10
Project Overview

• 599 Pacific Northwest firefighters
• Subjects randomized at station level
• Compared two methods of behavior change and a
  control conditions ability to promote healthy
  eating and exercise habits in participants

11
Conditions

• TEAM intervention
• Designed to increase task cohesion among team
  members such that the eating/exercise habits of
  others would increase personal intentions to
  improve diet and exercise (Moe et al., 2002).
• MI intervention
• Individual counseling designed for eliciting
  behavior change
• Control
• Subjects received pamphlet on healthy
  eating/exercise habits

12
TEAM Conceptual Theory
30 minutes Daily Physical Activity
Peer Influence (Modeling)
Decrease Fat consumed In Diet
Perceived Norms
Increase Daily Fruit/Vegetable Intake
Shared Behavioral Values
Decrease Body Fat And BMI
13
MI Conceptual Theory
Individual State of Change
Improved Diet and Exercise
14
TEAM Action Theory
Peer Influence (Modeling)
Peer-taught Weekly Meetings
Perceived Norms
Reward System for Meeting Goals
Shared Behavioral Values
Group Collaboration
15
MI Action Theory
Motivational Interviewing
Individual State of Change
16
PHLAME Mediation

• Change in mediator and outcome scores across
  measurement waves Time 1 and Time 2 were
  analyzed.
• Mediated effects for MI did not reach
  conventional levels of significance. Evidence for
  several mediators of the TEAM program as
  described below.

17
PHLAME Mediation TEAM
18
PHLAME Mediation Summary

• Analysis of first follow-up indicates that
  program effects on number of fruits and
  vegetables were mediated by changes in intentions
  to eat fruit and vegetables and ratings that
  co-workers eat fruit and vegetables.
• Future work will examine alternative longitudinal
  models with additional waves.

19
ATHENA Mediation
Athletes Targeting and Healthy Exercise and
Nutrition Alternatives Ranby, Elliot, MacKinnon,
Moe, Goldberg, McGinnis Funded by the
National Institute on Drug Abuse
20
Overview of ATHENA

• 1668 female high school athletes
• 807 ATHENA 861 Control at baseline
• 606 ATHENA 655 Control at follow-up
• Baseline measures before season
• 8 sessions throughout the season
• Health promotion and drug prevention
• Peer-led, coach facilitated in team setting
• Follow-up measures at end of season (3-4 months
  later)

21
Conceptual Theory
Mediators
Outcome
Social Norms
Perceived Susceptibility
Intentions to use body shaping drugs
Beliefs about self
Knowledge
Self-efficacy
Values
22
Action Theory
Program components
Mediators
Discussing healthy habits with peers and coach
Social Norms
Critically analyzing images in magazines
Perceived Susceptibility
Learning that bad effects of body shaping drugs
are common
Beliefs about self
Cognitive restructuring skills training
Knowledge
Learning about harmful effects of drugs
Self-efficacy
Teaching specific skills for healthy behaviors
Values
Spending a lot of time focused on healthy
behaviors
23
Single Mediator Model
Social Norm Belief magazine ads are
truthful (Baseline)
Social Norm Belief magazine ads are
truthful (Follow-up) R²
.174
.415
-.336
.157
ATHENA 1 treatment 0 control
.159
-.087
Intent to use body shaping drugs (Baseline)
Intent to use body shaping drugs (Follow-up)
R² .503
.724

• plt.05, plt.01, plt.001. N 1668.
• Clustering accounted for on the school level.
• Model fit ?² (4)13.360, p .0096 CFI .989
  RMSEA .037 SRMR .026.
• Indirect effect is significant. ß -.053
• Asymmetric confidence limits LCL -.08212 UCL
  -.02732

24
Mediated EffectsBaseline to Follow-up for 8
mediators
Note plt.05 plt.01 plt.001
25
Multiple Mediator Model
Self-efficacy (Baseline)
.492
Self-efficacy (Follow-up)
.256
Personal Value (Baseline)
Personal Value (Follow-up)
.313
.338
Social Norm Coach (Baseline)
Social Norm Coach (Follow-up)
.451
-.144
-.340
Social Norm Magazines (Follow-up)
Social Norm Magazines (Baseline)
.364
Perceived Susceptibility (Follow-up)
Perceived Susceptibility (Baseline)
-.220
.250
Knowledge alcohol (Baseline)
Knowledge alcohol (Follow-up)
.275
.011
.188
-.086
.988
Knowledge steroids (Baseline)
.046
Knowledge Steroids (Follow-up)
.284
Knowledge vomiting, etc. (Baseline)
Knowledge vomiting, etc. (Follow-up)
.178
.124
-.031
.268
.268
.003
ATHENA 1 treatment 0 control
-.027
-.026
Intent to use body shaping drugs (Baseline)
Intent to use body shaping drugs (Follow-up) R²
.558
.618
Note plt.05 plt.01 plt.001
26
Summary

• The effect of ATHENA on intentions to use body
  shaping drugs was mediated by
• Perceived susceptibility to diet pills
• - Social Norms
• Belief magazine ads are truthful
• Belief coach watches weight
• Self-efficacy for how to eat to be an athlete
• All 4 mediators lead to lower intentions.

27
Longitudinal models for the Mediterranean
Lifestyle Project

• Mediterranean Lifestyle Project (MLP) is an
  intervention for postmenopausal women diagnosed
  with type 2 diabetes. Funded by National Heart,
  Lung, and Blood Institute, 5 a day project
  (OBSSR).
• Randomized patients to physician (UC) usual care
  (N 116) or MLP (N163) assessed at baseline and
  6, 12, and 24 months following the initial
  assessment so four waves of data.
• Deborah J. Toobert (Oregon Research Institute)
  principal investigator. These results taken from
  Barrera, Strycker, MacKinnon, Toobert (in
  press, Health Psychology) and also Barrera et
  al., (2006, Journal of Health Psychology).
• Program delivered after baseline measurement. In
  general, change in mediators should be quick.
  Two outcomes, saturated fat consumption and
  physical activities, and one mediator
  social-ecological resources.

28
MLP Action and Conceptual Theory

• Action Theory Dietitian taught Mediterranean
  diet, advised to build up physical activity, yoga
  instruction, support groups led by professionals.
  
• Conceptual Theory social relationships are
  related to diet and physical activity. Social
  support considered broadly including friends,
  family, and neighborhood.

29
Analysis decisions

• Used steps From MacKinnon (2007)
• Step 1 Estimate longitudinal model for mediator
  and outcome separately to assess whether slopes
  significantly vary among patients. Code time to
  reflect that the measures are not equally spaced,
  e.g., 0, 6, 12, 24.
• Step 2 If slopes significantly vary among
  patients, estimate mediated effects within the
  LGM model following Cheong et al., (2003). If
  there was not evidence that intercepts and slopes
  differed among participants, then estimate an
  autoregressive mediation model (Cole Maxwell,
  2004). Generally, investigating mediation in the
  autoregressive model is more straightforward.

30
Latent Growth Model (LGM)

• LGM model change over time by estimating an
  intercept and slope for change in variables.
  These models can be used to investigate mediation
  by estimating change over time for the mediator
  and change over time for the outcome. The
  relation between the change in the mediator and
  change in the outcome represents the b path
  (Cheong et al. 2003).
• The causal direction of correlated change is
  ambiguous. Another LGM estimates change in the
  mediator at earlier time points and relates to
  change in the outcome at later time points
  providing more evidence for temporal precedence
  of the mediator.

31
Mediated effect in the Autoregressive Model

• Time-ordered mediation
• Path from X to M2 and the path from M2 to Y3
• Path from X to M2 and the path from M2 to Y4
• Path from X to M3 and the path from M3 to Y4
• Contemporaneous mediation
• Path from X to M2 and the path from M2 to Y2
• Path from X to M3 and the path from M3 to Y3
• Path from X to M4 and the path from M4 to Y4

32
Autoregressive mediation effects for the Diet and
Saturated Fat Model

• Program effect on diet at first followup to
  saturated fat at the first followup. Explanation
  of this was that change in diet occurred before
  the first followup and that led to change in
  saturated fat.
• Program effect on diet at the first followup to
  diet at the second followup to saturated fat at
  the second and third followups.
• Longitudinal mediation relation Program effect
  on diet at the second followup to saturated fat
  at the third followup.

33
MLP Mediation Summary

• There was evidence for mediation relations in the
  MLP for diet and physical activity.
• However, the pattern of mediation relations was
  complicated suggesting some patterns of mediation
  such that effects on mediators at earlier times
  had complicated relations with subsequent effects
  on outcomes at later times.
• Demonstrates the utility of the longitudinal
  mediation model for the analysis of real data.

34
HMC and 5 a day Mediation Opportunities

• Uniform sets of analyses can be conducted on data
  from different projects.
• Different domains but similar theoretical
  background across many outcome variables.
• Usually randomized interventions making
  interpretation of results more straightforward.
• Opportunity for different studies to provide
  replications.

35

• THE END