Assessing Mediation in HIV Intervention Studies Felicia Hardnett The findings and conclusions in this presentation are tho...
Mediator-Moderator Working Group <ul><ul><li>To review existing literature related to mediator/moderator analysis.  </li><...
What is Mediation?
What is Mediation? <ul><li>In behavioral intervention trials, mediation is the mechanism by which an intervention causes b...
Characteristics of Mediators <ul><li>“ In general, a given variable may be said to function as a  mediator  to the extent ...
Single Mediator Model Indirect effect: αβ Direct Effect:  τ ’ Total Effect: αβ +  τ ’ τ ’  Intervention Mediator Outcome α β
Mediation vs Confounding <ul><li>A mediator: </li></ul><ul><li>Can be changed by the intervention </li></ul><ul><li>Is spe...
Effect Suppression <ul><li>A suppressor is “a variable which increases the predictive validity of another variable (or set...
Characteristics of Suppressors <ul><li>Increase the association between the IV and DV. </li></ul><ul><li>Can be either a m...
Two Components of  Mediation Analysis <ul><li>Conceptual Component of Mediation </li></ul><ul><li>Behavioral Theory </li><...
Conceptual Components of an Effective Mediation Analysis  <ul><li>An established theory of behavior change </li></ul><ul><...
Conceptual Components of an Effective Mediation Analysis (cont’) <ul><li>Correct temporal ordering of modeling components....
Statistical Components of an Effective Mediation Analysis  <ul><li>Statistical Assumptions </li></ul><ul><ul><li>No signif...
Statistical Components of an Effective Mediation Analysis (cont’) <ul><li>Mathematical Modeling and Statistical Inference ...
Mediation in HIV Intervention Studies
Purpose of Mediation Analysis in HIV Intervention Studies <ul><li>Central question in HIV intervention studies are about m...
Common Mediators in Behavioral HIV Studies <ul><ul><li>Self-efficacy (e.g. for condom use) </li></ul></ul><ul><ul><li>HIV/...
How is Mediation Measured?
Baron and Kenny (1986) τ ’  Intervention Mediator Outcome α β
Baron and Kenny (1986)  Intervention Mediator Outcome α β τ ’
Baron and Kenny (1986) τ ’  Intervention Mediator Outcome α β
Baron and Kenny (1986) τ ’  Intervention Mediator Outcome α β
Baron and Kenny (1986) τ ’  Intervention Mediator Outcome α β
Causal Steps- Most Common <ul><li>According to Baron and Kenny (1986), a variable functions as a mediator when it meets th...
Controversy: Does the Intervention Effect Have to be Significant? <ul><li>A primary assumption of the Causal Steps approac...
Difference in Coefficients ( τ -  τ ’) <ul><li>Y =  β 01   +  τ X  +  ε 1   </li></ul><ul><li>Y =  β 02   +  τ ’ X  +  β Z...
Product of Coefficients ( αβ ) Indirect effect: αβ Direct Effect:  τ ’ Total Effect: αβ +  τ ’ τ ’  Intervention Mediator ...
Statistical Test of Mediation (I)  <ul><li>MacKinnon et al (1995) </li></ul><ul><li>z’= modified z statistic  </li></ul><u...
Statistical Test of Mediation (II) <ul><li>Asymmetric Confidence Limits </li></ul>Values of  δ α   and  δ β   are compared...
Example- SUMIT Trial <ul><li>SUMIT (Seropositive Urban Men’s Intervention Trial) </li></ul><ul><ul><li>Study Design: 2-arm...
Original SUMIT Analysis <ul><li>Association between IV and DV was not significant </li></ul><ul><li>Formal mediation analy...
SUMIT Re-analysis <ul><li>By applying less stringent criteria and constructing ACLs: </li></ul><ul><ul><li>Uncovered previ...
Conclusions <ul><li>Great opportunities for collaboration between behavioral scientists and statisticians </li></ul><ul><l...
Acknowledgements <ul><li>Sherri Pals </li></ul><ul><li>Craig Borkowf </li></ul><ul><li>Ann O’Leary </li></ul><ul><li>Jeffe...
Questions?
Upcoming SlideShare
Loading in …5
×

Assessing Mediation in HIV Intervention Studies

1,047 views
953 views

Published on

This presentation describes the use of asymmetric confidence limits to test for mediation when the direct effect was not significant and effect suppression was present.

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
1,047
On SlideShare
0
From Embeds
0
Number of Embeds
2
Actions
Shares
0
Downloads
13
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

Assessing Mediation in HIV Intervention Studies

  1. 1. Assessing Mediation in HIV Intervention Studies Felicia Hardnett The findings and conclusions in this presentation are those of the author(s) and do not necessarily represent the views of the Centers for Disease Control and Prevention.
  2. 2. Mediator-Moderator Working Group <ul><ul><li>To review existing literature related to mediator/moderator analysis. </li></ul></ul><ul><ul><li>To review a DHAP project that has used/will use this type of analysis. Consult with outside experts as needed. </li></ul></ul><ul><ul><li>Apply novel analytic approaches to determine whether additional information can be obtained to inform the design of future interventions </li></ul></ul>
  3. 3. What is Mediation?
  4. 4. What is Mediation? <ul><li>In behavioral intervention trials, mediation is the mechanism by which an intervention causes behavior change. </li></ul><ul><li>A mediator explains all or part of the intervention effect. </li></ul>
  5. 5. Characteristics of Mediators <ul><li>“ In general, a given variable may be said to function as a mediator to the extent that it accounts for the relationship between the predictor and the outcome…Mediators explain how external physical events take on internal psychological significance.” </li></ul><ul><li>Baron, R. M., & Kenny, D. A. The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 1986, 51 , 1173-1182. </li></ul>
  6. 6. Single Mediator Model Indirect effect: αβ Direct Effect: τ ’ Total Effect: αβ + τ ’ τ ’ Intervention Mediator Outcome α β
  7. 7. Mediation vs Confounding <ul><li>A mediator: </li></ul><ul><li>Can be changed by the intervention </li></ul><ul><li>Is specifically targeted by the intervention </li></ul><ul><li>Is an hypothesized component of the causal pathway between a predictor and an outcome. </li></ul><ul><li>A confounder: </li></ul><ul><li>Can be thought of as a nuisance variable </li></ul><ul><li>Is not targeted by the intervention </li></ul><ul><li>Usually cannot be changed by the intervention (e.g., age, gender, race) </li></ul><ul><li>Is not an hypothesized component of the causal pathway. </li></ul><ul><li>Falsely obscures or accentuates the association between an IV and DV. </li></ul>
  8. 8. Effect Suppression <ul><li>A suppressor is “a variable which increases the predictive validity of another variable (or set of variables) by its inclusion in a regression equation.” </li></ul><ul><li>Tzelgov. J., & Henrik, A. (1991). Suppression situations in psychological research: Definitions, implications and applications. Psychological Bulletin, 109, 524-536. </li></ul>
  9. 9. Characteristics of Suppressors <ul><li>Increase the association between the IV and DV. </li></ul><ul><li>Can be either a mediator or confounder </li></ul><ul><li>When the suppressor is a mediator (termed “inconsistent mediator”), the intervention actually has a negative effect on the outcome (i.e., the intervention causes an increase in risky behavior) </li></ul><ul><li>Can mask the association between the IV and DV in the presence of consistent mediators. </li></ul>
  10. 10. Two Components of Mediation Analysis <ul><li>Conceptual Component of Mediation </li></ul><ul><li>Behavioral Theory </li></ul><ul><li>Measurement Theory </li></ul><ul><li>Statistical Component of Mediation </li></ul><ul><li>Analytic assumptions </li></ul><ul><li>Mathematical modeling and statistical inference </li></ul>
  11. 11. Conceptual Components of an Effective Mediation Analysis <ul><li>An established theory of behavior change </li></ul><ul><li>A reliable and valid means of measuring behavioral constructs </li></ul><ul><li>Careful selection of potential mediators </li></ul><ul><ul><li>Often, behavioral studies are designed to include a large number of mediators </li></ul></ul><ul><ul><ul><li>Can result in a very complex analysis </li></ul></ul></ul><ul><ul><ul><li>How to handle multiple hypothesis testing problem? </li></ul></ul></ul><ul><ul><li>In design phase, study team should think through theoretical justification for inclusion of each mediator </li></ul></ul>
  12. 12. Conceptual Components of an Effective Mediation Analysis (cont’) <ul><li>Correct temporal ordering of modeling components. (i.e., X before M before Y) </li></ul><ul><li>Modeling components reflect true causal relations </li></ul>
  13. 13. Statistical Components of an Effective Mediation Analysis <ul><li>Statistical Assumptions </li></ul><ul><ul><li>No significant measurement error in variables. </li></ul></ul><ul><ul><li>Causal relationship between X, M and Y is correct. </li></ul></ul><ul><ul><li>No omitted variables. </li></ul></ul><ul><ul><li>No interaction between the intervention and the mediator (X and M). In other words, the association between the intervention and the outcome does not vary for different levels of the mediator. </li></ul></ul>
  14. 14. Statistical Components of an Effective Mediation Analysis (cont’) <ul><li>Mathematical Modeling and Statistical Inference </li></ul><ul><ul><li>Mathematical modeling </li></ul></ul><ul><ul><li>Statistical inference: </li></ul></ul>
  15. 15. Mediation in HIV Intervention Studies
  16. 16. Purpose of Mediation Analysis in HIV Intervention Studies <ul><li>Central question in HIV intervention studies are about mediating processes </li></ul><ul><li>Important for basic research on mechanisms of effect </li></ul><ul><li>Mediation analyses help to identify how an effective intervention works and why an ineffective one does not work </li></ul>
  17. 17. Common Mediators in Behavioral HIV Studies <ul><ul><li>Self-efficacy (e.g. for condom use) </li></ul></ul><ul><ul><li>HIV/AIDS Knowledge </li></ul></ul><ul><ul><li>Attitudes (related to protecting self, protecting partner, about condom use) </li></ul></ul><ul><ul><li>Intentions to use condoms </li></ul></ul><ul><ul><li>Outcome Expectancies (beliefs about the consequences of behavior) </li></ul></ul>
  18. 18. How is Mediation Measured?
  19. 19. Baron and Kenny (1986) τ ’ Intervention Mediator Outcome α β
  20. 20. Baron and Kenny (1986) Intervention Mediator Outcome α β τ ’
  21. 21. Baron and Kenny (1986) τ ’ Intervention Mediator Outcome α β
  22. 22. Baron and Kenny (1986) τ ’ Intervention Mediator Outcome α β
  23. 23. Baron and Kenny (1986) τ ’ Intervention Mediator Outcome α β
  24. 24. Causal Steps- Most Common <ul><li>According to Baron and Kenny (1986), a variable functions as a mediator when it meets the following conditions: </li></ul><ul><li>(a) variations in levels of the independent variable significantly account for variations in the presumed mediator (i.e., Path a ), </li></ul><ul><li>(b) variations in the mediator significantly account for variations in the dependent variable (i.e., Path b ), and </li></ul><ul><li>(c) when Paths a and b are controlled, a previously significant relationship between the independent and dependent variables is no longer significant, with the strongest demonstration of mediation occurring when Path c is zero. </li></ul>
  25. 25. Controversy: Does the Intervention Effect Have to be Significant? <ul><li>A primary assumption of the Causal Steps approach </li></ul><ul><li>Ignores the potential for suppressive effects </li></ul><ul><li>For long-term processes, power may be low to detect an X->Y effect </li></ul>
  26. 26. Difference in Coefficients ( τ - τ ’) <ul><li>Y = β 01 + τ X + ε 1 </li></ul><ul><li>Y = β 02 + τ ’ X + β Z + ε 2 </li></ul><ul><li>Y= Outcome </li></ul><ul><li>X= Intervention </li></ul><ul><li>Z= Potential Mediator </li></ul><ul><li>β 01 , β 02 = Intercepts </li></ul><ul><li>τ = coefficient relating independent and dependent variables (unadjusted) </li></ul><ul><li>τ ’ = coefficient relating independent and dependent variables adjusted for mediator effect. </li></ul><ul><li>ε 1 , ε 2 = Unexplained variability </li></ul>
  27. 27. Product of Coefficients ( αβ ) Indirect effect: αβ Direct Effect: τ ’ Total Effect: αβ + τ ’ τ ’ Intervention Mediator Outcome α β
  28. 28. Statistical Test of Mediation (I) <ul><li>MacKinnon et al (1995) </li></ul><ul><li>z’= modified z statistic </li></ul><ul><li>Empirical critical value for .05 significance is .97 </li></ul><ul><li>rather than 1.96 </li></ul>
  29. 29. Statistical Test of Mediation (II) <ul><li>Asymmetric Confidence Limits </li></ul>Values of δ α and δ β are compared with critical values in tables published by Meeker et al. UCL= αβ + M upper * σ αβ LCL= αβ + M lower * σ αβ
  30. 30. Example- SUMIT Trial <ul><li>SUMIT (Seropositive Urban Men’s Intervention Trial) </li></ul><ul><ul><li>Study Design: 2-arm randomized trial (“standard” intervention vs. “enhanced” intervention) </li></ul></ul><ul><ul><li>Outcome variables: any unprotected insertive anal sex with negative or unknown status partner </li></ul></ul><ul><ul><li>Potential mediator variables: HIV status assumptions, outcome expectations, personal responsibility, sexual compulsivity, anxiety, hostility, depression, drug use, peer norms, self-efficacy </li></ul></ul><ul><ul><li>Challenge: weak intervention effect </li></ul></ul>
  31. 31. Original SUMIT Analysis <ul><li>Association between IV and DV was not significant </li></ul><ul><li>Formal mediation analysis was abandoned in favor of assessing correlates of behavior change </li></ul><ul><li>O’Leary, A., Hoff, C. C., Purcell, D. W., Gomez, C.A., Parsons, J. T., Hardnett, F., Lyles, C. M. (2005). What happened in the SUMIT trial? Mediation and behavior change. AIDS, 19, S111-S121. </li></ul>
  32. 32. SUMIT Re-analysis <ul><li>By applying less stringent criteria and constructing ACLs: </li></ul><ul><ul><li>Uncovered previously unidentified mediating effects </li></ul></ul><ul><ul><ul><li>Serostatus Assumptions (participant’s tendency to make assumptions a potential partner’s serostatus) </li></ul></ul></ul><ul><ul><ul><li>Hedonistic Outcome Expectancies (belief that condom use will decrease sexual pleasure </li></ul></ul></ul><ul><ul><ul><li>Depression (marginally significant) </li></ul></ul></ul><ul><ul><li>Identified a marginally significant suppressive effect: </li></ul></ul><ul><ul><ul><li>Sexual compulsivity </li></ul></ul></ul><ul><ul><ul><li>Rather than decreasing reported UIAI, the intervention’s effect on this factor resulted in an increase </li></ul></ul></ul><ul><ul><ul><li>May partially explain the overall null intervention effect </li></ul></ul></ul>
  33. 33. Conclusions <ul><li>Great opportunities for collaboration between behavioral scientists and statisticians </li></ul><ul><li>Opportunities to analyze both new and existing data in a new way to answer questions about why interventions work/don’t work </li></ul>
  34. 34. Acknowledgements <ul><li>Sherri Pals </li></ul><ul><li>Craig Borkowf </li></ul><ul><li>Ann O’Leary </li></ul><ul><li>Jeffery Parsons </li></ul><ul><li>Cynthia Gomez </li></ul><ul><li>David MacKinnon </li></ul>
  35. 35. Questions?

×