Testing process hypotheses using more than just mediation analysis Johann Jacoby Leibniz-Institut für Wissensmedien / Knowledge Media Research Center Tübingen 20130903 14. Tagung der Fachggruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Public version 20130905 Disclaimer: Even though carefully checked, these slides may contain errors. If you identify them, please let me know: [email protected] 1 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Outline • What do we test using mediation analysis? • Challenges • Different approaches ‒ ‒ ‒ ‒ Causal chain (Spencer, Zanna, & Fong, 2005) Instrumentation (Angrist, Imbens, & Rubin, 1996) Controlling previous covariance of M and Y in (quasi-)longitudinal designs (Maxwell & Cole, 2007) TPIS (Jacoby & Sassenberg, 2011) • Conclusion: 2 Call for strategy diversity 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Preliminaries • Experimentally minded research, no large scale longitudinal observation studies • Manipulated X – mediation analysis without any manipulation not at focus • Mediation analysis may yield accurate estimates of the direct and the indirect effect if the model holds – mediation analysis cannot discover causal structure and indirect effects • → Mediation analysis tests hypotheses, it is not an exploratory technique 3 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses • "X affects Y via M" • = There is variance that X and Y share that is also shared by M ‒ ‒ ‒ 4 If M is to explain an effect of X, it surely must share variance with X If M is to explain an effect on Y, it surely must share variance with Y If M is to explain an effect of X on Y, it surely must share variance with both X and Y and the variances shared must be the same → triply shared variance 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses – an analogy Hagen Wuppertal map from http://www.openstreetmap.org 5 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Lüdenscheid Process hypotheses – an analogy • How many of the people going from Hagen to Lüdenscheid also came from Wuppertal? → How much variance is shared by X, M, and Y? • Necessary: Path a from M = aX ~ Wuppertal → Hagen 6 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses – an analogy a map from http://www.openstreetmap.org 7 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses – an analogy • How many of the people going from Hagen to Lüdenscheid also came from Wuppertal? → How much variance is shared by X, M, and Y? • People going from Hagen to Lüdenscheid: Path b.full from Y = b.full M 8 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses – an analogy b.full map from http://www.openstreetmap.org 9 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypotheses – an analogy • People going from Hagen to Lüdenscheid in general are of limited interest for the question of the process • Necessary: People who came from Wuppertal to Hagen and went on to Lüdenscheid: portion P • Portion P is represented by a × b, where b is the semipartial association of M and Y 10 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypothesis test – what we need • An accurate estimate of path a ‒ ‒ problematic with measured X unproblematic with manipulated X • An accurate estimate of the partial association b 11 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Process hypothesis test – what we need • Fundamental meaning of causality: "If it were not for the cause, the effect would not have happened" • Counterfactual ‒ ‒ ‒ with process: c (actually observed, factual) without process: c' (statistically simulated, counterfactual) indirect effect is the difference between factual and counterfactual states of affairs: c – c' = ab 12 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen • factual 13 • counterfactual 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Baron & Kenny (1986) Mediation analysis • Two regression models ‒ ‒ ‒ M = aX Y = c'X + bM [Y = cX] • Indirect effect = a×b • Testing a×b against 0: ‒ ‒ 14 a M X b Y c' (c) Sobel, Aroian, Goodman tests (normal distribution = necessarly false assumption) Bootstrap-based confidence intervals (Preacher & Hayes, 2004, etc.) 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Challenges: Ambiguity of path b •M and Y are measured • The association of two measured variables cannot be causally attributed in an unambiguous way ‒ ‒ ‒ ‒ Source: X → (M, Y) – common cause X Source: M → Y Source: Y → M Source: Unknown third variable → (M, Y) • Things get worse if we do not look for path b.full, but for the semipartial association path b 15 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Challenges: Ambiguity of path b • Isolating ‒ b and identifying ab is additionally problematic when a and b are moderated differently (see Bullock et al., 2010) Which conditional a and which conditional b should be used to asses the indirect effect ab? • There could be no single mediation case, but the test would detect mediation • There could be mediation for every single case, but the test would detect no mediation • ‒ b is moderated by X (sometimes incorrectly called "moderated mediation") a X ‒ 16 M c' (c) b Y M contains more than one single variance component (in addition to truely unsystematic error) 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Challenges: Uncontrollable interference with the process • Mediation analysis requires measurement of the mediator for all observations • If the measurement administration itself alters the process we will never know about it • No comparison "with measurement of mediator" vs. "without measurement of mediator" possible 17 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Alternatives to Mediation analysis • Causal Chain Design (Spencer, Zanna, & Fong, 2005) • Instrumentation (Angrist, Imbens, & Rubin, 1996) • Controlling previous covariance of M and Y in (quasi-)longitudinal designs (Maxwell & Cole, 2007) • TPIS (Jacoby & Sassenberg, 2011) 18 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Causal Chain Design (Spencer, Zann, & Fong, 2005) • Obtain estimate of the effect of M on Y in a separate experiments: ‒ ‒ Experiment 1: estimate M = aX (manipulated X, measured M) Experiment 2: estimate Y= b.full M (manipulated M, measured Y) • Requirement: M measurable and manipulable! • Variance in the measure of M must be the same variance as in manipulation of M • If the effect of Mmanipulated is not exactly the difference measured by Mmeasured, the Causal Chain Design will yield systematically incorrect conclusions. • Prototype: Word, Zanna, & Cooper (1974) 19 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Causal Chain Design • Gains ‒ ‒ Unambiguous estimate of path b.full Measurement of M cannot interfere with an assessment of the causal path • Challenges: ‒ ‒ ‒ ‒ 20 a measurement instrument M and a manipulation of M where M is exactly the same are required 2 × N observations Is the difference in Y reflecting the manipulation of M (i.e., the effect of M on Y) the same difference as the difference in X reflected in measured M? (M can be generally be influenced by many others variables, not only X) No straight forward representation of the indirect effect 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation • Identifying the unconfounded effect of M on Y Instrument Z z a X {X=1, X=2} 21 M c' (c) b Y 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation Instrument Z to M(X2) – Z increases M Y • Apply X=1 22 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation • Apply Instrument Z to M(X2) – Z increases M Y Effekt b.full * z X=1 23 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen X=2 & Z Instrumentation Instrument Z to M(X2) / Z decreases M Y • Apply X=1 24 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation • Apply Instrument Z to M(X2) / Z decreases M Y Effect b.full * z X=1 25 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen X=2 & Z Instrumentation Instrument Z to M(X1) – Z increases M Y • Apply X=1 26 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation Instrument Z to M(X1) – Z increases M Y • Apply Effekt b.full * z X=1 27 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen X=1 & Z Instrumentation Instrument Z to M(X1) / Z decreases M Y • Apply X=1 28 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen X=2 & Z Instrumentation Instrument Z to M(X1) / Z decreases M Y • Apply Effect b.full * z X=1 29 X=2 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen X=1 & Z Instrumentation Hello again, Mediation design and analysis. This time more restrictive. Instrument Z a M b X {X1, X2} 30 Y c' (c) 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Instrumentation • Gains ‒ ‒ Unambiguous estimate of path b.full Measurement of mediator not strictly necessary (albeit adding considerably to validity) • Challenges ‒ ‒ ‒ ‒ 31 Find Z with known size of effect on M (or measure M to assess z) Find Z ⊥ X that only influences Y via M, not directly and not via additional mediators No clear representation of the indirect effect 1.5 × N 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Longitudinal design with manipulated X t0 t1 M M a cov (M, Y) b "pure" X c' (c) Y 32 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Y cov(M,Y) ⊥ b Longitudinal design with manipulated X • Gains ‒ ‒ Estimate of the association between M and Y is adjusted for third pre-manipulation variables Traditional mediation design is preserved to a large degree • Challenges ‒ ‒ ‒ 33 M is featured as two different variables: trait vs. state Measure M and Y twice without introducing additional validity threats (e.g., consistency effects, changes of psychological meaning etc.) Measure M and Y twice while still avoiding interference 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen TPIS (Jacoby & Sassenberg, 2011) • Very old idea, but not much recognized as a test of process hypotheses anymore (Prototype: Zanna & Cooper, 1974) • Fundamental statement of a process hypothesis is that of counterfactuality: "If it were not for the hypothesized process, the effect of X on Y would be smaller." • TPIS: Create conditions B=0 and B=1, such that ‒ ‒ under B = 0: the intermediate process from X to Y can occur undisturbed under B = 1: the intermediate process between X and Y is broken • If the process hypothesis is viable, the effect of X on Y should be smaller under B=1 than under B=0 34 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen TPIS •B = 1: ‒ ‒ ‒ ‒ 35 Force M(X2, B=1) = M(X1, B=1) or at least M(X2, B=1) – M(X1, B=1) < M(X2, B=0) – M(X1, B=0) B is not the same as M, but a manipulated context variable independent of X Then even though X varies as in B=0, no process via M can occur Works best if M(X1) and M(X2) are near the end of the natural range of M (e.g., cognitive load, attribution errors) 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen TPIS • Gains ‒ ‒ ‒ ‒ Actual observation of the formerly counterfactual clause of the hypothesis ("If it were not for the process...") Differential moderation is a lesser problem No interference of measurement activity M needs not necessarly be measured (even though independent eveidence for the X → M link adds credibility) • Challenges ‒ ‒ ‒ ‒ 36 Manipulation that fixes M independently of X required Estimate of the magnitude of the indirect effect requires solid knowledge about the manipulation of B and its effects Sample size doubled Evidence for process by interaction ('moderation') – (seemingly) calls into question the wide spread strict distinction of mediation and moderation 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Y TPIS: Prototypical pattern – Evidence for hypothesized process by an interaction effect X=1 X=2 B=0 37 X=1 X=2 B=1 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Conclusion • Relying on mediation analysis exclusively will not do justice to psychological processes. • There are a number of viable and feasible alternatives. • None of the alternatives are entirely unproblematic. 38 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Conclusion • Observation of processes instead of experimentation on them often does not give solid evidence in noisy data environments as the social and human sciences are • Manipulation is not the easy answer, the process may easily be changed through intervention (even through observation) • Special challenges: ‒ ‒ ‒ 39 Asynchronous events and observations Strong requirements for definition of events and observations (variance partitionning) Mixed events 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Conclusion • The measurement of variable states and the events creating these states are two different things. • Processes measured. always need to be inferred, cannot be • Inference is rooted in design, not in analysis techniques or ways to derive confidence intervals. • Viable and solid understanding of processes will be grounded in good and diverse designs. 40 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Conclusion • Good ‒ ‒ ‒ ‒ ‒ designs comprise well defined, well distinguished, psychometrically sound variables (systematic variance + error), designing ahead, stating hypotheses ahead of time • No one single design of process testing will serve as a panacea, rather their versatile and systematically developed combination for each research domain will yield robust results. 41 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen Thank you! 42 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen References 43 • Angrist, J. D., Imbens, G. W., & Rubin, D. B. (1996). Identification of causal effects using instrumental variables. Journal of the American Statistical Association, 91, 444-455. • 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. • Jacoby, J., & Sassenberg, K. (2011). Interactions do not only tell us when, but can also tell us how: Testing process hypotheses by interaction. European Journal of Social Psychology, 41, 180-190. • Maxwell, S. E., & Cole, D. A. (2007). Bias in cross-sectional analyses of longitudinal mediation. Psychological Methods, 12, 23. • Preacher, K. J., & Hayes, A. F. (2004). SPSS and SAS procedures for estimating indirect effects in simple mediation models. Behavior Research Methods, Instruments, & Computers, 36, 717-731. • Spencer, S. J., Zanna, M. P., & Fong, G. T. (2005). Establishing a causal chain: why experiments are often more effective than mediational analyses in examining psychological processes. Journal of Personality and Social Psychology, 89, 845. • Word, C. O., Zanna, M. P., & Cooper, J. (1974). The nonverbal mediation of self-fulfilling prophecies in interracial interaction. Journal of Experimental Social Psychology, 10, 109-120. • Zanna, M. P., & Cooper, J. (1974). Dissonance and the pill: an attribution approach to studying the arousal properties of dissonance. Journal of Personality and Social Psychology, 29, 703. 14. Tagung der Fachgruppe Sozialpsychologie in der DGPs, 1.-4. September 2013, Hagen