# Classical and causal inference approaches to statistical mediation analysis/El enfoque clasico y el enfoque de la inferencia causal para el analisis de la mediacion.

The mediation model is one of the most popular procedures for studying the role of third variables involved in the relationship between an independent variable and a response/outcome variable (Ato & Vallejo, 2010). There are two main approaches for mediation analysis. The classical approach is the standard in psychology and the social sciences. It was developed in the 1980s (Judd & Kenny, 1981; Baron & Kenny, 1986), but has a more remote history, and is connected with Campbell's causation theory. Mediation estimation and inference with classical approach use a structural equation model (SEM) framework, from regression models (e.g., Anderson & Hunter, 2012; Sanchez-Manzanares, Rico, Gil, & San Martin, 2006) to path analysis with latent variables (e.g., Cava, Musitu, & Murgui, 2006). The causal inference (CI) approach is the standard in epidemiology and health sciences. It was developed in the 1990s (Robins & Greenland, 1992; Pearl, 2009) and it is connected with Rubin's causation theory (Shadish, 2010). Mediation estimation with causal inference approach is based on potential outcomes and contrafactuals.Dissatisfaction with how mediation is analysed and interpreted with classical approach has grown alarmingly in recent years (Bullock, Green, & Ha, 2010; Pardo & Roman, 2013; Spencer, Zanna, & Fong, 2005; Zhao, Lynch, & Chen, 2010). By far, the main problem is that applied psychological researchers have focused on the statistical model, ignoring the fact that a mediation model is essentially a causal model whose restrictive assumptions are not taken into account (Bullock & Ha, 2011; Roe, 2012).

In this paper, we compare the basic features of these approaches and we suggest complementing the classical with the causal inference approach. Although both produce similar equations when the variables are numeric, the second approach can be generalised to variables of any kind and to many more complicated scenarios. Moreover, it also has more sophisticated diagnostic tools and can be applied with programs that have been developed for the most popular computing platforms (R/STATA, SAS/SPSS and Mplus).

The classical approach

Recent research states that a mediation model (see MacKinnon, 2007, 2012; Hayes, 2013) is basically an structural model with two equations: one for explaining a mediator M of a treatment or exposure X (i.e., X [right arrow] M) and another to explain the outcome Y of a treatment, given the mediator M (i.e., X [right arrow] Y I M). The model may incorporate one or more covariates. The resulting regression equations, assuming for simplicity a single covariate C with c values, are:

[m.sub.i] = [[gamma].sub.0] + [[gamma].sub.1] [X.sub.i] + [[gamma].sub.2][c.sub.i] + [e.sub.i1] (1)

[y.sub.i] = [[beta].sub.0] + [[beta].sub.1] [x.sub.i] + [[beta].sub.2] [m.sub.i] + [[beta].sub.3][c.sub.i] + [e.sub.i2] (2)

where residuals [e.sub.i1] and [e.sub.i2] are assumed normally distributed with zero mean and variances [[sigma].sub.1.sup.2] y [[sigma].suib.2.sup.2], uncorrelated to each other and with treatment. These equations represent the classical approach when regression models are used, where coefficient [[gamma].sub.1] is a, [[beta].sub.1] is c' and [[beta].sub.2] is b (Figure 1), or when structural equation models are used. A combined model is often used to express Equation (2) in terms of Equation (1) as follows:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

In Equation (3), the effect of mediation or indirect effect represents the changes which X produced on Y transmitted through M and is usually estimated by the product of the coefficients [[beta].sub.2] and [[gamma].sub.1], while the direct effect is the effect of the treatment on the response at a fixed level of the mediator and is estimated by the coefficient [[beta].sub.1]. The inference is performed by dividing the indirect effect with a standard error using a formula proposed by Sobel (1982). The resulting Z test is very conservative because it assumes that the regression coefficients [[beta].sub.2] and [[gamma].sub.1] are independent and normally distributed, but it has been shown that the product distribution is highly skewed and leptokurtic (MacKinnon, Lockwood, & Williams, 2004). This is the procedure that is most commonly recommended in psychology.

To overcome some of the inferential problems posed by this procedure, several methods have been proposed to construct robust confidence intervals around the product [[beta].sub.2] [[gamma].sub.1], in particular, the delta method, the product distribution, the Monte Carlo method, and various forms of bootstrap resampling. The Monte Carlo method appears to behave better (see Preacher & Selig, 2012), but some authors recommend the bias-corrected bootstrap method for these (Hayes & Scharkow, 2013) and other cases (Vallejo, Ato, Fernandez, & Livacic, 2013). Various effect size measures have been proposed to assess the degree of mediation (see Preacher & Kelley, 2011) along with a power analysis for different mediation scenarios.

In addition to the linearity assumption (mediation analysis with the classical approach is not strictly applicable to models that include interactions or nonlinear terms), it is essential to consider three assumptions derived from the structural model, which represent situations where the error terms [e.sub.i1] and [e.sub.i2] covary.

Firstly, the mediation model should meet the assumption of temporal precedence of cause (X must precede in time to M and M to Y), a very complicated task with cross-sectional data (see Cole & Maxwell, 2003), and some have even suggested using the mediation model only with longitudinal data (see Maxwell & Cole, 2007; Maxwell, Cole, & Mitchell, 2011). If the treatment is a randomised variable, it is very unlikely to be caused by the mediator or the response. The mediator is rarely manipulated however, so the temporal precedence of the mediator on the outcome can only be justified measuring first the mediator and then the response or using appropriate control techniques, such as instrumental variables (Angrist, Imbens & Rubin, 1996) or propensity scores (Coffman, 2011; Jo, Stuart, MacKinnon, & Vinokur 2011).

Secondly, it is assumed that the treatment and the mediator are measured without error. If there is not a high reliability in the measurement of the variables, the coefficients become biased and the mediating causal effect will be affected. The consequences of measurement error are documented in Valeri (2012) and VanderWeele, Valeri and Ogburn (2012). There are several procedures to correct the bias due to measurement error in the variables, the most effective being the use of latent variables instead of individual indicators of treatment and mediator.

Thirdly, it is assumed that there are no omitted variables (also called spurious or confounder variables in this context) in relationships X [right arrow] M and M [right arrow] Y. Figure 1 shows a situation with two potential omitted variables ([Z.sub.1] and [Z.sub.2]), the first acting as a confounder in X [right arrow] M and the second acting as a confounder in M [right arrow] Y relationships. Randomisation of the treatment (e.g., using an experimental design) assures that no omitted variables will bias the X [right arrow] M relationship, but since it is not easy or not possible to randomise the mediator, the M [right arrow] Y relationship can potentially be affected by the presence of confounders. This is the most problematic situation that can often arise. Some solutions have been proposed, but the most effective way to circumvent the problem of omitted variables is to measure all potential confounders and control their effects.

The causal inference approach

An independent alternative to the classical is the causal inference approach (Pearl, 2010). This new approach focuses on the concept of the counterfactual, considering what would happen to an individual if instead of observing one feature (e.g., belonging to the experimental group) it were observed together with another (e.g., belonging to the control group). CI approach uses the theory of potential outcomes (Holland, 1986; Rubin, 2005), which defines a causal mechanism as a process by which a treatment (or exposure) causally affects an outcome given a mediator. Identifying a causal mechanism in this approach is formulated as a decomposition of the total causal effect into direct and indirect effects (Pearl, 2009).

Following Imai, Keele and Tingley (2010), Muthen (2011), Pearl (2012) and Valeri and VanderWeele (2013), let [Y.sub.i] (x) denote the potential outcome that would have been observed for i participant from a population (i = 1, ..., N) had the treatment [X.sub.i] been set at the value X. Note that [Y.sub.i] (x) is not an observed outcome so it may be counterfactual. Assuming a randomised binary treatment ([X.sub.i] = 0, for participants assigned to the control group and [X.sub.i] = 1, for those assigned to the experimental group), then for the participant 1, the causal effect of the treatment would be [Y.sub.i] (1) - [Y.sub.i] (0), but this result is not identified because participant i is observed for only one treatment. Nevertheless, for N participants of a large sample the average effect of treatment is identifiable and defined as the difference E[[Y.sub.i] (1) - [Y.sub.i] (0)].

Let [Y.sub.i] (1) and [M.sub.i] (1) now be values of the outcome and the mediator that would have been obtained for participant i if treatment [X.sub.i] were observed at level 1, and also let [Y.sub.i] (1,[M.sub.i] (1)) be the value of the outcome that would be obtained if both the treatment Xi and the mediator [M.sub.i] were observed at level 1. Additionally, introducing a covariate C (there can be more than one) observed at level c, the causal inference approach defines the Average Total Effect (ATE) by comparing the outcome variable between the two treatment levels conditional on the value c of the covariate,

ATE = E[[Y.sub.i] (1,[M.sub.i] (1)) - E([Y.sub.i] (0,[M.sub.i] (0)) | [C.sub.i] = c] (4)

This equation has two components: firstly, the effect of Average Causal MEdiation (ACME), which is the result of comparing the treatment at level 1 with level 0, fixing the mediator at level x, conditional on the level c of the covariate,

ACME = [[delta].sub.i] (x) = E[[Y.sub.i] (x,[M.sub.i] (1)) - [Y.sub.i] (x,[M.sub.i] (0)) | [C.sub.i] = c] (5)

and secondly the Average Direct Effect (ADE) which is the result of comparing the treatment at level 1 with level 0, setting the mediator at level x, conditional on the level c of the covariate

ADE = [zeta]i (x) = E[[Y.sub.i] (1,[M.sub.i] (x) - [Y.sub.i] (0,[M.sub.i] (x)) | [C.sub.i] = c] (6)

Equations (5) and (6) are average effects and assume that there is no interaction between treatment and mediator.

Identifying indirect (ACME) and direct (ADE) causal effects requires what is known as the Sequential Ignorability (SI) assumption (see Robins & Greenland, 1992; Pearl, 2009) which consists of two parts: (A) given one or more observed covariates, the treatment is ignorable, that is, independent of the potential values of the mediator and the outcome, and (B) given the treatment and one or more observed covariates, the mediator is ignorable, that is, independent of all potential values of outcome. More specifically, the SI assumption implies the following requirements: 1) there should be no latent confounders in the X [right arrow] Y path (all of the variables that cause both X and Y must be included in the model), 2) there should be no latent confounders in the M [right arrow] Y path (all the variables that cause both M and Y must be included in the model), 3) there should be no latent confounders in the X [right arrow] M path, and 4) treatment cannot be a cause of any confounder of the M [right arrow] Y path.

The first part of the sequential ignorability assumption (requirements 1 and 3) is satisfied if experimental designs or rigorous effective controls with non-experimental designs are used. But the second part of the sequential ignorability assumption (requirements 2 and 4) is not satisfied because it is virtually impossible to exclude the existence of omitted variables that confound the relationship between the mediator and the outcome. Under this assumption, Imai, Keele and Yamamoto (2010) proved that the effects of causal mediation are nonparametrically identified (i.e., they can be estimated without requiring a functional form and a known distribution), irrespective of the statistical model, of which the regression models of the classical approach are one particular case.

Then with the sequential ignorability assumption verified, the product of coefficients [[beta].sub.2][[gamma].sub.1] of the classical approach (equations 1 and 2, which must meet the assumptions of linearity and noninteraction) and the indirect effect of causal inference approach (Equation 5, assuming sequential ignorability) provide a valid estimate of the causal mediation effect if mediator and outcome are normally distributed variables. The main advantage of the causal inference over the classical approach is that the formulas for mediation (Pearl, 2012) can easily be generalised to a variety of models that do not require the assumptions of linearity and interaction to be met. Moreover, within the causal inference approach there are sophisticated procedures to assess the degree of compliance with the sequential ignorability assumption and the measurement error bias in the variables (Valeri, 2012). Additionally, Imai, Keele and Yamamoto (2013) have proposed specific experimental designs that can easily be implemented to facilitate the identification of causal mechanisms.

Sensitivity analysis

Even if the sequential ignorability assumption is not verified, it is possible to determine to what extent the estimates may be affected by the presence of confounders. Figure 2 shows a scenario where a latent confounder [Z.sub.2] (see Figure 1) is not included in the model creating a residual covariance between [e.sub.1] and [e.sub.2] whose magnitude depends on the mediation effect. Although the inclusion of a non-null covariance would render the model unidentifiable, Muthen (2011) proved that the removal of the M [right arrow] Y path does not affect the estimate of the causal effects and makes the model with a residual covariance included identifiable, as shown in Figure 3.

Imai, Keele and Yamamoto (2010) proposed to assess the presence of confounders through an analysis of sensitivity where the causal effects are calculated presetting values for the correlation between the residuals of equations (4) and (5). This procedure is useful in analysing existing data, because it allows us to assess the potential presence of confounders, but it is also used in planning new studies, because it allows us to estimate optimal sample sizes and the magnitude of the indirect effect so that the confidence bands do not include zero given a certain degree of confounding in the M [right arrow] Y path.

Sensitivity analysis is interpreted in terms of a range, and has a high degree of subjectivity, but it may be useful in assessing the degree to which the bias due to the inclusion of confounders may affect the interpretation of the effects (Imai, Keele, Tingley, & Yamamoto, 2011). The example in the next section illustrates this point.

An example with simulated data

Using a code previously proposed by Muthen (2011, pp. 3537), we conducted a study with a binary categorical randomised treatment (V3), a numeric mediator (V2) and a numerical outcome variable (V1) simulating 200 observations and 1000 repetitions. We used the model in Figure 2 as the population model, setting an indirect effect of 0.09 and an direct effect of 0.50, and as the analysis model we used the model in Figure 3, with three different fixed values of the correlation between residuals ([rho]): 0, 0.25 and 0.50. The remaining parameters were set to the appropriate values to produce the desired causal effects. For each value of p, the population values and the mean estimates of 1,000 repetitions were practically the same (with a maximum bias of .005 units), revealing that the indirect and direct causal effects were estimated correctly in the three cases. Then we randomly selected one of the 1,000 replicates registered for each value of p, and these data were analysed with the R-mediation package (4.2.4 version, see Tingley, Yamamoto, Hirose, Keele, & Imai, 2013). We follow the suggestions of Garcia, Pascual, Frias, Van Krunckelsven and Murgui (2008) for interpretation of power and confidence intervals.

The first part of the Output 1 (Table 1) shows the R code used and the estimates of the causal effects with quasi-Bayesian Monte Carlo confidence intervals, assuming that the correlation between residuals is p = 0. The two equations were estimated with the R-lm function, the mediation effects with the R-mediate function and the sensitivity analysis with the R-medsens function of the mediation package. For the random replicate used, we found a non-significant indirect effect (ACME = 0.0755, p = .08) and a mediated proportion of 0.115 (p = 0.08), so the presence of a mediating effect was discarded. The second part of Output 1 (Table 1) is a sensitivity analysis with the medsens function, which reveals that the confidence interval of indirect effects includes zero in the entire spectrum of the sensitivity parameter [rho] (from - 0.9 to + 0.9). In fact, sensitivity analysis is not required when the indirect effect is not significant. Figure 4 is a plot of sensitivity analysis and shows, together with the axes of indirect effect and [??], the observed mediating effect (dashed line) and the values that the indirect effect would reach varying the sensitivity parameter (solid curved line) along its range in steps of .10. Note that the confidence interval (limits represented with a grey background) always includes zero for the indirect effect whatever the value of [rho] is. This result is expected as a consequence of imposing a zero correlation between the residuals of Equations (1) and (2).

In contrast, the estimators of the causal effects assuming [rho] = 0.25 showed a moderate overestimation of the indirect effect, of 0.163 - 0.076 = 0.087 units (see Output 2 on Table 2). The indirect effect was now significant (ACME = .163; p = .04) and the proportion mediated jumped from 0.11 to 0.25. The sensitivity analysis shown in the textual output and the complementary plot of Figure 4 allow us to conclude that for the indirect effect to be zero, the correlation between the residuals of the regression models for mediator and outcome variables should be [rho] = .60 (the confidence interval in Figure 5 ranges from about 0.55 to 0.75). An algebraically equivalent form of sensitivity analysis is to use the product of the determination coefficients of both regression models. It is shown in the penultimate line of Output 2 and indicates that the indirect effect will be zero when the confounders of the mediator-response relationship together explain 36% or more of the residual variance (i.e., 0.60 x 0.60 = .36). The last line of the output refers to the total variance instead of the residual variance and indicates that the indirect effect will be zero when the total variance explained by confounding is greater than 17% (i.e., 0.40 x 0.43 = .17). In rigorously applied research where no important confounders are unchecked, the researcher should conclude that a correlation as high as that required for the ACME zero is unlikely and therefore he or she should interpret the causal mediation effect found with confidence, when in fact, the simulated data assume a correlation of [rho] = .25 and a causal effect was overestimated.

Finally, with [rho] = 0.5, the indirect effect estimator provided an important overestimation with respect to [rho] = 0, of 0.208-0.076 = 0.132 units (see Output 3 on Table 3), which was also significant (ACME = 0.208; p = .03), and the proportion mediated jumped in this case from 0.11 to 0.33. The textual output of the sensitivity analysis and Figure 6 show that the effect will be zero when the indirect effect is p = 0.80 or greater, or when the confounders in the response-mediator relationship explain at least 64% of the residual variance (e.g., 0.8 x 0.8 = .64), which would lead the researcher to conclude with absolute confidence on the relevance of the indirect effect if important confounders are not present.

This example with simulated data illustrates the overestimation of the indirect effect in the presence of confounders of the mediator-response relationship, from a non-significant indirect effect with p = 0 to significant effects with [rho] = .25 and [rho] = .50 using numerical variables with mediator and response. Unfortunately the existence of confounders is a fairly common situation in psychological research that results in an overestimation of the effects and frequently changes in the inference. The limited control of confounders in usual practice justifies the high degree of dissatisfaction with the mediation analysis applications within the classical approach and the need to supplement it with causal inference approach, where a better control on interaction mediator-response is available, the options to use non-interval variables for mediator and response and powerful diagnostic tools and its generalisation to more complicated research scenarios.

Software for mediation analysis with the causal inference approach

Four major programs of professional statistical computing (R, SAS, SPSS and STATA) can now be used to perform a mediation analysis with the causal inference approach. The main characteristics of these programs are summarised in Table 4 (see Valeri & VanderWeele, 2013). Firstly, the R-mediation package (Tingley et al., 2013), which was used in the previous section, is an easy to implement but powerful alternative. Among its main features, we emphasize the availability of linear and nonlinear models (including R functions such as lm for linear models, glm for generalised linear models and glmer for mixed models), the sensitivity analysis tools, the option to use multiple mediators and the generalisation to multilevel models. However, it does not work with latent variables. There is a similar program, albeit more modest, for STATA (Hicks & Tingley, 2011), which analyses linear and nonlinear models and also includes sensitivity analysis, but not multilevel models or multiple mediators.

Second, the macro mediation developed under SAS 9.2 and SPSS 19.0 versions (Valeri & VanderWeele, 2013) is an important alternative with many interesting options, but it does not include sensitivity analysis, latent variables or multilevel models.

And finally, in an unpublished document (Muthen, 2011), a Mplus code (version 6.2) is presented for mediation analysis within the causal inference approach, inspired by macros developed for SAS/SPSS (Valeri & VanderWeele, 2013). This code shows the versatility of the program and its potential to include interactive terms, nonlinear models, multilevel and multiple mediators, latent variables and even the option to program a sophisticated sensitivity analysis if desired.

Authors' Note

This work was supported by grants from the Spanish Ministry of Science and Innovation (Ref: PSI-2011-23395) and the Spanish Ministry of Economy and Competitiveness (Ref: EDU-201234433) awarded to the authors.

doi: 10.7334/psicothema2013.314

References

Anderson, S., & Hunter, S.C. (2012). Cognitive appraisals, emotional reactions and their associations with three forms of peer victimization. Psicothema, 24, 621-627.

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.

Ato, M., & Vallejo, G. (2011). Los efectos de terceras variables en la investigacion psicologica. Anales de Psicologia, 27, 550-561.

Baron, J., & 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.

Bullock, J.G., & Ha, S.E. (2011). Mediation analysis is harder than it looks. In J.N. Drukman, D.P. Green, J.H. Kuklinski & A. Lupia (Eds.), Cambridge Handbook of Experimental Political Science (pp. 508-521). New York, NY: Cambridge University Press.

Bullock, J.G., Green, D.P., & Ha, S.E. (2010). Yes, but what's the mechanism? (don't expect an easy answer). Journal of Personality and Social Psychology, 98, 550-558.

Cava, M.J., Musitu, G., & Murgui, S. (2006). Familia y violencia escolar: el rol mediador de la autoestima y la actitud hacia la autoridad institucional [Family and school violence: The mediator role of self-esteem and attitudes towards institutional authority]. Psicothema, 18, 367-373.

Coffman, D.L. (2011). Estimating causal effects in mediation analysis using propensity scores. Structural Equation Modeling, 18, 357-369.

Cole, D.A., & Maxwell, S.E. (2003). Testing mediational models with longitudinal data: Questions and tips in the use of structural equation modeling. Journal of Abnormal Psychology, 112, 558-577.

Garcia, J.F., Pascual, J., Frias, M.D., Van Krunckelsven, D., & Murgui, S. (2008). Diseno y analisis de la potencia: n y los intervalos de confianza de las medias [Design and power analysis: n and confidence intervals of means]. Psicothema, 20, 933-938.

Hayes, A.F. (2013). Introduction to mediation, moderation and conditional process analysis: A regression-based approach. New York, NY: The Guilford Press.

Hayes, A.F., & Scharkow, M. (2013). The relative trustworthiness of tests of the indirect effect in statistical mediation analysis: Does method really matter? Psychological Science, 24, 1918-1927.

Hicks, R., & Tingley, D. (2011). Causal mediation analysis. The Stata Journal, 11, 1-15.

Holland, P.W. (1986). Statistics and causal inference. Journal of the American Statistical Association, 81, 945-960.

Imai, K., Keele, L., & Yamamoto, T. (2010). Identification, inference and sensitivity analysis for causal mediation effects. Statistical Science, 1, 51-71.

Imai, K., Keele, L., & Tingley. D. (2010). A general approach to causal mediation analysis. Psychological Methods, 15, 309-344.

Imai, K., Keele, L., & Yamamoto, T. (2013). Experimental designs for identifying causal mechanisms. Journal of the Royal Statistical Society, 176, Part I, 5-51.

Imai, K., Keele, L., Tingley, D., & Yamamoto, T. (2011). Unpacking the black box of causality: Learning about causal mechanisms from experimental and observational studies. American Political Science Review, 105, 765-789.

Jo, B., Stuart, E.A., MacKinnon, D.P., & Vinokur, A.D. (2011). The use of propensity scores in mediation analysis. Multivariate Behavioral Resarch, 46, 425-452.

Judd, C.M., & Kenny, D.A. (1981). Process analysis: Estimating mediation in treatment evaluations. Evaluation Review, 5, 602-619.

MacKinnon, D.P. (2007). Introduction to mediation analysis. Mahwah, NJ: Erlbaum.

MacKinnon, D.P., Cheong, J., & Pirlott, A.G. (2012). Statistical mediation analysis. In H. Cooper, P.M. Camic, D. Long, A.T.Panter, D. Rindskopf, & K.J. Sher (Eds.). APA Handbook of Research Methods in Psychology, Vol 2: Research designs: Quantitative, qualitative, neuropsychological, and biological (pp. 313-331). Washington, DC: American Psychological Association.

MacKinnon, D.P. (2008). Introduction to statistical mediation analysis. New York, NY: Lawrence Erlbaum.

MacKinnon, D.P., Lockwood, C.M., & Williams, J. (2004). Confidence limits for the indirect effect: Distribution of the product and resampling methods. Multivariate Behavioral Research, 39, 99-128.

Maxwell, S.E., & Cole, D.A. (2007). Bias in cross-sectional analysis of longitudinal mediation. Psychological Methods, 12, 23-44.

Maxwell, S.E., Cole, D.A., & Mitchell, M.A. (2011). Bias in cross-sectional analysis of longitudinal mediation: Partial and complete mediation under an autoregressive model. Multivariate Behavioral Research, 46, 816-841.

Muthen, B. (2011). Applications of causally defined direct and indirect effects in mediation analysis using SEM in Mplus. Unpublished Mplus paper. URL: http://www.statmodel.com/download/causalmediation.pdf. Accessed: 2013-06-21 (Archived by WebCite[R] at http://www.webcita tion.org/6LIEfxWaJ).

Pardo, A., & Roman, M. (2013). Reflections on the Baron and Kenny model of statistical mediation. Anales de Psicologia, 29, 614-623.

Pearl, J. (2009). Causal inference in statistics: An overview. Statistics Surveys, 3, 96-146.

Pearl, J. (2010). An introduction to causal inference. The International Journal of Biostatistics, 6(2), Article 7 (DOI: 10.2202/1557-4679. 1203).

Pearl, J. (2012). The mediation formula: A guide to the assessment of pathways and mechanisms. Prevention Science, 13, 426-436.

Preacher, K.J., & Kelley, K. (2011). Effect size measures for mediation models: Quantitative strategies for communicating indirect effects. Psychological Methods, 16, 93-115.

Preacher, K.J., & Selig, J.P. (2012). Advantages of Monte Carlo confidence intervals for indirect effects. Communication Methods and Measures, 6, 77-98.

Robins, J.M., & Greenland, S. (1992). Identifiability and exchangeability for direct and indirect effects. Epidemiology, 3, 143-155.

Roe, R.A. (2012). What is wrong with mediators and moderators? The European Health Psychologist, 14, 4-10.

Rubin, D.B. (2005). Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American Statistical Association, 100, 322-331.

Sanchez-Manzanares, M., Rico, R., Gil, F., & San Martin, R. (2006). Memoria transactiva en equipos de toma de decisiones: implicaciones para la efectividad de equipo [Transactive memory in decision-making teams: Implications for team effectiveness]. Psicothema, 18, 750-756.

Shadish, W.R. (2010). Campbell and Rubin: A primer and comparison of their approaches to causal inference in field settings. Psychological Methods, 15, 3-17.

Sobel, M.E. (1982). Asymptotic confidence intervals for indirect effects in structural equation models. Sociological Methodology, 12, 290-312.

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, 845-851.

Tingley, D., Yamamoto, T., Hirose, K., Keele, L., & Imai, K. (2013). Mediation: R package for causal mediation analysis. Journal of Statistical Sofware (in press).

Valeri, L. (2012). Statistical methods for causal mediation analysis. Doctoral dissertation. Cambridge, MA: Harvard University DASH Repository.

Valeri, L., & VanderWeele, T.J. (2013). Mediation analysis allowing for exposure-mediator interactions and causal interpretation: Theoretical assumptions and implementation with SAS and SPSS macros. Psychological Methods, 18, 137-150.

Vallejo, G., Ato, M., Fernandez, P., & Livacic-Rojas, P (2013). Multilevel bootstrap analysis with assumptions violated. Psicothema, 25, 520-528.

VanderWeele, T.J., Valeri, L., & Ogburn, E.L. (2012). The role of measurement error and misclassification in mediation analysis. Epidemiology, 23, 561-564.

Zhao, X., Lynch, J.G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of Consumer Research, 37, 197-206.

Manuel Ato Garcia (1), Guillermo Vallejo Seco (2) and Ester Ato Lozano (1)

(1) Universidad de Murcia and (2) Universidad de Oviedo

Received: November 21, 2013 * Accepted: January 24, 2014

Corresponding author: Manuel Ato Garcia

Facultad de Psicologia Universidad de Murcia 30100 Murcia (Spain)

e-mail: matogar@um.es

Table 1 Output 1: R-mediation program for simulated data with [rho] = 0 > m1 = lm(V2~V3, data = sim1) > m2 = lm(V1~V2+V3, data = sim1) > med1 = mediate(m1, m2, treat = "V3", mediator = "V2") > summary(med1) Causal mediation analysis Quasi-Bayesian confidence intervals Estimate CI Lower CI Upper p-value 95% 95% ACME 0.0755 -0.00732 0.15906 0.08 ADE 0.5317 0.33333 0.73964 0.00 Total effect 0.6023 0.37395 0.82127 0.00 Prop. mediated 0.1147 -0.01425 0.25828 0.08 Sample size used: 200 Simulations: 1000 > asens1 = medsens(med1, rho.by = .1, effect.type = "indirect") > summary(asens1) Mediation sensitivity analysis for average causal mediation effect Sensitivity region Rho ACME 95% CI Lower 95% CI Upper [1,] -0.9 0.4599 -0.0424 0.9623 [2,] -0.8 0.3214 -0.0302 0.6730 [3,] -0.7 0.2545 -0.0244 0.5334 [4,] -0.6 0.2109 -0.0206 0.4424 [5,] -0.5 0.1782 -0.0180 0.3743 [6,] -0.4 0.1515 -0.0158 0.3189 [7,] -0.3 0.1284 -0.0141 0.2709 [8,] -0.2 0.1075 -0.0127 0.2277 [9,] -0.1 0.0879 -0.0115 0.1873 [10,] 0.0 0.0688 -0.0107 0.1484 [11,] 0.1 0.0498 -0.0106 0.1101 [12,] 0.2 0.0302 -0.0120 0.0723 [13,] 0.3 0.0093 -0.0189 0.0374 [14,] 0.4 -0.0138 -0.0442 0.0165 [15,] 0.5 -0.0405 -0.0920 0.0109 [16,] 0.6 -0.0732 -0.1574 0.0109 [17,] 0.7 -0.1168 -0.2470 0.0133 [18,] 0.8 -0.1837 -0.3859 0.0184 [19,] 0.9 -0.3223 -0.6747 0.0302 Rho R^2_M*R^2_Y* R^2_M~R^2_Y~ [1,] -0.9 0.81 0.6161 [2,] -0.8 0.64 0.4868 [3,] -0.7 0.49 0.3727 [4,] -0.6 0.36 0.2738 [5,] -0.5 0.25 0.1902 [6,] -0.4 0.16 0.1217 [7,] -0.3 0.09 0.0685 [8,] -0.2 0.04 0.0304 [9,] -0.1 0.01 0.0076 [10,] 0.0 0.00 0.0000 [11,] 0.1 0.01 0.0076 [12,] 0.2 0.04 0.0304 [13,] 0.3 0.09 0.0685 [14,] 0.4 0.16 0.1217 [15,] 0.5 0.25 0.1902 [16,] 0.6 0.36 0.2738 [17,] 0.7 0.49 0.3727 [18,] 0.8 0.64 0.4868 [19,] 0.9 0.81 0.6161 Rho at which ACME = 0: 0.3 R^2_M*R^2_Y* at which ACME = 0: 0.09 R^2_M~R^2_Y~ at which ACME = 0: 0.0685 > plot(asens1)--see Figure 4-- Table 2 Output 2: R-mediation program for simulated data with p = 0.25 > m3=lm(V2~V3,data=sim2) > m4=lm(V1~V2+V3,data=sim2) > med2=mediate(m3,m4,treat="V3",mediator="V2") > summary(med2) Causal mediation analysis Quasi-Bayesian confidence intervals Estimate CI Lower CI Upper p-value 95% 95% ACME 0.1625 0.0156 0.3036 0.04 ADE 0.4821 0.3276 0.6472 0.00 Total effect 0.6446 0.4213 0.8641 0.00 Prop. mediated 0.2508 0.0332 0.4210 0.04 Sample size used: 200 Simulations: 1000 > asens2=medsens(med2,rho.by=.1,effect.type="indirect") > summary(asens2) Mediation sensitivity analysis for average causal mediation effect Sensitivity region Rho ACME 95% CI Lower 95% CI Upper [1,] 0.5 0.0513 -0.0004 0.103 [2,] 0.6 0.019 -0.0118 0.0498 [3,] 0.7 -0.0241 -0.0575 0.0093 Rho R^2_M*R^2_Y* R^2_M~R^2_J~ [1,] 0.5 0.25 0.1206 [2,] 0.6 0.36 0.1736 [3,] 0.7 0.49 0.2363 Rho at which ACME = 0: 0.6 R^2_M*R^2_Y* at which ACME = 0: 0.36 R^2_M~R^2_Y~ at which ACME = 0: 0.1736 > plot(asens2)--see Figure 5-- Table 3 Output 3: R-mediation program for simulated data with p = 0.50 > m5=lm(V2~V3,data=sim3) > m6=lm(V1~V2+V3,data=sim3) > med3=mediate(m3,m4,treat="V3",mediator="V2") > summary(med3) Causal mediation analysis Quasi-Bayesian confi dence intervals Estimate 95% CI Lower 95% CI Upper p-value ACME 0.2080 0.0327 0.3921 0.03 ADE 0.4306 0.2788 0.5769 0.00 Total effect 0.6386 0.3978 0.8738 0.00 Prop. mediated 0.3269 0.0696 0.5179 0.03 Sample size used: 200 Simulations: 1000 > asens3=medsens(med3,rho.by=.1,effect.type="indirect") > summary(asens3) Mediation sensitivity analysis for average causal mediation effect Sensitivity region Rho ACME 95% CI Lower 95% CI Upper [1,] 0.7 0.0412 -0.0010 0.0833 [2,] 0.8 -0.0201 -0.0496 0.0093 Rho R^2_M*R^2_Y* R^2_M~R^2> [1,] 0.7 0.49 0.1669 [2,] 0.8 0.64 0.2181 Rho at which ACME = 0: 0.8 R^2_M*R^2_Y* at which ACME = 0: 0.64 R^2_M~R^2_Y~ at which ACME = 0: 0.2181 > plot(asens3)--Figure 6-- Table 4 Comparative features of programs for mediation analysis with the causal inference approach Mediation Mediation MPLUS (R/STATA) (SAS/SPSS) Estimates all causal effects NO (a) YES YES Interactions: Treatment-Mediator YES YES YES Mediator-Covariate/s NO NO YES Available models: Linear YES (lm) YES YES Generalised linear YES (glm) YES YES Non linear YES NO YES Works with latent variables NO NO YESb Research designs: Cross-sectional YES YES YES Cohort YES YES YES Case-control YES YES NO Standard errors: Bootstrap YES YES YES Delta YES YES (default) YES Bayesian Monte Carlo YES (default) NO YES Sensitivity analysis YES NO NOc Moderated mediation YES YES YES Multiple mediators YES/NO (d) NO YES Generalisation to multilevel analysis YES/NO (d) NO NO (c) (a): Natural causal effects only; (b): Not available yet; (c): Not available, but it may be programmable; (d): Is available with R but not with STATA

Printer friendly Cite/link Email Feedback | |

Title Annotation: | texto en ingles |
---|---|

Author: | Garcia, Manuel Ato; Seco, Guillermo Vallejo; Lozano, Ester Ato |

Publication: | Psicothema |

Date: | Apr 1, 2014 |

Words: | 6030 |

Previous Article: | Explicit and implicit assessment of gender roles/Evaluacion explicita e implicita de los roles de genero. |

Next Article: | Psychometric properties of the Looming Maladaptive Style Questionnaire (LMSQ-R) in young Spanish adults/ Propiedades psicometricas del Cuestionario... |

Topics: |