Limits...
Causal mediation analysis with multiple mediators.

Daniel RM, De Stavola BL, Cousens SN, Vansteelandt S - Biometrics (2014)

Bottom Line: Much progress has been made, mainly due to contributions from the field of causal inference, in understanding the precise nature of statistical estimands that capture such intuitive effects, the assumptions under which they can be identified, and statistical methods for doing so.These contributions have focused almost entirely on settings with a single mediator, or a set of mediators considered en bloc; in many applications, however, researchers attempt a much more ambitious decomposition into numerous path-specific effects through many mediators.We discuss the strong assumptions under which the effects are identified, suggesting a sensitivity analysis approach when a particular subset of the assumptions cannot be justified.

View Article: PubMed Central - PubMed

Affiliation: Centre for Statistical Methodology, London School of Hygiene and Tropical Medicine, Keppel Street, London WC1E 7HT, UK.

Show MeSH

Related in: MedlinePlus

With  (perfect correlation between  and  given ), all 24 possible decompositions of the total causal effect of heavy drinking on SBP into four path-specific components: a direct effect unmediated by BMI or GGT, an indirect effect via BMI alone, an indirect effect via GGT alone, and an indirect effect via both BMI and GGT. The numbers superimposed on the bars represent the code for that effect type (as defined in the caption of Table 2). The numbers along the x-axis represent the decomposition number, also defined in Table 2.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4402024&req=5

fig02: With (perfect correlation between and given ), all 24 possible decompositions of the total causal effect of heavy drinking on SBP into four path-specific components: a direct effect unmediated by BMI or GGT, an indirect effect via BMI alone, an indirect effect via GGT alone, and an indirect effect via both BMI and GGT. The numbers superimposed on the bars represent the code for that effect type (as defined in the caption of Table 2). The numbers along the x-axis represent the decomposition number, also defined in Table 2.

Mentions: The results are shown in Tables 4 and 5 and Web Figures 1 and 2. There is evidence of a total effect of heavy drinking on SBP, but the associated confidence interval is wide (mean difference 7.63 mmHg, 95% CI 3.89–11.37). Only a small proportion (1.7%) of the large variation in SBP across this sample of men is explained by the dichotomous heavy drinking variable. It is not surprising therefore that the estimates of the various path-specific effects are also imprecise. Examination of the residual distribution for each contributing associational model shows good agreement with the assumption of normality while evidence for the interaction terms was weak (see Web Table 5). There is evidence of a small indirect effect through GGT alone (mean difference ranging from 2.85 to 3.10 mmHg, lower 95% confidence limit ranging from 1.05 to 1.43, upper 95% confidence limit ranging from 4.31 to 5.06), little evidence of path-specific effects through either BMI alone or both BMI and GGT, with the remaining part of the total effect attributed to a direct effect via other pathways (mean difference ranging from 5.07 to 5.25 mmHg, lower 95% confidence limit ranging from 1.35 to 1.48, upper 95% confidence limit ranging from 8.76 to 9.03). There is little variation between the eight versions of each effect. As a consequence, when we depict the 24 possible decompositions in Figure 2, they are all similar, which suggests—in this example—that conclusions about the comparative strengths of different pathways could be drawn from just one particular decomposition.


Causal mediation analysis with multiple mediators.

Daniel RM, De Stavola BL, Cousens SN, Vansteelandt S - Biometrics (2014)

With  (perfect correlation between  and  given ), all 24 possible decompositions of the total causal effect of heavy drinking on SBP into four path-specific components: a direct effect unmediated by BMI or GGT, an indirect effect via BMI alone, an indirect effect via GGT alone, and an indirect effect via both BMI and GGT. The numbers superimposed on the bars represent the code for that effect type (as defined in the caption of Table 2). The numbers along the x-axis represent the decomposition number, also defined in Table 2.
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4402024&req=5

fig02: With (perfect correlation between and given ), all 24 possible decompositions of the total causal effect of heavy drinking on SBP into four path-specific components: a direct effect unmediated by BMI or GGT, an indirect effect via BMI alone, an indirect effect via GGT alone, and an indirect effect via both BMI and GGT. The numbers superimposed on the bars represent the code for that effect type (as defined in the caption of Table 2). The numbers along the x-axis represent the decomposition number, also defined in Table 2.
Mentions: The results are shown in Tables 4 and 5 and Web Figures 1 and 2. There is evidence of a total effect of heavy drinking on SBP, but the associated confidence interval is wide (mean difference 7.63 mmHg, 95% CI 3.89–11.37). Only a small proportion (1.7%) of the large variation in SBP across this sample of men is explained by the dichotomous heavy drinking variable. It is not surprising therefore that the estimates of the various path-specific effects are also imprecise. Examination of the residual distribution for each contributing associational model shows good agreement with the assumption of normality while evidence for the interaction terms was weak (see Web Table 5). There is evidence of a small indirect effect through GGT alone (mean difference ranging from 2.85 to 3.10 mmHg, lower 95% confidence limit ranging from 1.05 to 1.43, upper 95% confidence limit ranging from 4.31 to 5.06), little evidence of path-specific effects through either BMI alone or both BMI and GGT, with the remaining part of the total effect attributed to a direct effect via other pathways (mean difference ranging from 5.07 to 5.25 mmHg, lower 95% confidence limit ranging from 1.35 to 1.48, upper 95% confidence limit ranging from 8.76 to 9.03). There is little variation between the eight versions of each effect. As a consequence, when we depict the 24 possible decompositions in Figure 2, they are all similar, which suggests—in this example—that conclusions about the comparative strengths of different pathways could be drawn from just one particular decomposition.

Bottom Line: Much progress has been made, mainly due to contributions from the field of causal inference, in understanding the precise nature of statistical estimands that capture such intuitive effects, the assumptions under which they can be identified, and statistical methods for doing so.These contributions have focused almost entirely on settings with a single mediator, or a set of mediators considered en bloc; in many applications, however, researchers attempt a much more ambitious decomposition into numerous path-specific effects through many mediators.We discuss the strong assumptions under which the effects are identified, suggesting a sensitivity analysis approach when a particular subset of the assumptions cannot be justified.

View Article: PubMed Central - PubMed

Affiliation: Centre for Statistical Methodology, London School of Hygiene and Tropical Medicine, Keppel Street, London WC1E 7HT, UK.

Show MeSH
Related in: MedlinePlus