Funding

Causal models of complex interventions

Methodology for Development of Causal Models of Complex Interventions; Context and Mechanism of Action 

Background

It is important to understand not just which healthcare/public health interventions work but how they work. This is because science is inductive over place and time; the results in one place and time can be extrapolated to another only by understanding a theory that attempts to capture mechanisms of action of the intervention, and the nature of the context in which it intervened.  Causal modelling is one way to attempt this.

Causal models describe the causal relationships between variables in a system, which include exogenous variables (values determined by factors outside the model) and endogenous variables (values determined by factors within the model). The variables may be directly observed (or directly observable) variables, or latent variables that cannot be observed directly but are assumed to influence variables in the model. The variable set may include observational variables plus one or more experimental/interventional variables (manipulated by the investigator). A causal model may be specified as a set of equations or, equivalently, as a causal diagram.

Causal models describe the mechanism(s) by which variables in the model influence each other. They can, therefore, be used to understand and predict causal relationships between an intervention in a complex system and the context in which the intervention takes place. In this way, causal models can be used as a tool to aid intervention development, by specifying (i) the variables that need to be changed by the intervention in order to produce the desired change in outcome and (ii) the components of the intervention that target these variables (for an example, see Case Study 1 in the MRC Developing and evaluating complex interventions guidance).

However, methodological work is needed to enable extension of causal models beyond simple, single-level, linear cases, to multi-level and more complex systems that include feedback loops and other non-linear causal relationships.

A further challenge is in having the methodology to use causal models to understand and predict the interaction between complex intervention and complex system context.  Such methods would enable understanding and prediction of whether an intervention that has been shown to be effective in one context would be effective in another.

For example, Pearl and Bareinboim use a specific form of causal diagrams called “selection diagrams” to address the problem of the generalisability of empirical findings to new environments, settings or populations (which they call “transportability”).  There may also be useful heterogeneity between contexts which might give valuable information, e.g. in the UK between areas with different population demographics.

In developing a model there should be a balance between the use of empirical/real-world data to inform the model, and the utility of the model in predicting future outcomes in other cases; there is missing methodology for yoking together inductive and deductive reasoning in a model.

Complexity often comes from the fact that we have health systems where effects are emergent and context is not static – it continually changes.  With regards to the methodologies for studying dynamics and how a system emerges, there is a significant unmet need for causal modelling methods accounting for system dynamics.  Furthermore there are difficulties around appropriate methods for ways to achieve parsimony in selecting which variables to include in a complex model; there is a tension between the infinitely complex real world and the abstraction that models represent.  Sensitivity analysis, for example, may help to achieve parsimony in which variables are included in a model.

There are also methodological challenges in how models might be validated. Given sufficient data this can be accomplished by e.g. structural equation modelling or by Bayesian network modelling.

Highlight notice

MRC and NIHR invite applications, through the Methodology Research Programme panel, to conduct research into developing methodology for building causal models of complex interventions in complex healthcare and public health systems, enabling understanding of the interactions between context and mechanism of action.

Applications are particularly sought which propose development of methodologies for:

  • how to build and validate a causal modelling method which can induct from cases to specify a model, and then deduct from the model to predict future outcome in other contexts; enabling understanding of mechanism of action and interaction with context and predicting generalisability of interventions to other contexts
  • how to enable the model to account for dynamic changes in context over time
  • how to extend causal models beyond the simple, single-level, linear case, to multi-level and more complex systems that include feedback loops and other non-linear causal relationships
  • how to take account of the complexity of the real world context without overloading the model

Application process and schedule

Applications for projects are invited through the normal MRC funding grant schemes (Research Grant, New Investigator Research Grant) and must be submitted to the regular Methodology Research Programme Panel deadlines. Applications will be reviewed at the regular MRP panel meetings, and will be in competition with other applications received, but the Panel will be mindful of the strategic importance of this area.

Contact and guidance

The titles of all applications in response to this highlight should be prefixed with 'HCM:' when filling out the JES form, and on any attachments, e.g. “HCM: A method for…”

It is essential to discuss your proposals with MRC Head Office at an early stage. All applications must be approved by the Methodology Programme Manager prior to submission. Please contact:

Dr Samuel Rowley

E-mail: samuel.rowley@headoffice.mrc.ac.uk

Reference

Pearl J, Bareinboim E. External validity: From Do-calculus to transportability across populations. Statistical Science 2014;29:579-95.