This chapter addresses the question of whether causal models can themselves be justified. We outline strategies for justifying models on the basis of prior data and so empirically grounding beliefs about the probative value of clues.

We describe strategies for figuring out whether a model is likely doing more harm than good and for comparing the performance of different models to one another.

We provide a lay-language primer on the counterfactual model of causal inference and the logic of causal models. Topics include the representation of causal models with causal graphs and using causal graphs to read off relations of conditional independence among variables in a causal domain.

This chapter defines causal questions in the language of causal models. We describe major families of causal queries, including case-level queries about causes of effects, case-level attribution, population-level effects, and effects along causal pathways. We show how these can all be defined as queries about the values of nodes in a causal model.

This chapter integrates the notion of “theory” into the causal-model framework that we use in the book. We describe an approach in which theoretical claims are thought of as model justifications within a hierarchy of causal models. The approach has implications for the consistency of inferences across models and for assessing when and how theory is useful for strengthening causal claims.

When very young children are brought to a doctor or hospital with signs and symptoms consistent with head injury, it is important to determine the cause. For almost 50 years, the triad of subdural haematoma (SDH), retinal haemorrhage (RH) and encephalopathy has been regarded as an accurate predictor of deliberate shaking and widely used to diagnose shaken baby syndrome (SBS). Statistical analyses by Cardiff University researchers and others claim to show that certain combinations of findings are highly predictive of abuse and as a result of this conclusion, protocols such as mandatory reporting to police are invoked in the name of protecting the child. However, concerns have been raised about the circularity of approach used in the statistical analyses which requires each case to be classified explicitly as either abuse or non-abuse. By producing a causal model of the problem, we show that these findings are actually a poor predictor of SBS, even where there is some evidence of risk factors indicative of abuse.

Structural equation modeling (SEM) is a family of statistical techniques and methods for testing hypotheses about causal effects among observed or proxies for latent variables. There are increasing numbers of SEM studies published in the research literatures of various disciplines, including psychology, education, medicine, management, and ecology, among others. Core types of structural equation models are described, and examples of causal hypotheses that can be tested in SEM are considered. Requirements for reporting the results of SEM analyses and common pitfalls to avoid are reviewed. Finally, an example of evaluating model fit is presented along with computer syntax so that readers can reproduce the results.

In Chapter 3 I introduce the basics of causal modelling and Bayes nets. I show how formal causal models support various types of inference, including prediction, diagnosis, intervention-based reasoning and counterfactuals. The framework serves several purposes: helping us to build formal models of legal cases, giving us a benchmark against which to assess human inference, and suggesting how people can improve their reasoning.

The chapter shows that more sophisticated difference-making theories of causation that draw on so-called causal models can accommodate mental causation too. Causal modelling theories invoke more complex relations of difference-making than the simple principle about causation that was used in previous chapters. These relations of difference-making are represented by causal models. Accommodating mental causation – either in the non-reductive physicalist case or in the dualist case – calls for some heterodoxy in model-building. If the heterodox models are allowed, however, they prove useful not merely for explaining mental causation, but also for capturing the distinction between higher–level causes that are explanatorily relevant and higher-level causes that are not. The chapter also discusses the interventionist theory, an especially prominent member of the causal modelling family, in relation to mental causation.