Identifying causation: the role of the clinical expert


One of the most frustrating realities facing both service providers and researchers in the field of developmental disability is how to establish a clear connection between a putative ‘cause’ and an outcome or ‘effect’. The daily challenges of our clinical work concern, among many other dilemmas, how to identify what causes a condition and how to assess whether an intervention works (i.e. causes the intended effects). When a change is anticipated after intervention (or indeed over time) we need to be prepared to speculate – before the fact – about what we think are the active ingredients that are likely to ‘cause’ any improvements we observe.

It is a longstanding truism in clinical epidemiology that association is not causation.[1] We have been taught that cross-sectional studies that identify correlations among variables can at best illustrate the possible interrelationships among those factors, but (usually) not the temporal relationship, let alone the mechanisms that might link those elements. Furthermore – and we too seldom recognize this reality – even the elements that we initially choose to measure and correlate in cross-sectional studies reflect – consciously or implicitly – our ideas about the possible mechanisms that cause the association and lead us to choose the variables on which we focus our attention.

A classic example from the field of childhood disability is the destructive but once popular belief that ‘refrigerator mothers’ caused their children to become autistic. In the climate of beliefs about psychodynamic causality popular in the 1950s and 60s it was easy to make the leap – from the observation regarding puzzled and apparently ineffectual parents of children with autism spectrum disorders (ASDs) – to the causal inference that impaired parenting was the active ingredient in the genesis of these impairments. With the benefit of advances in our understanding of the underlying neuroimpairments in people with ASDs, and our recognition of the transactional nature of child-parent development, these old ideas are now clearly recognized to be rather bizarre. However, they serve as an important illustration that how we think determines what we see, and especially how we see it!

In a recent issue of DMCN, the commentary by La Bastide-Van Gemert and van den Heuvel[2] (on the paper by Faebo Larsen et al.[3]) made the argument that advanced statistical methods such as causal graphs may allow researchers to take a step beyond the caution about causal inferences to which most of us have been exposed. The arguments are interesting, if rather complex, and the references cited are even less accessible to the non-expert than is the commentary. It is worth noting, however, that even with the most experienced biostatistical expertise there remains an obligation on the part of clinical researchers to lead the charge.

Whatever causal pathways one might explore statistically need to be guided and informed by the best of our current understanding of possible pathways and temporal relationships among the factors (variables) being correlated. Returning to refrigerator mothers, one might expect that using today's statistical methods and 1960s concepts a strong relationship between measures of mothering and the presence of ASD might have been observed. It seems equally likely that, were these same imagined data explored with today's understanding of the nature of ASDs the pathways between the variables would be drawn differently – as would the conclusions emerging from the statistical analysis.

These comments are in no way meant to diminish the incredibly useful role that biostatistical colleagues and their advanced methods play in enabling clinical researchers to delve more deeply than ever before into the data. They are simply a reminder that clinical research depends, first and foremost, on the articulation of a clear research question, supported by arguments about both the rationale for that question and some a priori commitment regarding how the relationships being studied might work. Under those conditions the findings are likely to be more interpretable and perhaps less susceptible to the post hoc ‘Ahas!’ that so easily can creep into our thinking and bolster what we already believed.