Correlation (even multivariate analysis) is not causation

I’ve been reviewing some of my PhD proposal (it has to be submitted by the end of this week), and considering the topic of coping.  Coping refers to ‘the strategies people use to manage pain and
its impact.’  It is one of the two main topics researched in psychological contributions to the pain experience, with the other being beliefs.

Although coping has been studied extensively, it has primarily been studied in people who are seeking treatment for their pain – and conclusions drawn about what constitutes effective coping is drawn from outcome studies looking at disability scores and correlations between the strategies used.  My question is whether that is the most adequate way of viewing the use of coping strategies, maybe it would be good to look at what people who have never had formal CBT treatment for chronic pain and see what they use.  Anyhow, this editorial by Mark Jensen reviews the tendency for even really experienced researchers to be tempted to use language that suggests causation when a conclusion can only be drawn about the tendency for two (or more) variables to systematically vary in relationship to each other.

He is discussing the findings of Karsdorp and Vlaeyen who examine the associations among psychological variables in a large sample of patients with fibromyalgia. They conclude that two types of avoidance strategy independently predict disability in this cohort of patients.

Jensen states ‘Psychological models of pain often hypothesize causal and mediational associations between different psychological factors or domains’ – the problem with this is that given the close relationships between some of these psychological variables, a ‘multivariate analysis that (1) estimates the associations between these variables and important criterion variables that also (2) controls for other psychological variables, will likely underestimate the importance of the psychological factor(s) being examined.’

What this means is that, in this study, two variables – catastrophising and pacing – were found not to predict disability when each was controlled for the other.  The authors of the study concluded that neither were significant predictors of disability – but what if the two varied systematically with each other because of some unidentified third variable?  Jensen suggests that perhaps depression may influence the use of both catastrophising and pacing, making it seem as though neither were individually significant, but actually confusing the finding because depression was not considered to be a mediating variable.

Jensen suggests ‘In short, negative results from studies that use analyses that control for psychological variables should not be used to draw strong conclusions about the lack of importance of any one variable.’

He finishes by stating ‘their finding that active avoidance is uniquely important in the prediction of disability when controlling for other psychological factors is an important one… but at the same time, one must be constantly vigilant to avoid viewing the findings from correlational studies as suggesting the presence of causal associations’.

He goes on to say  ‘the true (causal) importance of a psychological variable is best identified by experiments that systemically alter the variable in question, and then determine the subsequent effect of a change in the variable on measures of important outcomes.’

This is a timely reminder for us as consumers of research – we really do need to know about research methodology, or we may well skip to the ‘Discussion’ and miss the important details of how the researchers drew those conclusions.  Only then can we decide how much weight we put on the findings.

Jensen, M. (2009). Research on coping with chronic pain: The importance of active avoidance of inappropriate conclusions Pain DOI: 10.1016/j.pain.2009.07.036