From observation to causation: Take-aways from a training course in Utrecht

A blog by Dr Sam Quarton

Does X cause Y? It is a key question in research, and yet can be surprisingly difficult to answer. We can measure something and observe that if it is higher, the likelihood of a particular outcome is greater or less. But teasing out the relationship between the two is not so simple. Does X influence Y, Y influence X, or are both impacted by other factors? Correlation does not equal causation, and a 5-minute internet search throws up any number of graphs showing spurious relationships between, for example, the per capita consumption of margarine, and the divorce rate in Maine.

As a researcher within the NIHR Birmingham BRC Infection & Acute Care theme, this is problematic. I am investigating hospital-acquired pneumonia, a complication of hospital care that is unfortunately common. However, when reviewing data for patients with hospital-acquired pneumonia I face this same problem – how can I disentangle the trends I see, to attribute cause and effect? For instance, I want to investigate which patients do better – those treated with antibiotics targeted at bacteria grown from that patient, or those treated with broader ‘best-guess’ antibiotics. Simply comparing outcomes between the two groups is likely to be confounded by multiple factors. For example, a frail patient who struggles to cough is less likely to produce a sputum sample that would allow us to identify bacteria causing infection. This in turn means they are more likely to be in the ‘broad antibiotics’ group. However, they are also more likely to have poor outcomes, regardless of the antibiotic they receive.

From problem to solution

This summer I was supported by the BRC to attend a week-long course hosted by Utrecht University on causal inference. This was a fantastic exploration of how to apply causal inference methods to data, hosted by world-leading experts in the field, and attended by researchers from around the world. Through lectures, discussion, and applied practical sessions, we developed skills in understanding and applying techniques to mitigate the problems discussed above and infer casual relationships within our data. These techniques, such as the potential outcomes framework, propensity score matching, and target trial development, provide a rigorous and statistically meaningful approach to identify the true impact of treatments.

Future applications

The opportunity provided for me by the BRC in attending this course means I will be able to apply the techniques learned to better understand what impacts prognosis for patients with hospital-acquired pneumonia. Excitingly, applying this may identify targets for intervention, to improve outcomes for what is a common and serious infection.

Beyond this, these techniques are relevant to any analysis looking at the impact of interventions on outcomes. As such this course has equipped me with tools I will be able to apply throughout my career. By helping me assess the impact of treatment across various research questions, with potential improved outcomes for patients, the causal impact of this course will hopefully be experienced for years to come.