Mediation Analysis

It’s not causal until it is

Mediation Analysis

Have you ever asked yourself “how long would I survive if I was diagnosed with cancer?” You would probably start thinking about what it would be like to be told that you had cancer, you may then think of living with the cancer, and maybe even surviving to see your cancer cured. Living with your cancer encompasses an infinte number of experiences that can impact positively, or negatively, on you surviving your cancer, for example you would experience how you responded to your treatment. This treatment is defined as a mediator because it can alter your survival depending on whether you had the treatment or not, or how well you responded to your treatment.

In most cases, your doctor may say “amongst patients with this cancer (like the one you are imagining), 90% will survive after 5 years” - this means that 90% of the patients with this cancer will be alive after 5 years. But these simple statistics very rarely incorporate different scenarios - it seems silly because every patient has a unique experience of their cancer. For example, we may have been diagnosed when we went to see our general practitioner or when we went into accident & emergency (this is another example of a mediator).

Our question now changes to “how long would I survive if I was diagnosed via my local GP or via A&E?“. It is a hypothetical question since we cannot be diagnosed through both our local GP and via A&E independently. Here is where mediation analysis becomes a useful tool.

Avatar
Matthew J Smith
Doctoral Student of Biostatistics

My research interests include survival analysis, causal inference and missing data.