Report from one of the sessions at the RSS 2017 Conference. More reports of conference sessions are listed here.
How do we know that smoking causes lung cancer? Does discrimination play a part in hiring decisions? Can we predict economic growth under a new tax policy?
These are all questions about causation, and as any statistician will know, correlation does not equal causation. It usually takes a randomised control trials (RCT) to make the case for this. However, observational data are more widely available and easier to acquire. Marloes Maathuis, professor of statistics at ETH Zurich who specialises in causal inference, discussed ways in which we might address causal questions using observational data during her Wednesday keynote session at the RSS conference.
Using an example of a table showing numbers of re-offending prisoners on release following a rehabilitation programme, she looked at what questions we might be able to ask regarding this observational data. Is it possible, for example, to make assumptions on whether the programme lowers the re-arrest rate? Could we predict re-arrest rate if the programme were made compulsory?
Marloes explored ways in which causality might be calculated, how confounding factors could be integrated and how total causal effects identified. She then showed how these methods could be applied (in this case, to multiple gene expression in a sample of plants) and how to make sense of the results.
In summary, Marloes recommended key points to remember:
- Specify the type of causal effect
- State causal assumptions
- Use causal methods
She stressed that RCTs are still important - but rather than not making any assumptions at all regarding observational data, use can still be made of it, provided it’s done in a principled way.
Marloes Maathuis’s talk was titled: Causality: using covariate adjustment to estimate total causal effects.