Comment on:

- Sariaslan, A. et al. (2016). Long-Term Outcomes Associated with Traumatic Brain Injury in Childhood and Adolescence: A Nationwide Swedish Cohort Study of a Wide Range of Medical and Social Outcomes

**Real abstract is too damn long, this is my TL;DR version:**

Sibling control study of TBI as a causal effect for various outcomes. Data = Swedish register data, n = very large. Findings: sibling comparisons consistent with causal effects of TBI, but there is some familial confounding.

*Before going further, I’d like to say that I believe in the causal effects of TBI for some outcomes.*

With that said: sibling comparisons are a natural experiment-type design that allows us to control for two things by design alone: shared ‘environmental’ effects — any non-genetic effects shared by siblings that make them more similar — and genetic effects. Let’s put on the JayMan hat: our prior is that similarity between family members is 100% due to genetics and 0% due to non-genetics. We can examine the results to see if they are consistent with a genetics + noise model, i.e. no causal effects of TBI. It goes like this: the raw association between TBI and outcomes is 100% due to genetics. When we use a sibling comparison, we adjust for 50% of the genetic confound, so we expect that the RR is reduced by 50%. If we had MZ data, we expect the RR to be 1.00. Are the numbers consistent with this?

Here’s the results from the paper:

(Model 3 is adjusted for some kind of educational variable. This seems fishy, especially for the educational outcome. Unfortunately, non-adjusted results are not reported.)

Then we calculate the predictions of a genetic + noise model:

We observe:

- A mean reduction of 54% of RR for raw to sibling. Genetics + noise model predicts 50%.
- Mean RR for hypothetical MZ of 1.07. Genetics + noise model predicts 1.00.

These data are either perfectly or very close to perfectly consistent with a pure genetic confound model. That being said, we can see some variation across outcomes, which may or may not be real variance or just sampling variation (n is large, but head injuries are rare). In any case, the most plausible (to me) causal effect is disability pension, which also has the largest RR for our hypothetical MZs of 1.22. That is, head injury causes a 22% increase in chance that one will be disability pension. Not a large effect, but one I’d like to avoid.

By writing this post, I feel a little bit like Fisher.