Monday, June 15, 2009

Inferring cause when there is a delay between cause and effect

Humans have evolved mental tools to seek causation from complex patterns. But the identification of causation is difficult when the effect is temporally distant from the many potential causes. For example, I've had a mildly sore piriformis m. for the last month. Friday morning, the soreness was much worse than the previous few days. My inference engine went to work to understand why, so I could avoid the cause and not repeat the pain. This is what humans do. It's what all animals do, it's just that the process rises to consciousness in humans (if not some other animals).

So what had I done to augment the soreness? Well, I pool ran for the first time in over a year on Wednesday, and pool ran again on thursday. To get a good workout pool running, you need to have very fast turnover. I ran Tuesday evening, 5 miles and easy, on grass. I hadn't run in 8 days prior to this. I went to the sports medicine doctor on Thursday morning, and in her series of diagnostic tests, she asked/facilitated various explosive jumps and deep stretches. And finally, I was reading about how posters had dealt with piriformis issues and I tried a few of the stretches that several posters swore by. There are many other antecedents that could effect pain (sleep patterns? stress? diet? sitting on the couch all day thursday? maybe just the way I slept or got out of bed that morning?) but this is a pretty good start. But how do I "know" which of these activities to avoid in the future to keep from aggravating my piriformis?

Humans will combine experience with biases (preconceived ideas or beliefs) to infer cause. If a belief is strong enough, it probably would not even occur to the person that there are multiple potential independent and interacting causes. The one prior event that matches the belief is recalled to memory but the other events are left in the dustbin of the brain. This is an example of the well-known and pervasive "confirmation bias". Confirmation biases reinforce prior belief and make alternative causes even less likely to be believed.

But I'm well trained to seek multiple potential causes - that's what I do for a living (although the discounting of some potential causes such as sleep, stress, and diet, and the complete refusal to even entertain planetary alignment & evil demons reflects my own set of beliefs). So I list out potential causal candidates, like I did above. But these causal candidates are a tiny fraction of the literally thousands of activities that I've engaged in over the last few days. The ones I listed reflect my own biases. Note that I put sleep, and stress, and diet in parentheses because I don't believe these as serious candidates for this case. Also note that I would never even entertain the possibility of an imbalance in my humors, spinal misalignment, or qi, much less something like planetary alignment or evil spirits.

So I turn to my experience (or maybe the experience of others). Of course it is these experiences that shape our biases, which in turn, shape our experiences (that is the strength of the memory of past events will be a function of our biases). But let's pretend my experience is objective. I've not had a sore piriformis before but I have had this injury wax and wane over the past month. What are the common activities that occurred prior to soreness feeling worse? Again, I simply cannot ask this objectively because the ability to recall activities in memory reflects my beliefs. So I turn to my logbook. But I write very little in my logbook with no notes at all on my diet, sleep, feeling of stress, etc., etc., etc. Regardless, certainly there are far too many common antecedents, most unrecognized, to infer causation with an confidence. My belief structure rejects most of these antecedents and narrows the list down to the few that fit my worldview (overuse and stretching) And now my inference engine looks for consistency; which of these consistantly preceded augmented soreness? That is our brain is searching for a pattern. Psychologists have a field day with freshman classes showing how humans frequently find patterns (associations between events) when non actually exist. There is too little data (coincidence between activity and increased soreness) and too much noise (all the spuriously correlated activities that have nothing to do with soreness and all of the un-recalled variables that might have contributed to the soreness). It's a futile exercise if the signal-to-noise ratio is low (I'll write about a case with a high signal-to-noise ratio later).

The point of all of this is, our ability to infer causation using our own experience when the effects are temporally far removed from the cause is utterly confounded by our belief structure and our proclivity to find pattern when none exists. If we want to know what works (avoiding injury, recovering from injury, performing at peak ability) in sport, then we have to exploit the methods of inference that overcome the many cognitive biases and illusions. These methods are the manipulative experimental methods of science.

Unfortunatly, we cannot do most of the manipulative experiments on humans that we'd really like to do. For example, I'd really like to know if running barefoot reduces injury. To do this, I'd need a population of humans that are cloned from a single genotype and raised in a common laboratory environment. They've each had the exact same diet, light, sleep, exercise, and education, all in their little cages. I'd then randomly assign 50 pairs to 50 different gyms with an indoor 200m track. Within each pair, one would always run with a modern running shoe and one barefoot. Same diet, same sleep schedule, same training schedule. And yes, they would never leave the gym. Then after two years, I'd stop the experiment and tabulate the rate of injury. Of course, maybe my shoe choice was particularly bad or maybe there is a genotype x shoe interaction (the shoe on the lab genotype caused more injury but wouldn't have on some other genotype). Or maybe there is no effect of shoes on a gym track but there is on a road or on a trail So I'd have to repeat the experiemnts using different genotypes, different shoes, and different surfaces. Not gonna happen.

Anything other than a controlled, manipulative experiment will have confounding factors. Without much ability to manipulate, what the science of sport (and most modern medicine) is left with is the attempt to control these confounding factors using statistics and, hopefully, large sample sizes. This is "evidence-based medicine" and it's the only game in town for inferring what works and what doesn't. There are huge problems with every evidence-based medicine study, but the best studies are better than our many experiments of one.

Note that Science Based Medicine has a really nice blog post this morning on, essentially, how the sort of confirmation bias that I've discussed above, can lead to pseudoscience quackery among professionals in health related fields. It's a long post and he only gets to the confirmation bias part near the end, but it's a worthy read.


  1. As a start for your barefoot study, why don't you just
    reuse the data from previous running shoe studies.

    They must have had some barefoot controls...right?

  2. Hi Scott - I'm hoping to post something on barefoot v. the shoe soon. I tend to lean toward the minimalism/barefoot world-view and always have, but I will be the first to admit this is largely because of wishful thinking (maybe it's the screw-the-man side of my personality!).