Predicting the future doesn't have to be hard

Here at c-WhIP we're especially interested in trying to predict who is going to recover in a reasonable and average time, and who is destined for a more bumpy road to recovery.  We attempt to do this through the creation of theoretical models, driven by sound empirical research where available, and then collecting data to evaluate the validity of those models.  For example, we are currently running a study built on the presupposition that activity in the stress-response system at or near the time of a whiplash injury influences the likelihood of recovery.  Since we can't directly see the stress-response system in action, the best we can do is evaluate markers of its activity (ie. cortisol in this case), and then see how acute cortisol response maps onto outcomes 3 or 6 months down the road.

One of the greater challenges in this line of research, one I didn't think would be as big a deal as it has turned out to be, is the definition of precisely what it is that we're trying to predict.  In other words, what is 'recovery' and how should we measure it?  That particular story is one for another post.  However, regardless of the definition, one thing that is almost universally true is that the best predictor of any chosen outcome will be scores on the same measure at baseline.  In other words, when creating multivariate regression models, I would say almost always (I've personally yet to see this not be true, but I'm sure it happens), the best predictor of outcome will be scores on the same scale captured 6 or 12 months earlier.

This has been shown once again in a new paper by Rasmussen-Barr and colleagues from the well-regarded Karolinska Institute in Stockholm, Sweden.  In this prospective evaluation of 71 subjects with recurrent low back pain, the authors sought to identify important baseline prognostic factors that could be used to identify people who are most likely (or most unlikely) to benefit from an 8-week rehabilitation program.  The outcome used in this case was the Oswestry Disability Index (ODI), with pain as a secondary outcome.  In the interest of brevity I'll spare the final details and cut to the chase - after multivariate regression, baseline ODI score, pain intensity and pain self-efficacy were the only variables, out of several collected, that explained significant unique variance in 12- and 36-month disability or pain scores.

There are several potential reasons for this phenomenon.  For example, when using regression, you are looking for the most parsimonious set of variables that can explain the variance (think: distribution) of scores at outcome.  It should come as no surprise that the variable with the most similar distribution of scores at baseline, the same scale, would explain the greatest variance at outcome.  Another mechanism through which this phenomenon occurs is probably due to internal calibrations - while two people may in fact have identical levels of disability (suspending disbelief in the interest of illustrating a point), they will very likely still score differently when asked about their perceived level of function.  For that reason, even if both were to recover at a similar rate - let's say I go from a 40 to a 30, and you go from a 15 to a 5 (assuming that our disability scale is in fact linear, likely requiring another suspension of disbelief), it could be said that we've both recovered the same amount - 10 points.  But compared to you, I started higher and I finished higher, so the correlation is going to be fairly strong and hence that scale will be retained when trying to predict outcomes.

So is there value in this type of knowledge?  Absolutely - from a policy-making perspective, I want to know that people presenting to our program with severe pain or disability ratings, or low self-efficacy, are less likely to respond to our program.  Therefore, perhaps we shouldn't dedicate the resources to your participation in program without addressing these other things first.  Good to know indeed.  But from a clinical decision-making standpoint, I get more excited by seeing models where a construct from a totally separate domain predicts what appears to be a distinct outcome.  This is why I get so excited by findings like those of my colleague Jim Elliott, who showed that one of the strongest predictors of fatty infiltrates in neck muscles at follow-up is scores on a post-traumatic stress symptoms scale at baseline.  Or even our recent finding that scores on the Neck Disability Index can be predicted from pressure pain threshold over the tibialis anterior at baseline.   Now we're getting into mechanisms - not only can we predict a person's outcome, but why is the person at risk (or not) of achieving that outcome?

This is not to say that work such as that of Rasmussen-Barr and colleagues is not important.  Quite the contrary, it is very important from a policy standpoint.  It's all a matter of the question you're asking and what you want to get out of the research.  As Doctor Phil likes to say 'The best predictor of future behavior is past behavior', and it seems as though he's usually right (please don't ask how I know that's what he likes to say).  However, this post also describes why you'll rarely see c-WhIP create regression models that include the same scale as both a dependent and independent variable - our concern is not only who will develop long-term problems, but why.

Don't get him mad folks, he'll come after you.