“Signal & Noise” is a bimonthly column by Brian Taylor, AuD.
During the 2008 election, data guru Nate Silver and his website FiveThirtyEight accurately predicted the winner of the presidential election in 49 of 50 states. After tweaking his data analysis methods he did even better in the 2012 presidential race by accurately predicting the winner of all 50 states. He also has demonstrated the same uncanny results for U.S House & Senate races – and even the Super Bowl. Time will tell, however, if his predictions for the 2016 race are correct: as of October 28, Hillary Clinton has a better than 80% chance of winning the general election, according to Silver’s analysis.
More Accurate Predictions
Regardless of your political persuasion, one is left to wonder, how does Nate Silver get so consistently accurate in his forecasts? The secret to his success is the creative use of Bayesian statistics. Named for English philosopher Thomas Bayes, a Bayesian approach combines knowledge about past events with the observational knowledge of the present to make decisions about the future.
In the case of the 2016 presidential election, Silver and his FiveThirtyEight colleagues applied two separate models to forecast the election – Polls-Only and Polls-Plus models. The Polls-Only model relied only on polls from particular states, while the Polls-Plus model was based on state polls, national polls and important endorsements.
For the presidential election at the state level, FiveThirtyEight produced probability distributions and averaged expected vote shares using both of these models. As polling data accumulates the probability of each candidate – Clinton & Trump – winning the race fluctuates. For example, since the summer, Donald Trump’s chances of winning the presidency, according to Silver’s model, has fluctuated between a high-water mark of around 47%, down to its current place of less than 20%. The reason for the fluctuation is simple: there are literally dozens of polls, continually updated with new information. Thus each candidate’s probability changes as new information becomes available.
Since Silver’s use of Bayesian statistics for predicting elections became famous, several other groups have applied the approach to predicting outcomes of this election. Each group uses the accumulating poll data a little differently, which, in the case of the 2016 presidential campaign, means Clinton’s chances of victory, when applying the varying Bayesian approaches, currently ranges between 80 to 99%.
Better Clinical Outcomes
So, what does a Bayesian approach to predicting elections have to do with audiology? When it comes to making decisions about diagnosing ear disease and predicting success with hearing aids, potentially – a lot. Think of it this way: each patient that you see in your clinic is kind of like the election, and the tests, case history and other things you learn about the patient modify the probabilities of the final outcome. The patient’s lifestyle, genetic make-up, test results, etc. all combine to provide you with a probability score for having an ear disease or being successful with hearing aids. In the case of ear disease, each piece of meaningful information that accumulates about that patient has the potential to improve the probability of identifying a medically treatable condition.
Historically, when it comes to making a referral for a medical evaluation for suspected ear disease, we have had a tendency to look at one or two tests at a single point in time. A patient with PI-PB rollover or asymmetrical word recognition scores, for example, gets referred for further testing without looking at other aspects of their history. (This would be akin to Nate Silver basing his prediction on a couple of polls, rather than dozens of polls collected over time.)
The more careful clinician, on the hand, would apply a Bayesian approach and look at the patient’s age and other risk factors before making the referral for further tests. A Bayesian approach – using accumulating knowledge to improve the probability of a future outcome – in audiology could have a dramatic impact on improving healthcare efficiency and cost. For example, under the current FDA guidelines, consumers can sign a medical waiver prior to obtaining hearing aids – the vast majority of patients fitted with hearing aids probably sign the waiver. Thus the burden of detecting medically treatable ear disease falls upon the consumer and the hearing care professional who asks the consumer to sign the waiver.
For those consumers who proceed with the audiological evaluation to rule out ear disease, they are relying on a test battery that is sensitive enough to identify a potentially treatable condition at the risk of wasting time money on unnecessary tests. In reality, both approaches to protecting the consumer– the signing of the medical waiver and the audiological test battery to identify ear disease – have surprisingly little evidence supporting their ability to protect consumers from unnecessary additional tests or to making hearing aids more accessible or affordable.
That’s where a Bayesian approach – the accumulation of past information to make decisions about the present – can help. Using Bayesian statistics, similar to Nate Silver & FiveThirtyEight, audiology researchers could build models using known risk factors for ear disease to determine probabilities for each individual who may be in need of further medical attention prior to using hearing aids. Putting an accurate Bayesian tool in the hands of the consumer or the audiologist has the potential to drive costs down and improve access & quality. In fact, one group of researchers seems to be in the process of developing such a tool.
The use of Bayesian statistics does not have to be confined to the more accurate and cost effective identification of ear disease. Such a tool could be developed that defines the probability of success with hearing aids or any another intervention. Let’s say you enter a patient’s information on hearing thresholds, SNR loss, LDLs, cognitive ability, age, physical ability, motivation and a constellation of other factors into a computer model, and in a matter of seconds you can share with the patient their probability of success with hearing aids. I suspect that experienced clinicians with great instincts already do something like this, but imagine their proficiency at forecasting hearing aid success when data is aggregated and used in the process. Of course, we need to get on the same page with the definition of “success” but once we get there, the application of Bayesian statistics could make our job much easier.
Calling All PhD Audiologists
We need PhD audiologists and data gurus like Nate Silver to build mathematical models allowing us to enter this data about every patient into a formula we can use to make probabilistic statements about the success of our interventions. Imagine how much quality of care could rise and costs could be reduced when we can stratify patients at-risk for poor outcomes before we fit them with hearing aids. A Bayesian approach may even help us raise the respect of our profession in the eyes of other health care providers and the public at large.
Brian Taylor, AuD, is Senior Director, Clinical Affairs, for Turtle Beach/Hypersound. He continues to serve as Editor of Audiology Practices, the quarterly publication of the Academy of Doctors of Audiology. During the first fifteen years of his career, he practiced clinical audiology in both medical and retail settings. Since 2005, Dr. Taylor has held a variety of leadership & management positions within the hearing aid industry in both the United States and Europe. He has published over 50 articles and book chapters on topics related to hearing aids, diagnostic audiology and business management. Brian has authored three text books: Fitting and Dispensing Hearing Aids(co-authored with Gus Mueller), Consultative Selling Skills for Audiologists, and Quality in Audiology: Design & Implementation of the Patient Experience. His latest book, Marketing in an Audiology Practice, was published in March, 2015. Brian lives in Golden Valley, MN with his wife and three sons. He can be reached at firstname.lastname@example.org or email@example.com.