Morgenstern, J. Empirical evidence of the availability bias in emergency medicine, First10EM, March 28, 2022. Available at:
https://doi.org/10.51684/FIRS.126555
I recently posted a textbook chapter that I wrote about availability bias. (I strongly recommend the textbook, if anyone is interested in the topic of bias and decision making in medicine.) However, most discussion of these cognitive biases in medicine has remained purely theoretical. The biases make sense, and sound important, but there is not great science demonstrating that they play an important role in medical outcomes, nor that we can do anything about them. (The topic of the other chapter I posted.) That is what makes the study we are going to discuss today so fascinating: it is a demonstration of availability bias in the real world.
The paper
Ly DP. The Influence of the Availability Heuristic on Physicians in the Emergency Department. Ann Emerg Med. 2021 Nov;78(5):650-657. doi: 10.1016/j.annemergmed.2021.06.012. Epub 2021 Aug 6. PMID: 34373141
The Methods
This is a retrospective study looking at a large national Veterans Affairs database from the United States.
Patients
Adults aged 21 and older who visited a VA emergency department with shortness of breath between 2011 and 2018.
Comparison/ Outcome
They compared the rate of PE testing (Ddimer and CTPA) ordered by individual doctors in the 60 days before that doctor made the diagnosis of a PE to the 60 days after.
The Results
They include 7,370 physicians who had seen 416,720 patients for shortness of breath across 104 hospitals. The population is what you would expect from a VA hospital: mean age 63, 90% male. 11% had a diagnosis of malignancy made in the prior 6 months, 6% had a past history of PE or DVT, and 2% had surgery in the last 4 weeks. (So the population is almost certainly higher risk than the average dyspnea patient in a community ED.)
Physician testing for PE increased by 1.4% (95% CI 0.4-2.3%) in the 10 days after making a diagnosis of PE as compared to the 10 days prior. This represents a 15% relative increase in PE testing compared to baseline. The rate of PE testing then returns to baseline after that first 10 day period.
They ran a control sample using the same analysis, but looking around the diagnosis of a pneumothorax, and there was no increase in PE testing.
My thoughts
This is a fascinating, and in my mind important, study. They demonstrate a clear increase in PE testing based not on patient characteristics, but on the recent experiences of the doctor. However, the experiences of the doctor are completely independent of the probability that a patient has a PE, so this increase in testing must represent harm to those patients.
There is an important distinction that needs to be made between heuristics and biases, although the distinction is admittedly very fuzzy. When a patient is rolled into my resuscitation room with a respiratory rate of 6 and pinpoint pupils, I immediately know the diagnosis. It would be silly to waste time sitting down and formally thinking through a complete differential diagnosis before acting. My decision making in this scenario is based on a heuristic, and it is incredibly powerful in emergency medicine. However, it quickly becomes a bias if I think that every patient with pinpoint pupils must be intoxicated, and refuse to expand my differential diagnosis to include things like trauma.
How do we know that this study demonstrates bias, and not just learning? Perhaps these physicians had never heard of the Well’s score, and only learned the risk factors for PE after making the diagnosis? Perhaps these physicians were improving with time, and the increase in testing simply represents improved recognition of PE symptoms with more experience? If this were the case, we would expect the increase in testing to be persistent. Instead, the increase only occurred in the 10 days after the diagnosis was made, and then reverted to baseline. Therefore, the increased testing appears to be the result of availability bias, not learning.
This data is all based on chart review, which introduces some obvious limitations. However, because the unit of study was the individual doctor, and the limitations on data should be equal in the before and after stage, I don’t think this is very likely to influence their results. (Although, doctors might be more explicit in their documentation of PE risk factors after seeing a PE, which would probably fit with their study hypothesis.) Despite the many limitations of observational data, I think this study is important, and hopefully represents the first of many using empirical data to demonstrate the impact of cognitive biases on medical decision making.
This might be the first paper I have ever covered that has just a single author. For such a creative and data heavy study, that is really incredible, and Dr. Ly deserves a lot of credit. On the other hand, it also means that there was no one internal to the study to double check his work, so errors and bias are probably more likely.
As a side note, not working in the United States, I didn’t know it was possible to avoid their notorious malpractice system. This author states that one of the benefits of using a VA database is that it “allows us to reduce the influence of malpractice concerns because VA physicians cannot be sued in civil court for a malpractice claim.” Such concerns certainly could alter practice after a missed diagnosis, although there is plenty of data from elsewhere that malpractice is not actually the big driver of over-testing, and we just hate being wrong, legal responsibility or not.
Unfortunately, although this paper clearly demonstrates that availability bias drives medical decision making in emergency medicine, it doesn’t provide us with any solutions. (Some theoretical, but unproven solutions are discussed in the availability bias textbook chapter.)
Bottom line
This retrospective database study provides real world evidence that availability bias influences physician decision making, to the detriment of our patients.
Other FOAMed
Decision Making in Emergency Medicine: Availability Bias
Decision Making in Emergency Medicine: We can’t escape bias
4 part series on Cognitive errors:
- Part 1: A brief overview of cognitive theory
- Part 2: Common cognitive errors
- Part 3: Possible solutions
- Part 4: Problems with cognitive theory
References
Ly DP. The Influence of the Availability Heuristic on Physicians in the Emergency Department. Ann Emerg Med. 2021 Nov;78(5):650-657. doi: 10.1016/j.annemergmed.2021.06.012. Epub 2021 Aug 6. PMID: 34373141
Morgenstern, J. Availability bias. In Raz M, Pouryahya P (Eds). Decision Making in Emergency Medicine. Singapore. Springer Singapore; 2021.
3 thoughts on “Empirical evidence of the availability bias in emergency medicine”
Of course this is availability bias and, being human, doctors are prone to it. But bias is often good and appropriate. Is the bias inappropriate? A reasonable interpretation is that physicians who are reminded about a disease in a visceral way, raise their pretest probability, and then (because humans are not good at changing behavior once a stimulus is removed) gradually revert to their old lower pretest probability estimate i. e. decide not to or simply forget to test. In this case the old estimate was clinically correct. But in other cases the new estimate will be correct: you have been missing cases or the incidence (probability) is actually increasing. Eliminating availability bias can be done by using a rigid protocol (think algorithm) i.e. by removing the human factor but that has negative as well as positive consequences.
Thanks for the comment Mark
You are correct that I did not emphasize clearly enough that this study has no way of knowing which state represented better care: the low or high testing states. I am almost certain in PE, the low testing state is better, as there is a massive amount of data that we significantly over-test for PE, not undertest. That might not be the same for all conditions. It is a really important, and related, but somewhat separate question. There is definitely much much more research that could be done in this area.
Actually in this case low testing proved as good as high testing. (And I’m pretty sure that even Wells picks up incidentalomas) .My point is that this is not generalizable.