Omission Bias in Therapeutic Medical Decision Making

Having studied so-called heuristics and biases, aka cognitive biases and System 1 & System 2 processes in medical decision making on and off for 25 years, and thinking constantly about them in the full-time practice of clinical medicine, I have come to the conclusion that they are hard to spot in the wild. Except one: Omission Bias.

Omission bias represents a preference for inaction that preserves the status quo over action which changes it, and it is insidious because each day we encounter countless status quo states in medicine, established earlier and perhaps by different providers, but often by ourselves. The status quo inures us to what is, obscuring the question of what ought to be.

In this recent paper, and an accompanying editorial, omission bias is (partly) explained and several recurring examples in the ICU are discussed. But this is the tip of the iceberg. I venture a guess that every day, there is a decision that is made, or not made, that instantiates this bias. In addition to those discussed in the linked papers, here I will point out some common scenarios outside of the ICU where this bias lurks, preying on the desultory physician.

The patient presented to the ER (or elsewhere) for pneumonia, and was started on Zosyn (or cefepime) and vancomycin. Despite having no risk factors for pseudomonas, anaerobes, or MRSA, the team continues those antibiotics rather than change them to standard CAP coverage. The prospect of not covering something, always nagging at the physician, now looms larger because if the patient does unexpectedly have a resistant organism causing pneumonia, the anticipatory regret of not leaving things as they were deters the change.

The patient was started on unfractionated heparin for intermediate (or low) risk DVT/PE, when LMWH is the preferred drug, because of the misguided worry that "an intervention may be needed." Even if that were a legitimate concern at the outset, the patient has made it through the night and has improved, yet UFH is continued, requiring those pesky aPTT checks and the unpredictable pharmacokinetics of the drug.

Sunk Kidney Bias: A Lethal Form of Sunk Cost Bias

Hal Arkes

The heuristics and biases program of Kahneman and Tversky, once an obscure niche of cognitive psychology, became recognized among lay persons with Kahneman's Nobel prize in economics in 2002.  The popularity of the program surged with Kahneman's book Thinking Fast and Slow several years ago and several among the scores of related books about behavioral economics became best-sellers.  This year, Richard Thaler was the Nobel laureate in economics for his work in behavioral economics.   I became aware of heuristics and biases just before Kahneman's Nobel and started looking for them in medicine in 2003.  We (Aberegg, Haponik, and Terry, Chest, 2005) indeed found evidence for omission bias, and have discovered other biases along the way, some which are very intriguing but we aren't even sure what to name them (Aberegg, Arkes, and Terry, Medical Decision Making, 2006).  My point here is that these biases are useful but difficult to identify as patterns systematically operating within medical practice in predictable ways - they pop up here and there only to recede and reappear years later, if they are recognized at all.

Then there are biases about the biases.  Highly cited expositions of biases in clinical care, such as those of the insightful emergency physician Pat Croskerry (Academic Medicine, 2003), among many others) very often surmise the presence of biases in clinical care, without the kind of empirical evidence that established the biases in the first place.  Sometimes, new and probably useful biases are proposed (such as "search satisfycing"), without any empirical evidence, at all in any domain, for their existence.  They are merely postulates.  (Granted, empirical evidence is very difficult to generate, this the reason I don't do this kind of research anymore.)  Finally, the descriptions of the biases applied to medicine are often strained, or just plain wrong.  My favorite is the bastardization of "anchoring and adjustment" into a description of any time a physician seizes upon a diagnosis and discounts disconfirming evidence or fails to consider alternatives.  This is not anchoring and adjustment.  Anchoring refers to a numerical anchor, and failure to adjust away from it when providing numerical estimates.  Here is a summary of the original descriptions, from the wikipedia entry on anchoring and adjustment:

The Number Needed Not To Treat To Harm (NNNTTTH): A Heuristic for Evaluating Trade-offs in Medical Decisions

A frequent conundrum of decision making that arises in medicine is when there is a generally indicated therapy, say, anticoagulation for atrial fibrillation, that poses unique risks in a particular patient.  CHADS2 and HAS-BLED scores are calculated, but don't quiet the hemming and hawing or quell the hand-wringing.  What is usually a simple dichotomous decision is now one laden with probabilities, risks and benefits, and compromise between competing objectives.  (See:  The Therapeutic Paradox:  What's Right for the Population May Not Be Right for the Patient.)  In order to restore nuance to the decision, we need to try to estimate the numerical values of the risks and benefits to determine if the net utility of anticoagulation is positive or negative, something the aforementioned calculators are intended to do in a semi-quantitative way.  But what if you opine that your patient has a specially enhanced risk of side effects and you're worried about falls or bleeding but ambivalent because of a concurrent fear of denying him of the benefit of stroke prophylaxis?  What if you think that he would have never been included in a trial of stroke prophylaxis and the results of those trials may have limited generalizability to him?  What if you think he has only a year to live?

The number needed not to treat to harm (NNNTTTH) is the number of patients whom you have to not treat with something beneficial in order to cause one harm from your omission.  It is numerically equivalent to the number needed to treat (NNT), but it reframes the decision from action to omission and from benefit to harm.  Ignoring bleeding altogether (because making relative utilities for bleeding and stroke is a fraught endeavor), you could ask yourself "how many patients can I withhold stroke prophylaxis from for one year before I statistically cause (or allow to happen, if you are prone to omission bias) a stroke?"  For most patients, withholding stroke prophylaxis has a NNNTTTH of about 25-30 per year (check the corresponding NNT from CHADS2 for a more "precise" estimate).  Reframing the question into "how much am I asking the patient to pay, in terms of statistical likelihood of stroke, to avoid anticoagulation and the particular side effects that cause me concern in his case?" can often provide some reassurance for the clinician and the patient alike.