Medicine and Risk Transfer NASSIM NICHOLAS TALEB INTRODUCTION M akridakis, Wakefield, and Kirkham (2019) provide a much-needed approach to iatrogenics, illness caused inadvertently by medical treatment or diagnosis, a problem that has been plaguing medicine since Hippocrates. The central obligation in any intervention, not just in medicine, is primum non nocere, “first do no harm.” And the central point in the paper is that "lowering risks creates more risks." Whether risk comes from a misalignment of incentives between patient and practitioner, a misunderstanding of the probabilistic structure and the dynamics of the problem, or psychological biases that favor overtreatment is a matter of long discussion. This note will put the medical exposures in the context of risk management and decision theory, and presents three mechanisms of risk transfer from visible to hidden risks. FIRST TYPE OF RISK TRANSFER— THE STANDARD CASE PATIENT-DOCTOR The Hippocratic injunction has been generalized (Taleb, 2018a) across fields where tampering with natural systems to “improve” them leads to a transfer of risk from the visible to the invisible, and from the practitioner to the patient. The transfer from the visible to invisible is more formally examined as moving exposures from the body (or center) to the tails of the distributions; that is, by transforming a likely small or medium harm into a less likely (and delayed) large one. This risk transformation is what, in economics, constitutes a principal-agent problem, where the conflict of interest between the two parties leads to a serious asymmetry that needs to be resolved. The core of the problem is that practitioners and patients both are incurring risks— not just the patient—and unless we identify the practitioner’s exposure and mitigate it, we cannot properly treat the patient. The details of risk transfer can be as follows. The doctor has the incentive to prescribe, say, statins in situations where the patient presents mild clinical symptoms. Statins may have delayed side effects, particularly from prolonged use. But the doctor has covered his or her legal risk in case the patient has a cardiovascular event—the delayed side effects will not show or be linked to that particular medicine for a long time. This is no different from bank lending, where the agent loads the balance sheet with long-term loans, knowing that their payoff will appear years after the bonus is disbursed, often after the banker’s retirement, when they’ve become someone else’s problem. Another example of risk transfer is provided in cancer treatments. Many mild in situ conditions do not satisfy a probabilistic breakeven between risks from treatment and benefits: treatments will patently reduce life expectancy but these effects are not visible. On the other hand, the risks of an untreated cancer are exceedingly visible and produce a legal liability. Hence the treatment favors intervention or what has been called via positiva over via negativa. Sometimes risk transfer is done via choice of treatment for the same disease and the shifting of exposures is done across time, from a short window into a decade-long one. Consider the case of stage one or two throat cancer for a 40-year-old, otherwise healthy patient. The options are laser surgery vs. radiation therapy. The former has no delayed side effects but clear visible ones (hoarseness). The latter shows fewer immediate clinical effects but can lead to delayed effects, such as another tumor (caused by radiation exposure) and hypothyroidism—these would appear later in the future, past the five years on which condition-free survival rates are accounted. Hence the five-year survival numbers would look more favorable for https://foresight.forecasters.org FORESIGHT 31 the institution under a policy of radiation than 20-year ones. fuzzy statistics (Freedman and Petiti, 2001). SECOND TYPE OF RISK TRANSFER— THE CASE OF PHARMA THIRD TYPE OF RISK TRANSFER— THE RESEARCHER CASE Pharmaceutical companies have an incentive to find mildly ill people and medicate them for a long time, as opposed to very ill people who have a shorter life expectancy. Intuitively, there are two-hundredthousand times more people in the one standard deviation range than the four standard deviations. This can be seen in the pressure and lobbying for the continuous lowering of the treatment threshold for a number of “preconditions,” such as “prehypertension” and “prediabetes.” Another situation of risk transfer produces a flow of risk from the researcher to society: the publication of shoddy yet peer-reviewed research that helps the researcher’s career though not society. Given that “everyone is doing it” and “we need to advance and progress,” the researcher isn’t particularly violating a strong ethical rule. Ioannidis’s seminal paper (Ioannidis, 2005) brought to light the defects in research that is taken more seriously than it should be. But it is easy to entangle the mechanisms. I have argued that p-values are stochastic, not deterministic variables, in the sense that a repetition of the exact same on statistical copies of the situation would produce different results. This point is not well known by researchers. A “true” p-value of .14 has a 50% probability of showing—in realizations—values below .05. Furthermore, the more the researcher tries, the more we see the upper bound of the results, particularly when done on epidemiology data. Hence many results that do not meet any standard of statistical significance are published via “p-hacking,” the gaming of p-values by repeating different experiments (without reporting failures) until a positive result is found. The method is done via statistical gerrymandering, by bundling the very ill and the mildly affected in the same category and producing results of the medical benefits of treatment for the joint group, not isolated ones. Yet, as Makridakis and DiNicolantonio (2014) show, the risks are convex (see Taleb, 2018b) in the sense that someone slightly hypertensive has, in the absence of other risk factors, a flat to slightly elevated risk of cardiovascular event compared to a healthy subject—not elevated enough to meet the therapeutic breakeven (breakeven of iatrogenics to treatment). The conclusion is bracing: since the therapeutic breakeven is not met in many such treatments, elevate the threshold of treatment and resist the pressure by lobby groups. The problem is that once a false statistical statement enters medical practice, it tends to stay—consider the link between salt and hypertension, generalized from Nassim Nicholas Taleb is Distinguished Professor of Risk Engineering at the New York University Tandon School of Engineering and co-editor-in-chief of the journal Risk and Decision Analysis. His 2007 book The Black Swan has been considered one of the dozen most influential books published since World War II. NNT1@nyu.edu. 32 FORESIGHT Spring 2019 REFERENCES Freedman D.A. & Petitti, D.B. (2001). Salt and Blood Pressure: Conventional Wisdom Reconsidered, Evaluation Review, vol. 25, no. 3, 267– 287. Ioannidis, J.P. (2005). Why Most Published Research Findings Are False, PLOS Medicine, vol. 2, no. 8, 124. Makridakis, S. & DiNicolantonio, J.J. (2014). Hypertension: Empirical Evidence and Implications in 2014, Open Heart, 1(1), e000048. Makridakis, S., Wakefield, A. & Kirkham, R. (2019). Predicting Medical Risks and Appreciating Uncertainty, Foresight, Issue 52 (Winter 2019), 28-35. Taleb, N.N. (2018a). Skin in the Game: Hidden Asymmetries in Daily Life, Random House. Taleb, N.N. (2018b). (Anti) Fragility and Convex Responses in Medicine, International Conference on Complex Systems, Springer, Cham, 2018.