Philosophers have been major contributors, along with economists and political scientists, to work on aggregating individual preferences in efforts to construct social choice functions — representations of what “society prefers”. The major theoretical challenges to this enterprise, starting from Kenneth Arrow’s famous “impossibility theorem” in the 1950s, have seemed worth tackling because the problem seems central to the ideal of democratic government. Elected officials, it is widely agreed, should enact policies and regulations that are aligned with the preferences of citizens. Even if, as in ideals of representative government, we think that leaders should try to also exercise superior judgement based on greater responsibility and access to information, democrats typically think that something has gone badly wrong if a government over-rides or ignores the preferences of majorities or large minorities of citizens.
The intractability of various problems in social choice theory has reflected the very stringent standards usually imposed on proposed solutions to individual preference aggregation. For example, theorists often require that a successful model cannot disregard the preferences of even a single citizen, or that it must identify social choices that never reverse themselves unless individuals change their minds. But stable, legitimate, democratic government can often be maintained on the basis of much looser ambitions. Political bargaining among interest groups can reliably produce approximate reflections of individuals’ averaged preferences as long as governments can be regularly changed through elections in which all adult citizens can participate and vote, and there is meaningful control of disproportionate influence from the wealthy. Many countries that aspire to democratic ideals in their constitutions and prevailing norms approximately satisfy the first criterion while obviously failing the second one. But we know that democratic representation is not a chimera because there have been numerous examples in history, and at least several of them flourish currently. Most people regard Iceland and Uruguay, for instance, as countries in which governments usually sincerely try to do what voters seem to want them to do.
A crucial enabling condition for this is that many policy preferences are reasonably clear and easy to discern at both individual and aggregate scales. We know, for example, that a large majority of Irish voters currently favour increased public intervention in the market for owned homes, and that the present coalition government has pledged meaningful intervention. These real cases resemble the artificial constructions used in social choice theory in one crucial respect: preferences are thought of as ordinal or cardinally weighted rankings over alternatives that citizens, government, and media describe and debate in broadly common language.
Where responses to the Covid-19 pandemic are concerned, however, it has been much harder to say whether democratic governments have been imposing the preferences of officials and their advisors, or reflecting those of their populations. Policies have varied substantially, ranging from strict and protracted lockdowns followed by cautious and gradual relaxation (Italy, Ireland, South Africa), to partial and quickly exited movement and activity restrictions (the UK, the USA), to near-total reliance on voluntary compliance with advice on individual self-management (Sweden, Japan). Some, but far from all, of this variation is statistically predicted by three exogenous variables: time taken to establish mass testing capacity, per capita availability of ICU beds and ventilators, and extent of experience with the previous most recent coronavirus alarm (SARS). But these relationships are relatively weak under cross-country regression analyses. This raises questions about the extent to which policy variation reflected accurately or inaccurately perceived variation in citizens’ preferences. Swedish authorities explicitly cited an alleged culture of autonomous civic responsibility in defending their decision to avoid general lockdown. The British government said at the beginning of the crisis that it expected its citizens to quickly reach compliance “fatigue” if it responded too quickly. Irish commentators cite their culture’s high level of health risk aversion while typically citing no specific evidence for this assumed fact. To what extent can we tell whether these narratives reflect real causal relationships or are ex ante rationalisations of decisions that were made, and are still being made, for concealed reasons, or as a result of accidents and confusion?
Where some aspects of Covid-19 policy and regulation are concerned, we can assess distributions of preference in the standard way, by using polls and surveys to track preference orderings over binary measures: within-country travel to be limited or not, restaurants and bars to be closed or not, mask-wearing in public to be mandated or merely advised. But there is no evidence that anygovernment made reference to systematic surveys of public opinion as part of its decision-making over these alternatives. Of course this is readily explained by the sudden onset of the emergency, and the general recognition that delay implied higher case and fatality rates and elevated danger of hospital capacity collapse. However, weeks and indeed months passed between the onset of initial responses in wealthy countries in mid-March and the beginnings of exit from lockdowns and restrictions. There has been plenty of time for public opinion research to be done, but policy decisions have not been justified by reference to it in any country of which we are aware.
I suggest that the main explanation for our current ignorance about relationships between adopted Covid-19 responses and aggregated social preference models is that the dimension of preference that the pandemic mainly implicated was risk preference (including risk perception). The pandemic confronted both governments and individuals with interacting domains of uncertainty. Evidence was insufficient, as it still is today, for specification of a consensus transmission model among epidemiologists. Infection fatality rates cannot be derived from case fatality rates because the majority of mild and asymptomatic cases are never detected or verified. Even to the extent that mortality risk can be estimated, serious morbidity rates are clearly significant but still unknown in general. In larger countries, prevalence rates show high regional variation, but time lags between local transmission outbreaks and their detection leave everyone in permanent suspense about their own risk until they actually find themselves in an epicentre. This model uncertainty inevitably interacts directly with subjective risk preferences. As lockdowns have eased or ended, individuals, households, schools, and business proprietors are called upon to decide how much risk of infection, morbidity, and mortality they want to take upon themselves, or impose upon others through their choices. Furthermore, people recognise that their risk tolerance interacts with the risk tolerance of everyone else. I might go to my health club if I expect most other members to have been avoiding bars, churches, and weddings, and to wear masks; otherwise I won’t.
I am a philosopher of science who specialises in the methodology of economics. I am also an applied experimental economist, and part of a global research institute devoted to the study of behavioural and institutional risk. From those perspectives I’ll offer some observations on the epistemology of social risk preference in the context of the pandemic, and report on online experiments in which my research team is currently engaged to get some traction on the coevolution of subjective risk perception with the dynamics of the disease.
Epidemiologists are ultimately the decisive experts on the statistical health risks associated with viruses. However, to assign numbers to these they need to have specified, on the basis of rigorously gathered and assessed empirical evidence, a general transmission model. Given the extent of current missing information, this target remains months, or perhaps even years, in the future. Epidemiologists have therefore tended to simply encourage behaviour most consistent with historically high risk and low risk tolerance. Such professional caution is easy to understand, but it doesn’t tell us anything about the alignment of policies and actual distributions of risk attitudes. In the meantime, there is a body of technical knowledge that is precisely designed to identify and assess policy options under conditions of empirical ignorance. The knowledge in question resides with economists.
The recurrent circumstance faced by economists of all types and specialisations is that they aim to estimate relative welfare effects of policy alternatives in conditions where they know there are more relevant variables than they can accurately measure, or indeed measure at all, and where causal relationships and their magnitudes are entangled, only indirectly observable, and laced with feedback. The body of special statistics economists use to minimise the epistemic and policy risks associated with these limitations is econometric theory and practice. It constitutes an advanced area of technical knowledge, notwithstanding the fact that it is and will always be incomplete. Thanks to the demand pressure driven by ever cheaper computational power and availability of large data sets, the past two decades have seen explosive development of new econometric techniques, particularly for handling and estimating sources of statistical bias.
Popular media have regularly cited economists’ views during the course of the Covid-19 crisis. However, almost every such engagement I have seen has concerned macroeconomic topics: the depths of national recessions we might expect from lockdowns of businesses and from public reluctance to buy services, travel, and go to work; short-term and long-term expected impacts on financial asset markets; and effects on longer-term unemployment, skills development, and earnings. However, as I explained above, central to modelling Covid-19 dynamics are subjective risk preferences, their relationship to behaviour in the face of uncertain health threats, and their contribution to the social welfare measures that democratic governments are implicitly tasked with optimising. Risk preference modelling and measurement are much more central to general microeconomic science than most people, in my experience, appreciate.
Microeconomic theory assumes that all people seek to consume baskets of different sources of utility, and trade these sources off against one another on margins that are determined by their budgets and by effective relative scarcities of each. This is why a typical utility function over a particular good or service is a concave curve from an origin of zero stock in the utility source in question. As this stock increases, further consumption of it decreases the set of feasible trade-offs, given a constrained budget, that are available to the person. Where most real choices are concerned, the rate of decrease involves some uncertainty. I might conclude, based on what I know today, that buying more home gym equipment is the best thing, from my subjective point of view, to do with a bunch of my money. Only time will tell if this conclusion was right ex post. Thus to the extent that an individual chooses in such a way as to optimise her expected utility, the curvature of her choice function over a source of utility is her risk preference with respect to her expectations about the utility stream it will actually deliver. Thus risk is not just one topic among many that economists sometimes study. As few choices of interest concern only certain prospects, risk is at the centre of almost all microeconomic analysis. Applications of this logic are routinely made to utility derived from a person’s health state, the utility source likely to be most salient when she wonders how best to respond to a pandemic outbreak.
Economists usually operationalise risk preference for measurement purposes by examining choices between lotteries over money rewards. This is not because they think that money is equivalent to utility. Where utility sources can be conveniently exchanged for money, however, it is typically the best available proxy for utility. But experiments have been run in which people choose between lotteries over other kinds of prospects. We can think of someone choosing whether to violate social distancing advice during the pandemic as choosing between lotteries that yield different probabilities of infection, illness and death, traded off against varying social enjoyment temptations.
It is more challenging to estimate and aggregate people’s risk preferences than to identify and aggregate their preference orderings over broad policy outcomes. Complications arise from several factors.
First, thinking of risk in terms of utility function curvature is useful for indicating how central a role risk plays in economic models of behaviour. But it is not a complete analysis because the assumption it depends on, that everyone aims to maximise expected utility, isn’t true empirically. In the many population samples from around the world whose risk preferences our group has measured experimentally over the past few years, substantial proportions choose in accordance with expected utility theory. But for majorities in almost all of these populations, we need to use models with additional parametric structure. (For reasons I’ll pass over here, we do not find the famous prospect theory of Kahneman and Tversky very useful.)
Second, “risk” is a complex idea that is not reducible to mere variance of outcomes. Distributions of outcomes also matter. When we say that a person is averse to risk, we might mean that she will accept a guaranteed reward of a magnitude less than the expected value (in the same currency of assessment) of a risky prospect. But this might mainly reflect aversion to skewed distributions of reward. Or perhaps it’s kurtosis she wants to avoid. These complications should make clear that we cannot discover people’s risk preferences by asking them survey questions, because few people have ever thought of, let alone performed, the relevant self-analyses.
To discover people’s risk preferences we must model, observe, measure, and econometrically analyse actual choices that they make, over real stakes. This is what our group does, using laboratory procedures in which people make sequences of choices between pairs of lotteries, which they know will be played out for real rewards. To yield predictions of behaviour, the lottery choice experiments must be supplemented with additional procedures. We ask our subjects to bet, again for real stakes, on probability distributions over future outcomes, where they know that their rewards will be based on the extent to which their forecasts turn out to conform to what actually transpires. When we analyse the results from this procedure jointly with results from the lottery choices, we can estimate a subject’s relative confidence in her own beliefs about the probabilities of the different outcomes. If risk is distributed over future times, as it is when people consider behavioural responses to the pandemic, we must also ask them to make choices between sequences of present and future rewards so we can factor in the extent to which they discount future welfare, and also detect possible “present bias”, that is, assignment of a premium to a reward just because it arrives right now. Atemporal risk preference, that is risk preference measured at a time but not over time, differs from intertemporal risk preference, that is, preference about how smoothly risk is distributed across intervals. The latter can be assessed through administering yet another kind of choice experiment with real incentives.
The procedures just described are being applied by our group to try to get purchase on relationships among atemporal and intertemporal risk preferences, subjective beliefs about Covid-19 prevalence and mortality, confidence in those beliefs, and time discounting during the course of the pandemic. We sampled 375 participants in the United States and 300 in South Africa, selected from pools of over 2,000 volunteers in each country, and spread them across 3 waves administered at least one month apart. Each participant works through the choice experiments described above, for real money rewards. Because morbidity and mortality risk vary greatly with age, we have them bet on distributions of these statistics for both their own national population as a whole, and for the population over 65 and over 60 in the USA and South Africa respectively (the difference reflecting use of different thresholds in the two countries’ official reports). Subjects forecast prevalence and mortality as of one month from their specific participation date, and as of mid-December, by which point the main surges of disease in both countries might have passed or an effective vaccine might be known to be on the horizon. Subjects complete surveys on their demographic data, their general health, their levels of anxiety, their sources of information about Covid-19 and their levels of trust in the reports they read in these sources, and on their own behaviours with respect to social distancing, mask wearing, and sanitation.
One general theoretical question that is difficult to directly study except during a major shared experience of preoccupation with a novel risk factor, is whether risk preferences and salience of risk mutually affect one another. The Covid-19 emergency provides us with a window on this. Do people confront varying levels of reported infection and mortality risk using atemporal and intertemporal risk preferences that stay approximately constant, or do the latter shift as a result of experience of the evolving crisis? The second case would make it almost impossible for public health authorities to make decisions that reflect citizens’ risk preferences, as consequences of policies would shift those very preferences.
We have now administered the first two waves of our study. Our data analysis to this point has not yet been integrated to generate general patterns and structural models, partly because our Wave 2 experiments were only performed a few days ago, but mainly because it would be premature to begin deep modelling until all data are in. But we can note three tentative observations from preliminary inspection. First, risk preferences have not shifted on average as the pandemic and information about it have evolved. Second, South African participants are more optimistic about future infection and mortality rates than American ones, perhaps because the former have not been shocked by what has been (falsely) represented in much US media as occurrence of a “second wave” of infections. Ironically, epidemiological modelling suggests that it is the South Africans whose current beliefs will prove less accurate. Third, our US data, but not our South African data, indicate increased discounting of the future from Wave 1 to Wave 2.
If our ultimate analysis affirms the first observation above, this will suggest that policy responses could, at least in principle, be made responsive to stable public risk preferences that can be aggregated in principled ways. But this would require much more deliberate cultivation of quantitative behavioural economic measurement, and of cutting-edge econometric analysis, than has figured in policy deliberations thus far.
