Causality and initiation of alcohol control policy. A response to Allamani

Abstract

In a recent commentary, Allamani asked how one can establish causality in epidemiological research, and specifically about causality as it relates to alcohol control policy. Epidemiology customarily uses a sufficient-component cause model, where a sufficient cause for an outcome is determined by a set of minimal conditions and events that inevitably produce the stated outcome. While this model is theoretically clear, its operationalisation often involves probabilistic elements. Recent advances in agent-based modelling may improve operationalisation. The implications for alcohol control policy from this model are straightforward: the so-called alcohol-attributable fraction denotes the cases of morbidity or mortality which would not have happened in the absence of alcohol use.

In his commentary to our overview of the various impacts of alcohol use on the course of liver disease [1], Allamani poses important questions on the link between the establishment of causality and alcohol policy or other actions [2]. We would like to answer these points first by illustrating the concept of causality used. Causality as customarily used in epidemiology [3,4] entails a ‘sufficient-component cause model’, where sufficient cause for an outcome is determined by a set of minimal conditions and events that inevitably produce the stated outcome. This implies that all of the minimal conditions or events are necessary for the outcome to occur. For example, drinking three glasses of wine together with driving a car plus night-time conditions and bad weather may be the elements that together produce a crash resulting in a traffic fatality. Each component cause in this example is a necessary part of the causal mechanism it contributes to; in other words, no one factor is stronger than any of the others. A specific component cause such as alcohol use may play a role in many sufficient causes for traffic injury fatalities. If a component is present in all sufficient causes, it is known as a necessary cause for the outcome; however, most identified component causes are neither necessary nor sufficient to cause a disease by themselves.

Going back to the original example of liver disease initiation and progression to liver cirrhosis, there may be thousands of such sufficient component-cause models involving various levels of alcohol use, and we need ways to quantify how many of the incident liver cirrhosis cases are caused by alcohol use. While we cannot predict if an individual will develop cirrhosis in a given year based on his or her drinking level, we are able to roughly estimate how many new cases of liver cirrhosis can be expected in a country using the attributable-fraction methodology based on the population distribution of alcohol use in this country, and the risk associated with different levels of drinking over time [1,5,6]. This will be a rough estimate based on a counterfactual scenario of no alcohol use [7], and somehow assuming that current levels of drinking are an approximation of drinking in the past (for more details on underlying methodology, see Reference [5]). Thus, while we have a theory of causality underlying individual cases and situations, the operationalisation and de facto use are statistical and probability based (in this sense going in the direction of Reference [8]; see also Reference [9]).

A recent direction for disease epidemiology is the use of empirically calibrated agent-based models (ABM) to reproduce and forecast trends in risk-exposing behaviours and related harms [10-12]. ABMs are simulation models that represent individuals (the ‘agents’) and their various life course trajectories, in terms of behaviours, socially situated interactions and events. An ABM explicitly represents the postulated causal mechanisms—configurations of interacting components—from which population trends emerge. If a particular ABM can reproduce observed trends then the configuration of mechanisms it encodes represents a candidate causal explanation—the so-called ‘generative sufficiency’ test [13]. By searching through the space of causal mechanisms—each realised as an ABM and subjected to the generative test—it is possible to identify a multiplicity of viable explanations and to seek commonalities in the components they contain [11]. Such an ABM ‘model discovery’ approach can be highly beneficial to epidemiology by identifying directions for future primary research and by enabling more robust forecasts of country outcomes. These models are generally not deterministic and will still include probabilistic elements. Model discovery may also close the gap between the postulated causal theories, and the operationalisations in epidemiology.

One of the advantages of the causal concept of epidemiology is its interpretability: as stated above, the sufficient-component cause model implies that each component is necessary for the outcome to happen [4]. That further implies, that against a counterfactual scenario of no alcohol use [14], the alcohol-attributable fraction measures the number of outcomes that could have been avoided if there had been no alcohol use [15]. And this implication is empirically testable and has clear meaning for decision makers on alcohol policy. We can predict the number of deaths that are avoidable by a policy based on the above, and we can compare these predictions with actual policy evaluations. For instance, taxation increases that reduce the affordability of alcohol will reduce level of consumption and subsequently level of harm. To give a recent example: the increase in excise taxes on beer, wine and spirits in Lithuania in March 2017 decreased consumption and saved lives. Each step here can be independently assessed: the impact on the price, the impact of the higher price on sales and the impact of sales on mortality. Alternatively, we can simply conduct interrupted time-series analyses and test if the mortality following the taxation increase deviated from the long-term trend in the direction and level predicted (it did, and it has been estimated that this measure saved more than 1400 in the following year [16]). Contrary to Allamani, we believe that all taxation that increases price enough to decrease affordability will result in less consumption, if unrecorded consumption is controlled [17], even in Southern European countries. There are clear cultural differences whether a country chooses such a policy, but if such a pricing policy is implemented it will work (for instance, see the effects of tobacco control policies in Southern Europe up to today and the potential effects of future implementation [18]). But this is not a matter of belief. To go back to the original article on alcohol and liver disease, all the predicted impacts of alcohol can be empirically tested. And the more the assumptions about the potential causal impact of alcohol use on disease outcomes are specified, the better we can test our knowledge in predicting real-world effects of changes in exposures, including changes via control policy interventions, the more precise our predictions will become, and the better we can advise policy. Which interventions societies will eventually choose, is not for scientists to decide.