Predictive risk scores estimating probabilities for a binary outcome on the basis of observed covariates are common across the sciences. They are frequently developed with the intent of avoiding the outcome in question by intervening in response to estimated risks. Since risk scores are usually developed in complex systems, interventions usually take the form of expert agents responding to estimated risks as they best see fit. In this case, interventions may be complex and their effects difficult to observe or infer, meaning that explicit specification of interventions in response to risk scores is impractical. Scope to modulate the aggregate model-intervention scheme so as to optimise an objective is hence limited. We propose an algorithm by which a model-intervention scheme can be developed by 'stacking' possibly unknown intervention effects. By repeatedly observing and updating the intervention and model, we show that this scheme leads to convergence or almost-convergence of eventual outcome risk to an equivocal value for any initial value of covariates. Our approach deploys a series of risk scores to expert agents, with instructions to act on them in succession according to their best judgement. Our algorithm uses only observations of pre-intervention covariates and the eventual outcome as input. It is not necessary to know or infer the effect of the intervention, other than making a general assumption that it is 'well-intentioned'. The algorithm can also be used to safely update risk scores in the presence of unknown interventions and concept drift. We demonstrate convergence of expectation of outcome in a range of settings and show robustness to errors in risk estimation and to concept drift. We suggest several practical applications and demonstrate a potential implementation by simulation, showing that the algorithm leads to a fair distribution of outcome risk across a population.
翻译:以观察到的共差为基础估计二进制结果的概率的预测风险分数在科学界是常见的。这些分数的制定往往是为了通过干预估计风险而避免有关结果。由于风险分数通常是在复杂的系统中开发的,因此干预通常采取专家代理人对估计风险作出反应的形式,在这种情况下,干预可能很复杂,其影响难以观察或推断,这意味着明确规定针对风险分数的干预措施是不切实际的。因此,调整总体模型干预计划以便优化目标的操作范围有限。我们提议一种算法,通过“刷新”可能未知的干预效应来制定模型干预计划。通过反复观察和更新干预和模型,我们表明这一方法会导致最终结果的趋同或几乎一致。我们的方法为专家代理人安排了一系列风险分数,并指示他们根据最佳判断来调整总体的逻辑。我们的算法仅使用“组合前”的观察,通过“吸收”可能未知的干预效应来发展模型干预效果;通过反复观察和最终的计算结果,我们显示在预测中可以显示一个不确定的预测值,通过输入结果来显示一个不确定的排序结果。