The Causal Continuum - Transforming Modelling and Computation in Causal Inference

科学和工程的一个核心方面是能够回答“如果”的问题。如果这个基因发生突变会发生什么？削减社会福利对公共健康有什么影响？我们能做些什么来减轻社会群体之间的差异？封锁在多大程度上有助于缓解大流行？如果在食品供应链等主要物流操作的这些点发生故障，将会产生哪些后果？这些都是因果关系问题。回答这些问题很难，因为这涉及到改变。历史数据可能无法捕捉变化的含义，将因果问题置于使用数据来告知决策的舒适区之外。像公共卫生官员或保险公司那样预测吸烟者的预期寿命是一回事。很难理解如果我们说服某人戒烟会发生什么，因为历史数据可能有大量案例表明，人们在死于呼吸系统疾病前不久因不适而停止吸烟。统计或机器学习方法忽略了这些因果解释，实际上可能会说戒烟对健康有害。理想情况下，我们希望进行随机对照试验，其中采取行动的选择是由抛硬币决定的，这样就可以消除因果之间的混淆因素。例如，为了令人信服地证明，在选择接种疫苗的人及其健康结果中，covid-19疫苗是由于生物过程而不是社会学混杂因素而起作用的，消除混杂因素是必要的。然而，在许多情况下，这样的试验可能非常昂贵（了解遗传网络涉及到很大的实验空间）或不道德（我们不能强迫某人吸烟或不吸烟），即使进行了对照试验，也可能无法完全控制感兴趣的因素（我们可以随机给病人分配药物或安慰剂，但如果他们呆在家里，我们可能没有办法让病人遵守治疗）。数据科学家并没有忽视这些问题，我们应该感谢流行病学家的辛勤工作，例如，他们提出了一个令人信服的案例，证明吸烟与肺癌之间存在有害联系。但在没有随机试验的情况下，“如果”问题的答案需要假设，否则就是不可知的。这意味着因果推理进展缓慢，容易出错。部分原因是，传统上，因果推理的方法很大程度上依赖于统计学家设计的方法所选择的预先定义的假设族，这些方法将提供明确的答案。然后，应用科学家根据他们对世界的理解（一个简单的例子：假设我们没有在数据中测量不到的共同原因！）选择采用一种特定的方法。虽然存在敏感性分析工具（如果假设在某些特定方面被违反了怎么办？），但它们并不能直接解决主要问题：领域专家应该有机会根据他们认为合适的方式指定预先假设，而不是人为地告诉他们一个单一的、方便的答案，而是可以从提供的信息的观察数据中解出什么。这种工作流不流行的原因之一是需要计算密集型算法来推断这些假设的结果。这个项目的目标是改变因果推理的普遍做法，提高透明度和速度，使我们能够理解我们知识的局限性，以及为了进步应该在哪里寻找。它将依靠先进的算法为领域专家提供一个灵活的沙箱，以非常灵活的方式表达他们的知识，同时也为所需的复杂计算方法提供后端支持。

A central aspect of science and engineering is to be able to answer "what if" questions. What will happen if this gene suffers a mutation? What are the public health consequences of having this social benefit cut? What can we do to mitigate disparities among social groups? To which extent are lockdowns useful to mitigate a pandemic? Which ramifications will take place if failures occur at these points of a major logistical operation such as food supply chains?These are cause-effect questions. Answering them is hard because it involves change. Historical data may fail to capture the implications of change, placing causal questions out of the comfort zone by which data is used to inform decisions. It is one thing to predict the life expectancy of a smoker, as done by public health officials or insurance companies. It is much harder to understand what will happen if we convince someone to stop smoking, as historical data may have a substantive number of cases where people stopped smoking shortly before dying of respiratory disease, due to discomfort. A statistical or machine learning method oblivious to these causal explanations may actually say that stopping smoking is bad for one's health.Ideally, we would like to perform randomised controlled trials where the choice of action to be taken is decided by the flip of coin, so that confounding factors between cause and effect are overridden. This removal of confounding is necessary to show convincingly, for instance, that a covid-19 vaccination works due to biological processes as opposed to sociological confounding factors among those who choose to be vaccinated and their health outcomes. However, in many cases such trials can be very expensive (understanding genetic networks involves a large experimental space) or unethical (we cannot force someone to smoke or not), and even when they take place, a controlled trial may not fully control the factor of interest (we can randomly assign a drug or placebo to a patient, but we may not have the means to make the patient comply with the treatment if they stay at home).Data scientists have not ignored these problems, and we can thank the hard work of epidemiologists, for instance, for presenting a convincing case establishing the harmful link between smoking and lung cancer. But without randomised trials, the answer to a "what if" question requires assumptions or otherwise it is unknowable. This means that causal inference progresses slowly and is prone to mistakes. Part of the reason is that, traditionally, methods for causal inference largely rely on pre-defined families of assumptions chosen by statisticians designing methods that will provide unambiguous answers. Applied scientists then choose to adopt a particular method according to what manages to be a good enough approximation to their understanding of the world (one simple case: assume we have no common causes that are not measured in the data!). Although there are tools for sensitivity analysis (what if assumptions are violated in some particular ways?), they don't address the main issue directly: a domain-expert should be given the chance of specifying upfront assumptions according to the way they see appropriate, and not be artificially told a single, convenient answer, but what indeed can be disentangled from the observational data given the information provided. One of the reasons this workflow is not popular is the need for computationally-intensive algorithms to deduce the consequences of such assumptions. This project has the ambition of changing the common practice for causal inference, increasing transparency and the speed by which we understand the limits of our knowledge and where to look for in order to progress. It will rely on cutting-edge algorithms for providing a flexible sandbox for domain experts to express their knowledge on a very flexible way, while offering also the backend support for the sophisticated computational methods needed.