Challenges in causal inference from Personal Monitoring Devices

Personal Monitoring Devices (PMDs) collect im- mense amount of data about health and wellness of hundreds of millions of people. One of the obstacles of the prevailing data analytics approaches to PMDs' data is limited value of correlation-based conclusions in a health context. Causal inference seems a natural solution, but general causal inference methodologies are difficult to apply to PMDs data due to size and complexity of observational data. Some methods, such as randomized trials, are largely infeasible in PMDs' context due to lack of control over the investigated population. In this paper, we overview existing approaches to causal inference including recent works that attempt to take advantage of time series data to automatically derive causality using extended difference- in-deference or Granger methods. We then outline challenges and opportunities for causal inference in the health context. Finally, we propose a following challenge: can we establish a new standard of evidence and a study design process that: (1) allows for drawing causal conclusions from large observational datasets and (2) can suggest interventions to enforce causal links discovered in the data.

---

1. Track 4: Software Systems Development & Applications

---

2. Technical Session: Joint 38th IEEE Software Engineering Workshop (SEW-38) and 5th International Workshop on Cyber-Physical Systems (IWCPS-5)