Classification and Regression are the lodestars of statistics and they come up again within the context of causal decisions. In this case the regression is really using the data to attempt to determine causal effects. In an analog to classification, when we define the optimal causal decision we are optimizing in a different way on the data than we do within the causal effect. This helps us to decide what we most want to determine from the data and what are the most important goals from the analysis. In this second week we continue our story for optimizing causal decisions by focusing on where best to locate a restaurant based on location data (anonymized) from cell phone user data.

Online advertising is both a challenge and an opportunity for businesses. From the marketing perspective, defining how much to spend and on who and when is the question they need answered. From a user perspective, the fewer the ads the better, but if there must be ads, can they be relevant and helpful. This is an example of a marketplace, based on online advertising. A simple example, one that we considered earlier, is A:B testing. In this week we return to this question of how to optimize the presentation of information, but now from a causal perspective. We will go over the causal forest route for how to determine an optimal causal decision. We will also go over randomized clinical trials and how the control of the treatment effect can make the analysis significantly cleaner.

David Blei's group contributed an intriguing route for defining causal decisions where confounding is still present, but can be treated in a way that still determines a causal understanding of the decision process. We will explore that paper, and its implications, along with other similar routes for working with data where the confounding issues may otherwise make the analysis a challenge. We open with how best to define the spend on a new movie project: hire that star actor or spend more on stunts? This is another example of a causally linked decision that does not have a final 'right' or 'wrong' but that does have major implications for a business.

How best to determine a decision for a large group versus an individual is our final motif for the specialization. In many ways this is the defining question for healthcare: how best to determine an optimal treatment for each individual? While there is a lot of work ahead to describe how this may be done in a detailed setting, there has been progress towards how to use information collected for many individuals that can then be a help in defining the best treatment for an individual. It is this current progress that has many people excited for the future of personalized medicine. At the same time it is a appropriate to realize the limits of what can be done, right now, with the question of an optimal individual treatment.