To all my loyal readers, I apologize for the hiatus, but life happens. About four months ago, I started a new position at Yodle leading the team of quantitative developers that work on our bidding algorithm. The first few months have flown and I have learned a ton, but have been very busy and have left little time for writing.
Or for attending meetups, but I made it to a pretty good one earlier this evening titled “Estimating The Causal Effect Of Online Display Advertising,” presented by Ori Stitelman of Media6Degrees. The main premise of the talk is that A/B testing, now an industry standard, can be costly or impossible, and it’d be nice if we could figure out causality based on something we already have lying around – our data. Of course, we are interested in display advertising at Yodle, but the use of the techniques Mr. Stitelman outlined have other applications: how do you measure what makes customers happy? how do you measure the effect organic and paid search have on each other?
He first started off by describing a methodology for doing this kind of quantitative analysis:
- State the Question: What is the business problem you are trying to solve?
- Define Causal Assumptions: Without getting into a statistical model, what is the causal relationship between events?
- Define Parameter: what is the parameter you actually care about?
- Estimate the Parameter.
Mr. Stitelman then described a few different way of estimating parameters:
- Inverse Probability-Weighted Estimates: a technique that weighs an event’s importance by how unlikely it was to happen on the particular example.
- Maximum Likelihood Estimator
- Targeted Maximum Likelihood Estimator
It’s obviously difficult to digest all this kind of material in a hour long talk, but the results presented show that this technique has promise.
What I have learned so far about meetup talks is that the ones that were worth attending add a whole new load of reading to your stack, and my stack is looking pretty high.
