Embarssingly Parallel Bayes

In this post I talk about recent developments in “embarrassingly parallel” Markov Chain Monte Carlo (MCMC) sampling. The post includes example MATLAB MapReduce code for a consensus algorithm for estimation of a logit model, as well as MATLAB code for an asymptotically exact parallel algorithm applied to a hierarchical logit model.      

Read More

(Un)intentional Targeting, Hidden Consequences for the Measurement of Marketing Effectiveness

Shopper presence, as well as their behavior, on media platforms is driven by their intrinsic nature and preferences. Many forms of media, especially digital forms, are delivered to maximize exposure. As a result, the measured effects of media, delivered on the focal platform or channel, may be over or under estimated.

Let’s consider an example. Sometimes digital marketers buy up the cheapest impressions to maximize exposure. These impressions are leftover as a result of other advertisers intentionally targeting high value shoppers, increasing exposure to the less attractive shoppers that remain. The implication is that an analysis of the effectiveness of digital display ads would underestimate the impact of the focal digital advertising channel, and subsequently the marketers would overly reduce digital display ad pressure in error, rather than optimally target the most responsive shoppers.

No matter which statistical methods or machine-learning algorithms are applied, if exposure propensity is unequal and non-random, biased media measurement is likely. Fortunately there is a long lineage of scientific research that proposes remedies for the probable bias. In essence this research reveals that if you weigh the observed dependent variable (click, conversion, etc.) by the probability the shopper is exposed to the media, you can carry out your analysis on the transformed dependent variable. The real challenge occurs when exposure is due to aspects that are not directly observed in the available data. If you can recover these unobserved aspects from the data, such as a the effectiveness of banner ads on a particular shopper, then you can jointly recover the dependency between exposure and the effectiveness of banner ads, and eliminate the bias due to unintentional banner ad targeting.  I can’t promise this is an easy thing to do, and unfortunately it cannot be achieved with standard statistical analysis packages.

Next time you are presented with digital media attribution results, or you are pitched a digital attribution solution be sure to ask yourself whether some unintentional form of targeting could be contaminating the results. Then ask what the attribution solution does to control for the bias that this unintentional targeting creates.

The Marketing Report Card

Did your parents ever let you get away with reporting the number of right answers you earned on a math or history exam, as opposed to the grade or % correct? Of course not. They wanted to know how well you did, given the possible.

Like the number of correct answers, investment ROI needs to be viewed in the context of the entire opportunity.  Otherwise, like the number of correct answers on a test, it is not very meaningful.

Some say that knowing ROI gives you a “sense” for how you might better invest your resources. However, ROI is an average return and in and of itself offers little insight into how you should adjust your investments.

Fear not, the return on the marginal investment is the unique guide for increasing ROI because it measures the return for each unit of marketing. If you earn more than you spend on the next unit of marketing then increase it, if you earn less, then decrease it, there’s nothing more to it.

If ROI neither indicates the quality of an investment, nor provides insight into how to improve an investment strategy, then it should not appear on the marketing report card.

Next time you produce your marketing report card there better be a grade based upon your attainment of potential ROI, and make sure the assessment clearly indicates the return on the marginal investments to guide improvement.


Welcome to the real data science blog hosted by Michael Cohen. This blog addresses issues and topics in data analytics and predictive modeling for marketing and media management insight and decision making. It hosts entries from both regular and guest bloggers, and covers topics ranging from technical issues in measurement to the challenges in the adoption of data-driven management within organizations. We look forward to interacting with you.