In a recent discussion with a prospect, I was asked if my company, OpenBI, had vertical specialization with our BI, analytics and big data consulting business. My answer was strategically no. Though we have a number of health care, educational and insurance customers, OpenBI promotes more of a horizontal orientation, with focus on finance, sales and marketing (especially digital). Often but not always, the work for sales and finance customers revolves on performance management – how’s the company doing -- while the marketing clients are bit more “data sciencey,” looking for business “lift” from data and analytics.

Several present customers are in the digital marketing world. For them, a central analytics objective is to determine which of many advertising encounters “drives” prospects to convert/purchase. Of all the digital stimuli, they seek to “attribute” cause and effect in some way, however that’s defined. With this “marketing attribution” information, they can then optimally allocate their advertising dollars across disparate channels.

Not surprisingly, there are few companies that take marketing attribution as seriously as Google, which defines MA simply as the practice of determining the role that channels play in influencing the customer journey. “Savvy marketers understand that you don’t capture your audience with just one message…It’s a complex process of planting the seed, nurturing it, and finally harvesting the fruits of your marketing efforts…. So when it comes to giving credit to the various elements of your marketing program…it’s essential to take stock of all the factors that affected your results.”

A study conducted for Google and released in 2012 affirms that while MA is currently adolescent, it’s starting to progress rapidly with the development of supporting technologies and analytics, the digital side leading the way. The research cautions that there’s not a single best MA approach: “Attribution proves that no one size fits all. Lessons from one company, product or campaign don’t necessarily port over to others.”

The challenges of imputing cause and effect from observational “designs” driven by customer journeys are indeed daunting. Some of the currently-popular, rule-based MA techniques that include 1) Last click gets full credit; 2) First click gets full credit: 3) All clicks share uniformly; 4) Time decay with clicks closer to conversion getting more credit – seem less than methodologically compelling, given the absence of experimental or quasi-experimental control.

In an intriguing Revolution Analytics’ webinar this week, European Services Director Andre de Vries put a different spin on marketing attribution, conceptualizing an approach that uses survival analysis statistical methods borrowed from biostatistics and engineering. SA is a modified regression analysis technique used for predicting lifetimes. In a medical clinical trial experiment, a researcher is interested in contrasting the survival times of patients as a function of the treatment they receive and other background variables. In engineering, the analyst is often concerned with the time to failure of machines as related to different manufacturing factors.

What differentiates survival analysis from traditional regression is that the ultimate lifetime-determining event – death or failure in the examples above – has not occurred in many cases. If the patient is still alive or the machine continues to function, the case is censored: “You do not know the exact lifetime, only that it is longer than a given value.” Of course the censoring must be accommodated in the models.

Simplified, SA combines the measured lifetime with the censoring indicator to form a proxy dependent variable, “regressing” it in turn on a host of independent measures that shed light on importance of various factors. Outputs of statistical packages like R’s “survival” include a survival function that estimates the probability of being “alive” or “non-failed” at a given time, and a hazard function that measures the risk of dying/failing within a very short period of time. The model coefficients are interpreted much like their standard regression counterparts to help “explain” the results.

The survival analysis theme need not be doom and gloom. Instead of death/failure, the outcome of interest could be churn or purchase, with censoring indicating still a customer or yet to buy. Indeed in the de Vries’ digital marketing attribution study, time to conversion is the ultimate measure, following the path: impressions→click→conversion, with every interaction included. Independent factors identified as “significant” as survival regressors include product type, event type and product supplier.

The use of survival analysis and other techniques developed for the sciences will assuredly increase to provide insight on business analytics questions in coming years. My take is that marketing attribution analytics presents very knotty challenges for untangling the causes (attribution) of conversion, and many of the existing approaches are simplistically biased. I suspect we’ll soon start to see more sophisticated models and designs.

In the meantime, MA analysts are advised to be skeptical of their models, contrasting the performance of various methodologies against future “test” sets to evaluate efficacy. And they shouldn’t be surprised if there’s no smoking gun “winner” technique identified.  Rather, they should expect contextual solutions that include the less-than-satisfying caveat “it depends.”