This post is my contribution to the discussion on Market Mix Modelling vs. Attribution Modelling and comparison of their trade-offs. Both methodologies were developed for measuring effectiveness of marketing activities, in particular advertising. While Market Mix Modelling (MMM) has been used over a number of decades - Attribution Modelling (AM) has appeared relatively recently. Digital agencies and software development companies have pioneered the trend. Many consultancies, media and research agencies jumped on the bandwagon.

Working in a company that offers both methods, I want to provide a more balanced view of the two approaches addressing their strengths, limitations and trade-offs.

What is Attribution Modelling?

Attribution modelling, as I refer to it, estimates probability of sales/conversion in particular time period taking into account all touchpoints and interactions. I do not include in this definition rule-based attributions based on the last or the first touch points.

In a nutshell, AM assigns to each touch point the weight based on its contribution to purchase or other conversion. The approach provides several benefits.

First and foremost, it allows accurate quantification of effectiveness of a wide range of touch points and customer interactions. It can be used to optimise the way the budget is allocated at a very granular level.

Second, AM provides base for good understanding of customer path to conversion. It can answer questions, such as

  • How many touch-points is an average customer exposed to?
  • How paid channels interact with each other and with producer/distributor own media (for example well-done website for a product or services).
  • When is the effect of individual exposure strongest – at the beginning of the path or right before conversion?
  • What are our main bottlenecks on target customers’ journey?

As such, AM, if used correctly, is very actionable. It is its level of detail and versatility that made some industry evangelists position AM as a ‘Holy Grail’ for marketers.

What Factors Have Fuelled AM Development?

There were several trends working in tandem that were the main driving force behind increasing popularity of AM and technology.

First, proliferation of digital presents a new universe for marketers. This trend includes rising e-commerce, innovations in online/mobile forms of advertising, own communication and PR and setting up innovative businesses operating fully online.

Second, the ubiquity of digital was magnified on one side, and enabled on the other by availability and affordability of electronic devices that allows us to be online on the go.

Third, technological progress in capturing customer level data and lower costs of data storage brought highly detailed data to analysts. At the same time, with increasing competition and rapidly changing consumer landscape, the need to understand how consumers form their purchase decisions was never more urgent. The purchase path consists of multiple touch points across owned, earned, and purchased media. This interplay, unique for every brand and product, requires sophisticated methods of analysis that has given rise to the practice of marketing attribution.

How Attribution Modelling Differs From Market Mix Modelling?

MMM is a regression analysis that isolates and quantifies the effect of individual marketing activities and all other factors affecting sales. It is a leading approach for evaluating the effect of media when only aggregate data series are available.

While MMM was pioneered by big FMCG companies to measure the effectiveness of their large advertising budgets invested mainly in traditional media channels, AM development was driven ty measuring effectiveness and optimisation of digital spend.

The two methods approach the measurement task from very different angles. In MMM, a modeller regresses aggregate series of sales or conversions on aggregated time series of advertising exposures and other factors. To put it simply, correlations between the series determine how big effect is attributed to each factor. When several activities take place at the same time - separation of their incremental impact poses a significant challenge for the MMM approach.

Advances in capturing all exposures and interactions at an individual level presents a big opportunity. Through AM modelling, one can achieve a very good understanding how advertising affects (potential) consumer at different stages of their purchase path.

This difference has implications for extracted insights:

AM is instructive for optimising digital channels, including owned and earned media. It picks up the effect of low budget activities and can incorporate word-of-mouth and social network effects. Through giving them a fair credit for their contribution, AM helps justify a higher proportion of budget to support effective but so far undervalued activities.

The fact that AM is done at the individual level eliminates the problem of multi-collinearity caused by overlapping events with very similar patterns (it is typically present in MMM). As such, AM can drill into much more detail while not sacrificing accuracy and efficiency of its output.

MMM is mainly looking at an incremental - 'above the base line' - effect of modelled activities. As a result, it is more likely to underestimate the overall impact of advertising. AM analyses purchase path of all sales and has potential to provide a more accurate picture what drives sales.

MMM is a robust method providing valuable top-line recommendations for budget splits across channels, markets and product campaigns. If some major activity does not do much to sales or other KPI,the mix model will detect it. It also provides guidelines for campaign execution, e.g. TV flighting, campaign timing, what formats are more effective, what channel combinations work better.

Which One Is Preferred?

This depends on our objectives and data constraints. If complete data is available at an individual level, i.e. we have a single view of each individual customer, AM provides more detailed actionable insights.

What Makes AM More Actionable?

AM has potential to answer questions such as

  • What is the effect of different creatives, ad formats and websites?
  • How the effectiveness of channels vary at different stages of purchase path or in a different sequence of exposures?
  • What combinations of exposures and in what order are most effective?

This is ideal for managing resources that can be moved around quickly based on their ROIs. Some activities can be made more effective with better timing. The exposures can be delivered at the right level and frequency to avoid diminishing returns. It allows optimising at a very granular level, e.g. discarding least effective key words in SEM or affiliate websites, tweaking formats and creatives for online display.

MMM is based on past 2-3 years of weekly data. Some of the insights and recommendations, though more robust, might quickly become obsolete. In comparison, AM results are based on individual consumer journeys over several months or weeks.

While MMM has been developed to measure the effect of paid channels (and is good at this),the approach is not as suitable for measuring owned and earned media (so called ‘pull’ channels such as the attractiveness of own website, organic and paid search, newsletter, social network activity). AM method captures well all types of exposures and their synergies.

What Is the Catch?

Data, data, data…

Attribution modelling requires collecting data on, ideally, all touchpoints and customer interactions at an individual level. This means tracking all actions of every customer across all devices the person uses and ideally all offline exposures, such as TV, radio, print advertising, store visit, posters in the shop, out of home or even relevant information sourced from friends.

This might sound like a line from a futuristic movie, but this is what it takes to build a good attribution model that well represents reality.

And of course, though may seem obvious, one needs to have the right attribution modelling methodology.

So For Whom Is Attribution Modelling Useful?

The brands and companies that make most of their sales and advertising online and/or are able to capture their customers’ touch points and exposures can benefit most from AM.

Does AM Need To Be Complemented by MMM?

It depends. If AM can explain significant part of brand sales and the recommendations and optimisations based on the model drive further improvements, regularly updated AM is sufficient.

If only some offline touch points are missing, we can approximate them and include in the attribution model.

If we have customer-level information only for a subset of sales, attribution model can still be built in addition to the mix model. However, the less information is available at an individual level, the weaker are the grounds for AM.

How to Use Mix Model to its Full Potential?

For many companies MMM stays as the main approach for measuring the media impact. Instead of giving up on its measurement, the focus should be on how to get most out of MMM. This requires transparency when it comes to used variables and their transformations. Common standards should be set for data sources used in a model.

In addition, there will always be areas where broader mix model can provide very relevant findings. These include the impact of pricing, promotions and other external factors, long-term impact of brand tracking indicators on the bottom line. MMM is also more effective for future sales projections.

To Sum Up

The companies need to decide what approach allows the most holistic and accurate measurement Completeness and granularity of their individual-level data should be one of the main factors in this decision.

Even if AM will be posing increasing competition to MMM, its data requirements and the state of methodological framework indicate that the transition will be gradual.

The goal of the modelling exercise is to understand which activities primarily contribute to sales and how cost-effective they are. What matters is that we measure how marketing performs and are getting better at it.

Sydney, 09 May 2014,

Elena Yusupova, PhD