Attribution and the Marketing Mix Model

Inconsistent user-identifiers and the walled garden policies of dominant social media players, together with the (imminent) abolition of third-party cookies has led to renewed interest in the marketing mix model as an attribution tool. However, to be useful in a post-MTA world any ‘next generation’ MMM framework needs to deliver on three fundamental business issues.

Firstly, to serve as a true attribution solution, MMM needs to focus on causal estimation methods. Too often we see reliance on consumer journey solutions to address the problems of last-touch attribution. However, these ignore the critical issues of selection bias endemic in much online media – leading to endogeneity bias and misallocation of the marketing mix. The growing popularity of automated machine learning approaches to the mix model only serve to exacerbate this problem, where the focus is on prediction not causation.

Secondly, MMM needs to quantify the long-term (base building) effects of marketing and so inform brand-building strategy. Standard approaches are simply not set up to measure these effects, with fixed baselines and a focus on short to medium-term lag structures or Adstocks. Alternative time series structures are required that can quantify both short and long-term (base) variation – coupled with dynamic network models that can explain the causes of base variation and the economics of brand-building.

Finally, next-generation MMM needs to fill the gap left in a cookie-less world to deliver granular and swift insights on marketing ROI and optimal budget allocation. Suitably identified high dimension mix models – across consumer cohorts by day or hour – can fit the bill. This can provide many of the claimed benefits of MTA such as granular online media effectiveness ranking by publisher and placement with the added benefit of quantifying the contribution of pricing and offline media, controlling for the wider economic environment and the ability to analyze brand building.

Modelling short and long-term effects in the consumer purchase journey

The role of advertising in consumer demand has received much attention in the marketing literature, where the marketing response model has emerged as a dominant analytical framework. In recent years, the basic model structure has evolved in three directions. Firstly, the incorporation of online paid, owned and earned media to accommodate the off-online path-to-purchase. Secondly, the introduction of consumer mindset metrics to capture the emotional foundations of brand-building. Finally, combinations of both to offer a complete view of the consumer purchase journey. In this paper, we argue that no one development successfully provides a fully holistic representation of the role of advertising. We thus propose a more comprehensive structural approach: one that combines search, paid, owned and earned media with a cointegrating mechanism for brand-building, linking mindset metrics directly to long term base sales. The result is a practical modelling system for marketing mix practitioners with a rigorous economic and statistical foundation for short and long-term advertising effects. The proposed approach is applied to a global electronics manufacturer in the US market, where results fall into three broad areas. Firstly, we demonstrate how off and online marketing interact in the journey structure, the part played by search, owned and earned media and how specific media types play different roles. Secondly, we illustrate how consumer touchpoint experience is a key element of the process, improving measurement of in-store marketing. Finally, we show how consumer mind-set metrics and social media play an important word-of-mouth role, with paid and earned media working together to drive long-term demand evolution.

Brand management and the marketing mix model

Brand management is typically defined as the way in which brands are positioned in the marketplace, both in terms of tangibles such as price, packaging and the marketing mix and intangibles such as consumer perceptions and brand equity. The conventional marketing mix model is often used to inform the tangible elements, but is lacking into two key aspects. Firstly, it ignores the role of intangibles. Secondly, the focus is solely on individual brands in isolation. This ignores the wider competitive context, where the decision to choose one brand is the simultaneous decision not to choose another. Successful brand management, however, requires a simultaneous holistic view of all players.

To address both issues, this paper argues for a dynamic time series version of the discrete choice attraction model. Firstly, the demand system structure treats the entire category as a single unit, capturing competitive steal, cannibalisation, halo and category expansion effects of brand specific marketing. This provides accurate marketing ROI and budget allocation, facilitating the manufacturer-retailer relationship. Secondly, the time series approach allows us to quantify the evolution and drivers of consumer brand tastes – critical to understanding brand intangibles. This enables managers to set marketing strategy for optimal long-term brand performance.

Dynamic MMM and Digital Attribution

Digital media attribution aims to identify the combination of online marketing activities and touchpoints contributing to online sales conversion. Given the availability of unique user-identifiers, analysis conventionally traces the actions of single individuals. Traditional media attribution, on the other hand, evaluates the offline sales impact of offline marketing investments. Measurement is typically carried out at an aggregated level, using marketing mix analysis applied to groups of consumers, either at store, chain or market level. With the advent of multi-channel marketing, comes the need to measure the sales impact of inherently micro-focused digital media alongside more macro-oriented traditional advertising. Consequently, any analytical approach that aims to incorporate both elements inevitably involves a degree of data aggregation or disaggregation depending on whether we adopt a macro or a micro route.

This article presents an aggregate modelling framework for traditional and digital marketing attribution. The model structure is based on a theory of consumer purchase behaviour that naturally combines off and online marketing touchpoints, with response parameters estimated using appropriate dynamic econometric techniques. Outputs provide many managerial benefits, ranging from accurate ROI and media planning inputs through to simulation and demand forecasting.

Advanced Methods in Marketing Econometrics

Econometrics is broadly defined as the application of statistical and mathematical models to test and quantify economic theory. Marketing econometrics specifically aims to quantify the role of marketing in driving consumer purchase behaviour and is often called Marketing Mix Modelling (MMM). All marketing mix models are essentially price demand curves augmented to include a wide range of media and economic variables and routinely offered by consultancies and media agencies to assist clients in three main areas:

1. Price elasticity, marketing performance and ROI

2. Optimal allocation of marketing resources 

3. Forecasting consumer demand to aid supply chain planning

The economic foundations of marketing mix models are typically approximated with simple single-equation regression models of product demand. However, these approaches are ill-equipped to solve many client business issues. In this article, we look at how marketing analytics has evolved to provide more accurate representations of consumer demand.

Marketing Mix Modeling & Big Data

In the marketing sector the growing wealth of data on individual purchase behavior, online activity, social media and socio-demographic profiles is changing the face of media buying and analytics. Firstly, by consolidating the information into single data management platforms, media agencies can achieve more granular segmentation of viewing audiences, leading to increasing efficiency in digital media buying and targeting. Secondly, the sheer volume and complexity of available data has prompted an increasing use of machine learning methods to generate marketing insights. This typically covers correlation-based model building, network analysis, consumer segmentation, classification and forecasting.

Against this backdrop, more conventional ‘small data’ analytics such as econometrics and marketing mix modelling have taken something of a back seat. This is unfortunate since big data should not necessarily be viewed as more accurate with no need for conventional interpretation. Data mining techniques for example are ideal for association rule learning, where customer purchasing patterns on frequent and jointly purchased products can help guide marketing strategy. However, such methods shed little light on underlying data generation processes, marketing ROI and the causal impacts central to simulation and scenario planning. Consequently, despite the significant role that unstructured big data can play in marketing analytics it should not overshadow the importance of traditional analysis of large structured data sets. The key is to build analytical frameworks that can harness the value of increasing data size, yet retain the benefits of sound economic theory and valid causal inference. This is the domain of Big Data econometrics.

Marketing mix modelling and return on investment

This article provides an overview of the commercial marketing mix model (MMM), from economic foundations, through statistical estimation to commercial outputs. Against the background of the conventional approach, we put forward alternative theoretical and econometric frameworks for improved short-term ROI evaluation, together with techniques for evaluating the long-term effects of marketing investments and how these may be combined with short-term results to provide total ROI. We conclude with a discussion of the managerial benefits of the mix model and the total returns on marketing.

Limitations of Conventional Marketing Mix Modelling

The primary function of all marketing mix models is to quantify the sales contribution of marketing activities, with a view to calculating Return on Investment (ROI). To accomplish this, all such models employ econometric techniques to decompose product sales into base and incremental volume. Models that focus solely on incremental volume often recommend a marketing budget allocation skewed towards promotional activity: short-run sales respond well to promotions, yet are less responsive to media activity – particularly for established brands. This, however, ignores the long-run view: that is, the potential brand-building properties of successful media campaigns on the one hand and the brand-eroding properties of heavy price discounts on the other. Acknowledging and quantifying these features is crucial to a complete ROI evaluation and a more strategic budget allocation.

This article puts forward a unique approach to resolving this issue. Measuring the long-run impact of marketing investments essentially amounts to quantifying their impact on the trend component of the sales series: that is, on the evolution in base sales over time. However, this is not possible in conventional models since base sales are essentially fixed by construction. To deal with this problem, the marketing mix model needs to be re-structured as an Unobserved Component Model (UCM) to accommodate both short-run and long-run variation in the data. The former is used to calculate ROI on marketing investments in the usual way. The latter measures the evolution of brand preferences over time. This generates an evolving baseline which, when combined with marketing investments and consumer tracking information, allows a quantification of long-run ROI.

Modelling and forecasting brand share: a dynamic demand system approach

A key feature of imperfectly competitive markets is the presence of differentiated products that cater specifically for the heterogeneity of consumer tastes. Identifying substitution patterns between such products in response to changes in price and other marketing variables is central to an understanding and estimation of consumer demand. A popular modelling framework is the continuous choice Almost Ideal Demand System (AIDS) of Deaton and Muellbauer (1980). In this paper, we develop the static functional form of the AIDS to include additional marketing mechanics and provide a dynamic generalisation based on the Unobserved Component Model (UCM) and state-space modelling techniques. 

The proposed model improves on conventional dynamic econometric share models in three key ways. Firstly, the model structure explicitly separates the transitory and permanent components of each brand share. This allows a formal analysis of their time series properties that is statistically superior to the standard unit root tests generally employed in the literature. Secondly, the parameters of the marketing variables are directly interpretable as short-run demand effects: parameter estimates in standard short-run models formulated in first differences are not readily interpretable in this way, whereas parameters based on short-run models using lagged dependent variables are biased. Finally, the extracted permanent components can be used to assess the long-run impact of marketing activity.