Augmented Intelligence
Jul 07, 2016

Innovations in drawing insights from marketing ROI modeling

Companies in the US spent an estimated $161 billion on advertising in 2014[1]. Digital advertising spend in the US for 2013 was $43.1 billion and is expected to reach nearly double that amount—$82.2 billion—by 2018[2]. By the end of 2016, global mobile advertising spend is expected to reach $24.5 billion[3]. In 2012, pharmaceutical companies spent $15 billion in the US on sales force efforts and another $5.7 billion on samples distributed by that same sales force[4], with some estimates suggesting this will increase by at least 10% per year over the next five to eight years. All major verticals in the US spend billions of dollars on marketing each year. Since investments in media and marketing are so large and must compete for funding from other initiatives, there is a continuously growing need to measure, track, and optimize the return on these investments.

Some of the tools used for quantifying the role of marketing include marketing ROI models, which are classes of statistical techniques that determine the relationship between changes in investments and subsequent changes in sales, attrition, lifetime value, loyalty, and other dependent KPIs. These are powerful methodologies since they inform the amount of sales due to each marketing driver, from which ROI response curves are determined, budget scenarios tested, and marketing optimizations derived. Thus, ROI modeling helps meet the goals of measurement, tracking and optimization.

All companies in all industries, regardless of data maturity or analytic capability, can only truly optimize their marketing budgets if they know the nature and magnitude of the return generated from the spend. The subtle difference between investing and spending is not semantic but relates directly to the generation and expectation of incremental business due to those dollars. In order to use ROI models to quantify the return from media and marketing, decisions must be made about ROI modeling options to reflect market realities and consumer dynamics. Hence, for businesses to gain maximum accuracy and value from their ROI models several key assumptions must be calibrated.

Lag impact
Media and marketing are expected to drive impact over time, and with time the impact of prior media and marketing is expected to erode. This lag impact is typically captured in marketing ROI models with a function called Adstock.

This standard approach, however, assumes that the consumer has the greatest response to marketing immediately following exposure, and that, as time goes on, response diminishes. In consumer packaged goods (CPG), this is often considered reasonable since the majority of the consumer base is regularly and frequently in-market, and thus exposure to marketing coincides with purchase activity. However, in pharmaceuticals, patients cannot simply get medication without the intermediary steps of scheduling physician appointments, the physician prescribing the medication, and the pharmacy thereafter filling the prescription. Since the impact of marketing on prescriptions may not initially spike given expected lags associated with the intermediary steps, the lag between marketing and prescriptions may instead first build before peaking and thereafter decaying.

Like in pharmaceuticals, the response to marketing in financial services may take a considerable amount of time to build before the incremental lift in business is observed. The build-before-decay could be considerably long, and thus the marketing ROI model would be most accurate if most the realistic lag structure for the business is used, rather than just the industry standard assumption of Adstock.

Diminishing returns and threshold
The impact of diminishing returns is well understood in marketing budget decision making and captured in marketing ROI models with industry standard methodologies such as Log-Log modeling.

Increasing spend brings diminishing returns. Spend reduction is common, for instance, when margins are at risk, when cost savings are required, or simply when shifting spend between marketing vehicles. An assumption of diminishing returns is that at low levels of spend one should still expect incremental business to be generated. In the above example, reducing TV support from 120 GRPs to 40 GRPs is still expected to drive incremental sales. In fact, the diminishing returns structure imposes the assumption that the ROI always climbs when reducing spend.

The problem is that in certain industries, and in certain marketing programs, low levels of investment may not be sufficient to break through the noise of the market.

In these common scenarios, lower levels of spend may be too low to generate consumer awareness or behavior. Per the red curve in the above graph, a reduction from 120 GRPs to 40 is expected to generate no return whatsoever, a diametrically opposite business implication from the diminishing-returns-only approach to marketing ROI models. Hence, instead of assuming only diminishing returns are in play in the market, analysts should also test minimum thresholds in the marketing ROI models as the implication on budget decision-making is vast.

Assisted impact
The standard reports generated from marketing ROI models include contribution, ROI, the effectiveness of each marketing driver, and the optimal distribution of spend across the drivers. However, as competitive pressures increase, marketing budgets decrease, and customer loyalty to brands evolves, additional insights are required for marketers to gain customer attention.

Common needs among these additional insights include:

  • The synergistic effects between different marketing drivers
  • Optimal sequencing of messaging across channels
  • Short-term, long-term, direct, and synergistic ROI of marketing
  • Impact of marketing changes on non-marketing activities, and vice versa

Each of these needs can indeed be built into marketing ROI models, even though, as an industry-standard practice, analysts generally exclude them. To address these more sophisticated business needs requires replacing the standard linear model with a path model. The linear model quantifies the lift each marketing driver has on sales directly. The path model quantifies both the direct lift and the lift one driver has on another. For instance, TV may drive sales, but TV may also drive search, which in turn drives sales. Hence, path modeling enables insights for both strategic and tactical marketing decision-making.

Advanced users of ROI model results often ask questions only addressable by path modeling, such as, “Do banner ads lead customers to search about my product? If so, how much of the search effectiveness is due to banner ads?" Since Path modeling tests all combinations and permutations of paths between drivers and sales, the insights are often quite revealing, especially when a path is unexpectedly weak. The extent to which each marketing channel assists other channels provides an opportunity for marketing to optimize more than just spending levels, but also:

  • Sequencing of activities and messages
  • Budget prioritization
  • Enhanced strategic investment
  • Improved “in-market" decision making
  • Support for improved customer engagement in an increasingly complex multichannel environment

Thoughts on future improvement and innovation

a) Data granularity
Most marketing ROI models are constructed at a high level of aggregation, such as at a national or regional level. The requisite data is usually available at aggregated levels, and CMO and CFO dashboards are often built from data at these high levels. However, despite the ease and convenience of constructing high-level models, there are many opportunities for insights from models built with disaggregated data.

  • CPG – store-level modeling provides optimization opportunities for each store, along with the ability to determine which stores are less responsive to marketing for strategic and tactical intervention and solutioning
  • Pharmaceutical – physician-level models provide opportunities to supplement sales force analytics and optimization with the ability to determine which marketing levers can move each physician, especially key influential physicians in each sales territory
  • Financial Services – models by asset class and by geography provide insight into how different segments of the consumer base behave, and specifically insight into propensities to move from one asset class to the next
  • Automotive – dealership-level modeling provides insights into switching behaviors between competing brands in joint dealerships (e.g., a joint Honda/Volvo dealership) and the influence and interplay between nearby dealerships
  • Restaurant – models by store and by day of week provide insights into the unique impact of wait times, days of week, and menu nuances, as influenced by overall media and marketing messages

Convenience-based analytics, such as using higher levels of aggregation, leave out many local-level nuances, distinctions, and factors, and thus leave out too many analytic insight opportunities. Hence, to maximize value of the investment in marketing ROI modeling, the most granular the data is available should be considered in the decision for the level of the modeling to be created, not strictly the convenience of getting the data at aggregated levels.

b) Impact evolution and multi-audience targeting
Although many companies leverage marketing ROI models, they too often fail to see tremendous benefit. This often happens when the assumptions made by the model do not align to the business realities, such as using a lag structure unrealistic to the market, assuming a diminishing only returns behavior, or assuming no paths or synergies between drivers. Companies should ensure all ROI modeling assumptions are validated to drive real value and benefit from the models, including other important assumptions such as:

  • Impact Evolution: The impact of media and marketing often changes over time as messaging gets old, awareness levels rise, or competitive messaging changes. Hence, the model must allow the coefficients of media and marketing to change over time. Standard mix models assume the coefficients do not change over time
  • Multi-Audience Targeting: For some products and services, there may be several different target audiences, each of which have unique responses to media and marketing. Hence, the model should allow for multi-audience response curves. Standard modeling assumes only one primary audience

To gain the most accurate — and thus actionable — ROI models, statistical modelers must work with subject matter experts in the business to ensure the assumptions and underlying models are accurate representations of the business. Failure to do so leads to models whose results are not reflective of the market, consumer, or brand.

Maximizing the value of ROI modeling in business decision making
Knowledge and assumptions are often difficult to distinguish. One might firmly believe marketing is working, yet an unbiased statistical analysis may lead to a different conclusion. Likewise, one might believe a marketing program would not drive business, but an unbiased analysis may suggest otherwise. Since marketing decision-making relies on a combination of knowledge and assumptions, understanding the limits of any and all assumptions becomes crucial. To ensure ROI analytics provide maximum accuracy and value to decision makers, all assumptions must be validated, and the analytics modified if any assumptions are found invalid. It is thus incumbent upon the analyst and modelers to articulate all assumptions made and important for consumers of ROI models to review all assumptions, in order to drive the balance toward knowledge and away from assumptions. At the end of the day, marketers compete in a variety of ways, not the least of which is on decision-making. The best way to do this is to arm the decision-makers with the best possible data, most actionable insights, and the most realistic models.

  1. The Statista Portal, “Advertising spending in the United States from 2011 to 2014",
  2. The Statista Portal, “Digital advertising sin the United States from 2012 to 2018",
  3. The Drum, “Global mobile advertising spend set to reach $24.5 billion in 2016",
  4. eMarketer, “US pharma marketing spend falls for most channels",

About the authors

David Hauser, PhD

David Hauser, PhD

Chief Science Officer, CPG and Specialty Analytics

Follow David Hauser, PhD on LinkedIn

Jingfen Zhu

Jingfen Zhu

Chief Scientist, Analytics, Chief Science Officer

Follow Jingfen Zhu on LinkedIn