Quantifying both the short-term activation and long-term brand-building effects of advertising is a central part of commercial marketing mix modelling (MMM). To directly isolate each effect, a common practice includes both short and high-retention Adstock variables in the estimating equation. However, if, as is often claimed, activation and brand-building are different mechanisms, such a ‘dual-Adstock’ approach is inconsistent and inherently flawed. Not only does it characterize each effect in terms of the same type of distributed lag process, it approximates long-term advertising effects with an ad hoc drift term. This can lead to spurious correlation issues and tells us nothing about the dynamics of long-term brand-building.
In this paper we demonstrate the theoretical and statistical pitfalls of the dual-Adstock model with two empirical case studies. We argue that the only credible way to separate short-term activation and long-term brand building effects is to first identify whether a permanent (long-term) component is actually present in the sales data. If so, observed sales can be meaningfully decomposed into short-term variation and long-term (base) evolution. Brand-building effects then equate to the impact of advertising on base evolution, whereas activation effects correspond to transitory variation around the baseline. To illustrate the point, we survey three extant approaches, each focusing on an appropriate time series treatment of the long-term component.
Contributions to the marketing literature are twofold. Firstly, we demonstrate both theoretically and empirically how a popular industry approach can lead to spurious claims of long-term effects, highlighting the need for more credible measurement techniques in practical MMM work. Secondly, we show how long-term effects can mean different things depending on the time series properties of the data. Consequently, if commercial MMM practice is to provide meaningful industry generalisations, long-term advertising effects need to be clearly defined and categorised. We conclude with managerial implications followed by avenues for future research.