18 Modeling in Marketing

18.1 Definitions

  • Structural in quantitative marketing: An estimation strategy where we assume and impose a structure to the consumer’s maximization problem and these parameters are of the consumers’ utility functions (Tülin Erdem and Keane 1996).

    • These parameter are policy-invariant.
  • Reduced form brand choice model: functions of marketing strategy variables (e..g, marketing mix). Hence, reduced-form models’ parameters are variant to policy. So it’s hard to study outcome of policy implementation may be reliable.

Uncertainty about product characteristics + learning behavior affect brand choice.

Two models (good in cases where consumer learning is vital to the choice process):

  • Dynamic structural model with immediate utility maximization

  • Forward-looking dynamic structural model (i.e., consumers don’t choose products based on their current utility, but also future utility).

18.2 Quasi-Experimental

(Goldfarb, Tucker, and Wang 2022) for a review

  • Settings: Weather, border, contract changes, firm policy, life or regulatory changes.

  • Nine Stages to go through in a quasi-experimental design

    1. Research Question

    2. Data Question

    3. Identification Strategy

    4. Empirical Analysis

    5. challenges to research design

    6. Robustness

    7. Mechanism

    8. External validity

    9. Unproven or Caveats

18.3 Transformation

18.3.1 Log-transformation

(Manchanda, Rossi, and Chintagunta 2004; Wies et al. 2019) used 1 in place of 0 for log-transformation. And also use 0.5 and 0.0001 for sensitivity analysis.

To control for firm size effects, (Wies et al. 2019) scale advertising investments by the firm’s total assets in the given year

18.4 Endogeneity

Check Background in strucutral Models section under marketing mix models for more up-to-date academic practice.

18.4.1 Control Function

In the context of consumer choice model (Petrin and Train 2010)

18.5 Variance Info Factors

Sometimes you can decrease the maximum variance inflation factor by removing the squared terms and centering all non-dummy regressors.