Endogeneity and marketing strategy research: an overview

Our discussion on endogeneity focuses on empirical research with firm data. For researchers interested in endogeneity issues in survey data, Sande and Ghosh (2018) provide an overview.

A strong caveat is the assumptions needed for a structural model to be identified. Often, these result in models that will not fit a marketing strategy application.

For ease of exposition we do not include any exogenous variables in the model formulations.

The key idea is to leverage variation independent of the variable of concern. For example, online spend (ad costs) for Portland in our case when the focal market is Seattle.

For 2 Stage Least Squares (SLS) please see: https://​www.​rdocumentation.​org/​packages/​AER/​versions/​1.​2-5/​topics/​ivreg. For simultaneous estimation please see https://​cran.​r-project.​org/​web/​packages/​ivmodel/​ or https://​cran.​r-project.​org/​web/​packages/​ivpack/​.

The R package for 2SLS estimation called “ivreg” (see Footnote 2) also provides a Hausman test.

The R package “ivreg” provides the Sargan test for 2SLS estimation. For Generalized Methods of Moments (GMM) estimation the R function “sargan” calculated the Hansen-Sargan test, please see: https://​www.​rdocumentation.​org/​packages/​plm/​versions/​1.​6-5/​topics/​sargan.

The key idea is to leverage variation independent of the variable of concern. For example, online spend (ad costs) for Portland in our case when the focal market is Seattle.

Please note that Kenneth Train as software available on his website, albeit only for the Mixed Logit Model: https://​eml.​berkeley.​edu/​~train/​software.​html.

Please see: https://​cran.​r-project.​org/​web/​packages/​REndo/​index.​html.

Park and Gupta (2012) give more details how to non-parametrically estimate the density of the marginal distribution of the endogenous variable, p.570–572.

In more detail, \( \tau ={\sigma}_{\varepsilon}\sqrt{1-{\rho}^2}{\varpi}_2 \) where ρ is the correlation coefficient, ϖ₂~N(0, 1) and σ_ε is the standard deviation of ε.

The goal of a field experiment is to be able to compare outcomes directly between treated and non-treated groups to determine the effect of the treatment, e.g. a given marketing strategy. In non-experimental data, the researcher cannot distinguish between a treatment and control group as the use of a particular strategy is generally non-random, but rather self-selected by the firm to implement, making the choice to implement a strategy, and thus the “effect” of the strategy, endogenous. To remedy potential self-selection bias in observational data, it may be possible to “match” firms exhibiting a given strategy with firms that do not, but otherwise appear very similar on multiple criteria through propensity score matching. This idea is to determine the probability of treatment assignment, for example, the propensity to start a loyalty program, conditional on observed baseline characteristics, e.g. firm size, customer demographics, and so forth. This allows the researcher to design and analyze an observational study to mimic the characteristics of a randomized controlled trial between the observed treated group and a matched non-treated group (Rosenbaum and Rubin 1983). In practice however, this may not always be feasible due to data constraints on which to produce a propensity score for each observation (see Austin 2011, p.414–415). For further reading on methodology, see Austin (2011), Heckman (1979), Rosenbaum and Rubin (1983), Guo and Fraser (2010), and the PSMATCH2 STATA module from Leuven and Sianesi (2003). For recent substantive applications and extensions in marketing research see Kumar et al. (2016), Ballings et al. (2018), and de Haan et al. (2018).

Please see: https://​stat.​ethz.​ch/​R-manual/​R-devel/​library/​stats/​html/​00Index.​html.

We thank an anonymous reviewer for raising this point.