We examine issues associated with various operational measures and model specifications in marketing-duality research that focuses on either a balancing or a combining perspective. In Study 1, we use archival data to demonstrate various operationalizations and models that subscribe to either perspective to test the impact of the exploration::exploitation duality, which produce mixed results. We then show that the multidimensional response surface approach constitutes a superior omnibus test that coalesces the two perspectives. Study 2, a simulation-based study, affirms the severity of biases in the existing approaches and the comprehensiveness and precision of the response surface approach. The cumulative evidence shows that the existing perspectives employ transformed measures that fail to distinguish firms with different approaches to the two activities underlying a duality. Thus, they suffer from many conceptual blind spots, produce mixed statistical conclusions, and are sensitive to changes in mean values. Our findings provide guidelines for the study of marketing dualities.
Note that the use of difference score is prevalent but heavily criticized in the domain of marketing (e.g., Peter et al. 1993). Nevertheless, the use of ‘gap’ scores and ratios remain prevalent in marketing (e.g., expectation disconfirmation gap). The observed interaction is spurious when the coefficient of the product term in the regression is significant even when there is no true interaction. The observed interaction is misleading when the observed interaction term is positive (or negative) while the true interaction is negative (or positive). The use of only the product term without the lower-order term can be found in research that relies on the product term as an index of ambidexterity (e.g., Gibson and Birkinshaw 2004; Jasmand et al. 2012).
We recognize that firms may be able to achieve higher levels of performance even when their approach to ambidexterity is highly imbalanced. Our polynomial approach can accommodate contextual variables (e.g., resource slack, firm age) by incorporating a moderator, either through split group analyses or through interaction models.
The median R-squared value is 0.43, ranging between 0.39 and 0.48 across all the iterations. Our conclusions do not change if we increase or decrease the variance explained. In terms of the condition index that may lead to unstable estimates, the highest number reported across all of our iterations is 10.86. The average across all of our simulation specifications is 4.821. Specifically, the average condition index for the polynomial model is 10.86, the sum total is 3.27, the conventional interaction is 6.09, raw difference is 2.48, the ratio is 2.49, and the absolute difference is 3.74. By “average condition index,” we mean the average overall condition index number across all the different iterations (meaning changing values of the mean and 1000 runs per mean value) of the simulation per model specification. These condition index values are all below 30, indicating that multicollinearity is not a major concern (Kennedy 2008).
We can also store the mean value of the beta coefficients, but we find that these become skewed, such that the betas do not follow a normal distribution when the model features an operationalization of ambidexterity that biases the estimator. Since the coefficients the models predict should follow a normal distribution (Cohen et al. 2003; Wooldridge 2010), we store the medians to ensure our results are not distorted.
We recognize that the combining or balancing perspectives might be appropriate for specific research contexts. For example, younger firms are less likely to have resource slack allowing them to pursue concurrent ambidexterity.