Me, myself, and I: influence of CEO narcissism on firms’ innovation strategy and the likelihood of product-harm crises

A classification scheme for radical and incremental innovations is provided in Table 2.

We thank the Area Editor for suggesting that we explore the role of customer orientation in the link between CEO narcissism and innovation outcomes.

These industries include SIC 1 (construction), SIC 2 (light manufacturing), SIC 3 (heavy manufacturing), SIC 4 (transportation and communications), SIC 5 (retail), SIC 6 (finance and insurance), SIC 7 (personal, business, and recreation), and SIC 8 (health and education). We considered it important to test our conceptual framework across multiple industries in order to increase the generalizability of our results.

The average CEO tenure in S&P 500 firms was 9.2 years between 2006 and 2014, with CEO turnover being 9.9 years in 2014 (Aguilar et al. 2015). These statistics suggest that restricting our sample to firms whose CEOs were appointed between 2003 and 2005 and remained CEOs for at least 5–7 years (i.e., remained CEOs through 2010), did not result in a highly skewed sample of very stable CEOs. Nevertheless, we repeated our analysis after dropping firms whose CEOs had the longest tenure (CEOs who were appointed in 2003 and remained CEOs through 2010) with our overall conclusions remaining the same (see Appendix 1, Table 8).

Chatterjee and Hambrick (2007) conduct a series of validity tests which show that their index has strong construct validity. These tests include having a panel of experienced securities analysts rate the degree of narcissism of 40 of the CEOs in their sample. The analysts exhibited strong interrater reliability and their ratings of the CEOs were very much in line with Chatterjee and Hambrick’s (2007) unobtrusive index scores (correlation = .82, p < .01). We believe the unidimensional measure of narcissism developed by Chatterjee and Hambrick (2007, 2011) that we employ is also theoretically justified because narcissism, by definition, is a unitary superordinate construct that has multiple elements (Edwards 2001). Furthermore, as explained by Chatterjee and Hambrick (2007, 2011), the four indicators together capture the four facets of narcissism identified by Emmons (1987): superiority/arrogance, exploitativeness/entitlement, self-absorption/self-admiration, and leadership/authority. All of the four indicators we used can reasonably be thought to align with one or more of Emmon’s (1987) facets. Prominence in annual report photographs/ prominence in press releases align with superiority/arrogance, self-absorption/self-admiration, and leadership/authority, while the two compensation items (CEO relative cash/non-cash pay) align with superiority/arrogance, and exploitativeness/entitlement (Chatterjee and Hambrick 2007).

The proportion of radical NPIs was not defined for firm-years with zero NPIs. These firm-years were dropped while analyzing the link between CEO narcissism and proportion of radical NPIs.

We used negative binomial regression rather than a poisson regression because while both types of regression can model count dependent variables, a poisson regression assumes that the mean and variance of the dependent variable are equal (Kennedy 2003), which (as Table 3 shows) was not true for our dependent variable, i.e., NPIs. We used random effects rather than fixed effects in equations (1) and (3) as our focal independent variable (CEO narcissism) was time-invariant, and using a fixed effects would have necessitated that our focal independent variable be dropped. Nevertheless, for Models 2 and 4 in Table 4, we carried out a Hausman test and failed to reject the null hypothesis that random effects are consistent and efficient. The footnote of Table 4 provides details of these tests.

Our results remained robust to the use of a GLS random effects regression and to a two-limit Tobit regression bounded between 0 and 1. However, we present the results of a fractional logit regression given that a GLS random effects regression ignores the zero lower bound (Papke and Wooldridge 1996), while a Tobit regression employed in our case resulted in an excessive amount of censoring (values below 0 or above 1).