Leaving in Droves: Exit Chains in Network Attrition

Abstract

The article examines the emergence of “exit chains”—temporal clusters in attrition, which are expected but rarely documented. Studying attrition in an industry peer network (IPN), we compare the three modes of leaving: as an initial exit in a chain (“leader”), a subsequent exit in a chain (“follower”), and a stand-alone exit (“loner”). Combining regression and simulation techniques, the analysis affirms the role interdependence between leader and followers, whose rationales for leaving are distinct but complementary, one internal based on exchange imbalances, the other external based on exposure to peer influence. Exit chains are depicted as a by-product of social embeddedness and the inherently high costs of relationship termination.

ACKNOWLEDGMENTS

I would like to thank Olivier Fourcadet for his valuable research help and two anonymous reviewers for their useful comments.

NOTES

Notes

¹ Given the exclusive nature of leadership, it is possible that only those at the very high end of the status order are able to start an exit chain, suggesting that the status effect may also display a nonlinear form.

² Please see Zuckerman and Sgourev (2006) for a detailed discussion of industry peer networks.

³ Respondents were given a group roster and asked to specify with respect to each name on it whether: “You turned to that person for help or advice in solving a difficult business problem”; “You interacted with that person informally such as one would with a friend”; “You looked to that person as a source of motivation or inspiration.”

⁴ Using membership satisfaction, instead of expectation of leaving, produces substantively identical results.

⁵ We do not control for gender, race, or age, as the respondents are overwhelmingly white, middle-aged males.

⁶ We use two months as a boundary here because there is only one exit between 61 and 80 days, but more importantly, it is difficult to make a case for interdependence of exits that are more than two months apart.

⁷ One case posed a particular coding difficulty, as five members left on the same day, followed by four more a week later and another one two weeks later. We managed to interview one of the initial five, who asserted it was not an individual but a joint decision to leave. To maintain the maximum possible number of cases, we decided to include all five as leaders in the analysis. To check if it biased our results, we reestimated the analysis, excluding two and then three of the five. The substantive results were the same.

⁸ Endogeneity is not a problem here because the outcome—leaving—takes place sometime, from a few days to a full year, after the collection of the survey and performance data used for the construction of the independent variables. The duration of membership is taken into account, alleviating censoring concerns.

⁹ We also estimated a mixed logit model with a random coefficient but detected no significant group-level impact on the individual-level results. This was confirmed by including density and network centralization measures in the multinomial model, with both coefficients insignificant. These are excluded from Table 2.

¹⁰ The substantive interpretation of the coefficients is based on the results from the multinomial logit, as the probit coefficients are unwieldy to such interpretation. The results of the two models are very similar.

¹¹ About 8 percent of members have an indegree of zero, nominated by no one, while a similar proportion are nominated by over 70 percent of their peers.

¹² We checked whether external ties to BN nonmembers had an impact on “leadership” and we found none, the same for the other two roles. We have no systematic data on other potential external pulling forces.

¹³ In additional analyses not presented here, but available upon request, we checked whether it is exposure to influence by the subgroup or the BN group that matters for followers. The lack of significance for the group measure indicates that the true reference set for followers is their subgroup, not the group as a whole.

¹⁴ Additional analyses (available upon request) show that exposure to influence both by the subgroup and the group predict leaving as a loner, but the latter effect is stronger. A one-unit increase in mean group expectation to leave raises three times the odds of a loner exit, while the same unit increase in the mean subgroup expectation to leave raises these odds about two times.

¹⁵ It is limited-information because it uses the observed number of exits, E, but not the observed exit date.