Decision biases and entrepreneurial finance

We cannot in this paper infer whether the entrepreneurs engage in self-deception or not. For an experimental paper on the role of overconfidence, self-deceptive beliefs, and persuading others, see Schwardmann and Van der Weele (2016).

An exception is a working paper by Dushnitsky (2009), which at a more aggregate level of analysis correlates country-level optimism with the country-level valuation of start-ups by venture capitalists. Also, Dushnitsky (2010) and Van den Steen (2004) have formally modeled the potential effects of entrepreneurial optimism on potential investors’ valuations of an entrepreneur’s venture ideas.

The CIC evaluates potential entrepreneurs and their venture ideas at an early stage and provides diagnostic feedback. For more information about the CIC and its evaluation process, see, for example, Åstebro and Koehler (2007).

All of the data except for the IAP evaluation were collected through a telephone survey. We developed a list of 6405 inventor-entrepreneurs who had submitted ideas for IAP review between 1994 and 2001. Of this number, we were able to trace 1352 current addresses. Of these, 1272 addresses led to actual contacts. The adjusted response rate was calculated by the Center as the contact rate (1272/1352) multiplied by the cooperation rate (830/1272). The Center follows the statistical methods and best practices of the American Association of Public Opinion Research, http://​www.​aapor.​org. For further information about the survey procedure, please contact the authors.

We imputed the missing item responses five times, assuming data were missing at random (MAR), using a switching regression approach that is described in van Buuren et al. (1999). For an introduction to multiple missing data imputation, see Graham and Hofer (2000).

To better understand the composition of the entrepreneur sample, we further draw a comparison sample from the general Canadian population. Using random digit dialing, we queried a sample of 300 Canadians from the general population based on sampling quotas for province, employment, and gender, to reflect the similarities in the aggregate with the entrepreneurs on these three variables. Comparisons are then made on background characteristics (results available upon request). The combined samples from the general population matched with the entrepreneurs contain unusually a high proportion reporting that they are self-employed (63 %) or that they have owned a business (60 %). However, the rate of entrepreneurship is much higher for the entrepreneur sample than it is for the general population sample.

Sales data are truncated by survey date for some observations. If data truncation is correlated with independent variables, coefficient estimates may be biased. In an alternate analysis, we therefore forecasted future sales that were conditional on truncation using exogenous parameters. This implies that the forecast was uncorrelated with covariates and should not produce biased regression parameters, but it may introduce more noise, thus increasing the standard errors. We used the Bass diffusion with exogenous parameters to forecast sales. Results of this analysis are consistent with those reported in text. Contact the authors for detailed methods and results.

Pretests results indicated respondents’ difficulty with the items “I feel blue’ and ‘I dislike myself.’ After careful deliberations on the possible effects of the two items on the questionnaire response and reliability, we decided to exclude these items and two other items that are matched with them—hence the 6 items out of the original 10 items.

While following the basic principles for establishing mediation as implemented for OLS (Baron and Kenny 1986), our tests are performed within a more general statistical estimation framework (see Clogg et al. 1995) as follows. The correlation between the focal independent variable and the dependent variable is estimated without the mediator, while including all covariates. The model is then re-estimated adding the mediator. This generates two sets of parameter vectors and variance–covariance matrices. A Chi-square test on the difference in the target coefficient between the two estimations, given the differences in the parameter vectors and variance–covariance matrix between the two estimations, is computed to test whether there is mediation. The suest routine in Stata is used for this purpose (Weesie 1999).