What determines entry? Evidence from Sweden

As mentioned, however, the self-employed can decide to withdraw part of their income as wages. Furthermore, to avoid or at least reduce tax evasion, Sweden has special tax legislation providing that dividends and capital gains for closely held limited companies (fåmansbolag) above a certain level are to be taxed as employment income (Stenkula 2012).

To normalize for different sizes of spatial areas, many previous studies defined entry using rates of new firm formation, t. But dividing by the existing stock of firms in the previous period causes endogeneity problems, as the denominator becomes a product of a lagged value of the numerator (Ashcroft et al. 1991). This also suggests that new businesses spring from existing ones, which is a “difficult conceptual position to sustain” (Love 1995, p. 156). Love instead proposes a normalization by working population, which allows for entry to arise from “the entrepreneurial acts of individuals, both employed and unemployed” (see also Audretsch and Fritsch 1994a). Some studies instead operationalize entry as absolute numbers of new firms, usually with a logarithmic model to avoid problems with normality. While computationally convenient, this means that the dependent variable is only well defined for industries showing positive entry. Two other solutions to deal with zero-inflated count data are square-root transformation, and to dichotomize the data. But both of these approaches also have major problems (Karazsia and van Dulmen 2008).

All tests were undertaken post-estimation. Results at all levels of aggregation are available from the author upon request.

If present, spatial autocorrelation could bias the results from these models and underestimate the SDs of the estimated coefficients (Anselin and Rey 1992). For this reason, Moran’s I (Moran 1950) was computed for each year. Results are available in Table 8 in the “Appendix.” Moran’s I can range from \(-\)1 to 1, with higher magnitudes indicating more spatial dependence. In our case values fluctuate around 0.02 and are generally not significant at the 95 % confidence level, i.e., the null hypothesis of spatial dependence can be rejected. The exceptions are the years 2000 and 2008, but even then the values are very low. Hence, it appears that spatial correlation is not a serious problem.

Estimations were also made using the Huber/White/Sandwich SEs in order to control over heteroskedasticity (Davidson and MacKinnon 1993, p. 552). This had very little effect on the SEs of the results.