Extract

In selecting an appropriate multivariate statistical method, I must take into account the nature of the data. Specifically, using a truncated count measure as a dependent variable (in the case of patent applications and product innovations) poses some econometric and measurement difficulty (Greene, 2003). The dependent variable can only take non-negative integer values. Further, many of the values assumed by the dependent variable are bunched close to zero. Employing an ordinary least squares (OLS) specification with a count measure as the dependent variable violates several of the basic assumptions of OLS in known ways (Kennedy, 1998). In particular, the errors become non-normally distributed. However, due to the central limit theorem, the test statistics and confidence intervals hold asymptotically. Asymptotically, then, the least squares estimator remains best, linear and unbiased, regardless of how the error is distributed (Kmenta, 1997). Therefore, I first ran regressions using an OLS specification. As the point of departure, equation (A.1) presents the standard OLS specification. Here I specify innovation as a linear function of the vector X of independent variables for firm i at time t and an error term, which I label uit: InnovationϭXit␤ ϩuit. (A.1) Given the panel data structure with several observations per firm, the possibility arises that uit in equation (A.1) is not independent across time within firms. Should I be unable to identify and measure all systematic effects on innovation not included in X, a systematic component would be embedded in the error....