Exploring The Diffusion Of Innovation Among High And Low Innovative Localities

Abstract

Berry and Berry (1999, 2007) argue that diffusion of policy innovations is driven by learning, competition, public pressure or mandates from higher levels of authority. We undertake a first time analysis of this whole framework and present three sub-studies of innovation. First, we examine the drivers of total innovation. Second, we assess whether the factors influencing the most innovative localities are similar to or different from the factors impacting the low localities. Finally, we disaggregate total innovation into three different innovation types. Our findings, undertaken on a panel of English local governments over four years, reveal that a majority of the diffusion drivers from innovation and diffusion theory are indeed positively significant for total innovation. However, local authorities that adopt higher and lower levels of innovation than predicted do things differently while the framework has limited applicability to types of management innovation. We concluded that the Berry and Berry model is best suited to the analysis of total innovation, but not as well suited to the analysis of different types of innovation. We also outline a research agenda that might better explain the diffusion of public policy and public management innovation types than is captured by current literature.

Key words:

• Diffusion
• empirical
• high and low innovation localities
• innovation

Notes

1 In this analysis we do not have the dates of innovation adoption, so we cannot test the early adopters vs the laggards as many policy innovation studies do when studying single policy adoption. But with a limited time series and with variation in types and degrees of innovation across the English local governments, we can study whether they respond to the same innovation drivers. For example, the analysis may demonstrate that authorities adopting a higher than predicted level of innovation are more susceptible to drivers, such as learning than lower innovators who may be more susceptible to drivers such as peer pressure rather than policy and management factors.

2 The Cronbach alpha for organizational innovations is .64, slightly below the alpha of .70 that is usually considered adequate for scale reliability. However it is considered permissible to accept a lower alpha on scales with a small number of items and for new scales (Nunnally 1978; Hull and Nie 1981).

3 All the control variables were logged because of long tails. Correlations between the control variables were low bar that between population and service diversity (r = .71, p .05). We retain these measures in the model because they are theoretically expected to influence the adoption of innovations, and as we note below, there are not problems of multicollinearity in the dataset.

4 While the fixed-effects estimator takes account of the fact that intercepts and/or slopes vary across units (localities in our case), the random-effects estimator assumes that they vary randomly across units (Johnston and DiNardo 1996). Moreover, unlike the fixed-effects regression, the random-effects estimation allows us to include time invariant variables – three of our four control variables do not vary across time.

5 In order to examine if there are trade-offs between the adoption of different types of innovation (based on the assumption that the adoption of innovation types is intertwined and therefore correlated with one another), we also ran seemingly unrelated regression, as it assumes that the errors from the five equations are correlated (SUR). That is, instead of running independent regressions for each innovation type, SUR models a regression in which all our dependent variables (six) are predicted by the same set of predictor variables. SUR produces more efficient estimates because it weights the estimates by the covariance (if any) of the residuals from the individual regressions (Greene 2003: 340–51; Smith 2006). The SUR results did not suggest that there was covariance across the five innovation types, thus we proceed with the random-effects model.

6 For each model, we ran the influence and leverage diagnostics. This test allows us to identify whether there are localities whose extreme values influence the total estimation. Whereas a few localities were influential, they were isolated (not accumulated over years); therefore, we included all observations. On the other hand, there also exists the potential of having independent variables that are correlated, which leads to inefficient regression coefficients because an independent variable(s) can be totally (perfect collinearity) or partially (multicollinearity) predicted by other independent variable(s). However, the mean variance inflation factor (VIF) for each model shows no significant multicollinearity (VIF is 1.61 for the total innovation model).

7 That is, the split is not conditional on the mean value (above and below) of the dependent variable but on the residuals from the pooled regression with the entire sample. Therefore, these residuals also embody other potential influences from the organizational context, which are not explained by the independent variables included in the model.

8 This analysis is equivalent to estimating the model in the whole sample (pooled dataset) after adding dummies variables for high and low innovative localities and their interaction terms with all the independent variables. In doing this, however, several degrees of freedom are lost. Given that, we opted for the splitting sample and conducting the Chow test statistic to determine whether there are differences in coefficients across the two samples.

9 We ran models with and without control variables, and removing one control variable at a time. Results from these models show neither statistical nor substantive variation. We ran a joint F-test on the set of control variables. According to the result (F = 22.6, p 0.000), we are able to reject the null hypothesis that the coefficients on the control variables are equal to zero, justifying their inclusion in the model.

10 The survey data we employed in this study did not pair specific diffusion questions with specific innovations. If this were undertaken in future research it would allow a more complete test of the differences across types of innovation, and may produce stronger results.