Common pool size and project size: an empirical test on expenditures using Danish municipal mergers

The Danish municipal mergers deviate from the model by Weingast et al. (1981) in two respects: Firstly, the decision-making process did not follow a norm of universalism, but the collective decision materializes as the product of the merging municipalities’ decisions. Secondly, the model assumes that only a single political representative is elected from each district while multimember districts are used in Danish local government. However, this is arguably less important as it is already assumed that politicians care exclusively for their local electorate’s interests.

One might argue that the latter will act as a proxy for the former. However, a merging municipality’s population can account for a small or a large share of the newly merged municipality’s population without the number of municipalities in the merger being correspondingly lower or higher. Such deviations occur frequently in the case examined in this paper (r=0.50; n=217).

The number of municipalities varied between 270 and 275 from 2003 to 2006. Of these, 21 to 26 were excluded based on three criteria: if a municipality has had both municipal and county tasks, has been part of a merger not directly associated with the 2007 merger reform, or has been put under administration in the time period or a number of years before.

The assumption of a common trend in the DD design makes it well suited for analyses of subnational governments such as the Danish municipalities, which are all subject to the same overall legal and political framework.

The specifications of Eqs. (1) and (2) can be extended to account for persistence over time. Firstly, the period-specific specification in Eq. (1) is less sensitive to persistence over time as collapsing the data into a pre- and post-treatment period produces consistent standard errors when serial correlation is present (Bertrand et al. 2004). Secondly, all models are estimated with standard errors clustered at the municipal level. Finally, fixed effects specifications of all models, estimated using deviations from means, are also considered.

The explanatory power of the models with controls for changes in final accounts is a respectable 14–19 %. For budget overruns, however, the explanatory power is lower at 4–6 %. The explanatory power of the subsample design (examined in the next section) is similar at 11–15 % and 3–7 %.

For both outcome measures, excluding cases of doubt leads to marginal increases in the treatment effects. Furthermore, models (1)–(16) were reestimated using robust regression to investigate the influence of outliers. The treatment effects retain their significance except for model (2), and the estimates tend to be reduced, but decrease by no more than about a third. All models were also re-estimated with White heteroscedasticity-consistent standard errors, which make the treatment effect in models (2) and (6) turn insignificant. This does not lead to a different conclusion for the availability or size of the common pool. In addition, fixed effects specifications of all models were also estimated, which likewise produces very similar results. Models (2), (4), (10), and (12) were also reestimated with a binary indicator of population size (below, equal to or above 20,000 inhabitants, cf. the agreement on the reform) instead of logged population size. However, this yields only small changes in the treatment effects. As a final note, the identification strategy calls for the use of interactions which can create problems with multicollinearity. This can result in too large standard errors that increase the chance of Type II error, and hence, overly conservative test statistics (Gujarati 2003: Chap. 10). Multicollinearity is therefore more problematic in cases with nonfindings. The largest variance inflation factors are around ten in the models estimated on the full sample, and at the most six in the models for subsample #2. The highest values are found primarily for the interactions.