Electoral accountability and local government efficiency: quasi-experimental evidence from the Italian health care sector reforms

Our estimates report that a 1 % increase in real per capita regional tax revenues increases efficiency by between 0.013 and 0.005 %.

Friuli-Venezia Giulia, the autonomous province of Trento, the autonomous province of Bolzano, Aosta Valley in the north, and the two main islands of Sardinia and Sicily in the south.

Northern regions: Piedmont, Lombardy, Veneto, Liguria, Emilia Romagna. Central regions: Tuscany, Marche, Umbria, Lazio, Abruzzo. Southern regions: Molise, Campania, Apulia, Basilicata, Calabria.

Law no. 43/1995.

In 1993, the direct election of the mayor and the president of the province revolutionised the local electoral systems. In 1994, after the referendum of the previous year, the election of members of the national parliament was held under a new majority system, which replaced the previous proportional system.

During those years, the entire political class underwent a change as a result of “Tangentopoli”, the system of corruption that was uncovered by a nationwide judicial investigation of political corruption named “Mani pulite” (Italian for “clean hands”). Mani pulite began in February 1992 and led to the demise of the so-called First Republic, resulting in the disappearance of the parties that had led the political scene since the post-war period.

Constitutional law no. 2/2001.

Imposta Regionale sulle Attività Produttive (Regional Tax on Productive Activities).

It is levied on the net value of the production of each region; its standard rate at present is 3.9 % and regional governments can increase or decrease it by up to 1 % (Longobardi 2005).

At present, its standard tax rate is 0.9 % and each regional government can increase it up to 1.4 %.

Law no. 56/2000.

As a result, there is very little overlapping between the electoral reform and fiscal reform, since most of the special regions after 2001 started to adopt the same electoral system of normal regions.

In 2003 the centre-right government, in order to fulfil its programme of cutting the general tax burden, banned regional governments from raising IRAP and additional personal income tax rates. The ban was then removed in 2006 by the new centre-left government.

The average 2005 per capita GDP in the south of the country is around 60 % of the average GDP per head in the rest of the country.

Electoral accountability is defined in Seabright’s 1996 seminal paper (Seabright 1996) in terms of the probability that welfare levels of a given jurisdiction determine the election of the government. Recently, Lockwood (2006) proposed to characterise this concept more precisely, either by the degree to which institutions allow the government to divert rents or by the degree to which institutions allow special interest groups to distort government decision-making by lobbying.

Government efficiency in the provision of goods and services is usually measured by the difference between the actual level of the output and the maximum level of output achievable, given the inputs employed in the production process. Alternatively, government efficiency can be measured by the difference between the actual cost of local services and the minimum cost attainable, given the actual output and the inputs prices. However, this second approach has been discarded in this paper in order to avoid the problems related to the scarce availability of input price data.

DEA was first developed by Charnes et al. (1978); a complete survey of data envelopment analysis can be found in Ali and Seiford (1993).

Stochastic frontier models have been developed simultaneously by Aigner et al. (1977) and Meeusen and Broeck (1977); a complete survey of panel data models is provided by Kumbhakar and Lovell (2000).

One of the first applications of this procedure was by Timmer (1971) in an attempt to explain interstate variation in technical efficiency in US agriculture. A two-stage approach has been used also by McCarty and Yaisawarng (1993) to investigate efficiency in New Jersey public school districts. Worthington and Dollery (2002) compare different methods to account for the effect of environmental factors on the efficiency of 73 New South Wales local governments in Australia. Afonso and Aubyn (2006) considered a two-stage approach in relation to the health production process of OECD countries by regressing efficiency scores on a set of variables such as GDP per head, education level, and health behaviour (such as obesity and smoking habits). Recently, Adam et al. (2014) have used the same methodology to estimate the effect of decentralisation on the efficiency of the public sector using a panel of 21 OECD countries over the period 1970–2000.

Debreu (1951) and Farrell (1957).

It is important to stress that input-based and output-based approaches to the evaluation of efficiency do not need to produce the same results. In particular the input and output approach indices of efficiency are equivalent only in the restrictive case of constant returns to scale. Hence, in our analysis, the use of both indices can be considered a sort of mutual robustness check.

2001 is the year in which special regions could start adopting the new electoral system, therefore the impact of the electoral reform is evaluated up to this point.

An alternative estimation strategy would be that of assuming a specific distribution for the two error component, usually a standard normal distribution for \(v_{it}\) and a truncated normal distribution for \(u_{i}\). The impact of the policy variables, then, can be estimated through maximum likelihood assuming that the policy variables are a specific component of the mean of the inefficiency error component (see Battese and Coelli 1995 for more details about this approach). The advantage of this approach is that \(u_{i}\) can be time variant, but the consistency of the maximum likelihood estimator depends on two crucial assumptions: the independence between the inefficiency error term and the inputs, and the correct specification of the distributions of the composite error term. In this paper it has been preferred not to rely on results based on these two strong assumptions because the focus is not on the measurement of the efficiency.

In the case of correlation between the input variables and the idiosyncratic error term, lagged input variables will be used as instrumental variables.

It is important to stress that the region Trentino Alto-Adige has been divided into the autonomous province of Trento and the autonomous province of Bolzano, since they have different health care systems.

Life expectancy is the average number of years of life remaining at a given age and is computed separately for men and women.

Infant mortality is the number of babies who die during the first year of life per 10,000 new-borns.

Neonatal mortality is the rate of new-borns who die during the first day of life per 10,000 new-borns. In some cases it is also measured in relation to the first 6 or 28 days of life.

The empirical evidence obtained using life expectancy as a measure of output (not shown for brevity in the paper but available on request) is only partially in line with the results reported in the paper. However, given the lower reliability of life expectancy as a measure of output this discrepancy does not undermine the robustness of the results discussed in the paper.

It is also important to note that most of the variables linked directly to lifestyle, such as, for example, the consumption of tobacco or the level of pollution, are not available disaggregated at the regional level for the entire time series dimension. Instead data about variables that might influence mortality rates, like the mother’s age at the birth of her first child, the birth rate, and the percentage of caesarean births can be easily collected.

Although current public expenditure is included among the inputs, either DEA or SFM will be used to estimate technical efficiency because current health expenditure is only a proxy of current physical inputs.

Remember that, as explained in Sect. 2, the tax reform was introduced in 1998 but became effective in 2001 after a transitionary period of 3 years.

Yardstick competition occurs when voters can compare tax policies and levels of public-good provision that have been adopted by officials in other regions with those offered in their own jurisdiction and then use their ballots as votes on the performance of their incumbents.

Tax competition occurs when local governments compete over tax rates in order to attract more tax-payers to expand their tax base. In the spirit of the Leviathan hypothesis (Brennan and Buchanan 1980), tax competition should reduce local government taxing power, improving voters welfare, only if officials are rent-seeking. Otherwise, the ‘race to the bottom’, (reduction of the tax rate) would very likely reduce social welfare due to an undersupply of public goods.

The impact of the health care deficit has been estimated separately for the two group of regions because this last source of financing followed a different pattern in the two groups (see Fig. 2).

Most of the empirical works (for example, Goudriaan and de Groot 1991; Bjurek et al. 1992; McCarty and Yaisawarng 1993; Lovell et al. 1993; Borger et al. 1994; Vitalianno 1998, and more recently ) that use a two-stage approach usually account for the fractional nature of the dependent variable using a Tobit model in the second-stage regression (one-limit or two-limits). The only case similar to Bernoulli quasi-MLE has been found in Worthington (1999), who uses a second-stage logistic model. Essentially, as suggested by Papke and Wooldridge (2008), the possible choice of a two-limit Tobit model for the second-stage regression is not suitable in this context because, although our response variable is bounded from below by zero, there are no observations at zero.

The reduced number of observations is due to the endogeneity of the regional tax revenues, because this problem has been taken into account using as instrumental variables lagged values of the tax revenues and of regional GDP.

The same results are obtained, including the control and policy variables among the regressors.