Is there room for containing healthcare costs? An analysis of regional spending differentials in Italy

ORIGINAL PAPER

Is there room for containing healthcare costs? An analysisof regional spending differentials in Italy

Maura Francese • Marzia Romanelli

Received: 19 January 2012 / Accepted: 22 January 2013 / Published online: 20 March 2013

� Springer-Verlag Berlin Heidelberg 2013

Abstract This work aims at identifying the determinants

of health spending differentials among Italian regions and

at highlighting potential margins for savings. The analysis

exploits a data set for the 21 Italian regions and autono-

mous provinces starting in the early 1990s and ending in

2006. After controlling for standard healthcare demand

indicators, remaining spending differentials are found to be

significant, and they appear to be associated with differ-

ences in the degree of appropriateness of treatments, health

sector supply structure and social capital indicators. In

general, higher regional expenditure does not appear to be

associated with better reported or perceived quality in

health services. In the regions that display poorer perfor-

mances, inefficiencies appear not to be uniformly distrib-

uted among expenditure items. Overall, results suggest that

savings could be achieved without reducing the amount of

services provided to citizens. This seems particularly

important given the expected rise in spending associated

with the forecasted demographic developments.

Keywords Government � Expenditure � Health � Regional

inequalities

JEL Classification H51 � I1

Introduction

Health spending is significant in Italy as in the majority of

developed countries. Public health expenditure accounts

for about 7 % of the GDP and over the last decades it has

been growing faster than income. Given the expected

population ageing, it is likely that this trend will continue

in the future as well [11], raising a serious challenge to

fiscal policy and the sustainability of public finances.

The provision of healthcare in Italy is the responsibility

of the regional governments, within guidelines set at the

central level and aimed at ensuring that all citizens access

similar amounts and quality of healthcare. Notwithstanding

a significant redistribution of resources, there remain

spending differentials between regions. These could be due

to structural differences (such as demographic or epide-

miological characteristics) or they could reflect differences

in the quality of the services provided to patients and in the

degree of efficiency in the use of public resources.1 In the

latter case, spending differentials could highlight potential

scope to improve the value for money in healthcare.

The objective of this work is to examine whether saving

margins exist by identifying the main drivers of regional

spending differentials that are not due to structural differ-

ences in population or other determinants of healthcare

demand.

In this work we follow an aggregate approach. Instead

of focussing on specific categories of treatments or par-

ticular public programmes, we consider per capita healthThe views expressed in this paper are those of the authors and do not

necessarily reflect those of the Bank of Italy.

M. Francese (&) � M. Romanelli

Research and International Relations, Banca d’Italia-Economics,

via Nazionale 91, 00184 Rome, Italy

e-mail: [email protected]

M. Romanelli

e-mail: [email protected]

1 A similar argument has been put forward also in other analysis. For

example, for the US, a CBO study argued that ‘large differences

across the country in spending for the care of similar patients could

indicate a health care system that is not efficient as it could be,

particularly if that higher spending does not produce a commensu-

rately better care or improved health outcomes’ (see CBO [6]).

123

Eur J Health Econ (2014) 15:117–132

DOI 10.1007/s10198-013-0457-4

expenditure in the Italian regions2 and study the drivers of

unexplained differentials. In doing so, we start with an

approach similar to that used in the literature to compare

health spending across countries. In this respect our setting

presents several advantages, given that it is reasonable to

assume a similar structure of preferences, input prices and

broad institutional arrangements throughout the country.

Italian regions have a common legal and institutional

framework and, as explained above, are mandated to pro-

vide homogeneous healthcare coverage. They anyway

operate in a context of administrative autonomy: each

region can adopt, up to a certain extent, different organi-

sational architectures3 and within limits vary the degree of

financing through regional taxation and co-payments. To

be noted is that regions share common financing mecha-

nisms. Notwithstanding the many changes and reforms, it

can be taken that the resources made available at the

regional level are based on the overall need for healthcare

in each region and past spending,4 independently of the

source of financing and regional fiscal capacity.5 In prac-

tice, the central government has been responsible for filling

the gap between the financial needs of each region and the

actual funding derived from revenue directly collected by

the regions (e.g. a regional tax on productive activities and

a personal income surtax) and through an equalising cen-

tralised procedure that uses general fiscal revenues. Given

that the financing mechanism and equalising procedure are

common to all the regions,6 their incentives/disincentives

are likely to be the same, plausibly impacting on the overall

level of spending but less on the regional differentials.

In order to set the background for the empirical analysis,

we start by providing a short overview of the available

literature and by describing health expenditure dynamics

and their regional distribution. The fourth section presents

the model from which we derive our estimating equations;

the fifth and sixth respectively describe the analysis of per

capita regional spending determinants and identify the

drivers of observed differentials and discusses the results.

The final section draws some conclusions.

The results of previous studies

The comparison of spending levels among jurisdictions or

production units is not new. Other works have explored this

issue in order to ascertain whether differentials reflect dif-

ferences in the quality and/or access to care (for the US, see

[33] and [12]), either focussing on certain types of services

(for example hospital care as in [34]) or on spending for

particular public programmes and categories of treatments

(for example Fisher et. al. [13], consider end-of-life spending

for Medicare beneficiaries in the US). These works suggest

that higher spending does not systematically reflect better

quality or access, but that it is mainly linked to care practices

(such as a greater orientation towards inpatient or specialist

care) or to targeting specific care practices to the appropriate

reference population (see for example Bodenheimer and

Berry–Millet [2]). Furthermore, higher spending does not

appear to be systematically related to improved health out-

comes or care satisfaction [14, 15]. The latter result also

generally holds when the analysis between health care

expenditure and health outcomes is run at the international

level (for an overview of the available recent research on

international and within countries comparisons, see [4, 29]).

Health expenditure developments and regional

differentials

Since the establishment of the Italian national health system

(NHS), total public spending on healthcare has increased

significantly (from 5.1 % of GDP in 1979 to 7.3 % in 2010).

The considerable effort made to curb expenditure in the first

part of the 1990s, owing to the crisis at that time and also in

view of the need to meet the Maastricht criteria in order to

qualify for admission in the Eurozone7 was the only real

exception. However, the decline was achieved mainly

2 In particular, we consider all the Italian regions, both ordinary and

special statute plus the two autonomous provinces of Trento and

Bolzano.3 Within given guidelines, regions are allowed to define some

characteristics of the health system, such as the structure of the

hospital network, the share of private providers and the extent to

which they resort to direct distribution of pharmaceuticals.4 Also recent measures on fiscal federalism maintain the principle of

equalising resources among the regions on the basis of the popula-

tion’s need for healthcare. The challenge of the reform lies in defining

the criteria (the ‘standard costs’) to compute such needs. For a

description of the evolution of the financing mechanisms of the NHS,

see for example Caroppo and Turati [5].5 The fiscal capacity of the Italian regions is strongly (if not mainly)

affected by the uneven distribution of the tax bases across the country

(see De Matteis and Messina [10]), which reflects the degree of

economic development of the regions themselves and can be

marginally changed by the economic policies adopted within each

region.6 Different arrangements are in place between the ordinary statute

regions and the special statute ones (Valle d’Aosta, Trento, Bolzano,

Friuli Venezia Giulia, Sardegna and Sicilia). However, this affects

mainly the composition of financing (share of own resources vs. funds

drawn from national general taxation). The principle of providing

comprehensive and homogeneous coverage through a national health

system applies to all parts of the country.

7 For example, Bordignon and Turati [3] suggest that the reason

behind the expenditure reduction in the mid 1990s was the temporary

changes in the expectations by the regions, which, at the time

perceived central government’s threat not to bail them out in the event

of spending overruns to be a credible one. Another explanation is the

presence of flypaper effects and the temporary reduction in transfers

from the central to local governments in the early 1990s.

118 M. Francese, M. Romanelli

123

through spending postponements and generalised cuts in cash

outlays. The positive results of the first half of the 1990s were

more than offset by the increase over the ensuing decade.

The strong dynamics of health expenditure in Italy

might reflect not only structural factors (such as the rising

costs of new expensive technologies and population age-

ing) but also the lack of effective use of public resources.8

Given the expected population ageing and the costs of the

economic and financial crisis, it is of growing importance

to identify appropriate measures to increase efficiency in

the use of public funds.

In this respect, understanding the determinants of the

gap between expenditure levels among the different Italian

regions can be a useful starting point. A first glance at per

capita public current health expenditure highlights non-

negligible variation (with spending in 2006 ranging from

1,543 euro in Calabria to 2,137 in Bolzano). However, it is

possible to identify at least one cluster formed by the

southern regions, which generally have: (1) lower than

average per capita expenditure and (2) higher growth rates.

The estimation of a growth path of per capita health

expenditure over the period 1996–2006 suggests the pres-

ence of convergent dynamics among the regions.9

This evidence is questioned when adjusting expendi-

ture to take into account patient mobility among regions10

and the different age structure of the population. In par-

ticular, we observe many southern patients travel to the

Centre and North for medical care so that per capita

spending for residents in the northern regions is higher

than spending per treated person. This pattern is stable

over time. Furthermore, as to the impact of the age

structure, on average southern regions have a younger

population. We therefore adjust per capita expenditure to

take into account these two features.11 In this case, per

capita expenditure appears to be higher on average in the

South.12 Moreover, variability increases across the coun-

try and the estimated growth path does not support con-

vergence in the expenditure pattern.

The level of within-country variation in Italy appears

lower than that observed for UK (Fig. 1), notwithstanding

the similarities in the organisation of the NHS system. It is

close to that in Spain where health expenditure is a

responsibility of regional governments as in Italy and to

that in Canada where healthcare is financed jointly by the

local and central government (via transfers and an equali-

sation programme). In the US, where the system is very

different and the role of the private sector significantly

more relevant, even public spending (in particular Medi-

care and Medicaid) shows much higher within country

variability (CBO [6]).

The reference model

We consider a benevolent regional government (i), which

wants to maximise the following:

maxxh;xu

p(KiÞUihðxi

h;KiÞþð1� p(KiÞÞUiuðxi

u;KiÞ ð1Þ

where

Uig : 9exg\1 such that Ui0

g ðxÞ\08 x [ exg g = h,u

Uhi is the utility of a healthy individual from the con-

sumption of healthcare (x) and Uui is the utility of a sick

one; xih and xi

u are the quantities of healthcare consumed

respectively by a healthy and a sick individual in region i; p

is the fraction of healthy individuals in the region and 1-p

the fraction of sick ones; Ki is a vector of economic, social,

demographic and health status characteristics that affect

both the probability of being sick and the preference

structure for the consumption of healthcare. We assume

that the budget constraint is either not binding13 or that the

regional government can draw additional resources from

the residents in the region to finance the production of the

optimal/desired health spending.

The optimal level of healthcare provision is:

x�i ¼ pxi�h þ ð1� pÞxi�

u ð2Þ

where xhi * and xu

i * solve problem (1).

8 For example, regarding the hospital sector, Iuzzolino [23] observes

that, compared with the other OECD countries, Italy has a larger share

of public hospitals of small size, probably because of poorer

outpatient care and an unbalanced allocation of resources between

different types of services. This is reflected in the unit costs of

hospitalisation and the share of outlays for hospital in total spending,

which are higher than the OECD average.9 For the details of the model estimation, see [17].10 Patient mobility among regions is allowed under the national

public healthcare system: healthcare costs are covered by the NHS

independently of the region actually providing the service.11 In order to compute per capita public current health expenditure

adjusted for patient mobility and the age structure of the population,

we have used population data (broken down by region and age) and

the coefficients published by Ministero della Salute [27], which

capture differences in the consumption of pharmaceuticals and

hospital services between individuals at different ages. We have then

adjusted expenditure to take into account population mobility (using

expenditure data for non-residents published yearly by Ministero

dell’Economia e delle finanze for each region) and finally computed

spending per capita values. For more details, see [17].

12 In particular, taking into account patient mobility and the age

structure of the population, the per-capita expenditure of the southern

regions in 2006 increases from 1,601 euros to 1,703 euros, while in

the Centre and the North it decreases from 1,735 euros to 1,680.13 This assumption is consisent with the literature on soft budget

constraints and with the evidence of systematic spending overruns and

deficits in the health sector.

An analysis of regional spending differentials in Italy 119

123

Given the cost function Ct(x), assumed to be the same in

all regions, in each period t per capita public health

expenditure will be:

sit ¼ Ctðxit � ðKitÞÞ þ wi 8i ð3Þ

where wi (‘waste’) captures the distance of region i with

respect to the cost frontier.14 Assuming that each

region’s relative inefficiency is invariant over time, we

can derive the following estimating equation for per

capita spending:

sit ¼ aþ wi þ k0itbþ eit ð4Þ

with eit being the error term.

This framework assumes that under different financing

settings (i.e. resources drawn from a central budget or

through local taxation) the regional governments would

anyway be able to finance spending sit, hence ensuring the

production of x�it, whatever the region’s fiscal capacity and

efficiency level.

For this to be possible:

1. the objective of the central government (CG) must be

to ensure the provision of x�i ;

2. CG must be able to collect enough resources to finance

the production of x�i in all the regions;

3. CG and the regions observe spending levels and the

composition of financing but not waste levels ðwiÞ.

In a setting where the above conditions hold and

where healthcare is financed via both transfers received

from the central government and using revenue from

local taxes, regions have the incentive to set the latter at

the minimum if they believe that the CG will cover the

remaining costs. In such a framework each region will

be able to produce x�i and tax revenue collected by CG

in rich regions will be used to finance spending in poor

ones as well. Tax rates at local level will be correlated

neither with efficiency in the production of healthcare

nor with income differentials.

A similar framework, where the CG announces fixed

flat transfers but only some regions believe in its com-

mitment, will have an analogous outcome. Ex post, CG

will find that to stick to its commitment is not optimal

(otherwise in some regions it will not be possible to

produce x�i ) and low local tax rates will not necessarily

reflect high productive efficiency. They could also be due

to regional beliefs (about the lack of credibility of CG

commitment) or low demand levels. Again the ranking of

regions according to the level of local tax rates will not

necessarily reflect the ranking in productive efficiency. In

this setup regions that overspend with respect to the

resources set by the central government are not neces-

sarily the most inefficient ones.15

We can therefore proceed with the estimation of Eq. (4)

disregarding fiscal capacity, local taxation or regional

deficit indicators. The aim in the first step is to infer dif-

ferentials in spending among regions (wi), which are not

related to structural differences in the demand for health-

care (aþ k0itb), and in the second step to analyse the pos-

sible drivers of such gaps.

Regional differentials in health expenditure. The

empirical analysis—first step: the impact of demand

factors

The first step considers regional per capita spending levels

and their determinants in order to control for the demand

structure (such as demography, education, etc.), which can

be assumed, at least in the short and medium term, to be

independent of healthcare policies.

The empirical analysis at the international level has

pinpointed the correlations between health spending levels

and demand factors such as income, the demographic

structure and the health status of the population (see, for

0.04

0.05

0.06

0.07

0.08

0.09

0.10

0.11

0.12

0.13

1992

1993

1994

1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

Italy Spain UK Canada US

Fig. 1 Per capita public health expenditure-coefficient of variation

Sources: our calculations on: Italy: public health expenditure

corrected for mobility (see footnote 11); Spain: Ministerio de

Sanidad, Servicios Sociales e Igualdad, INCLASNS-DB database,

‘‘Gasto sanitario publico territorializado, por habitante protegido’’;

UK: HM Treasury, Pesa 2008, per capita public health expenditure;

Canada: Canadian Institute for health information (excluding North

West Territories and Nunavut); US: public (Medicare and Medicaid)

expenditure-Centers for Medicare and Medicaid services

14 wi can be interpreted as the cost of inefficiency for region i.

15 For a more detailed discussion of the impact of the financing

mechanism in our framework, see [17].


123

example, [18, 28]). Usually, most of the empirical literature

also controls for labour market and education indicators or

variables capturing institutional differences between

healthcare systems.16

We use a similar approach to investigate regional per

capita spending levels and in turn identify ‘‘unexplained’’

differentials. We use a panel of the 21 Italian regions for

14 years (1993–2006), which includes variables on per

capita health expenditure and its breakdown, per capita

GDP, health status indicators, variables relating to the

labour market and education (for descriptive statistics, see

Table 1).

In particular, we estimate the following semilog

expenditure model:17

ln rexp pcitð Þ ¼ aþ g ið Þ þ x0itbþ eit ð5Þ

where rexp_pcit is the real per capita expenditure

observed in region i at time t (corrected to take into

account patients mobility), g(i) is the region specific

effects that we initially assume to be composed of fixed

effects,18 i.e. g(i) = ai, and x includes: real per capita

GDP; the dependency ratios of the elderly and children (to

capture the U-shape of the age-expenditure profile) and

female life expectancy (the latter two variables as indi-

cators of the demographic structure of the population); a

proxy for the health status of the population (this proxy

takes into account the following conditions: high blood

pressure, heart attack, angina, other heart diseases, dia-

betes, ictus, together with male and female prevalence

rates for malignant cancer); a proxy19 for habits that affect

health (such as tobacco and alcohol);20 the female

employment rate, as an indicator for the provision of

informal care; the level of education (% of people with at

least a bachelor degree); proxies for the regions’ devel-

opment measured as the percentage of employees

respectively in agriculture and in manufacturing.21

We also consider a time-specific effect in order to sep-

arate the years before and after 1995.22 Finally, eit is the

error term.

In particular, we consider three specifications (Table 3),

which simply differ as to the denominator used to compute

the dependent variable (population, equivalent population for

the consumption of hospital services or equivalent population

for the consumption of pharmaceuticals; in the latter cases the

variables accounting for the demographic structure of the

population are not included among the regressors). We esti-

mate (5) using a panel fixed effect estimator.

Results23 are broadly in line with those generally found

in the literature on international health expenditure com-

parisons. Per capita expenditure is positively correlated

with income (even though the variable is not always sig-

nificant) as well as with life expectancy. As for labour

market indicators, we consider female employment rates,

since the more women are employed in the labour market,

the more often formal care substitutes informal care.

However, the employment rate turns out to be not statis-

tically significant.

The dependency ratios for children (0–4 years old) and

for the elderly (75 years of age or older) allow us to capture

the effect of the age structure of the population. The age

profile of per capita expenditure highlights how the

demand for healthcare is greater at very young ages and for

older age cohorts [24]. The results are partially unexpected:

while the elderly dependency ratio has a positive and sig-

nificant sign, the dependency ratio of the very young is not

significant. This finding may be driven by considering too

broad an age class (0–4 years of age) as the expenditure

profile is higher at younger ages, particularly in infancy or

early childhood [8].

16 In our analysis we do not include the latter variables as the regions

share a common institutional framework.17 The specification used reflects the nature of the regressors. All of

them, apart from per capita GDP, are ratios or dummies.18 An alternative estimation strategy would be to implement a

multilevel model that simultaneously considers both the first and the

second step. Even though this procedure could be more efficient, it

has two main drawbacks: (1) it requires stronger distributional

assumptions on the error terms; (2) it would make it less straight-

forward to work out an estimate of the inefficiency levels and

decompose and analyse them as reported in the following paragraph.

We have estimated such a model as a robustness check, however;

estimated coefficients of the variables from the first and the second

step and their significance are in line with those obtained following a

two-step procedure.19 Both the proxies for the health status and ‘bad’ habits are obtained

via a principal component analysis (see Table 2), which allows us to

reduce the dimension of the set of regressors and exploit the

correlation among the variables.20 In particular, the followings are taken into account: the share of

smokers (variable ‘smoke’), the share of the population drinking at

least half a litre of wine a day (variable ‘wine’) and the share of those

who drink alcohol at least twice a week (variable ‘alcohol’).

21 Including measures of development as demand factors may raise

some concerns about possible correlation between those variables and

regional supply-side characteristics, such as wage levels and the

quality of human capital. This problem however is quite minor in the

Italian setting. Indeed, (1) as concern wages, national labour contracts

have been used and are used in the Italian NHS so that there are no

systematic differences in this respect across the country; (2) as

regards human capital quality, hiring procedure are based on the same

requirements (in terms of medical school degrees, specialisation

patterns) across the whole country.22 In particular, we include a dummy variable equal to 1 for the years

after 1995.23 Since the modified Wald suggests the presence of heteroskedas-

ticity and the Wooldridge test for autocorrelation in panel data [35]

rejects the null hypothesis of no first order autocorrelation, the

standard error estimator used is robust to both hetoreskedasticity and

serial correlation.


123

The health status and bad habit indicators are both sig-

nificant and have the expected sign (‘poorer’ health and

‘worse’ habits are associated with higher expenditure).

The dummy variable aimed at capturing the general cost

containment effort exerted in the early 1990s is highly

significant and has the expected sign (the intercept of the

expenditure curve is shifted downwards after 1995).

Regional fixed effects are shown in the lower panel of

Table 3. They are strongly consistent among the three

specifications, with the correlations among the series

ranging from 0.95 to 0.99. The geographical pattern in

Fig. 2 shows the expected North-South differences.

First step: robustness checks

To check the robustness of our estimates we perform a

series of exercises described in Appendix 1. We re-estimate

the model using a stochastic frontier approach, considering

both time invariant and time varying specifications. All of

the estimated models give very similar results; the correlation

between the regional effects series ranges between 0.95 and 1.

The time-varying models, in which unexplained expenditure

is a function of time, indicate that, other things being equal,

differentials in spending between regions are increasing over

time, i.e. more efficient regions are also at an advantage to

contain expenditure growth (not only levels).

In order to address concerns that controls for economic

development could overcontrol for differences in the

quality of providers (instead of only accounting for demand

side characteristics), we also re-estimate the model

including only demographich variables, health status and

habits among regressors in the first step. Estimates (sign,

magnitude and significance of the coefficients) are in line

with those presented in Table 3 and discussed above.24

Table 1 Descriptive statistics

Variable Obs Mean SD Min Max

Real per capita health expenditure (corrected for mobility)a 294 1,222 203 884 1,810

Real per capita health expenditure (corrected for mobility and structure of the population)a 294 1,283 214 923 2,030

Real per capita GDPb 294 19,818 5,102 10,836 28,504

Dependency ratio for children (0–4)c 294 21.22 3.66 15.05 30.22

Dependency ratio for elderly (75?)c 294 12.36 3.02 6.38 21.08

Life expectancy (female)c 294 82.38 1.28 78.87 85.00

Bad habit

Alcohol 294 7.02 2.55 3.65 11.92

Wine 294 5.72 1.49 2.12 8.94

Smoke 294 23.50 2.47 17.33 30.68

Health statusc

High blood pressure 294 125.32 22.42 75.10 179.20

Heart attack 294 14.62 3.79 6.90 24.80

Angina 294 8.51 2.80 3.90 17.60

Other heart deseases 294 33.70 5.18 22.90 49.70

Diabetes 294 40.34 8.49 19.20 57.90

Ictus 294 11.43 2.69 6.80 18.50

Malignant cancer (males) 294 21.84 6.97 9.84 38.67

Malignant cancer (females) 294 26.25 8.61 11.71 47.75

Participation rate (female)d 294 40.81 11.15 18.82 61.54

Tertiary educationd 294 7.64 2.16 3.65 14.20

Share of employed in agriculture (%)d 294 7.28 4.05 1.53 21.43

Share of employed in manufacturing (%)d 294 29.56 6.96 17.96 43.97

a Computations on Ministero dell’Economia e finanze, Relazione Generale sulla Situazione Economica del Paese (RGSEP) (various years)b Istat, Conti Economici Regionali (2007); Istat, Conti Economici Regionali (2008)c Istat, Health for All (2009)d Data on labour market participation, employment rates and education levels of the population (broken down by region and sex) have been

computed using individual data provided by Istat, Rilevazione sulle Forze di lavoro (various years). We thank Federico Giorgi for having

provided such computations

24 Second step results (to be discussed in the sixth section) are

confirmed by using fixed effects obtained by this restricted model,

too. These robustenss check results (both for step 1 and step 2) are

available upon request.


123

The regional fixed effects series derived by this restricted

model is close to the one obtained in the general

specification.25

Following a similar approach to Schmidt and Sickles [32],

we approximate ‘waste’ (or unexplained differentials) for each

region with the estimated fixed effects, that is wi ¼ ai. In this

way, however, any time invariant heterogeneity between the

regions is also captured by the waste term (Greene [21] and

[22]). Even though we acknowledge this identification issue,

we also believe that it is not as relevant in this study as in other

applications. In particular, the common legal and institutional

framework and the shared rules set at the central level signif-

icantly mitigate the problem of heterogeneity among regions

potentially associated with institutional factors. Secondly, at

this stage we are interested in isolating unexplained (by

demand factors) differentials whatever the source might be

(time invariant heterogeneity or ‘technical’ inefficiency). In

the second step of the analysis we will try to see whether the

waste levels are correlated with indicators aimed at capturing

the quality of the public administration (supply side indicators)

or other structural regional characteristics.

First step: a disaggregated analysis by spending

category and the relation with quality

Before moving to the second step we estimate regional

fixed effects using (5) (according to the specification

adopted for Model 1) also breaking down the dependent

variable into its main sub-categories. We consider the

breakdown by economic category (in which we distinguish

Table 2 Principal component analysis

a: ‘Bad’ habit

Component Eigenvalue Proportion of total variance

1 1.349 0.450

2 1.148 0.383

3 0.503 0.168

Eigenvectors

Variable 1 2 3

Alcohol 0.709 -0.333 0.621

Wine 0.705 0.345 -0.620

Smoke -0.007 0.878 0.479

b: Health status

Component Eigenvalue Proportion of total variance

1 3.216 0.402

2 1.855 0.232

3 1.420 0.178

4 0.605 0.076

5 0.464 0.058

6 0.273 0.034

7 0.158 0.020

8 0.009 0.001

Eigenvectors

Variable 1 2 3 4 5 6 7 8

High blood pressure 0.492 -0.091 0.141 0.019 0.455 0.001 -0.722 -0.036

Heart attack 0.406 0.085 0.195 -0.587 -0.619 0.227 -0.101 0.012

Angina 0.086 0.569 -0.399 -0.002 0.247 0.665 0.063 0.047

Other heart diseases 0.243 0.569 -0.089 -0.287 0.205 -0.667 0.197 0.000

Diabetes 0.224 -0.008 0.715 0.024 0.349 0.229 0.514 -0.011

Ictus 0.310 0.351 0.156 0.749 -0.431 -0.079 -0.054 -0.013

Malignant cancer (males) 0.440 -0.340 -0.318 0.082 0.021 -0.039 0.260 0.717

Malignant cancer (females) 0.432 -0.312 -0.373 0.060 0.006 0.006 0.302 -0.694

25 The correlation among the two fixed effects series is 0.94.


123

Ta

ble

3S

tep

1—

Pan

eles

tim

atio

nw

ith

fix

edef

fect

s

(Mo

del

1)

(Mo

del

2)

(Mo

del

3)

Rea

lto

tal

exp

end

itu

re(c

orr

ecte

dfo

r

pat

ien

tm

ob

ilit

y)

Rea

lto

tal

exp

end

itu

re(c

orr

ecte

dfo

rp

atie

nt

mo

bil

ity

and

equ

ival

ent

po

pu

lati

on

a)

Rea

lto

tal

exp

end

itu

re(c

orr

ecte

dfo

rp

atie

nt

mo

bil

ity

and

equ

ival

ent

po

pu

lati

on

b)

Co

eff.

pv

alu

eC

oef

f.p

val

ue

Co

eff.

pv

alu

e

Per

cap

ita

GD

P0

.00

00

10

.24

80

.00

00

10

.43

60

.00

00

2*

0.0

78

Dep

end

ency

rati

ofo

rch

ild

ren

(0–

4)

-0

.00

64

50

.27

9

Dep

end

ency

rati

ofo

rel

der

ly(7

5?

)0

.04

33

7*

**

0.0

10

Lif

eex

pec

tan

cy(f

emal

e)0

.04

13

3*

**

0.0

03

0.0

54

61

**

*0

.00

00

.05

89

0*

**

0.0

00

Bad

hab

it0

.02

17

2*

*0

.01

50

.02

64

5*

**

0.0

02

0.0

14

79

*0

.06

6

Hea

lth

stat

us

0.0

17

63

**

*0

.00

80

.03

10

0*

**

0.0

00

0.0

22

98

**

*0

.00

0

Par

tici

pat

ion

rate

(fem

ale)

0.0

01

54

0.7

03

0.0

05

59

0.1

50

0.0

05

44

0.1

81

Ter

tiar

yed

uca

tio

n0

.00

31

70

.57

50

.01

23

40

.17

80

.01

39

30

.13

5

Sh

are

of

emp

l.in

agri

cult

ure

(%)

-0

.00

67

8*

0.0

89

-0

.00

98

6*

*0

.04

5-

0.0

11

21

**

0.0

20

Sh

are

of

emp

l.in

man

ufa

ctu

rin

g(%

)0

.00

12

70

.25

40

.00

21

1*

0.0

78

0.0

02

16

*0

.07

5

Du

mm

yfo

rth

ey

ears

afte

r1

99

5-

0.1

17

47

**

*0

.00

0-

0.0

96

03

**

*0

.00

0-

0.1

15

64

**

*0

.00

0

Co

nst

ant

3.1

12

31

**

*0

.00

32

.28

08

9*

*0

.01

21

.72

26

5*

0.0

51

No

.O

bse

rvat

ion

s2

94

29

42

94

F(1

1,

20

)1

15

.76

12

3.7

21

23

.96

Pro

b[

F0

.00

00

0.0

00

00

.00

00

R-s

qu

ared

0.9

30

40

.92

39

0.9

24

4

Ad

jR

-sq

uar

ed0

.92

22

0.9

15

60

.91

61

Ro

ot

MS

E0

.04

53

0.0

46

20

.04

77

Reg

ion

alfi

xed

effe

cts

(ai)

(Mo

del

1)

(Mo

del

2)

(Mo

del

3)

Reg

ion

sR

eal

tota

lex

pen

dit

ure

(co

rrec

ted

for

pat

ien

tm

ob

ilit

y)

Rea

lto

tal

exp

end

itu

re

(co

rrec

ted

for

pat

ien

tm

ob

ilit

y

and

equ

ival

ent

po

pu

lati

on

(1))

Rea

lto

tal

exp

end

itu

re

(co

rrec

ted

for

pat

ien

tm

ob

ilit

yan

d

equ

ival

ent

po

pu

lati

on

(2))

1P

iem

on

te(N

)-

0.1

44

-0

.17

0-

0.1

75

2V

alle

d’A

ost

a(N

)0

.06

50

.05

60

.00

7

3L

om

bar

dia

(N)

-0

.13

5-

0.2

30

-0

.28

6

4B

olz

ano

(N)

0.2

63

0.2

21

0.1

52

5T

ren

to(N

)-

0.0

34

-0

.07

1-

0.1

08


123

Ta

ble

3co

nti

nu

ed

Reg

ion

alfi

xed

effe

cts

(ai)

(Mo

del

1)

(Mo

del

2)

(Mo

del

3)

Reg

ion

sR

eal

tota

lex

pen

dit

ure

(co

rrec

ted

for

pat

ien

tm

ob

ilit

y)

Rea

lto

tal

exp

end

itu

re

(co

rrec

ted

for

pat

ien

tm

ob

ilit

y

and

equ

ival

ent

po

pu

lati

on

(1))

Rea

lto

tal

exp

end

itu

re

(co

rrec

ted

for

pat

ien

tm

ob

ilit

yan

d

equ

ival

ent

po

pu

lati

on

(2))

6V

enet

o(N

)-

0.1

31

-0

.19

9-

0.2

14

7F

riu

liV

enez

iaG

iuli

a(N

)-

0.2

51

-0

.25

2-

0.2

33

8L

igu

ria

(N)

-0

.27

0-

0.2

01

-0

.15

9

9E

mil

iaR

om

agn

a(N

)-

0.2

78

-0

.30

5-

0.3

03

10

To

scan

a(C

)-

0.2

60

-0

.25

4-

0.2

49

11

Um

bri

a(C

)-

0.2

45

-0

.26

1-

0.2

21

12

Mar

che

(C)

-0

.22

7-

0.2

59

-0

.24

2

13

Laz

io(C

)0

.10

40

.04

5-

0.0

00

14

Ab

ruzz

o(S

)-

0.0

34

-0

.01

8-

0.0

04

15

Mo

lise

(S)

0.0

17

0.0

79

0.1

23

16

Cam

pan

ia(S

)0

.41

20

.43

40

.43

7

17

Pu

gli

a(S

)0

.25

90

.30

40

.31

5

18

Bas

ilic

ata

(S)

0.1

39

0.1

88

0.2

23

19

Cal

abri

a(S

)0

.22

30

.25

90

.30

0

20

Sic

ilia

(S)

0.2

85

0.3

52

0.3

72

21

Sar

deg

na

(S)

0.2

43

0.2

80

0.2

66

*,

**

,*

**

Sig

nifi

can

ceat

10

,5

and

1%

lev

el(s

tan

dar

der

rors

rob

ust

toh

eter

osk

edas

tici

tyan

dau

toco

rrel

atio

n);

(N)

No

rth

,(C

)C

entr

e;(S

)S

ou

tha

Usi

ng

wei

gh

tsfo

rh

osp

ital

serv

ices

bU

sin

gw

eig

hts

for

ph

arm

aceu

tica

lco

nsu

mp

tio

n


123

between wages, purchases of goods and services, spending

for private hospitals recognised within the NHS, primary

care, pharmaceuticals, and specialised care).26

Given the estimates for each expenditure category we

can recover aggregate regional ‘waste levels’. Assuming

that each expenditure item h presents a relationship as in

(5) and using the same set of regressors, i.e. E chit

� �

¼ex0 itb

hþahþahi ; h ¼ 1; . . .;H (H being the total number of

expenditure items), we have that:

ex0 itbþaþai ¼X

H

h¼1

ex0 itbhþahþah

i ð6Þ

From this it follows that:

eai�aimin

|fflfflffl{zfflfflffl}

ni

¼X

H

h¼1

ex0itbhþah

ex0itbþa

|fflfflffl{zfflfflffl}

shi

eahi

eaimin|ffl{zffl}

~nhi

ð7Þ

The distance to the minimum cost for each region i can be

represented as a weighted average of the inefficiency scores

for each expenditure item ð~nhi Þ, with the latter normalised

with respect to the minimum total regional effect. Estimated

fixed effects obtained by aggregating the results for the

expenditure items are in line with those obtained estimating

aggregate expenditure.27 Figure 3a shows for each region the

contribution of each expenditure item to the difference

between its own inefficiency score and the minimum one

Di ¼ ðni � nminÞ.28 Figure 3b instead shows the contribution

of each expenditure item to the difference between the actual

inefficiency score and the score that would have occurred in

region i had all the ~nhi been equal to 1.29

This decomposition allows us to point out some interesting

patterns. Low levels of unexplained expenditure (characterised

by Di equal or close to 0) seem compatible with different

choices as regards spending composition. Furthermore, in

regions with higher spending the distribution of inefficiency

among expenditure items does not seem to be uniform. In

particular, in the regions with a positive total fixed effect

(meaning, ceteris paribus, higher than average unexplained

expenditure), the management of the outlays that account for

relatively smaller shares of the budget seems in need of par-

ticularly careful monitoring. This is evident if we consider the

distribution of the inefficiency among expenditure items

computing #hi ¼

~Dhi

Di� sh

i , an indicator that compares the share

of the deviation from minimum spending because of a given

expenditure item and the reference share of that same item in

total outlays (Fig. 3c).30 For the southern regions with a

positive total fixed effect, the distortion in spending on phar-

maceuticals is particularly relevant, while the weight of the

inefficiency in hospital spending is less pronounced than

implied by their reference share of total spending.31 This might

from -0.28 to -0.25from -0.25 to -0.03from -0.03 to 0.14from 0.14 a 0.41

Fig. 2 Regional fixed effects—geographic pattern. Data are taken

from Table 3, lower panel, Model 1; for Trentino Alto Adige an

average of the estimated values for Trento and Bolzano is used

26 As to the magnitude of the expenditure categories considered, it

should be noted that in 2005, for the Italian average, wages accounted

for 33 % of total expenditure, the purchase of goods and services for

27 %, primary care for 6 %, pharmaceuticals for 13 %, private

hospital care for 9 % and specialised care for 4 %.27 Given (7), regional effects can be obtained by aggregating the

results of the estimation of the expenditure items equations as:

aidisag ¼ lnP

H

h¼1

shi eah

i

� �

. The correlation between ai and aidisag is 0.99.

28 The difference can be decomposed as follows: Di ¼ ni � nmin ¼

P

H

h¼1

ðjhi � jh

iminÞ ¼P

H

h¼1

Dhi where jh

i ¼ shi~nh

i is the contribution of each

expenditure item and variables with the subscipt min refer to the

region with the lowest ai. Values shown in Fig. 3a are computed

considering the average values of shi over the estimated period.

29 If ~nhi ¼ 18h (the term of comparison this time is the case in which

inefficiency is uniformly distributed among the expenditure items and

equal to aggregate minimum ai min) then jhi ¼ sh

i . It follows that the

described difference can be decomposed as: Di ¼ ni � niref ¼P

H

h¼1

ðjhi � sh

i Þ ¼P

H

h¼1

~Dhi .

30 Taking a region i, values of #hi

�

�

�

� that are close to zero for every h

suggest that in region i the level of inefficiency is similar among the

various expenditure items. A large value of #hi for a given h implies a

particularly high value of ~nhi for that item with respect to the others.

An analytical discussion of indicator #hi is provided in Appendix 2.

31 This means that for the regions under consideration the ineffi-

ciency level for this item, although large, is less pronounced than the

inefficiency in other spending categories.


123

reflect several features, such as national guidelines and the

application of national labour contracts for the employees of

the NHS. The breakdown suggests that in some regions, in

particular Lazio, the level of distortion in reimbursements for

services provided by private hospitals plays a non-negligible

role in overall spending overruns. As regards the distortion in

the purchases of goods and services, it seems in line with

average inefficiency levels in each of the regions. For this item

the large impact on overall unexplained spending seems to

reflect mainly its large incidence on total expenditure. It has to

be noted, however, that regions with lower health spending

seem to have been particularly effective in containing outlays

in this category.

Given the identification problem discussed above,

before moving on to disentangle some of the characteristics

that might be related to the regional fixed effects, it is

worth checking whether the differentials isolated in our

first step reflect differences in the quality of services pro-

vided to the public. To do that we consider two types of

indicators: (1) the mobility of patients between regions; (2)

declared patient satisfaction.

If we look at patient mobility between regions32 as an

indirect indicator of the quality of health services (or at

least of the degree of quality ‘perceived’ by patients), we

observe that patients consistently tend to move from the

South towards the Centre and North of Italy. The correla-

tion between the regional fixed effects and the average

attraction index observed over the period considered is

highly negative (-0.64) and significant so that the higher

the ‘perceived’ quality, the lower the ‘unexplained’

spending. Similarly, indicators that take into account

declared patient satisfaction33 for hospital services display

a clear North-South pattern with patients being more sat-

isfied with hospital services in the North. In particular, the

correlation between average patient satisfaction indices for

nurses, physicians and sanitary services and the estimated

regional effects varies between -0.48 and -0.31. Finally,

the complexity of the treatments is also on average higher

in the regions characterised by lower regional effects.34

In brief, the higher spending in the southern regions

compared with the Centre and North does not seem to be

ascribable to perceived or reported higher quality. So, what

causes the difference? Are there any factors that policy

makers could control for in order to cut spending without

reducing the amount of services provided?

-0.1

0.1

0.3

0.5

0.7

0.9

wages goods & services private hospitals

primary care pharmaceuticals specialised care

-0.2

0.0

0.2

0.4

0.6

0.8



(c)

-0.5

-0.3

-0.1

0.1

0.3

0.5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21

2 4 13 15 16 17 18 19 20 21



(b)

(a)

Fig. 3 Regional fixed effects a breakdown of deviation with respect

to the region with lowest unexplained expenditure; b breakdown of

deviation with respect to the reference expenditure levels and

composition; c hih: share of deviation from minimum spending due

to a given item (h) minus its weight on total outlays (only regions with

positive fixed effects, i.e. higer than average unexplained expenditure)

32 In particular, we consider the attraction index, measured as the

ratio between the inflow and outflow of patients into a given region

(data from Ministero della Salute [26]). An index value[1 means that

the inflow of patients from other regions is greater than the outflow to

other regions and hence the region attracts patients from other areas of

the country.33 The satisfaction with hospital services is drawn from surveys

included in the Indagine multiscopo regularly performed by Istat.34 The correlation between the average case mix (measured as the

standard hospital stay multiplied by the normalised applicable DRG)

over the observed period and regional fixed effects is -0.89.


123

The empirical analysis—second step: the drivers

of regional differentials

The second step of our analysis tries to address this issue.

We investigate the correlation between regional effects and

some supply factors and regional characteristics that could

affect the unexplained variability observed in spending

levels. Italian regions, which operate in a context of

administrative autonomy, have the possibility to choose

different organisational arrangements within a common

legal and institutional framework that mandates them to

provide homogeneous healthcare coverage.35

The analysis exploits the result of the expenditure levels

estimation in step one. In particular, we consider the

covariates of the regional fixed effects estimating the fol-

lowing linear model:36

ai ¼ q0idþ fi i ¼ 1; . . .;N ð8Þ

where fi is the standard i.i.d. error term and qi is a

vector of j regressors and includes indicators for the

following categories: (1) appropriateness of treatments;37

(2) supply structure characteristics;38 (3) quality in public

administration39; (4) social capital.40 In order to

corroborate the idea that variables used in the second

stage reflect supply side characteristics, and not a com-

bination of supply and demand factors, we checked

whether there is any correlation with demand side driv-

ers (such as population age structure and health status,

using the same proxies employed in step 1 of our

analysis). Results suggest that no significant correlation

can be found.

Given the small number of observations, we started by

considering one (or at most two) regressor for each cate-

gory: the index for Caesarean sections, the index for total

employment and its composition; the indicator for general

practitioners with a large number of patients; the ratio

between the number of beds in private and public hospitals;

the share of the use of generic pharmaceuticals and proxies

for social capital (Table 4). To check the robustness of the

results, we also considered each category at a time and

substituted the proxy/ies included in the estimated speci-

fication. Table 4 also shows the results of a SURE model

having as dependent variables the regional effects for the

categories of hospital spending, primary care, outlays for

pharmaceuticals and specialised care.41

The appropriateness index turns out to be significantly

related to unexplained spending differentials; the higher the

inappropriateness is, the higher the regional effects. Inap-

propriateness characterises healthcare treatments that could

be performed ensuring at least the same effectiveness for

the patient, but incurring lower risks and/or employing a

lower amount of resources. Following the literature and

policy practice (see OECD [30], WHO [19], Francese et al.

[16] and references therein) as an indicator for appropri-

ateness, we have used Caesarean section rates at the

regional level. Results are confirmed even when we sub-

stitute Caesarean sections with inappropriate surgical ward

use index.42

As to supply indicators, we include both indexes related

to hospital organisation and to primary care, which are

directly established by policy makers according to national

and regional rules. As concerns the former, we find that a

35 Compensation of employees represents an exception; wages are set

by national labour contracts (however regions can decide on the

number and composition of employees).36 This is consistent with the assumption of regional effects being

time invariant. As seen above, the time invariant component

estimated using time varying frontier models or fixed effects gives

equivalent results. Apart from the common impact due to the

decaying factor, estimated waste levels are strongly correlated using

the four estimation procedures discussed in Appendix 1.37 The indicators considered under this category are the average over

the observed period of the incidence of Caesarian sections and of the

incidence of patients discharged by a surgical ward with a medical

DRG (data from Ministero della Salute [26]).38 In this category we considered the averages of the incidence of

NHS employees on residents; composition of employees (ratio of

medical staff over total employees); incidences of public hospital

beds on residents; incidences of private hospital beds; ratio between

number of beds in private to public hospitals; share of expenditure for

care provided by private hospitals; average stay in public hospitals;

average stay in private hospitals; number of general pratictioners;

number of patients per general pratictioner; number of general

pratictioners with more than 1,500 patients. All the data are drawn

from ISTAT, Health for All.39 To proxy the quality of administration we consider: the share of

pharmaceutical expenditure for generic (off-patent) pharmaceuticals

and the share of expenditure for pharmaceuticals delivered directly by

the NHS in total pharmaceutical expenditure. We also consider an

indicator for the use of IT in other Italian local governments in each

region; in particular, we use a municipality IT index, which is given

by the share of municipalities whose general registry office has been

computerised.40 In particular, we consider several proxies for social capital

commonly used in the literature: incidence of blood donors per

1,000 residents; incidence of recyclable waste collection; average

turnout at referenda; interest in politics and morality. The indicators

Footnote 40 continued

for solidarity/‘morality’ and participation/‘interest in politics’ are

drawn from Giordano et al. [20]. In particular, solidarity/‘morality’ is

a composite index summarising self-reported pro-social attitudes and

objective measures of altruistic behaviour (such as blood donation);

similarly, interest in politics is built from self-reported answers and

more objective measures (such as referendum participation).41 All the estimated models exclude the constant term. In fact, the

mean value of the fixed effects is zero by construction. To check for

this, however, we also replicate the estimation including the constant

term, which turns out to be not significant. Moreover, neither the sign

nor the magnitude of the estimated coefficients are affected by

including/excluding the constant.42 In this case, however, it interferes with the significance of the

number of employees indicator.


123

Ta

ble

4S

tep

2—

Dri

ver

so

fre

gio

nal

effe

cts

Reg

ion

alfi

xed

effe

cta

To

tal

exp

end

itu

re

Co

eff.

pv

alu

e

%C

aesa

rian

sect

ion

s0

.01

79

**

*0

.00

0

Inci

den

ceo

fN

HS

emp

l.o

nre

sid

ents

0.0

48

5*

**

0.0

02

Co

mp

osi

tio

no

fem

pl.

(med

ical

staf

f/em

pl.

)-

0.0

06

60

.23

2

GP

sw

ith

mo

reth

an1

50

0p

atie

nts

0.0

08

8*

**

0.0

00

#b

eds

inp

riv

ate/

#b

edin

pu

bli

ch

osp

ital

s0

.00

21

0.1

37

%g

ener

icd

rug

s-

0.0

54

5*

**

0.0

04

Inte

rest

inp

oli

tics

-0

.01

55

**

*0

.00

1

So

lid

arit

y0

.00

81

**

0.0

18

No

.O

bse

rvat

ion

s2

1

F(8

,1

3)

22

.07

Pro

b[

F0

.00

00

q y,y

20

.93

32

Reg

ion

alfi

xed

effe

ctb

Ho

spit

alex

pen

dit

ure

Pri

mar

yca

reP

har

mac

euti

cals

Sp

ecia

lise

dca

re

Co

eff.

pv

alu

eC

oef

f.p

val

ue

Co

eff.

pv

alu

eC

oef

f.p

val

ue

%C

aesa

rian

sect

ion

s0

.01

02

**

*0

.00

00

.00

17

**

*0

.00

00

.00

33

**

*0

.00

00

.00

10

**

0.0

27

Inci

den

ceo

fN

HS

emp

l.o

nre

sid

ents

0.0

48

9*

**

0.0

00

0.0

05

1*

**

0.0

00

0.0

04

6*

0.0

61

-0

.00

07

0.7

31

Co

mp

osi

tio

no

fem

pl.

(med

ical

staf

f/em

pl.

)0

.00

43

0.2

52

0.0

00

00

.95

40

.00

17

*0

.09

20

.00

01

0.9

18

GP

sw

ith

mo

reth

an1

50

0p

atie

nts

0.0

07

6*

**

0.0

00

0.0

00

4*

**

0.0

03

0.0

00

5*

*0

.03

50

.00

01

0.5

04

#b

eds

inp

riv

ate/

#b

edin

pu

bli

ch

osp

ital

s0

.00

16

*0

.09

50

.00

03

**

0.0

20

0.0

00

6*

*0

.02

90

.00

03

0.1

24

%g

ener

icd

rug

s-

0.0

39

9*

**

0.0

01

-0

.00

17

0.3

16

-0

.00

62

**

0.0

44

-0

.00

24

0.3

34

Inte

rest

inp

oli

tics

-0

.01

14

**

*0

.00

0-

0.0

01

4*

**

0.0

00

-0

.00

27

**

*0

.00

0-

0.0

00

80

.15

6

So

lid

arit

y0

.00

79

**

*0

.00

00

.00

05

*0

.10

00

.00

12

**

0.0

48

0.0

00

9*

0.0

58

No

.O

bse

rvat

ion

s2

12

12

12

1

F(8

,1

3)

55

9.1

41

84

.66

24

4.2

20

.15

Pro

b[

F0

.00

00

0.0

00

00

.00

00

0.0

00

0

q y,y

20

.99

71

0.9

91

30

.99

34

0.9

25

4

*,

**

,*

**

Sig

nifi

can

ceat

resp

ecti

vel

y1

0,

5an

d1

%le

vel

aO

rdin

ary

leas

tsq

uar

eb

Max

imu

mli

kel

iho

od

seem

ing

lyu

nre

late

dre

gre

ssio

n.

Sm

all

sam

ple

stat

isti

csar

eco

mp

ute

d


123

higher number of employees is related to higher spending,

as expected. Instead, the coefficient of the proxy for the

composition of the staff is not significant. Proxies (either

the ratio of beds in private to public hospitals or the share

of spending for private hospitals in total spending) for the

relevance of the private sector display a positive but not

significant coefficient in the regression for aggregate

regional effects; the coefficient is significant, however, in

the estimation of spending categories, in particular hospi-

tals, primary care and pharmaceuticals.

As regards primary care, the workload of general prac-

titioners (GPs) is significant; in particular, the share of

practitioners with a very large number of patients is related

to higher spending (the same happens when considering the

average number of patients for GP). Interestingly, the

workload of GPs also affects other spending items (besides

primary care); in fact, it is found that doctors workload has

a strong positive impact on hospital and pharmaceutical

spending as well as spending for primary care. This sug-

gests that too many patients to follow might make it more

difficult for general practitioners to avoid unnecessary

access to hospital services and use of pharmaceuticals.

Furthermore, it highlights that the overall organisation of

the system matters, since outlays in different categories are

related. As noted above when discussing first step results,

different expenditure compositions seem to be compatible

with overall efficiency.

To proxy the quality of administration we consider

indicators of practices that are aimed at reducing the

resources absorbed by consumption of pharmaceutical

without reducing the quantity provided or the effectiveness

of the treatments. These variables are intended to capture

good management by public authorities since they reflect

attention in providing effective services at reduced costs

(by reducing, for example, transaction costs or those

associated to branded drugs). In particular, we focus on the

share of pharmaceutical expenditure for generic (off-pat-

ent) drugs and the share of expenditure for pharmaceuticals

delivered directly by the NHS in total pharmaceutical

expenditure. As expected, indicators for good practices (we

considered alternatively the two indexes) do display a

negative relation to unexplained spending.

Finally, we also included in our model proxies for

informal rules and habits that might play a role in the

performances of different public administrations. The

economic literature often highlights the impact of history

and economic institutions on the development of different

regions. Institutions also include informal rules and con-

straints that are often summarised in the concept of social

capital.43 In this context, social capital captures the level of

trust and habits of cooperation shared among members of a

local community. As to our results, social capital indicators

turn out to be significant when splitting the two aspects of

solidarity/‘morality’ and participation/‘interest in politics’,

the first one being positively related to spending—probably

reflecting a higher propensity to share fundamental ser-

vices–the second one being negatively related to regional

effects—suggesting that the more public attention is given

to policy makers’ activity, the stronger their incentive to

manage the available resources efficiently.

Although the second step of the empirical analysis is

based on a small sample size, the correlations we detect are

sufficiently robust to derive some general policy consid-

erations. Regional spending differentials seem to be related

to both persistent (sticky?) regional features (such as social

capital) and characteristics that in the medium term can be

controlled by policy makers (such as the appropriateness of

health treatments, the workload of general practitioners or

the use of generics). Spreading administrative best prac-

tices and enhancing the degree of appropriateness of

treatments would be an effective way of reducing spending

differentials and increasing efficiency in the use of public

funds. Even relatively limited improvements could turn

into substantial savings: given our results, an increase in

the appropriateness level or in the use of generics by 1

percentage point44 implies on average a reduction of

respectively 1.8 and 5.3 % in total expenditure (€20.9 and

€62.3 per capita).

Concluding remarks

Several studies have documented the presence of regional

differentials in public health spending. The aim of this

work is to contribute to this discussion by analysing the

determinants of per capita health expenditure focussing on

identifying the drivers of inefficiency for the panel of

Italian regions and the existence of potential room for

savings. The article presents an empirical analysis of per

capita expenditure levels that takes into account the

regional differences in the need for health services. The

remaining differences in spending (the ‘‘unexplained dif-

ferentials’’) are then studied in relation to structural char-

acteristics and policy variables.

First, the results confirm the existence of potential

margins for savings that are related at least in part to dif-

ferences in the effectiveness of public administration.

However, with respect to previous studies, our methodol-

ogy not only allows us to evaluate the overall inefficiency

43 See Putnam [31], de Blasio and Nuzzo [9].

44 These computations use the result of the estimation of Model 1 in

Table 3 and the coefficents reported in the first column of Table 4.

Values are computed using variable averages over the regions and the

years 1993–2006.


123

level for every region, but also to estimate the contribution

of each spending item (e.g. wages, pharmaceuticals, etc.) to

that inefficiency. For all the regions, the analysis further

breaks it down into two factors: the first accounts for the

budget share of each spending category; the second cap-

tures its relative inefficiency. In our framework it is thus

possible to single out the categories that need careful

monitoring in the regions with poorer performances, such

as outlays for pharmaceuticals.

The second step of the study suggests that there are

policy tools (e.g. the use of generic drugs or the workload

of general practitioners) that might help to keep ineffi-

ciency under control. Moreover, some supply structure

indicators and proxies for good practices seem to be par-

ticularly important for certain spending categories.

Acknowledgments We thank two anonymous referees for their

comments, the discussants and the participants at conferences and

seminars where previous versions of this work were presented, and

the colleagues at the Public Finance Division of the Bank of Italy. The

usual disclaimers apply.

Appendix 1: Robustness checks

To check the robustness of our estimates, we run a series of

exercises. We estimate model 1 using alternative estima-

tion techniques. We start by considering a stochastic

frontier model45 of the type:

cit ¼ f ðxitÞnit ð9Þ

where nit ¼ 1 for the best unit in the sample and nit [ 1 for

the others.

We specify the efficient expenditure function as

f ðxitÞ ¼ econsþx0 itb ð10Þ

and hence we can write the estimating equation as

ln cit ¼ consþ x0itbþ uit þ mit ð11Þ

where mit iid Nð0; r2mÞ is the error term and

uit iid Nþðl; r2uÞ captures each region’s distance with

respect to the efficient frontier46:

nit ¼ euit ð12Þ

Initially, we assume that the stochastic frontier is invariant

over time.47 We then consider a time varying version a la

Battese and Coelli [1]:

uit ¼ e�gðt�TÞui with uitiidNþðl; r2uÞ ð13Þ

where the distance from minimum cost can be decomposed

into two factors, a given level of inefficiency ui charac-

terising each region and a time varying component, which

is a function of time and a decay factor g.

We also estimate a fixed effect model with time varying

regional effects. In particular, we derive an estimating

equation equivalent to the time varying stochastic frontier

described above (13). In this case the expenditure equation

becomes48:

ln cit ¼ x0itbþ e�gt/i þ eit with

/i ¼ ½egTai� ð14Þ

Using the approximation

e�gt ffi 1� gt þ g2 t2

2ð15Þ

we derive the following estimating equation49:

ln cit ¼ x0itbþ /i � gt/i þ g2 t2/i

2þ eit ð16Þ

From the estimates of (16) we can recover the parameters we

are interested in: the regional (time invariant) effects ai and

the decay factor g. In both time-varying models the estimated

decay factor is negative and significant. This might reflect a

common increasing pattern in health expenditure that could

be due to the impact of technological developments in the

production of healthcare, in the organisation at national level

of the health system or in the structure of preferences.

Appendix 2: Indicator for the distribution

of inefficiency among expenditure items

The indicator #hi can be rewritten as follows:

hhi ¼

shi~nh

i � shi

ni � niref

� shi or

hhi ¼ sh

i

~nhi � ni

ni � niref

ð17Þ

45 A drawback of this approach, however, is that using stochastic

frontiers requires explicit assumptions in terms of the distributions

involved.46 In general, model (11) can be related to model (5). In particular,

rearranging the estimating equation for model (5) with regional fixed

effects, we have that: ln rexp pcitð Þ ¼ aþmin aið Þ½ � þa�min aið Þ½ � þ x0itbþ eit with 0� ai �minðaiÞ ¼ ainorm� þ1,

which can be directly compared to (11) where 0� uit � þ1captures the inefficiency levels.

47 So that in terms of model (5) and rearranging the time invariant

fixed effects, the estimation of ai-min(ai) is analogous to the

estimation of ui.48 In terms of rearranging model (5) it is like writing: aitnorm ¼e�gðt�TÞainorm ¼ e�gðt�TÞai � e�gðt�TÞðminðaiÞÞ with ainorm ¼ 0 for the

best unit in the sample.49 This specification is analagous to those discussed in Cornwell et al.

[7] and Kumbhakar [25], where the authors suggest substituting a

constant inefficiency parameter with a quadratic specification as in

(16).


123

when ~nhi ¼ ni 8h then hh

i ¼ 0 8h. All the items have the

same level of inefficiency. The distortion might be very big

but it is uniformly distributed among the expenditure

items. On the other hand, consider a case in which all

the first h–1 items are characterised by ‘full efficiency’ so

that ~nhi ¼ 1 8h ¼ 1; . . .;H � 1. In this case with

niref ¼P

H

h¼1

shi ¼ 1, we will have hh

i ¼ �shi 8h ¼ 1; . . .;

H � 1 while the overall inefficiency will reflect the

distortion in the last expenditure item: hHi will be

[ 0 [ hhi 8h ¼ 1; . . .;H � 1. Given (7), we will in fact

have that: ni ¼P

H�1

h¼1

shi

�

þ sHi

~nHi . Finally it should be

noted that for items characterised by ~nhi [ ni; h

hi will be

positive, while for ~nhi \ni; h

hi will be negative. This means

that a large distortion in a relatively small spending item

might be compensated by a relatively smaller distortion in

a spending item that accounts for a large fraction of total

outlays.

References

1. Battese, G., Coelli, T.: Frontier production functions, technical

efficiency and panel data: with application to paddy farmers in

India. J. Prod. Anal. 3, 153–169 (1992)

2. Bodenheimer, T., Berry-Millet, R.: Follow the money—control-

ling expenditure by improving care for patients needing costly

services. N. Engl. J. Med. 361, 1521–1523 (2009)

3. Bordignon, M., Turati, G.: Bailing out expectations and public

health expenditure. J.Health Econ. 28, 305–321 (2009)

4. Cantanero Prieto D., Lago-Penas S.: Decomposing the determi-

nants of health care expenditure: the case of Spain. Eur. J. Health

Econ. 13, 19–27 (2012)

5. Caroppo, M.S., Turati, G.: I sistemi sanitari regionali in Italia.

Milano, Vita e Pensiero (2007)

6. Congressional Budget Office: Geographic variation in health care

spending. CBO Paper 2978. Washington DC. (2008)

7. Cornwell, C., Schmidt, P., Sickles, R.: Production frontiers with

cross sectional and time series variation in efficiency levels.

J. Econom. 46, 185–200 (1990)

8. Cutler D. M., Meara E.: The medical costs of the young and the

old: a forty year perspective, NBER Working Papers 6114.

Cambridge (MA). (1997)

9. de Blasio G., Nuzzo G.: The legacy of history for economic

development: the case of Putnam’s social capital. Temi di Dis-

cussione, 591. Banca d’Italia, Roma. (2006)

10. De Matteis, P., Messina, G.: Le capacita fiscali delle Regioni

italiane. Rivista economica del Mezzogiorno 3, 363–396 (2010)

11. Economic Policy Committee, European Commission: The 2009

ageing report: economic and budgetary projections for the EU-27

member states (2008–2060). Eur. Econ., 2 (2009)

12. Fisher, E.S., Bynum, J.P., Skinner, J.S.: Slowing the growth of

healthcare costs—lessons from regional variation. N. Engl.

J. Med. 360, 849–852 (2009)

13. Fisher, E.S., Wennberg, D.E., Stukel, T.A., Gottlieb, D.J., Lucas,

F.L., Pinder, E.L.: The implication of regional variations in

Medicare spending. Part 1: the content, quality, and accessibility

of care. Ann. Intern. Med. 138(4), 273–311 (2003)

14. Fisher, E.S., Wennberg, D.E., Stukel, T.A., Gottlieb, D.J., Lucas,

F.L., Pinder, E.L.: The implication of regional variations in

Medicare spending. Part 2: health outcomes and satisfaction with

care. Ann. Intern. Med. 138(4), 288–322 (2003)

15. Fowler, F.J., Gallagher, P.M., Anthony, D.L., Larsen, K., Skin-

ner, J.S.: Relationship between regional per capita medicare

expenditures and patient perceptions of quality of care. J. Am.

Med. Assoc. 299, 2406–2412 (2008)

16. Francese M., Piacenza M., Romanelli M., Turati G.: Under-

standing inaprropriateness in health care. The role of supply

structure, pricing policies and political institurions in caesarean

deliveries. Working paper, 1. Department of Economics and

Statistics, Universita degli Studi di Torino, Torino (2012)

17. Francese M., Romanelli M.: Healthcare in Italy: expenditure

determinants and regional differentials. Temi di Discussione,

828. Banca d’Italia, Roma (2011)

18. Gerdtham, U.G., Jonsson, B.: International comparisons of health

expenditure: theory, data and econometric analysis. Handbook of

Health Economics. Elsevier, Amsterdam (2000)

19. Gibbons L., Belizan J. M., Lauer J. A., Betran A. P., Merialdi M.,

Althabe F.: The global numbers and costs of additionally needed

and unnecessary caesarean sections performed per year: Overuse

as a barrier to universal coverage. World health report back-

ground paper 30 (2010)

20. Giordano R., Tommasino P., Casiraghi M.: Behind public sector

efficiency: the role of culture and institutions, European Com-

mission Occasional paper, 45. Brussels (2009)

21. Greene, W.H.: Distinguishing between heterogeneity and ineffi-

ciency: stochastic frontier analysis of the World Health Organi-

zation’s panel data on national healthcare systems. Health Econ.

13, 959–980 (2004)

22. Greene, W.H.: The econometric approach to efficiency analysis.

In: Fried, H., Lovell, K., Schmidt, S. (eds.) The measurement of

efficiency, pp. 92–250. Oxford University Press, Oxford (2008)

23. Iuzzolino G.: Domanda e offerta di servizi ospedalieri. Tendenze

internazionali, Quaderni di Economia e Finanza (Occasional

Papers), 27. Banca d’Italia, Roma (2008)

24. Jacobzone S.: Healthy ageing and the challenges of new technolo-

gies. Can OECD social and health-care systems provide for the

future? In: OECD (ed.) Biotechnology and Healthy Ageing, Policy

Implications of New Research, pp. 37–53. OECD, Paris (2002)

25. Kumbhakar, S.: Production frontiers and panel data, and time

varying technical inefficiency. J. Econom. 46, 201–211 (1990)

26. Ministero della Salute: Rapporto sull’attivita di ricovero ospe-

daliero. Roma (various years)

27. Ministero della Salute: Rapporto nazionale di monitoraggio dei

livelli essenziali di assistenza-anno 2004. Roma (2007)

28. Newhouse, J.P.: Medical care expenditure: a cross-national sur-

vey. J. Human Resour. 12, 115–125 (1977)

29. Nixon, J., Ulmann, P.: The relationship between health care expen-

diture and health outcomes. Eur. J. Health Econ. 7, 7–18 (2006)

30. OECD: Caesarean sections. In: OECD: Health at a Glance 2009:

OECD Indicators. OECD Publishing, Paris (2009)

31. Putnam, R.: Making democracy work: civic tradition in modern

Italy. Princeton University Press, Princeton (1993)

32. Schmidt, P., Sickles, R.: Production frontiers with panel data.

J. Business Econ. Stat. 2(4), 367–374 (1984)

33. Skinner J. S., Fisher E. S., Wenneberg J.: The efficiency of

Medicare. NBER Working Paper 8395. Cambridge (2001)

34. Yasaitis, L., Fisher, E.S., Skinner, J.S., Chandra, A.: Hospital

quality and intensity of spending: is there an association? Health

Aff. 28(4), 566–572 (2009)

35. Wooldridge, J.M.: Econometric analysis of cross section and

panel data. MIT Press, Cambridge (2002)


123

Documents

Is there room for containing healthcare costs? An analysis of regional spending differentials in Italy