what does the world spend on criminal · pdf filewhat does the world spend on criminal...

No. 20 2004

WHAT DOES THE WORLD SPEND ON CRIMINALJUSTICE?

Graham Farrell and Ken Clark

WHAT DOES THE WORLD SPEND ON CRIMINAL JUSTICE?

Graham Farrell and Ken Clark

Graham Farrell is Professor of Criminology and Director of the Midlands Centrefor Criminology at Loughborough University.

Ken Clark is Senior Lecturer in the School of Economic Sciences at Manchester University.

The European Institute for Crime Prevention and Control,affiliated with the United Nations

Helsinki, 2004

HEUNI Paper No. 20

This document is available electronically from:http://www.heuni.fi

HEUNIThe European Institute for Crime Prevention and Control,affiliated with the United NationsP.O.Box 444FIN-00531 HelsinkiFinlandTel: +358-103665280Fax: +358-103665290e-mail:[email protected]://www.heuni.fi

ISSN 1236-8245

Contents

Acknowledgements.......................................................................................... 7Summary ........................................................................................................... 9Introduction .....................................................................................................112. Data Sources and Preparation .................................................................. 122.1 The UN Crime Survey ............................................................................................................ 122.2 Sampling and Representativeness of Countries .................................................................... 142.3 Other data sources ................................................................................................................. 143. Analysis ....................................................................................................... 154. Discussion .................................................................................................. 205. Conclusion .................................................................................................. 22References ...................................................................................................... 22TECHNICAL APPENDIX.................................................................................. 24Supplementary Data Sources........................................................................ 26Notes ............................................................................................................... 26

Acknowledgements

We thank Daniela Peterka for assistance with data collation and cleaning, and Brandon Webster for assistancein identifying related literature. For advice and comments on earlier drafts we are grateful to Mitchell Chamlin,John Eck, Ken Pease, John Roman, and the editor of this series, Kauko Aromaa. Some of the preparatorywork for this paper was undertaken while Ken Clark was a visiting professor of the Department of Economicsat the University of Cincinnati - we thank Professor Nicolas Williams for his guidance and hospitality infacilitating that stay.

7

9HEUNI Paper No. 20

Summary

This study produces an estimate of the global costs of public expenditure on criminal justice. The globalestimate is an extrapolation from data provided by the governments of seventy countries. At the country-levelthere is a strong relationship between the level of available public money and expenditure upon public policing,courts, prosecution and prisons. The relationship is explored using six regression models, and criminal justiceexpenditure in other countries is estimated using the best models. Global criminal justice expenditure in 1997is estimated at $360 billion (the equivalent of $424 billion in 2004 prices), of which 62 percent was spent onpolicing, 3 percent on prosecutions, 18 percent on courts, and 17 percent on prisons. As the first systematicempirical estimate of global criminal justice expenditure, it is hoped that the present research may spur betterdata collection practices and further research.

10HEUNI Paper No. 20


Knowledge relating to the overall well-being of theplanet, and cross-national variation within it, isincreasingly used to inform national andinternational policy. Some global indicators arecounts, their attraction being that they werepreviously unknown: In 1999, the earth’s humanpopulation exceeded six billion for the first time - apsychological watershed if quantity measureshumanity’s success. Countries are frequently rankedon key socio-economic indicators such as literacy,mortality and gross domestic product. Compositeindicators such as the Human Development Index(see e.g. United Nations Development Programme2002) and the Corruption Perceptions Index(Transparency International 2002) are derived. Suchindicators serve as benchmarks against which tomeasure variations in time and place. Their politicalimportance is such that they can provokeinvestigation of good and bad practice in countriesthat are statistically deviant. Yet, while the globalvillage requires global indicators, they are relativelyfew and far between, and still often at a fledglingstage of development. Further, while there is agrowing body of research highlighting theimportance of estimating the costs of crime (seeCohen 2000 for a review), there is little, if any, onaspects of the global cost of criminal justice. It is inthis context that the present study produces whatwe believe to be the first systematic empiricalestimate of the direct cost of public expenditure onthe global criminal justice system.

Previous research has identified a strongrelationship between a country’s economic welfare,measured as GDP, and its expenditure on criminaljustice (Newman and Howard 1999). This studycontinues work begun in Farrell et al. (2001) whichexamined only expenditure on policing. Notsurprisingly, on average, richer countries spend moreper capita on criminal justice than poor countries.In this study, the relationship between GDP andspending on criminal justice is examined using datafor seventy countries. To examine the extent to whichGDP can be used to predict criminal justiceexpenditure, six regression models are developed.The ‘best’ model of the relationship is identified foreach of policing, prosecution, courts and prisons.

These models are then used to predict criminaljustice expenditure in other countries, which aresummed to produce a global estimate. It should benoted at the outset that neither expenditure onprivatised nor informal community aspects ofcriminal justice are included in the estimates.

The bulk of this study is concerned withensuring transparency of method with regard to datapreparation and statistical technique. A summary ofthe key finding is as follows: It is estimated that theworld spent $360 billion on criminal justice in 1997.Of this total, 62 percent ($222.5 billion) was spenton public policing1, 3 percent ($11.2 billion) onprosecutions, 18 percent ($63.5 billion) on courts,and 17 percent ($62.5 billion) on prisons.

What is the point of estimating global criminaljustice expenditure? Primarily, the estimate is onecomponent of the costs of crime that has not beenestimated to date. Estimating the costs of crime isan increasingly important area of criminal justiceresearch. While such an estimate may prove usefulfor future research and analyses, it is also necessaryto be cautious about the potential abuse of suchestimates. Mark Cohen notes some of the benefitswhile warning of the potential for abuse of aggregatecost estimates:

“[W]hat are we to do with thisinformation? If we are successful in fullyestimating the cost of crime we cancompare this total cost estimate with thatof other social problems (e.g. cancer, autocrashes, homelessness). Whether oneagrees that this is a useful exercise or not,various advocacy groups do compare‘costs of crime’ estimates with the cost ofother social ills in an effort to affect policydecisions. Unfortunately, misuses of thesedata occur on both sides of the politicaldebate.” (Cohen, 2000; 269)

While it is possible that any research findingscan be wilfully abused, this should not become areason to hinder the accumulation of knowledge.However, the possibility does highlight the need formethodological transparency, and this is one of thekey facets of the present study. In addition, thefindings of this study are presented with a range ofcautions and caveats.

Introduction


To our knowledge there is only one previousresearch study with a primary focus upon theexpenditure data from the UN Crime Survey. JonSpencer (1993) examined trends over time inexpenditure between 1982 and 1986. He foundexpenditure had increased generally and thatchanging levels of expenditure tended to becorrelated between stages of the criminal justicesystem. If spending on one stage increased then,generally speaking, it increased for the other stages.Spencer concluded that “This analysis suggests thatgovernments allocate resources to criminal justicewith little or no attention to outcomes” (p.1).Spencer’s analysis was of trends and did not includeany estimates of spending for countries where thedata was not available.

The remainder of this study is structured asfollows. The next section describes data sources anddata preparation. The regression model analysis andfindings are then presented. The discussion sectionsuggests uses of the present research, makessuggestions for future research, and is followed bythe conclusion.

2. Data Sources and Preparation

2.1 The UN Crime SurveyThis section discusses survey methodology and dataquality control. The primary data source was reportsby national governments to the sixth United NationsSurvey of Crime Trends and Criminal JusticeSystems (hereafter the UN Crime Survey), for whichthe most recent data relates to 1997. The release ofthe survey used here was that of 27 June 2001.Survey responses had been received from eighty-three countries of which seventy provided full orpartial information on criminal justice expenditures.

Methodological issues have been extensivelyexplored in the literature relating to the UN CrimeSurvey. Cross-national differences reflect differencesin the composition of criminal justice systems,variation in definitions and terms, as well asdifferences in the extent, capacity, logisticaldifficulty inherent to, and emphasis upon, national-level data reporting and recording practices.

Countries differ in purchasing power, so they canbuy different amounts for the same dollar price.Differences in legal definitions may mean that, forexample, country A might include military policeas part of policing whereas country B does not. Suchdefinitional differences may or may not beidentifiable depending on whether they are reflectedin the terminology of the report from a government.There are also differences in practice: Country Amay prosecute all offences whereas country B doesnot. Procedural differences may produce spendingdifferences if, for example, country A definedprosecution as only those cases that reach court,whereas country B included warnings issued by theprosecutor among its prosecutions. Counting rulesand statistical classification are a source of variationas some countries may include minor localexpenditures in national estimates whereas othersdo not. Certain factors remain unknown relating toinformal and other justice procedures, such aswhether the costs of Sharia courts are included wherethey exist.2 Similarly, where there is a divisionbetween central and local government, it is notalways clear whether local activities are includedin national budgets. There is also the possibility oferror during the various stages of reporting andrecording the data: A country cannot accuratelyreport expenditure levels if the data does not exist,if national data collection is not coordinated, if dataare out of date, if the survey went to the wronggovernment department or respondent, or if clericalerrors were made in the completion of the surveyform. Finally, the data is fragmentary in part becausesome countries simply do not respond to the surveywhile others return it incomplete withoutexplanation. It has been suggested that response ratesand completion rates could reflect secrecy relatingto issues of national security issues, or variousaspects of international politics, and all of theseissues have been extensively reviewed elsewhere(Nalla and Newman (1994), Newman and Howard(1999a, 199b), Neapolitan (1996), Kangaspunta etal. (1998a, 1998b), Marshall (1998), Spencer (1993),Pease and Hukkila (1990; Tseloni and Pease (1994),Joutsen (1998); Aromaa and Joutsen (2003).


The quality control measures incorporated inthe UN Crime Survey data collection process are,like most research, far from perfect, but include anumber of aspects. First, and perhaps key amongthese is the clear definitions of terms on the surveyquestionnaire sent to governments, plus a separateset of instructions for each component part (police,courts, prosecutions, prisons). The following are theinstructions given on the survey for respondentscompleting the section on policing expenditures (butthe wording is essentially the same for the otherparts, substituting the relevant term for police):

“Total police budget/financial resourcesshould include all monies allocated to thecivil police function at the national level,including salaries and fixed assets. Whencalculating salaries, please include totalmonies spent on every individualemployed in the given area. Whencalculating fixed assets, please include allmonies invested in non-personnel assets,such as buildings, automobiles and officeequipment.” (Centre for InternationalCrime Prevention, 1999; 8)

A second aspect of quality control is the effortmade to track each individual person who isresponsible for completing the survey in a particularcountry. This reduces anonymity, generatesaccountability, and improves communicationbetween respondents and the survey team.Respondents working for individual governmentscan contact the UN data coordinators with questions,and UN data coordinators can contact respondentsto verify data where necessary. A third key controldiscussed by Newman and Howard (1999a; 8), ishow political pressures relating to the release ofnational-level data promote accountability:Governments generally do not release data into theinternational arena without having first closelyscrutinized it. Some countries may, of course, overtlyabuse or misrepresent their statistics for politicalgain, but with increasing democracy in EasternEurope and the ex-Soviet block this type of abusehas undoubtedly declined in recent years. Otheraspects of quality control for the UN crime surveyhave improved over the quarter century that thesurvey has been conducted (see Burnham 1999 foran overview of the survey’s origins and history). A

fourth quality control strategy is the maintenance ofinstitutional memory through the continuedinvolvement of key researchers. Even as newadministrators or clerical staff are introduced intothe process, this institutional memory reduces thelikelihood of basic errors in data collection andprocessing. Fifth, the survey format has been revisedto incorporate new and innovative approaches as partof a learning process.3 Sixth, the small ‘core’ groupworking on the survey at the UN are informed by aninformal ‘friends of the survey’ international networkof survey researchers. Seventh, the survey processand speed of communication (facilitating thechecking of data) has been helped by the Internetduring the last decade. Eighth, the increasing amountof published academic research based on the survey,and its resultant gradual permeation of media andacademic debates on cross-national issues, and itspublic availability on the Internet (viawww.unodc.org) means that its utility is increasinglyrecognized - another form of institutional memory.It should also be borne in mind that, relative to manynational-level surveys such as the US NationalCrime Victimization Survey or the British CrimeSurvey which are sponsored by nationalgovernments, the UN Crime Survey runs on aremarkably small budget. To our knowledge, thesurvey is typically undertaken largely by oneprofessional researcher who later employs the skillsof a data entry clerk. While other researchers areconsulted as discussed above, no independent surveycompanies, data collectors or researchers arecontracted. By most western research standards, thesurvey team are shockingly under-resourced. Subjectto such constraints, the survey is arguably quiteremarkable, while most of its critics derive from abetter-resourced tradition.4 This is not intended tosuggest that the survey is perfect or might not benefitfrom improved data collection practice and furtherreliability and validity checks: Indeed, as discussedlater, it is hoped that the present study might be usedto stimulate better reporting of data by countries.However, it remains the case that, when approachedwith the necessary caution, the survey is the mostcomprehensive existing resource for many areas ofcross-national research using criminal justice data.


The general quality control issues discussedabove were supplemented with a further setundertaken specifically for the present study. Thesixth survey requested information for 1995 to 1997inclusive. Generally, those countries that reporteddata did so for each year. Data for 1997 was cross-checked for consistency with that for previous years.In a small number of cases where 1997 data wasabsent it was substituted with data for 1996 whereavailable. Preliminary descriptive data analysis wasundertaken to eyeball the expenditure data forobvious errors. Errors in reporting and recording aswell as at the data entry stage were the most likelyexplanation for the small number of extreme valuesthat were found. For example, an examination ofexpenditure upon Canadian prisons revealed that thesource data set contained a value in Canadian dollarsthat was in error by two decimal places. This datumwas corrected. However, to ensure a conservativeapproach, a small number of data items wereexcluded where their true value could not be verifiedeven by cross-checking. Most of these were not fromhighly developed countries, so it is arguably morelikely that such discrepancies reflect variation in thequality of data reporting and recording systems. Apossible error source could be discrepanciesintroduced via exchange rates during the conversionof currencies into US dollars, particularly if localcurrencies were extremely volatile. The country-level specifics of the study’s quality control processare detailed in a technical appendix.

Despite the various quality controls, the presentanalysis should be viewed with caution. Even thoughthey are the best available, the data are imperfect.Indeed, if the data were perfect there would be noneed for estimation of missing values: globalcriminal justice expenditure could be found bysimply summing the values reported by individualcountries. However, we proceed from the reasonableposition that the data cup is half full rather than halfempty.

2.2 Sampling and Representativenessof CountriesRicher countries can afford to spend more upon datacollection efforts. They are consequently over-represented in this data set as in most cross-nationalanalyses. The sample of countries is therefore notrepresentative of the world. However, the key issuefor present purposes is not whether the sample ofcountries is representative. The key issue is whetherthe relationship between GDP and criminal justiceexpenditure that derives from the data isrepresentative. Since, as will become evident, thegreatest divergence in this relationship occurs amongricher countries, it is a positive attribute of the datasetto have an over-representation of richer countries(those toward the right-hand side of Figures 1 to 4).The majority of missing data points relate to poorercountries which, in absolute dollar terms, contributefar less to global expenditure totals.

A second aspect of the study overtly seeks toaddress the possibility that the relationship betweenGDP and criminal justice expenditure is differentfor different sets of countries. Specifically, thestatistical models that is split or kinked (Model 4) isincluded to capture the fact that there may be adifferent GDP-spending relationship among poorercountries than richer countries. In short, once againwe conclude that the present data set is imperfectbut that there is good reason to believe this is not anoverwhelming obstacle.

2.3 Other data sourcesThe database of the International Monetary Fund(IMF) was used to complete the list of worldcountries and populations, and was the source ofthe bulk of information on international exchangerates. 5 The specifics of this procedure are detailedin the technical appendix. Using the 1997 exchangerates, expenditure data for all countries wasconverted to 1997 US dollar equivalents. All analysisdescribed below uses US dollars.


3. Analysis

As an analytical quality control, statistical outlierswere excluded from the main analysis. They wereidentified as national expenditures where thestandardized residual resulting from the regressionof per capita GDP upon per capita expenditure (foreach stage) was greater than the 5 percent value froma two-tailed Student’s distribution.6 Eight data pointswere excluded, three of which related to NorthernIreland for each stage of its criminal justice systemexcept prosecutions. Bahrain and Malaysia wereoutliers for police and prison expendituresrespectively. Scotland, Colombia and Panama wereoutliers for prosecutions. All of the outliers werefar higher per capita expenditures, relative to theper capita GDP, than the remaining countries.

1) Six regression models of the relationshipbetween GDP and criminal justice expenditure weredeveloped. The aim of the modelling was to identifythe model which yielded the most accuratepredictions for each of police, prosecution, court andprison expenditure. Each model represents amodification to the basic relationship between percapita expenditure and per capita GDP. That modelwas selected, for each stage of the criminal justicesystem, which best satisfied the criteria of goodnessof fit and statistical efficiency as detailed below. Foreach stage, that model was then used to predictexpenditure for countries for which the data wereunavailable.

2) The first model is the simple linear per capitamodel. It reflects the possibility that richer countriesspend more on criminal justice. The precise form ofthe relationship may, however, be more complexthan the simple linear per capita model. It may differacross stages of the criminal justice system, acrosstime, across country type and so on. The subsequentmodels investigate whether the estimates of globalcriminal justice expenditure are sensitive to variouschanges in the functional form of the regressionmodel.

3) Figures 1 to 4 show plots of the data pointsfor per capita GDP and per capita criminal justiceexpenditure, with a linear best fit line (Model 1)

superimposed. The chart for policing includes thequadratic fitted curve (Model 2) for illustrativepurposes. Countries are labelled to the extentpossible on the charts. The clusters of countries atthe bottom left of each chart are not labelled if itwas not feasible. On the charts, readers should notethe variation in the vertical axes, which reflectgreater per capita spending on policing, followedby prisons, courts, and prosecutions. Countries thatwere statistical outliers as described above, are notshown in the charts. Expenditures for countries withmissing values were estimated using the regressionequation, and global expenditure estimated as thesum of national values. In the global total, thereported values were used for countries that hadpreviously been excluded as outliers in the model’sdevelopment. The 95 percent confidence intervalswere calculated according to the following formulafor the variance of the prediction of total expenditure,based on the method described by Judge et al. (1988:251):

var = ′ ′ ′ +−σ 2w [X (X X) X I ]w01

0 m (1)

where σ2 is the regression mean square error,w is an m x 1 vector of population values for the mcountries we are predicting, X0 is an m x 2 matrix ofexplanatory variables for the countries we arepredicting, X is an n x 2 matrix of explanatoryvariables for the countries in the regression sample,and Im is an m-dimensional identity matrix.


Figure 1: Policing Expenditure and GDP

0

50

100

150

200

250

300

0 5000 10000 15000 20000 25000 30000 35000 40000

GDP per capita

Expe

nditu

re p

er c

apita

($U

S)

Switzerland

Japan

Denmark

United States

Singapore

Iceland

Sweden

Hong KongEngland & Wales

Belgium

Finland

Australia

Ireland

Canada

New Zealand

Israel

Andorra

Spain

Bahamas

Greece

Cyprus

Portugal

Korea Rep.

Slovenia

Czech Rep

Malaysia

Croatia

S. Africa

Mauritius

Figure 2: Prosecution Expenditure and GDP

0

2

4

6

8

10

12

14

16

18

0 5000 10000 15000 20000 25000 30000 35000 40000

GDP per capita (US$)

Expe

nditu

re p

er c

apita

($U

S)

Switzerland

Japan

Denmark

United States

Singapore

Sweden

Germany

England and Wales

Finland

Ireland

New Zealand

Israel

Spain

Cyprus

PortugalBahrain

Korea Rep.

Slovenia

ArgentinaCzech Rep.

Lithuania

Latvia

Fiji


Figure 4: Prison Expenditure and GDP

0

10

20

30

40

50

60

70

80

90

100

0 5000 10000 15000 20000 25000 30000 35000 40000


Expe

nditu

re p

er c

apita

(US$

)

Switzerland

Japan

Denmark

United States

Singapore

Iceland

Sweden

Hong Kong

Germany

England and WalesNetherlands

Finland

Australia

Ireland

Canada

Italy

Scotland

New Zealand

Israel

Andorra

Spain

Greece

Cyprus

Portugal

Argentina

Czech Rep.S. Africa

Hungary

Figure 3: Court Expenditure and GDP

0

10

20

30

40

50

60

70

80

90

100

0 5000 10000 15000 20000 25000 30000 35000 40000


Expe

nditu

re p

er c

apita

(US$

)

Switzerland

Japan

Denmark

United StatesGermany

SingaporeIceland

Sweden

Hong Kong

England and Wales

Finland

Ireland

Israel

Cyprus

Portugal

Bahrain

Argentina

Czech Rep.

Malaysia

ColombiaCosta R ica


To illustrate the potential importance ofmodifications to the per capita model, consider againpolicing expenditure shown in Figure 1. Visualinspection of the scatter plot and fitted line suggestthat there is some systematic pattern to the extent towhich actual police expenditure per capita divergesfrom the fitted value (the value represented by theregression line)7. In particular, at low levels of percapita GDP (below $10,000), the linear modeloverpredicts per capita police expenditure forvirtually all countries, while at higher levels of percapita GDP, the opposite is the case with mostcountries having lower fitted than actual values. Itseems that, in the data, the slope of the relationshipbetween per capita police expenditure and per capitaGDP is higher at lower levels of per capita GDP andvice versa. Alternatively put, the marginal propensityto spend on policing is lower at higher levels ofincome. Such a relationship may fit with anexplanation of criminal justice expenditure as anecessity whereby poorer countries spenddisproportionately more on criminal justice thanricher countries. This suggestion leads to the secondstatistical model of per capita police expenditure:

(X/P)i = α + β(G/P)i + γ(G/P)i2 + εi (2)

i = 1, …, n.

where X - police expenditureP - populationG - GDPε - a random error termi - country indexn - number of countries which report

police expenditure (excluding outliers).

Through the quadratic term, Model 2 allowsfor a declining regression slope at higher levels ofper capita GDP. The fitted regression line from thismodel shown in Figure 1 yields a noticeableimprovement in fit upon the linear per capita model.There is a modest improvement in the model R2 (theproportion of the variation in police expenditureexplained by the model) from 0.54 to 0.59. Inaddition the coefficient on the quadratic term isstatistically significant (two-tailed p-value = 0.030).

Clearly, introducing a quadratic term to themodel is only one way whereby the per capita modelcould be modified. Other forms of non-linearrelationship are possible, as are othertransformations of the dependent and independentvariables in the regression. Different stages of thecriminal justice system might be best representedby quite different specifications. To investigate thisissue fully, a series of regression models wasestimated for each stage of the criminal justicesystem to find the one which fitted best. Threecriteria were used to choose and refine the regressionmodels. First, since the over-arching goal of thepresent study is to provide estimates of total worldcriminal justice expenditure, the class of regressionmodel was selected using the criterion that it hadthe best in-sample forecasting performance. Second,in the interests of statistical efficiency, standardtechniques of statistical inference were used to refinethe resulting model. Third, we were at pains toensure that the resulting predicted values passedwhat Hamermesh (1999) calls the “sniff test”, thatis, that they appeared to make economic sense.

TABLE 1: Results of Linear Per Capita Model

CJ Stage Ncountries

#outliersexcluded

R2 Coeff. Constant EstimatedGlobalExpenditureUS$ Billion

Lower95% CI

Upper95% CI

Police 47 2 0.5405 0.0052 24.003 $286.560 80.220 492.901Prosecution 35 3 0.6343 0.0003 0.3967 $12.055 4.447 19.663Courts 34 1 0.483 0.0013 1.4604 $59.152 -10.073 128.376Prisons 54 2 0.6181 0.0015 0.5946 $58.693 24.656 92.730Total $416.460 198.088 634.834


Table 2 describes the various models that wereevaluated and summarizes their in-sampleforecasting performance. Models 1 and 2 werediscussed above – the linear per capita model andquadratic per capita model, respectively. Models 3and 4 are variants of the per capita model whichattempt to capture differences in the slope of therelationship between per capita expenditure andGDP at different levels of per capita income. Model3 is a double log model that allows for non-linearitythrough the specification of the dependent andindependent variables in natural logarithmic (the lnfunction) form. This is a commonly used model inapplied statistical research. Model 4 is effectively apiecewise linear model which estimates separatelinear regression lines for ‘poor’ countries (definedas those with less than $10,000 per capita GDP) and‘rich’ countries. Thus Model 4 also overcomes theissue of the over-representation of rich countries inthe sample, a point discussed earlier. Closeexamination of Figures 1 to 4 suggests that the clumpof countries at the bottom left (low per capita GDP,low criminal justice expenditure, many of which arenot labelled due to lack of space in the charts) couldbe qualitatively different from other countries. Byallowing for the possibility of a kinked expenditurecurve, this model ensures the integrity of findingsfor ‘poorer’ countries if their relationship betweenGDP and criminal justice expenditure is somewhatdifferent than that for ‘richer’ countries.

Models 5 and 6 are based around the notionthat, since the aim is to predict total expenditure ineach country, the appropriate dependent variable isthe level of total criminal justice expenditure ratherthan per capita expenditure. Uncoupling populationfrom the financial variables allows population itselfto enter as a separate independent variable. HenceModels 5 and 6 are multiple regression models.Model 6 is a double log version of the relationshipto account for potential non-linearities.

Table 2 reports the root mean square forecasterror for each of the six models examined. This isdefined as

(3)RMSFE =

where Xi is expenditure on the relevant stage ofthe criminal justice system by country i and

iX̂

is the predicted or forecast level from theregression model under consideration for the samecountry. Note that iX̂ is a transformation of theregression fitted value for models1-4 and 6; theper capita models require the multiplication ofthe fitted values by population while the doublelog models require the antilog transformation tobe taken.8

TABLE 2: The Model Selection Process

Root Mean Square Forecasting ErrorModelNo.

Name EquationPolice Pros. Court Prisons

1 Linear PerCapita

(X/P)i = α1 + β1(G/P)i + εi 3695.03 205.84 1975.33 1759.40

2 Quadratic PerCapita

(X/P)i = α2 + β2(G/P)i + γ2(G/P)i2 + εi 3490.07 322.66 1930.18 1767.41

3 Double LogPer Capita

ln(X/P)i = α3 + β3ln(G/P)i + εi 3740.37 187.70 2042.51 2098.39

4 FullyInteracted -Per Capita

(X/P)i = α4 + β4(G/P)i + γ4Di +δ4Di(G/P)i + εi

3614.74 248.33 1964.19 1773.85

5 MultipleRegression

Xi = α5 + β5Gi + γ5Pi + εi 3659.00 148.39 1334.63 1425.55

6 Double LogMultipleRegression

ln(Xi) = α6 + β6ln(Gi) + γ6ln(Pi) + εi 4638.97 239.11 2035.86 2466.01

( )n

XXn

i ii

2

1ˆ� =

−


Perusal of the columns of root mean squaredforecast errors for each of the stages shows that,broadly speaking, the linear per capita and multipleregression models perform better than the doublelog models. For policing, RMSFE was minimizedby the quadratic per capita model (Model 2)discussed above. This model also seemed to givereasonable predictions of the missing values. Forthe other three elements of criminal justiceexpenditure the linear multiple regression model hadthe best in-sample performance. However, thismodel had the unfortunate property of producingnegative predictions of expenditure for a significantnumber of countries. One approach to solving thisproblem would be simply to set the predicted valuesto zero for such countries. As well as being ad hocthis is clearly unrealistic and so a different routewas identified: since investigation suggested thatstatistically insignificant explanatory variables werethe cause of the negative estimates, such variableswere excluded from the regression models. ThusModel 5 was re-estimated for each of prosecution,court and prison expenditure with the restriction a5= g5 = 0 imposed. This restriction was supported bythe data in each case (the p-values were 0.3171 forprosecutions, 0.4575 for courts and 0.4134 forprisons).9

With a ‘best’ model in hand for each of thestages of the criminal justice system it was a simplematter to compute estimates of global expenditurefor each category and to find the associatedconfidence interval. The results are displayed inTable 3. The 95 percent confidence intervals forpolice expenditure were computed as in equation 1while for the other stages of the criminal justicesystem, the vector w was replaced by an m-dimensional unit vector. This is because, unlikemodels based on per capita expenditure, there wasno need to weight the predicted values by the levelof population.

4. Discussion

The use of several statistical models was more thanjust a changing room exercise of trying on varioushats to see which fits best. Although it took placelargely as an iterative and empirically-drivenprocess, the models derive some theoreticaljustification from the literature on public budgeting.Lynch’s classic text concludes that

“It is money. The one common subject inany budget discussion is money. Othersubjects are important, but they arementioned in relationship to money or aretranslated into money. Budgeting involvesdollars and cents often expressed in themillions of dollars.” (Lynch 1979; 2,emphasis added)

TABLE 3: Results of Refined Estimation Procedure

CJ Stage EstimatedGlobal ExpenditureUS$ Billion

Percent of Total Lower 95% CI Upper 95% CI

Police $222.530 61.87 16.990 428.070Prosecution $11.163 3.10 6.950 15.376Courts $63.472 17.65 26.216 100.727Prisons $62.528 17.35 26.171 98.886Total $359.693 100.00 149.550 569.836

Note to Table 3: Total in percentage column does not equal 100 due to rounding.


In practice, society’s ideal of a welfare-maximization goal is transgressed by privateinterests from sources including pressure groups,bureaucrats and voter preferences as well as theinfluence of previous budget levels that produce pathdependence (Steiner, 1971; Rubin, 1997; Peacock,1978; Tullock, 1989). More specifically in relationto national spending on criminal justice, the majorconclusion of the only other previous systematiccross-national study was that “governments allocateresources to criminal justice with little, or no,attention to outcomes.” (Spencer 1993; 7). Even acritic of our parsimonious approach to variableselection cannot overlook its possibilities as afoundation upon which future research might build.Such future research would be testing additionalhypotheses of the predictability of national-levelcriminal justice expenditure. We anticipate however,that such research may well experience negative orrapidly diminishing returns to the investment.

The introduction to this study began to addressthe utility of the present type of cost estimationexercise. It was suggested that global-level costestimates may, in time, become increasingly useful- just as at one time it may not have been intuitivelyapparent that regional, national, city, or village-levelestimates were of any utility. It was also suggestedthat the potential for wilful abuse of the estimates isinsufficient reason to stop the pursuit of knowledge.However, some aspects of the potential utility ofthe present study will benefit from furtherelucidation. The present exercise is intended as apreliminary step towards better estimates in the hopethat future research and data collection efforts canimprove upon them. However, it may still bepossible to develop some tentative comparisons toother global expenditures: If 1997 criminal justiceexpenditures were inflated to 2004 prices, globalexpenditure on criminal justice would be estimatedas $424 billion.10 This is around half the estimated$812 billion global expenditure on armaments, fivetimes the estimated $80 billion spent on education,and roughly double the turnover of the globaltobacco industry.11 Other policy-relevant researchquestions could include: What are the main between-country and within-country variations in criminaljustice expenditures? How do they vary across stagesof the criminal justice system? What is the income

elasticity of demand for criminal justice? Why dosome countries spend proportionally more on someparts of the criminal justice system than others? Arethere particular ‘types’ of countries that spend moreon some stages of the criminal justice system thanother? If prosecution and courts represent the stagesof ‘justice’, are there some countries that pay arelatively greater attention to justice than others?How do these and other variables change over time?How do the income elasticities of demand forcriminal justice compare to those for other publicand non-public goods? This is not intended as anexhaustive list but, rather, is suggestive of thepotentially broad and policy-relevant researchagenda that could emerge.

As the first systematic empirical estimates ofglobal spending, the current estimates may have thepotential to instigate an improvement ofinternational data collection. Improvements in dataquality rather than refinements in statisticalregression method are the preferable long-termsource of improved confidence in global criminaljustice expenditure estimates. The relevant UnitedNations bodies should consider using the presentestimates to encourage countries to improve theirdata collection and reporting practices (a listing ofactual and predicted expenditures by country andstage of the criminal justice system is available fromthe corresponding author upon request but, at thesuggestion of several commentators, was excludedhere for brevity). This is not meant tocondescendingly suggest that poor countries willimprove data collection when faced with the threatof estimates produced by foreigners, but it maystimulate greater effort in data collection andreporting (particularly a reduction in unnecessarynon-responses and partial responses) in somemarginal cases. In addition, the use of predictivetechniques for the imputation of missing values, suchas those used here, may have broader implicationsfor the use of cross-national data other than that oncriminal justice expenditure.


5. Conclusion

The study found that criminal justice expenditurelevels are significantly tied to levels of availablepublic monies as determined by the strength of anational economy. Despite acknowledgedlimitations of data and method, the study producedthe first systematic empirical estimate of globalexpenditure on criminal justice. The confidenceintervals are soberingly wide, the estimates shouldbe interpreted with caution, and the study viewedas an incremental step towards improved datacollection and estimation method. It is hoped thatthe present study provokes better data collection andencourages the investigation of methods for theestimation for missing data in this field.

Money makes the world go round, or so theysay. However, while money does not necessarily takeresponsibility for the sun’s gravitational pull uponthe earth, it does facilitate the allocation of scarceresources in human society. Whether or not it provesto be a robust theory of global expenditure uponcriminal justice remains to be tested by futureresearch.

ReferencesAromaa, K. and M. Joutsen. 2003. ‘Introduction’

in K. Aromaa, S. Leppä, S. Nevala, and N.Ollus (Eds.) Crime and Criminal Justice inEurope and North America 1995-1997.HEUNI Report 40. Helsinki: United NationsEuropean Institute for Crime Prevention andControl.

Berk, R. A. (1990) A primer on robust regression.In Modern Methods of Data Analysis, ed. J.Fox and J. S. Long, 292-324. Newbury Park,CA: Sage Publications.

Burnham, R. W. 1999. ‘A short history of thecollection of UN crime and justice statisticsat the international level’ in G. Newman (Ed.)Global Report on Crime and Justice. NewYork: Oxford University Press.

Centre for International Crime Prevention. (1999)Questionnaire for the Sixth United NationsSurvey of Crime Trends and Operations ofCriminal Justice Systems, covering theperiod 1995-1997. Vienna: United NationsOffice for Drug Control and CrimePrevention.

Cohen, Mark. (2000) ‘Measuring the costs andbenefits of crime and justice’ inMeasurement and Analysis of Crime andJustice, volume 4 of Criminal Justice 2000.Washington D.C.: Department of Justice,National Institute of Justice.

Goldberger, Arthur S. (1968) The Interpretationand Estimation of Cobb-Douglas Functions.Econometrica, Vol. 36, Issue 3/4, 464-472.

Farrell, Graham, Erin Lane, Ken Clark andAndromachi Tseloni. (2001) ‘What does theworld spend on policing?’ InternationalJournal of Comparative Sociology, XLII, 1-2, 59-73.

Granger, Clive and Paul Newbold. (1977)Forecasting Economic Time Series. NewYork: Academic Press.

Hamermesh, Daniel S. (1999) The Art ofLabormetrics. National Bureau of EconomicResearch Working Paper 6927. [http://www.nber.org]

Joutsen, Matti. (1998) ‘Introduction’, pp.1-19 inKristiina Kangaspunta, Matti Joutsen andNatalia Ollus (Eds.) 1998a. Crime andcriminal justice systems in Europe and NorthAmerica 1990-1994. Monsey, NY: EuropeanInstitute for Crime Prevention and Control;Criminal Justice Press.


Judge, G., R.C. Hill, W.E. Griffiths, H. Lutkepohland T.C. Lee. (1988) Introduction to theTheory and Practice of Econometrics. NewYork: Wiley.

Kangaspunta, Kristiina; Matti Joutsen; NataliaOllus (eds.) (1998a) Crime and criminaljustice systems in Europe and North America1990-1994. Monsey, NY: European Institutefor Crime Prevention and Control; CriminalJustice Press.

Kangaspunta, Kristiina, Matti Joutsen, NataliaOllus and Sami Nevala (eds.) (1998b)Profiles of Criminal Justice Systems inEurope and North America 1990-1994.Monsey, NY: European Institute for CrimePrevention and Control; Criminal JusticePress.

Lynch, Thomas D. (1979) Public Budgeting inAmerica. Englewood Cliffs, New Jersey:Prentice-Hall, Inc.

Marshall, Ineke Haen. (1998) ‘Operation of theCriminal Justice System’, pp. 54-114 inKristiina Kangaspunta, Matti Joutsen andNatalia Ollus (Eds.) 1998a. Crime andcriminal justice systems in Europe and NorthAmerica 1990-1994. Monsey, NY: EuropeanInstitute for Crime Prevention and Control;Criminal Justice Press.

McDonalds. (2000) ‘McDonalds Reports RecordGlobal Results’ McDonalds Financial PressRelease 26, January 2000.

Nalla, Mahesh K. and Graeme R. Newman.(1994) Crime in the U.S and the FormerU.S.S.R.: a comparison of crime trends fromthe third United Nations Survey.International Journal of Comparative andApplied Criminal Justice 18 (1), pp. 85-94.

Neapolitan, Jerome L. (1996) ‘Cross-nationalcrime data: some unaddressed problems’Journal of Crime and Justice, XIX, 1, 95-112.

Newman, Graeme (ed.). (1999) Global Report onCrime and Justice. Published for the UnitedNations Center for International CrimePrevention. Oxford and New York: OxfordUniversity Press.

Newman, Graeme, and Greg Howard. (1999)‘Introduction: data sources and their use’ inG. Newman (ed) Global Report on Crimeand Justice, published for the United NationsCenter for International Crime Prevention.Oxford and New York:.Oxford UniversityPress.

Peacock, Alan. (1978) ‘The Economics ofBureaucracy: An Inside View’ in TheEconomics of Politics. IEA Readings 18.London: Institute for Economic Affairs.

Pease, Ken and Andromachi Tseloni. (1994)‘Juvenile-adult differences in criminaljustice: evidence from the United NationsCrime Survey’ Howard Journal of CriminalJustice, 35, (1): pp. 40-60.

Rubin, Irene S. (1997) The Politics of PublicBudgeting: Getting and Spending, Borrowingand Balancing, 3rd. Chatham, New Jersey:Chatham House Publishers, Inc.

Spencer, Jon. (1993) ‘Criminal JusticeExpenditure: A Global Perspective’ HowardJournal of Criminal Justice, 32, 1, 1-11.


Tullock, Gordon. (1986) ‘Wanted: New PublicChoice Theories’ in Martin J. Anderson (Ed.)The Unfinished Agenda: Essays on thePolitical Economy of Government in Honorof Arthur Seldon. London: Institute ofEconomic Affairs, pp. 13-28.

Transparency International. (2002) TheTransparency International CorruptionPerceptions Index 2002. Berlin:Transparency International (atwww.transparency.org ).

Steiner, Peter O. (1971) ‘The Public Sector andthe Public Interest’ in Robert H. Havemanand Julius Margolis eds. Public Expendituresand Policy Analysis. Chicago, Illinois:Markham Publishing Company.

Van Horne, Sheryl and Graham Farrell. (1999)Drug Offenders in the Global CriminalJustice System. HEUNI Paper #13. Helsinki,Finland: The European Institute for CrimePrevention and Control, affiliated with theUnited Nations.

United Nations Development Programme. (2002)The Human Development Report 2002. NewYork: Oxford University Press.

TECHNICAL APPENDIX

The first section of this appendix containsinformation relating to the UN Crime Survey data.The short second section relates to supplementarydata sets.

The UN Crime Survey DataThe Bulgarian government reported expendituredata on courts, prosecution and prisons but not onpolicing, noting that “The data on the numericalstrength of the police personnel and financial

resources of the police in Bulgaria are a state secret.”Partial data on national policing expenditure wasprovided by the Estonian government with the riderthat “[The] Security police board is excluded as theirpersonnel figures are not public.” The CzechRepublic reported that “The budget of the CzechRepublic police covers total for current and capitaldata, including the Czech Republic InvestigationOffice. Data for the Czech Republic InvestigationOffices within these years were not separatelysurveyed and investigated and that is why we cannotpresent them separately.”

The response from the Israeli government noted thatthe police budget “[i]ncludes the budget of policeprosecutors” while with respect to prosecution,“[t]he budget figures are for governmentalprosecutors within the Ministry of Justice. ” Thereport of the Romanian government noted in relationto prosecutions that “The budget of the PublicMinistry is different from and independent of thebudget of the Ministry of Justice and that of thePolice. The structure of the annual budget comprisesin principle the following sections: expenses forstaff; maintenance (buildings, repairs, vehicles andother assets, fuel, office equipment, inventory items);investments (furniture, vehicles).”

In relation to courts in Bulgaria it was observed that”Included are financial resources for wages andsalaries and for social insurance contributions of themagistrates and the personnel of the Ministry ofJustice and Legal Eurointegration.” A comment inrelation to Danish courts noted that “Finances ofDanish law courts, capital summary and edp inregistration of title to land, etc. are included in thiscategory”

The sixth UN Crime Survey was accompanied by aCrime Guide.12 The Guide contained comments fromresponding governments, translated as appropriate.Based on the comments we made five adjustmentsto a specific data item. Each adjustment involvedan assumption or extrapolation aimed at reducingthe amount of error: The Danish government notedthat “police and prosecution data could not beseparated”, and that the same value had beenincluded for both in the database. This value wassplit proportionally between police and prosecutions


to reflect the proportional spending levels ofcountries which reported the data separately. Forother countries, 6.08% of the combined police andprosecution budget was spent on prosecutions, and93.92% on policing, and the Danish police andprosecution budget was split to reflect theseproportions; A similar revision was made in relationto courts and prosecutions for Germany andSwitzerland. In relation to Germany, the CrimeGuide noted that a “[d]istinction between court andprosecution budget is not available”, while theresponse of the Swiss government noted that thecourt budget “includes the prosecution budget”.Countries that reported both prosecution and courtdata separately spent an average of 16.24% of thecombined total on prosecutions and 83.76% oncourts. The German and Swiss budget figures weresplit to reflect this average; The Finnish governmentnoted that “Lay judges in circuit courts were financedby municipal resources until 1998. The cost, about30 million marks is not included in [the] state budget[the figure given].” Consequently, the reportedFinnish court budget of 931 million marks wasadjusted to 961, an increase of 3.22%. The fifthchange to the data was based upon a comment fromthe Canadian government that their prison budgetincluded “[a]dult data only”. Separate budget datafor adult and juvenile prisons was not available forany country, and so the proportions going to each inother countries could not be calculated and appliedto Canada. However, data on the number of beds inadult and juvenile prisons was available elsewherein the UN survey, although not for Canada. Onaverage, the number of juvenile beds was 5.53% ofadult beds. On the assumption that proportionalexpenditures may mirror the proportion of beds, theCanadian prison budget was inflated by 5.53% as acrude estimate of the additional cost of juvenileinstitutions.

Twenty additional data items were excluded fromthe analysis. Four exclusions were made in relationto policing: Kazakhstan, Kyrgyzstan, Scotland andTanzania. The lowest given value for policingexpenditure in any of these countries was six timesgreater than expenditure in any other country. Sixexclusions were made in relation to prosecutions:Bulgaria, Denmark, Georgia, Kyrgyzstan, Malaysiaand Tajikistan, whose values were at least ten times

those for other countries. Three exclusions weremade in relation to court expenditure: Georgia,Kazakhstan and the Republic of Korea, each ofwhich had expenditures at least six times greaterthan all other countries. Three exclusions were madein relation to prison expenditure: Kyrgyzstan, Jordanand Panama, each of which had values that wereseveral orders of magnitude greater than those ofany other country. An effort was made to trace thevalues to determine the origins, but they were in theUN crime survey database. It is likely that some ofthe errors arose during the reporting stages –government officials who completed thequestionnaire may have misinterpreted an aspect ofthe question, or perhaps may not have noticed thatdata were to be reported in millions of local currencyunits. The UN database is compiled in local currencyunits and so such discrepancies would be unlikelyto be noticed if the currencies were unfamiliar.Hence it is reasonable to expect that expendituredata reported incorrectly by a responding countrywould go unnoticed prior to the release of the dataset.It is notable that seven of the exclusions were fromthe three countries Georgia, Kazakhstan, andKyrgyzstan, suggesting that a general error was madein relation to those countries. It is possible that oneor more of the criminal justice expenditures thatwere excluded as extreme outliers were not errors.For example, it could be that these three formerSoviet states undertook massive capital expenditureson their criminal justice systems in 1997 - a pointthat would require further investigation. For presentpurposes however, such extremes of expenditurewould have been excluded as statistical outliers bythe method detailed in the Analysis section. Withthe exceptions of Scotland, Denmark and possiblyBulgaria, the other six exclusions related to countriesthat are not in the upper-echelons of developmentrankings and which are therefore arguably less likelyto have high quality reporting and recording systemsfor their aggregate-level data relating to expendituresacross the country.


Supplementary Data Sources

The list of world countries and their populations wascompiled from the UN crime survey and the databaseof the International Monetary Fund (IMF). Wherepopulation data could not be found in these sourcesit was supplemented with data from, in order ofpreference, the 1998-based National PopulationProjections13, the UN Population Division (WorldPopulation Prospects Population Database), and theCIA World Factbook.14

The primary source of data on national grossdomestic product (GDP) was the database of theIMF, supplemented with other sources as necessary.Eight other sources were used, and in a few instancesGDP data for 1996 was used when 1997 data wasnot obtained. The other sources of GDP data werethe CIA World Factbook 2001, the Northern IrelandFinancial Times, the UN Human DevelopmentReport 1999, The Scottish Parliament15, UNCTAD16,Geographic’s Country Profiles17, Yahoo! Finance18,and Freiburg University19. There was no reason tobelieve that these sources contained bias in thereporting of GDP data, and every reason to believethat the inclusion of the data produced anincremental improvement in our database that waspreferable to the exclusion of that data element.

Expenditure data was reported in local currencyunits. Exchange rates for 1997 that were not in theUN crime survey database were obtained from thedatabase of the International Monetary Fund.

Where there was a difference in specific data itemsbetween the various possible sources, the differenceswere typically minor discrepancies of a fewpercentage points, almost certainly reflectingdifferences in the reporting and recording ofnational-level data. In such instances, the UN crimesurvey data was used if available. There were onlya handful of cases where a decision had to be maderegarding a non-trivial difference between sources.In such instances the two data items were comparedand it would usually be apparent that one containedan obvious error (probably due to reporting,recording or data entry error at different stages andin different sources). Where it was not clear which

data item was preferable, additional qualitativeinformation was sought to allow the identificationof the most appropriate datum. If available, a thirdsource of data was utilized as an arbiter. The endresult was a database with complete GDP andpopulation data for 204 countries.20

Notes

1 Private policing and security expenditures are notincluded since, while they are substantial, they donot utilize government funds.

2 We thank Kauko Aromaa for comments on thissection in particular.

3 It is, as with any change, possible to dispute someof the revisions: Recent streamlining of the surveyhas removed overlapping years of data from eachiteration of the survey, leaving the survey open tosuggestions that fewer validity checks can beundertaken.

4 For discussion on such issues we thank AdamBouloukos, formerly of the United Nations Officeof Drugs and Crime.

5 We thank Matthew Fleming, then of theInternational Monetary Fund, for assistance inlocating the relevant exchange rates. The onlinecurrency converter available at www.oanda.com/convert/classic was used to complement the IMFdata in only a handful of instances.

6 An alternative approach to extreme observationswould have been to use a robust regression technique(see, for example, Berk, 1990)

7 Such a pattern in the residuals might reflect theinfluence of omitted explanatory variables, discussedfurther below.


9 The restriction increased the root mean squareforecasting error but never by enough to prevent thismodel being ‘best’ in its column. The figures were154.39 for prosecutions, 1369.87 for courts and 1451.48for prisons.

10 Inflation adjustment using the inflation calculator ofthe US Bureau of Labor Statistics at http://www.bls.gov/cpi/home.htm.

11 These comparative global spending estimates weretaken from sources of varying degrees of integrity,relevance and comparability, and are included only tobe suggestive of possible future applications of thefindings - the intention is to put criminal justiceexpenditure in a broad comparative perspective ratherthan to provide definitive estimates for other items.The source of the estimates is as follows: Globalmilitary spending of $812 billion in 2001 from theCenter for Defense Information at www.cdi.org/issues/wme , retrieved on 2nd October 2002; Annual globalspending on education from United Nations Children’sFund (UNICEF) report, The State of the World’sChildren, 1999 (see http://www.penpress.org/docs/PEN84.pdf); Annual sales turnover of the tobaccoindustry from www.didyouknow.cd/fastfacts/money.htm., page 1, retrieved 2nd October 2002.

12 Our thanks are due to Sami Nevala for making theCrime Guide available.

13 Series PP2 no22; retrieved Jan 31, 2002 fromstatbase: www.statistics.gov.uk/statbase/

14 Available online at http://www.cia.gov/cia/publications/factbook/

15 http://www.scottish.parliament.uk/

16 www.unctad.org

17 www.geographic.org

18 http://biz.yahoo/

1 9w w w. z m k . u n i - f r e i b u r g . d e / E u r o p e /yugoslavia.htm

20 We note that 204 countries is not a definitivecount or one that is necessarily approved by theUnited Nations. However, it is hoped that we didnot exclude any major populations that wouldsignificantly influence the key findings of thestudy.

8 For the double log models there is an issue concerningbias correction when forming the prediction of totalexpenditure (see Goldberger, 1968). We applied thebias correction suggested in Granger and Newbold(1977) however this rarely improved the in-sampleforecasting performance and even when it did, theimprovement was not enough to significantly increasethe ranking of the relevant model. We therefore do notpursue the issue further.

Helsinki 2004

what does the world spend on criminal · pdf filewhat does the world spend on criminal...

Documents