Redundancy and Performance in Public Service Delivery Networks

Sean Nicholson‐Crotty
Department of Political Science and
Harry S Truman School of Public Affairs
University of Missouri

Jill Nicholson‐Crotty
Harry S Truman School of Public Affairs
University of Missouri

Robert Shireman
Department of Political Science
University of Missouri
 Scholars studying policy implementation in traditional hierarchical organizations have

long been interested in the impact of redundancy on program performance. Specifically, they

have asked if a duplication of effort among agencies performing similar or competitive functions

is justified by an improvement in the quality of programmatic outcomes. Early work in this area

suggested that redundancy increases program effectiveness by reducing the errors and omissions

that might occur in single agency production. More recent work has been less sanguine,

suggesting that redundancy may actually decrease performance as all agencies gravitate toward

the standards of the least effective, or that the increased costs of redundancy are only justified by

improved performance under a relatively limited set of circumstances.

During the same period that scholars have been debating the impact of redundancy in

traditional policy implementation, we have learned that fewer and fewer programs are actually

implemented within such structures. Research now suggests that many (most?) public programs

and services are delivered through collaborative or networked arrangements between

government, private, and third-sector organizations, rather than through singular (or redundant)

hierarchical government agencies. Thanks to a large and growing body of research, we have also

learned a great deal about the impact of these network structures on the performance of public

organizations and programs, as well as the factors that condition that impact.

Interestingly, however, the impact of redundancy in network implementation has yet to

be explored.1 This oversight is particularly surprising because redundancy is at least as likely (if

not more likely) to occur in nonhierarchical and decentralized network settings, relative to a

more traditional intra-governmental implementation scenario. It is also surprising because of the

1
Redundancy as we describe it is different than the concept of network redundancy described in the private sector
management literature. That concept describes the degree to which an organization is a part of two or more
networks with overlapping or redundant members or purposes (see Echols and Tsai 2005) and not the degree to
which two organizations within a network are providing the same services to clients.

1
potential consequences of redundancy for network performance. As in traditional settings,

duplication of effort in networks inevitably produces increased costs that can only be justified by

a commensurate or super-commensurate increase in output quality. There are reasons to believe,

however, that such performance gains are likely to occur only under certain conditions.

This manuscript explores the impact of redundancy on organizational performance in

network settings and, more specifically, the network, organizational, and individual

characteristics that moderate that impact. We test assertions regarding the relationship between

redundancy and network performance in analyses of reentry/recidivism-prevention services

delivered to clients by organizations in 15 community corrections networks around the state of

Missouri in 2009/2010.

Networks and Effectiveness

Over the past several decades, public management scholars have become increasingly

focused on the implementation of public programs via interorganizational networks (see for

example Agranoff and McGuire 1998; O’Toole 1997). An important dimension of this work has

focused on the creation, structuring, size, and management of, as well as the interactions within,

these networks (see for example Agranoff 1991; Graddy and Chen 2006; Silva and McGuire

2010; Alter and Hage 1993; Sydow 2004; Teisman and Klijn 2002). Another body of work has

focused on the impact that networks have, relative to one another or traditional hierarchical

arrangements, on the organizations within them and the programs they implement (see for

example O’Toole and Meier 1999; Provan and Milward 1995). Because we are ultimately interested in

questions of effectiveness, we will focus primarily on this latter body of work.

The predominant assertion in this literature has been that collaboration is generally better than

isolated action and that coordination among service providers increases effectiveness through the

2
dispersion of information and expertise, the attainment of economies of scale in resource use,

and the ―catching‖ of clients who might be missed or fail to take advantage of programs offered

by a single provider (see for example Alter and Hague 1993; Provan and Milward 2001). The

majority of work in this area has focused on organizations, or nodes, within networks when assessing

performance. This work has suggested that, when holding inputs and other characteristics (including

internal management activities) constant, interaction with more nodes in the network increases the

performance of public organizations (see for example Meier and O’Toole 2001; Nicholson-Crotty and

O’Toole 2004; Rethemeyer and Hatmaker 2007). More recent work, in public management, has

posited a nonlinear relationship between network activity and effectiveness, suggesting that staff

resources and managerial quality moderate the impact of networking on organizational performance

(Hicklin et al. 2008).

Scholars have also suggested that the character of interaction between nodes, rather than simply

the volume, determines the impact of networking on the outcomes for individual organizations. The

private-sector management literature has long recognized for some time that ―embeddedness‖ within a

network also correlates with firm performance (see Borgatti and Foster 2003; Moran 2005).

Embeddedness differs from network activity, or ―centrality,‖ in that it is the strength of connections with

a subgroup of organizations in one part of the network, rather than the total number of connections that is

assumed to determine outcomes for an individual node (Wasserman and Faust 1994). Strong ties and

consistent interactions among a small group of organizations aids in the development of trust and

reputation, which are ultimately attributes that quality organizations can leverage to receive more

resources from the network. The public management literature has also explored the impact of

embeddedness on organizational performance within networks. Provan and Sebastian (1998)

identified the performance advantages of small group membership, similar to embeddedness, in

the mental health networks the studied. Schalk et al. (2010) find higher levels of satisfaction among

3
the graduates of colleges that have strong ties to cohesive subgroups of other organizations within the

network of Dutch teacher training colleges.

A smaller body of work has examined the performance of networks (or the service delivery

nexus) as the unit of analysis (Knoke et al. 1996; Provan, Fish, and Sydow 2007). Provan and

Milward (1995) find that integration and centralization among network nodes, external control of

network activities, and the stability of network partners all influence the perceived effectiveness

of mental health networks among clients. Provan and Kenis (2008) also argue that network

governance matters, but suggest that its impact on effectiveness is moderated by the fit between

structure and other network characteristics. Specifically, they suggest that more decentralized

forms—such as shared governance among nodes—is most effective in networks marked by low

membership, high trust, and relatively little need for network level competencies. Alternatively,

more centralized arrangements—with management by an external network authority at the far

end of that continuum—is best suited to larger networks, where nodes have lower levels of

experience/trust, and which must perform tasks that require higher levels of inter-node

coordination.

Redundancy and Effectiveness

Scholars of implementation first addressed the issue of redundancy more than 50 years

ago. The interest in the subject was driven by the recognition that the duplication of effort among

agencies existed in myriad corners of the federal government and, more importantly perhaps, that

such duplication was frequently encouraged through formalized interagency competition and

review procedures. Downs’ (1967) argued that such competition was likely to produce increased

organizational effectiveness when 1) both agencies were funded by the same budgetary

authority and 2) when the agencies were sufficiently distant organizationally to ensure that they

had distinct loyalties and did not temper their behavior for fear of retribution. Landau (1969)

4
similarly suggested that intentionally overlapping services could increase the performance of the

duplicate agencies, primarily due to the increased reliability and quality that arises from having

one agency check the outputs of another. Others suggested that redundancies might also increase

efficiency and lower costs by reducing the level of errors or mistakes (Landau 1979).

Initial accounts of potential benefits of redundancy for reliability have been empirically

tested and theoretically extended. Bendor (1985) formalized the relationship between

redundancy and reliability and provided empirical tests in the area of metropolitan transportation,

which largely confirmed the existence of a positive relationship. Others demonstrated empirical

support for the benefits of redundant systems in air traffic control and naval operations (LaPorte

and Consolini 1991; Rochlin, LaPorte, and Roberts 1987). Heinman (1993) extended the concept

of redundancy by considering the different types of error that public organizations might try to

avoid through redundant systems. Through an examination of the Challenger disaster, his work

demonstrates that redundant systems are better at preventing ―type II‖ administrative errors,

where an agency fails to take an action that is necessary, but are actually more likely to create

―type I‖ errors, where an agency chooses to act when it should not. Heinmann (1995)

demonstrates that redundancy is a superior choice for increasing reliability when agencies must

make nonprogrammatic decision, where variability and uncertainty are high. Alternatively, he

suggests that serial strategies are preferable when decisions are repetitive and routinized.

Others have also argued that redundancy only increases effectiveness under certain

conditions. Krause and Douglas (2003) suggest that gravitation toward performance of the least

effective organization eliminates the advantages of redundancy in many parallel systems.

Drawing on the theoretical perspective of Adverse Reputational Herding, they argue that, ―when

a parallel (redundant) agency performs at the same or lower level compared to an initial agency

5
did before the former bureau’s creation, the initial bureau will play to the level of its competition

by either exhibiting no change or a reduction in task performance quality‖ (Krause and Douglas,

2003). Ting (2003) also illuminates a number of conditions under which redundancy may not

increase effectiveness. In a game theoretical approach which allows for strategic action by

agents, he demonstrates that that redundancy can help the principal achieve a policy goal, which

is assumed to be effectiveness, when the that principal and the agents have divergent preferences.

The positive impacts of redundancy decrease markedly, however, when the principal and at least

one of the agents share preferences regarding outcomes. Redundancy may also not be

unnecessary if the principal can take away a single agent’s right to produce a good in the event of

poor performance.

Redundancy and Network Effectiveness

The literatures on network effectiveness and redundancy both offer conditional

expectations about the relationship between these concepts and effectiveness. In other words,

both suggest a potential positive impact on organizational performance under the right

conditions. Interestingly, however, there has been relatively little work attempting to integrate

these approaches to program effectiveness. Landau (1991) did envision the possibility for

redundancy in ―multiorganizational systems‖ 20 years ago, extolling the potential benefits of

organizing systems of ―loosely coupled‖ horizontally linked organizations with duplicate

jurisdictions and functions. He suggested that this type of system would be better at recognizing

and correcting errors in the provision of public services, making the system better than any of the

component parts. Landau’s (1991) account is suggestive and intriguing, but, to date, no one has

built upon it to offer a more systematic examination of the impact redundancy in network

implementation settings.

6
 Doing so requires that we 1) define network redundancy and 2) attempt to reconcile the

disparate literatures discussed above to understand conditions under which such redundancy

should improve network effectiveness. We define network redundancy narrowly as a condition

where a client receives a similar good or service from more than one organization within a

network of nonhierarchically linked providers within six months, when they could be receiving

that service from a single provider. There may be performance effects from simply having two or

more organizations offering the same services to different clientele within the network, but

restricting the definition to redundant services delivered to the same client better matches the

concept of parallel processing used to define redundancy in the more traditional implementation

literature. As a final definitional note, we are interested primarily in redundancy in consequential

(Graddy and Ferris 2006) or goal-oriented (Kilduff and Tsai 2003) networks, where

organizations purposefully come together and coordinate some degree of their action in order to

produce a particular outcome or accomplish a particular goal. Redundancies likely exist in what

have been described as ad-hoc or ―serendipitous‖ networks (see Provan and Kenis 2008), but,

because these often are not organized to produce a specific outcome it may be more difficult to

identify a metric against which ―effectiveness‖ can be measured.

With network redundancy defined, we move now to integrating the disparate literatures

on these topics. In order to do so, we take what we believe is a parsimonious and intuitive, if not

particularly sophisticated, approach. If the expectations of the network and redundancy

literatures regarding the factors that condition their impact on performance move in the same

direction, we expect redundancy to have a positive impact in a network setting. Alternatively, if

they move in opposite directions, we expect a null impact. In other words, if a characteristic

increases the likelihood that redundancy will improve effectiveness and increases the likelihood

7
that a network arrangement increases performance, then we expect redundancy within a network

to have a positive impact when that characteristic is present. Alternatively, if a characteristic

increases the positive impact of networks, but decreases the advantage of parallel systems, then

we expect those things to cancel one another out, producing a null impact for network

redundancy. In those cases where either literature is conflicted on the impact of a particular

variable on effectiveness, we do not offer an expectation about the direction or significant of the

relationship network redundancy and effectiveness under those conditions. By using the

individual receiving services from organizations in the network as the foundation for our

definition of redundancy, it allows us to specify moderating characteristics at the individual,

organizational, and network levels.

We begin with propositions regarding the moderating impact of network structure and

activity on redundancy. The literature on networks suggests that those organized around a central

node or authority tend to be more effective (Provan and Milward 1995), particularly when goal

consensus and the need for coordination among nodes is relatively high (Provan and Kenis

2008). The benefits of these structures accrue because they result in greater integration and

coordination of network activities (Graddy 2008). In a similar vein, the network literature

consistently suggests that ―centrality,‖ or the frequency of contacts with other nodes, is an

important predictor of success for individual organizations within a network (Nicholson-Crotty

and O’Toole 2004; O’Toole and Meier 2004; Schlak et al. 2010). Centrality (as opposed to

integration) increases performance by increasing the organization’s access to resources and

expertise and improving its ability buffer key operations from environmental shocks.

The literature on redundancy in traditional implementation scenarios also addressed the

issues of integration and communication among organizations, and it has evolved regarding the

8
expected impact of these factors. Downs (1967) suggested that coordination among units

eliminated the benefits of redundancy. Alternatively, Bendor (1985) demonstrated formally that

integration does not necessarily reduce the positive effects of parallel processing. Finally,

Landau (1991) provides some case evidence that coordination among ―loosely coupled,‖

redundant service providers is what allows that type of system to better identify and solve

problems relative to hierarchical or serial production systems. Because it seems as though the

expectations in the literatures on redundancy and networks move in the same direction, we offer

the following propositions related to redundancy and network effectiveness:

Proposition 1: All else being equal, the impact of redundancy on network effectiveness will be

more evident when a centralized authority works to integrate network activities.

Proposition 2: Redundancy will have a larger impact on effectiveness when an organization has

a higher level of centrality or communication with other nodes in the network.

There are also places where the literatures on redundancy and networks overlap in terms

of their expectations regarding effectiveness of organizations within networks. The first of these

that we treat deals with the issues of organizational capacity and/or quality. The literature on

network effectiveness originally argued that contact between network nodes was sufficient to

increase performance relative to hierarchical implementation (See O’Toole and Meier 1999).

Recent work suggests that the benefits of network activity may not accrue equally to all

organizations, and demonstrates that the staff resources of an organization and the quality of its

manager moderate the relationship between network activity and performance (Hicklin et al.

2008).

The literature on redundancy also emphasizes the importance of high quality/capacity

organizations. Indeed, work in this area consistently suggests that increasing the capabilities of

9
individual organizations is necessary to reduce error and improve reliability (see Landau 1969;

Heinman 1993). It simply argues that this approach is not sufficient to guarantee reliable outputs

and that overlap and duplication are, therefore, warranted. Interestingly, even work challenging

the desirability of redundancy indicates the need for high quality organizations. Krause and

Douglass (2003) argue that performance in redundant systems sometimes regresses to the level

of the lowest performing organization. This suggests that such system will be most effective

when both organizations have higher individual performance levels. Combining these arguments

from the literatures on redundant systems and networks, which again seem to move in the same

direction, we expect that:

Proposition 3: Redundancy will have a larger impact on effectiveness when the organizations

providing redundant services have higher capacity or quality.

Finally, we believe that similar predictions in the literatures on networks and on

redundancy suggest at least one individual level factor that may influence the impact of

redundancy within networks. The work on networked implementation suggests that it is

preferable to hierarchical arrangements in a variety of settings (Agranoff and McGuire 2008).

This seems to be particularly true, however, in the area of human service delivery, not only

because the incentives of providers are more properly aligned in such settings (Provan and

Milward 2001), but also because these are ―wicked‖ problems that require a multifaceted

approach (O’Toole 1997) and because they often serve the most ―vulnerable‖ clients (Aday

2001). This latter point reappears regularly in the literature with assertions that network

implementation is better at catching those difficult clients who are most likely to fall through the

cracks in traditional settings (Alter and Hague 1993; Provan and Milward 2001).

10
 Interestingly, though it was never explicitly been interested in individual client outcomes,

the literature on redundancy makes a similar argument regarding difficult problems. Initial work

suggested that redundant systems were preferable for most problems faced by government

(Downs 1967). Subsequent work demonstrated formally and empirically, however, that

redundancy was best suited to deal certain types of problems and less able to provide

effectiveness gains in others. Heinman (1993) argues that Type II errors, where government

failed to act when action was necessary, are particularly amenable to correction by parallel

systems. This becomes more true the greater the necessity of action. So, combining the findings

regarding problem type from the literatures on networks and redundancy, we expect that:

Proposition 4:Redundancy will have a larger impact on effectiveness for the most difficult

clients.

Empirical Test

We test hypotheses drawn from these general propositions in data from community

corrections networks in Missouri. Missouri maintains a relatively large incarcerated population

relative to its population (Pew Trust 2008) and in 2009, there were more than 100,000 persons on

parole or probation within the state (MODOC 2010). The challenge of reintegrating these

persons into society is well documented. Within 3 years of their discharge from the system,

approximately 67% of convicted felons will be rearrested. Forty-four percent will come in

contact with the justice system within 1 year of release. Of course, these rates vary dramatically

by offense severity and criminal history. A study by the U.S. Sentencing Commission found that,

among those convicted in federal court, the most serious offenders had a 1-year rearrest rate of

55% , while the lowest risk individuals reoffended at approximately 14% (CRS 2008). Despite

11
this variation, few dispute that offender reentry is a significant problem, particularly for

governments with punitive justice regimes.

In response to the scale and scope of the reentry problem, the state of Missouri has taken

a progressive approach to reintegration. It was among the first 8 states chosen by the U.S.

Department of Justice to develop a demonstration community corrections system modeled on the

Transition from Prison to Community Initiative (TPCI) developed by the National Institute of

Corrections and Abt Associates (Parents and Barnett, 2004). Lawmakers, judges, and

corrections administrators have taken a variety of steps creating this system. The first of these

was the creation of Transition Accountability Plans (TAP) (Parents and Barnett, 2004). TAPs

provide valuable information about offender needs to organizations that administer services post-

release. Assessment is done through a series of different programs that are administered in the

beginning of an offender’s time on probation, or at admission to prison and 6 months prior to

release.

The second step was the creation of Missouri Reentry Process Teams to coordinate the

activities of local public, private, and nonprofit organizations that might deliver reentry services.

This began with the formation state-level MRP Steering Team, which included representatives

from the Department of Corrections, Department of Mental Health, Department of Revenue,

Department of Elementary and Secondary Education, Department of Social Services, Office of

the State Court Administrator, Department of Economic Development, Department of Public

Safety, Department of Transportation and the Department of Health and Senior Services. This

steering committee oversaw the creation of 34 local MRP teams that are responsible for

coordinating reentry activities in 1 to 5 counties each. These teams work with parole officers and

service providers to coordinate reentry activities with the goal of decreasing recidivism rates in

12
the community. As an example these Teams’ activities, the District 11 team, which is responsible

for five counties, said in its annual report that it ―worked with the Central Ozarks Private

Industry Council Inc. (COPIC) in order to increase the effectiveness, efficiency and availability

of client employment referrals and opportunities.‖ These teams resemble, though are not

identical, to what scholars have identified as Network Administrative Organizations (See Provan,

Issett, and Milward 2004; Provan and Kenis 2008). They do not deliver services themselves, but

instead focus on building overall network capacity. The act primarily as a coordinator of network

activities, but may become more or less involved with individual programs as needed to produce

network goals.

The third step that the state took produced the data for this project. The Community

Reentry Funding Initiative (CRFI) was launched in 2007 with the supervision fee collected by

the Department of Corrections and the Division of Probation and Parole. The Initiative is

designed to provide funding to innovative community organizations and programs to assist

offenders as they reenter the community. The initiative focuses on programs that assist offenders

with housing, transportation, employment, mental health and substance abuse, education, basic

essentials, training, and a variety of other needs. The CRFI is intended to support local efforts

that will reduce the risk presented by offenders to commit new crimes and return to prison. In

2009, CRFI provided funding to 36 reentry providers across the state. These organizations

provide employment, housing, transportation and basic skills assistance, as well as cognitive

behavioral counseling. These organizations were funded with the understanding that they would

fit into community corrections networks already being coordinated (in most places) by a local

MRP team.

13
 As a component of an evaluation conducted for the Missouri Department of Corrections

and the National Institutes of Justice in 2010/2011, the Institute for Public Policy at the

University of Missouri collected data from the 36 funded organizations and the 4276 clients they

served. Organizational data include information on staff, budgets, funding sources, clients

served, and services offered, as well as communication with other reentry service providers,

parole officers, and MRP teams. Client data include information on employment, housing,

perceived needs, perceived support, and services received. We also collected information from

the MDOC on clients and all other parolees and probationers in the state in 2010, which included

demographics (race, sex, age, and marital status), risk assessment (from a standardized Field

Risk Reduction Instrument), offense type (violent, drug, etc…), sentence type (parole or

probation), recommended supervision level, and reoffense.

In terms of network structure, our data collection allows us to see pieces (smaller and

larger) of 15 community corrections networks around the state of Missouri. We know the basic

structure of these networks based on the guidelines established by the state-level MRP team.

Typically, each network includes multiple service providers, the parole and probation office

associated with the local district court, and the regional MPR team tasked with coordinating

community corrections activities in the area. We can observe the degree of integration at the

organizational level with data on MRP involvement in each organization’s delivery of services.

We can observe the activity of the 36 organizations we studied within these networks with data

on their contacts with other service providers, municipal government agencies, community

groups/members, and parole/probation officers. The greatest limitation of these data is that we

do not observe all community reentry providers within each network. This limits the questions

we can ask, but does not impair our ability to address the propositions discussed above, which

14
are all at the level of the organization or individual. Because we do not have data on every

possible organization from which an individual could receive services, there may be

redundancies that we cannot observe. These should, however, bias the results against our

hypotheses and, therefore, be unproblematic for the analysis. Nonetheless, we urge readers to

exercise the appropriate amount of caution given the imperfect nature of our data.

Dependent Variable

The dependent variable in all of our analyses is a dichotomous indicator of whether an

individual was rearrested within one year of entering parole or probationary custody. We choose

this measure because recidivism prevention is the primary stated objective of the Transition from

Prison to Community Initiative, the Missouri Reentry Project, and the Community Reentry

Funding Initiative. As such, it is the best measure of the effectiveness of the network of providers

created and/or coordinated by these efforts. The overall reoffense rate in our data is .14, though

that figure varies dramatically depending on the type of offender. The one-year recidivism rate is

only .05 for persons with a recommended supervision level of 1. Alternatively, it is .33 among

classified as Level 3, which indicates the most serious and/or repeat offenders.

Independent Variable

The primary independent variable in subsequent analyses measures whether an individual

received redundant employment, housing, counseling, or basic skills services. There are several

subcategories within each of these service types, including for example training, resume

development, GED acquisition, and job search within Employment, or short-term shelter versus

rental assistance within Housing. We do not directly observe which of these subcategories of

service a client receives, but we do know the range of services provided by the organizations

with which they are enrolled. In other words, we do not know whether Client A received resume

15
or training assistance from Organization X, but we do know that she could have received all 4

employment services from that organization.

The score on our measure of redundancy is determined by the services delivered by the

organizations that an individual is enrolled with, as well as the date on which he or she began

receiving services from them. If a person is enrolled contemporaneously with two organizations,

redundancy is coded 1 if they receive services in the same category from both, when either one

could have provided the service. As an example, if Client A simultaneously enrolls in

employment services from Organization X and Y when X is a comprehensive provider (all

employment services) and Y provides job training, we code that as redundant because they could

have received all services from Organization X. Alternatively, if Client B simultaneously enrolls

in employment services with Organization Y, which provides search assistance, and

Organization X, which provides training, that is coded as nonredundant because they could not

have received all services from either of the organizations individually.

When individual enrolls in two programs noncontemporaneously (but within 6 months),

the coding of redundancy depends most heavily on the services offered by the initial provider. If

the first provider offers all services offered by the second provider, we code this as redundant.

Alternatively, if the second provider with which a client enrolls offers a different basket of

subservices within a category, this is not coded as redundant. Using these decision rules, 477

individuals, or approximately 10% of those served, received redundant services in our sample.

As noted above, we code redundancy only for housing, employment, basic skills, and

counseling services. Transportation services are not treated as redundant or not because

numerous organizations assist clients with transportation to the services they provide. So, even if

16
an individual received transportation services from two providers it would not be redundant in

the same way that receiving rental or job search assistance from two organizations is.

Moderating Variables

We have identified 4 factors that should moderate the relationship between redundancy

and effectiveness in networks including integration, communication, capacity, and need. We

measure integration, at least to the degree it affects our organizations, with a question to service

providers regarding the ―involvement‖ that the local MRP team has with their program. Answers

range from 1 for no involvement to 5 for very high involvement. To provide some substance to

these numerical categories, one organization that marked a 1 said in an available comments field

that ―We do not know our MRP team.‖ Alternatively, an organization that marked a 5 said, ―The

local MRP team is very active and engaging. Attendance has remained high and many segments

of our community are represented.‖ We assume that higher involvement from the MRP team is a

good proxy for the level of integration and expect it to enhance the effect of redundancy on

recidivism prevention.

We measure communication between the organizations we study and other network

members (or centrality) by asking if they communicate regularly with other community

organizations, members of the community, government officials, and the parole/probation office.

The measure ranges from 0 to 4, with higher values indicating more communication. We expect

higher levels of communication to increase the impact of redundant service provision on

effectiveness.

We measure capacity with the number of paid staff. Most of the organizations we study

are 501(c)(3) service providers and there is evidence that staff resources correlate with capacity

in these types of organizations (Eisinger 2002; Hall 1998). Nonprofit organizations can serve a

17
relatively large number of clients relying heavily on volunteer labor. Volunteers certainly

provide valuable resources to the organizations they serve, but they are interchangeable with paid

staff for only a limited proportion of tasks (Handy, Mook, and Quarter 2008). In part, this is

because volunteers have significantly higher levels of turnover which affects continuity,

institutional memory, and agency morale (Fischer & Schaffer, 1993). We expect redundancy to

have a greater positive impact on effectiveness when staff capacity in the organizations

providing redundant services is high.

Finally, we measure the need of individual clients—or the relative importance that the

network act to help prevent their reoffense—with the supervision category. Based on the

characteristics of current and past offenses, as well as the Field Risk Reduction Instrument score,

which is comprehensive behavioral assessment designed explicitly to measure risk of recidivism,

each offender in the Missouri system is assigned a Supervision Category between 1 and 3. The

highest value represents those offenders who, by the state’s assessment, are at the highest risk of

reoffense and need the most significant intervention to prevent that from occurring. We suggest

that these offenders represent the greatest possibility for a Type II error in Heinman’s (1993)

conceptualization, where government must act in order to produce a desired outcome. According

to the author, these are the problems that redundancy is best suited to address. We expect that

redundancy will have a greater positive impact on effectiveness as the supervision level of the

individual offender increases.

Control Variables

Finally, subsequent models include a set of variables designed to control for alternative

explanations for recidivism. The first of these is race. Nationally, Hispanics are slightly more

likely to come in contact with the criminal justice system than Whites relative to their

18
population, while African Americans are significantly more likely to do so. More importantly for

our purposes, the recidivism rate among African American offenders is also significantly higher

(CRS 2008). Evidence suggests that these differences become far less evident when controlling

for other factors (see for example Florida Department of Corrections 2010), but we, nonetheless,

include a dichotomous indicator of whether an individual is white or nonwhite. The second

control variable measures sentence type. It is coded 0 for previously incarcerated offenders

currently on parole and 1 for those who were sentenced to probation rather than a prison term.

Even after controlling for other factors, incarceration dramatically increases the risk of reoffense

relative to those offenders who are allowed to remain in their communities. All models also

include the supervision score discussed above as a control. In our theoretical story the variable

moderates the impact of redundancy, but it is also the single best available predictor of

recidivism (U.S. Sentencing Commission 1992).

Subsequent models also control for age, based on the recognition that the likelihood of

criminal activity declines precipitously as an individual grows older (Florida Department of

Corrections 2010). We also control for offense type, based on evidence that certain types of

offenders are most likely to come back in contact with the criminal justice system. Nationally,

drug offenders have among the highest recidivism rates, because physical and mental addiction

make it very difficult for these individuals to abstain from illegal behavior (New York State

Commission on Drugs and the Courts 2000). All models discussed below include a dichotomous

indicator of whether or not an individual was convicted of a drug crime. Finally, we control for

the ―dosage‖ of treatment received by offenders from organizations in the community corrections

network. Prior to the data collection, organizations were asked to standardize their treatment

measurement based on a metric that we developed and all providers agreed to. So, as examples,

19
all organizations reported 1 hour of job readiness classes as 1 unit of employment, or 1 day of

rental assistance as 1 unit of housing, or 1 hour of substance abuse treatment/counseling as 1 unit

of counseling. All models include a measure of the total treatment units.

Methods

The community corrections organizations that we study treated 4276 of the more than

100,000 persons on parole or probation in the state of Missouri in 2009/2010. Unfortunately we

do not observe random selection of offenders into these organizations. In fact, sample selection is

based on a mix of self-selection by offenders, suggestion/assignment by parole or probation

officers, and recruitment by the organizations themselves. It is this decentralized selection

process that we suggested at the outset of this paper could produce redundancy in network

settings. Despite the seemingly stochastic nature of selection into these programs, however, we

know that certain offender characteristics, such as age, race, sex, sentence type, and supervision

level, predict selection into these programs. This complicates the analyses because many of these

characteristics also predict the likelihood of reoffense. This selection effect biases the findings in

most standard analytic techniques because any observed impact from a program or service is

likely to be a function of the characteristics of those selected into the program. In other words, a

program populated by older lower supervision offenders will correlate negatively with the

likelihood of reoffense, but the result simply reflects the fact that older nonviolent offenders are

less likely to reoffend.

The most appropriate way to deal with such bias is via a selection model and we employ

a Heckman approach. Generally speaking, the Heckman procedure corrects for sample selection.

More specifically, the model estimates the two equations simultaneously and takes advantage of

information from the individuals that select out of the sample in order to produce asymptotically

20
efficient and unbiased estimates for those that remain. In the case of reentry programs, the model

uses information about individuals that are not treated by our grantees to more accurately

estimate the probability of recidivism among those that are. In other words, it reduces the

likelihood that differences in the characteristics between offenders treated and untreated

offenders do not drive the findings. Because the dependent variable here is categorical, we

employ a Heckman estimator that fits a maximum likelihood probit model with sample selection.

Because we are interested in the impact of organizational and individual factors on

recidivism, as well as the interaction between these factors, we have an inherently multi-level

question. Typically this would suggest the use of a Hierarchical Linear Model or some other

multilevel estimator. Unfortunately, we are not aware of (and are not skilled enough to derive) a

model that can deal with sample selection in the first stage and estimate an HLM in the second.

As an alternative, we estimate a model with both organizational and individual variables and

cluster the standard errors at the organizational level. This does not provide all of the information

that an HLM estimation would, like the proportion of variance explained by different levels, but

it should produce unbiased results.

Findings

The results from Heckman probit models testing the propositions outlined above

presented in Tables 1 through 4. Collinearity, as well as the difficulty of getting the selection

models to converge when estimating multiple interactive parameters, compels us to test these in

4 separate models. We focus most of the discussion on the second stage results, presented in the

bottom panel of the tables. Before moving to those we can briefly discuss the first stage models

predicting selection out of our analytic sample, which are presented in the top panels. Because of

missing data in the information collected from the MDOC, we have 85, 185 usable observations

21
in the total sample. The sample in the second stage is 4276, indicating 80, 955 censored

observations. We predict censoring with race, age, sex sentence type, and supervision level. The

model suggests that parolees and those with higher recommended supervision levels were more

likely to select into, or be selected into, the programs funded by Community Reentry Funding

Initiative. In all 4 models, a significant Wald test suggests that the first and second stage

equations are not independent of one another and, thus, that the selection model is appropriate.

The upper panel of Table 1 presents the second stage results from the model of the

interaction between MRP team involvement and redundant service provision. Because we are

measuring the impact of integration in both organizations that provide a redundant service, the

model contains 2 measures of MRP involvement and 2 interaction terms, along with the indicator

of redundancy. The null coefficients on MRP involvement suggest that, in nonredundant

systems, increased integration also does not have an independent impact on the likelihood of

reoffense. The insignificant coefficient on redundancy suggests that, when both measures of

MRP involvement were 0, it has no effect on the probability of recidivism, though it is important

to remember that this condition is impossible because the lowest value of the MRP measure is 1.

The significant coefficient on the first interaction term indicates that redundancy becomes

significant as the level of MRP involvement in the first organization’s program increases. The

second interaction term is not significant either alone or jointly, indicating that, after controlling

for integration in the first organization, the level of MRP involvement with the second does not

moderate the impact of redundancy on recidivism.

In order to assess the direction of the impact of redundancy at higher levels of MRP

involvement it is necessary to examine the predicted probability of recidivism at different values

of the two variables and their interaction, with all other variables set to their means or modes.

22
We do so by calculating the change in predicted probability of going from nonredundant to

redundant services in conditions of low and high MRP involvement. The results suggest that,

when the level of MRP involvement is at the lowest level switching to redundant service

provision from nonredundant decreases the probability of reoffense by .004. Alternatively,

switching to redundant provision at the highest level of the MRP variable decreases recidivism

by .008. This represents a substantial change, and given that the baseline probability of offense in

this model is .072.

The control variables in this first model performed largely as expected. Parolees were

more likely to reoffend relative to probationers, with former incarceration increasing the

probability of recidivism by .04. Those with higher supervision scores also had a higher

probability of recidivism, with the likelihood of reoffense increasing by .042 from the lowest to

the highest category. Drug offenders were more likely than individuals that committed other

crimes to come back in contact with the criminal justice system, but the substantive effect was

inconsequential. Finally, total units of treatment correlated with a reduced probability of

recidivism, suggesting that even after controlling for other factors, a higher dosage of the

services offered in the community corrections network help to reduce the probability of

reoffense. On the other hand, when other factors were controlled for, race of the offender did not

have a significant impact on the likelihood of recidivism.

The upper panel in Table 2 contains the second stage model examining the moderating

impact of network activity on the relationship between redundancy and recidivism. The

significant coefficients on networking suggest that increased activity correlates negatively with

the likelihood of reoffense when an offender is not receiving redundant services. The null

coefficient on redundancy suggests that, in the absence of any network activity it does not have a

23
discernable impact recidivism. One of the interactions between networking and redundancy is

significant suggesting that redundancy begins to have an impact on recidivism as network

activity in at least one of the organizations providing redundant services increases.

Calculating the change in predicted probabilities suggests that the moderating impact of

network activity is quite substantial. With at least one organization that engages in no network

activity, redundancy actually appears to increase the likelihood of recidivism, though the change

is not statistically significant. Alternatively, if that organization engages in the highest level of

activity, redundancy decreases the probability of reoffending by a very large .038. Not

surprisingly, the control variables performed identically, with parolees, those that received fewer

treatment units, those with a higher risk score, and drug offenders having the highest probability

of recidivism.

The upper panel in Table 3 presents the findings from the analysis of staff resources and

redundancy in networks. In this case, both interaction terms are significant, suggesting that

greater staff resources in the second organization providing redundant services moderate the

impact of that redundancy even after controlling for the moderating impact of staff in the first

organization. The positive and significant coefficient on the measure of redundancy indicates

that it would actually increase the likelihood of reoffense if both organizations delivering

services had 0 paid staff. Of course, that condition does not exist in our sample, but even if we

calculate the change in the predicted probability of recidivism with both staff levels set to 10

employees, switching from nonredundant to redundant services increases that probability by

.004. Alternatively, if each organization has 50 full and part time staff, moving from singular

provision to redundancy decreases the likelihood of recidivism by that same amount.

24
 Finally, the results from the second stage model of supervision level are presented in the

upper panel of Table 4. The significant interaction term suggests that redundancy is moderated

by supervision, but the direction of the effect was different that we expected. At the lowest level

of supervision, indicating offenders with the lowest risk of reoffense, a change from

nonredundant to redundant services decrease the likelihood of recidivism by .006. However, at

the highest level, receiving services from two rather than one provider increases probability that

an offender will come back into contact with the criminal justice system.

Discussion

As noted above, the result regarding supervision level is contrary to our expectations. Our

assertion here was that the benefits of redundancy should be particularly evident for the highest

risk offenders, where the necessity of government action is highest. The findings suggest,

however, that it is these offenders who are actually harmed by redundant services. The direction

of this result holds, though not always the significance, if we use alternative measures for

―need.‖ More research is needed to better understand this result, but it may arise because of the

opportunity costs that redundancy imposes. Receiving the same service from two providers helps

to increase reliability, but it may also mean that an individual is forgoing another service that

might benefit them. It appears as if the more serious offenders benefit from exposure the varied

basket of services available in the community corrections network, rather than from the more

reliable provision of one of those services.

The remainder of the results provide support for the propositions outlined earlier in the

paper. Integration, captured at the organizational level by the involvement of the local MRP team

in program delivery, increases the positive impact of redundancy. We suggested that this should

be the case because integration has been shown to have a positive impact on network

25
performance (Graddy 2008) and because the redundancy literature has come to support some

communication and integration among loosely coupled providers in multiorganizational

networks (Landau 1991).

The benefits of integration in the case we study are particularly unsurprising because of

the nature of the networks we examine in this study. The majority of these networks are in

nonurbanized portions of the state and, typically, cover offenders in several towns across more

than one county. This produces a condition that should increase the benefits of more centralized

integration. Despite their distance from one another, offenders regularly receive services from

providers in different communities. Because of those distances, however, the need for a

coordinating entity that has a sense of the network as a whole, including which services are being

offered by which providers and on what schedule, is particularly high. Under these conditions,

we believe that redundancy may have greater benefits in more integrated networks because

offenders (or their POs) that are unsatisfied with a service being received from one organization

are more likely to be aware of programs in neighboring areas that can provide them with a

slightly different version of that service. Even in cases where clients ―stumble‖ into redundant

services, they may be more effective in integrated networks because the MRP team has worked

to make them complimentary.

The findings also provide strong support for our assertion that greater network activity by

an organization increases the benefits of redundancy. At the lowest levels of networking,

redundancy actually correlates positively with the probability of recidivism, but as such activity

increases it becomes negatively associated with reoffense. This result likely arises because more

active organizations have a better sense of not only what other organizations in the network are

doing, but also what community members and local government officials believe are the biggest

26
challenges facing offenders trying to gain employment, housing, or the skills and temperament

necessary to stay out of prison. If redundancies in multiorganizational networks are generally

better suited to identifying and solving service delivery problems relative to serial production

systems (Landau 1991), this should be particularly true when those organizations communicate

more regularly with others in the network about the nature of and solution to those problems.

Finally, our proposition that organizational capacity should increase the positive impact

of redundancy was also supported by the analysis. Network scholars have begun to argue that

capacity may moderate the positive impact of network activity and redundancy scholars have

always suggested that parallel processing among high skill agencies was the best approach.

Similarly, we find that redundant service provision by organizations that do not invest heavily in

human resources actually produces more negative outcomes for clients, while redundancy in

organizations with higher staff capacity reduces the likelihood that an individual will reoffend

(relative to provision by a single organization). In many ways, this result is intuitive simply

because having two organizations produce the same service when both have limited capacity to

provide it is unlikely to produce measurable performance gains. Indeed, in our model, it is not

just one, but rather both, redundant providers that must have above average staff capacity before

the benefits of redundancy begin to accrue.

Conclusion

We began this essay with the assertion that the characteristics of policy implementation

networks make redundancy and duplication among providers relatively likely and that it is

important to understand the impact of that redundancy on network effectiveness. We explored

that impact in an assessment of community corrections networks and demonstrated that

redundant service provision improves outcomes for clients when integration, network activity,

27
and staff capacity are high. We believe the findings suggest a need for the literature on networks

to seriously consider the topic of redundancy. Numerous questions persist about the impact of

redundancy on effectiveness and answering these is important for scholars who study

performance in network systems, but also for the policy makers who design them and the

organizations that deliver public goods and services within them. There are also myriad

questions for those that study the structure of and interaction within networks including: the

extent of redundancy in service delivery networks, the variation that might exist based on the

service being delivered or the size and structure of the network and the degree to which

redundancies appear by design, to name just a few.

While we believe the results of our analyses do warrant a closer examination of the role

of redundancy in networks, we also caution readers about drawing definite conclusions because

of the many shortcomings in this study. First and foremost, these data were collected for a

slightly different purpose than a study of network effectiveness and, therefore, do not always

provide perfect measures of the concepts we are interested in. This is also a study of only a

single service type delivered via a single type of network structure. Future research will

determine if the results we report here are robust to more precise measurement, different service

types, and more variability in network structure and governance.

28
 References

Aday, Lu Ann. 2001. At risk in America: The health and health care needs of vulnerable
populations in the United States. 2nd ed. San Francisco, CA: Jossey Bass.

Alter, Catherine, and Jerald Hage. 1993. Organizations Working Together. Newbury Park, CA:
Sage Publications.

Agranoff, Robert. 1991. ―Human services integration: Past and future challenges in pubic
administration.‖ Public Administration Review 51:533–42.

Agranoff, Robert, and Michael McGuire 1998. ―Multinetwork Management:

Collaboration and the Hollow State in Local Economic Policy.‖ Journal of Public
Administration Research and Theory 8(1): 67-91.

Agranoff, Robert, and Michael McGuire. 2008. ―Enhancing Performance Through Public Sector
Networks: Mobilizing Human Capital in Communities of Practice. Public Performance
and Management Review 31(3): 320-347.

Bendor, Jonathan B. 1985. Parallel Systems: Redundancy in Government. Berkeley: University

of California Press.

Borgatti, Stephen P., and Pacey C. Foster. 2003. ―The network paradigm in organizational
research: A review and typology.‖ Journal of Management 29 (6): 991–1013.

Downs, Anthony. 1967. Inside Bureaucracy. Boston, M.A.: Little, Brown.

Graddy, Elizabeth A. 2008. ―The Structure and Performance of Inter-organizational

Relationships within Public Service Delivery Networks.‖ presented at Consortium on
Collaborative Governance Mini-Conference, Santa Monica, California.

Graddy, Elizabeth A., and Bin Chen 2006. ―Influences on the Size and Scope of
Networks for Social Service Delivery.‖ Journal of Public Administration Research
and Theory 16(4): 533-552.

Graddy, Elizabeth A., and James Ferris 2006. ―Public-Private Alliances – Why, When,
and to What End? In Institutions and Planning, Current Research in Urban and
Regional Studies Series. ed. N. Verma U.K, Oxford, 175-185.

Heimann, C. F Larry. 1993. "Understanding the Challenger Disaster: Organizational Structure

and the Design of Reliable Systems." American Political Science Review 87(2):421-35.

Heimann, C. F Larry. 1995. ―Association Different Paths to Success: A Theory of Organizational

Decision Making and Administrative Reliability.‖ Journal of Public Administration
Research and Theory 5(1): 45-71.

29
Hicklin, Alisa, Laurence J. O’Toole Jr., and Kenneth J. Meier. 2008. ―Serpents in the Sand:
Managerial Networking and Nonlinear Influences on Organizational Performance.‖
Journal of Public Administration Research and Theory 18: 253-273.

Kilduff, Martin, and Wenpin Tsai. 2003. Social Networks and Organizations. London: Sage
Press.

Knoke, Davis, Franz Pappi, Jeffrey Broadbent, and Yutaka Tsujinaka. 1996. Comparing policy
networks. Labor politics in the U.S., Germany, and Japan. Cambridge, UK: Cambridge
Univ. Press.

Krause, George A., and James W. Douglas. 2003. ―Are Two Heads Always Better than One?
Redundancy, Competition, and Task Performance Quality in Public Bureaus.‖ presented
at the 15th Annual Conference of the Association of Budgeting and Financial
Management, Washington DC.

Landau, Martin. 1969. "Redundancy, Rationality, and the Problem of Duplication and Overlap."
Public Administration Review 29(4):346-58.

Landau, Martin. 1991. "On Multiorganizational Systems in Public Administration." Journal of

Public Administration Research and Theory 1(1):5-18.

LaPorte, Todd R., and Paula M. Consolini. 1991. "Working in Practice but Not in Theory:
Theoretical Challenges of High-Reliability Organizations." Journal of Public
Administration Research and Theory 1(1):19-47.

McGuire, Michael., and Chris Silvia. 2010. ―The Effect of Problem Severity, Managerial and
Organizational Capacity, and Agency Structure on Intergovernmental Collaboration:
Evidence from Local Emergency Management. Public Administration Review 70 (2):
279-288.

Meier, Kenneth J., and Laurence J. O’Toole Jr. 2001. ―Managerial strategies and behavior in
networks: A model with evidence from U.S. public education.‖ Journal of Public
Administration Research and Theory 11:271–95

Nicholson-Crotty, Sean, and Laurence J. O’Toole Jr. 2004. ―Public management and
organizational performance: The case of law enforcement agencies.‖ Journal of Public
Administration Research and Theory 14(1):1–18.

O’Toole, Laurence J. Jr. 1997. Treating networks seriously: Practical and research-based
agendas inpublic administration. Public Administration Review 57:45–52.

O’Toole, Laurence J. Jr., and Kenneth J. Meier. 1999. ―Modeling the impact of public
management: The implications of structural context.‖ Journal of Public Administration
Research and Theory 9:505–26.

30
O’Toole, Laurence J. Jr., and Kenneth J. Meier 2004. ―Desperately seeking Selznick: Cooptation
and the dark side of public management in networks.‖ Public Administration Review
64:681–93.

Provan, Keith G., Amy Fish, and Jeorg Sydow. 2007. ―Interorganizational networks at the
network level: A review of the empirical literature on whole networks.‖ Journal of
Management 33 (3): 479–516.

Provan, Keith G., and H. Brinton Milward. 1995. ―A preliminary theory of interorganizational
network effectiveness: A comparative study of four community mental health systems.‖
Administrative Science Quarterly 40(Mar): 1.

Provan, Keith G., and H. Brinton Milward 2001. ―Do networks really work? A framework for
evaluating public sector organizational networks.‖ Public Administration Review
61:414–23.

Provan, Keith G., and Juliann G. Sebastian. 1998. ―Networks within networks: Service link
overlap, organizational cliques, and network effectiveness.‖ Academy of Management
Journal 41 (4): 453–63.

Provan, Keith G., Kimberley R. Isett, and H. Brinton Milward. 2004. ―Cooperation and
compromise: A network response to conflicting institutional pressures in community
mental health.‖ Nonprofit and Voluntary Sector Quarterly 33:489–514.

Provan, Keith G., and Patrick Kenis 2008. ―Modes of Network Governance: Structure,
Management, and Effectiveness.‖ Journal of Public Administration Research and
Theory 18(2): 229-252.

Rethemeyer, R., Karl, and Deneen M. Hatmaker. 2007. ―Network management reconsidered: An
inquiry into management of network structures in public sector service provision.‖
Journal of Public Administration Research and Theory 17:1–30.

Rochlin, Gene I., Todd R. LaPorte, and Karlene H. Roberts. 1987. "The Self-Designing High-
Reliability Organization: Aircraft Carrier Flight Operations at Sea." Naval War
College Review (Autumn):76-90.

Schalk, Jelmer, Rene´ Torenvlied, and Jim Allen. 2010. ―Network Embeddedness and Public
Agency Performance: The Strength of Strong Ties in Dutch Higher Education.‖ Journal
of Public Administration Research and Theory 20: 629-653.

Sydow, J. 2004. ―Project-Based Organizations, Embeddedness and Repositories of Knowledge.‖

Organization Studies 25(9): 1475-1489.

Teisman, Geert R., and Erik-Hans Klijn. 2002. ―Partnership arrangements: Governmental
rhetoric or governance scheme?‖ Public Administration Review 62:197–205.

31
Ting, Michael. 2003. ―A Strategic Theory of Bureaucratic Redundancy.‖ American Journal of
Political Science 47(2): 274-292.

Wasserman, Stanley, and Katherine Faust. 1994. Social network analysis: Methods and
applications. Cambridge, NY: Cambridge Univ. Press.

32
Table 1 Interactive Effect of MRP Involvement and Redundancy on Recidivism

Coef. s.e.
Stage 2: Reoffense
Race -0.129 0.135
Sentence Type -1.140 **** 0.173
Supervision 0.285 **** 0.061
Age 0.000 0.004
Redundancy -0.117 0.161
MPR org. 1 0.020 0.023
MRP org. 2 -0.007 0.022
Redun X MRP org. 1 -0.070 ** 0.037
Redun X MRP org. 2 0.077 0.048
Drug Offense 0.046 * 0.025
Total Units 0.000 ** 0.000
Intercept -0.708 0.665

Stage 1: Selection
Race 0.187 0.128
Sex -0.096 0.099
Sentence Type -0.198 **** 0.040
Age 0.005 0.004
Supervision 0.268 **** 0.056
Intercept -1.941 0.579

Log pseudolikelihood = -17037.58

Wald test of indep. eqns., chi2(1) = 40.77 Prob > chi2 = 0.0000
N = 85231
Uncensored = 4276

Standard Errors Clustered on Organization. * p<.1, ** p<.05, *** p<.01, **** p<.000

33
Table 2 Interactive Effect of Network Activity and Redundancy on Recidivism

Coef. s.e.
Stage 2: Reoffense
Race -0.127 0.128
Sentence Type -1.151 **** 0.173
Supervision 0.287 **** 0.061
Age 0.000 0.004
Redundancy 0.013 0.041
Nework org. 1 -0.052 ** 0.026
Network org. 2 -4.072 **** 0.282
Redun X Network org. 1 0.019 0.030
Redun X Network org. 2 3.986 **** 0.281
Drug Offense 0.049 * 0.026
Total Units 0.000 ** 0.000
Intercept -0.604 0.680

Stage 1: Selection
Race 0.187 0.128
Sex -0.098 0.098
Sentence Type -0.198 **** 0.040
Age 0.005 0.004
Supervision 0.268 **** 0.056
Intercept -1.939 0.579

Log pseudolikelihood = -17035.38

Wald test of indep. eqns. chi2(1) = 52.18 Prob > chi2 = 0.0000
N = 85231
Uncensored = 4276

Standard Errors Clustered on Organization. * p<.1, ** p<.05, *** p<.01, **** p<.000

34
Table 3 Interactive Effect of Staff Resources and Redundancy on Recidivism

Coef. s.e.
Stage 2: Reoffense
Race -0.127 0.137
Sentence Type -1.143 **** 0.173
Supervision 0.286 **** 0.062
Age 0.000 0.004
Redundancy 0.083 ** 0.040
Staff org. 1 0.000 0.000
Staff org. 2 0.000 0.000
Redun X Staff org. 1 0.000 **** 0.000
Redun X Staff org. 2 -0.002 ** 0.001
Drug Offense 0.045 ** 0.024
Total Units 0.000 * 0.000
Intercept -0.666 0.670

Stage 1: Selection
Race 0.187 0.128
Sex -0.099 0.100
Sentence Type -0.198 **** 0.040
Age 0.005 0.004
Supervision 0.268 **** 0.056
Intercept -1.937 0.581

Log pseudolikelihood = -17036.32

Wald test of indep. eqns.chi2(1) = 45.54 Prob > chi2 = 0.0000
N = 85231
Uncensored = 4276

Standard Errors Clustered on Organization. * p<.1, ** p<.05, *** p<.01, **** p<.000

35
Table 4 Interactive Effect of Supervision Level and Redundancy on Recidivism

Coef. s.e.
Stage 2: Reoffense
Race -0.130 0.135
Sentence Type -1.147 **** 0.174
Supervision 0.276 **** 0.062
Age 0.000 0.004
Redundancy -0.200 0.136
Redun X Supervision -0.091 * 0.053
Drug Offense 0.048 ** 0.025
Total Units 0.000 * 0.000
Intercept -0.625 0.684

Stage 1: Selection
Race 0.187 0.128
Sex -0.097 0.097
Sentence Type -0.198 **** 0.040
Age 0.005 0.004
Supervision 0.268 **** 0.056
Intercept -1.940 0.579

Log pseudolikelihood = -17038.27

Wald test of indep. eqns. chi2(1) = 40.55 Prob > chi2 = 0.0000
N = 85231
Uncensored = 4276

Standard Errors Clustered on Organization. * p<.1, ** p<.05, *** p<.01, **** p<.000

36