Encounter data in resource management and ecology: pitfalls and possibilities

Authors

  • Aidan Keane,

    Corresponding author
    1. Centre for Environmental Policy and Department of Life Sciences, Imperial College London, Silwood Park Campus, Ascot, Berkshire, UK
    Search for more papers by this author
    • Present address: Department of Anthropology, University College London & Institute of Zoology, Zoological Society of London, London, UK.

  • Julia P. G. Jones,

    1. School of the Environment, Natural Resources and Geography, Bangor University, Bangor, Gwynedd, UK
    Search for more papers by this author
  • E. J. Milner-Gulland

    1. Centre for Environmental Policy and Department of Life Sciences, Imperial College London, Silwood Park Campus, Ascot, Berkshire, UK
    Search for more papers by this author

Correspondence author. E-mail: aidan.keane@ucl.ac.uk

Summary

1. Simple indices based on the number of encounters with a study object are used throughout ecology, conservation and natural resource management (e.g. indices of abundance used in animal surveys or catch per unit effort (CPUE) data in fisheries management). All forms of encounter data arise through the interaction of two sets of behaviours: those of the data generators and those of the data collectors. Analyses of encounter data are prone to bias when these behaviours do not conform to the assumptions used to model them.

2. We review the use of CPUE indices derived from patrol data, which have been promoted for the study of rule-breaking in conservation, highlighting potential sources of bias and noting how similar problems have been tackled for other forms of encounter data.

3. We identify several issues that must be addressed for analyses of patrol data to provide useful information, including the definition of suitable measures of catch and effort, the choice of appropriate temporal and spatial scales, the provision of suitable incentives for ranger patrols and the recording of sufficient information to describe the spatial pattern of sampling. The same issues are also relevant to encounter data more generally.

4.Synthesis and applications. This review describes a common conceptual framework for understanding encounter data, based on the interactions that produce them. We anticipate that an appreciation of these commonalities will lead to improvements in the analysis of encounter data in several fields, by highlighting the existence of methodological approaches that could be more widely applied, and important characteristics of these data that have so far been neglected.

Introduction

Data on the numbers of ‘encounters’ with a subject of interest are widely used in formal ecological studies to monitor spatial and temporal patterns of abundance (e.g. population census, distance sampling, Williams, Nichols & Conroy 2002). Similar data are also collected opportunistically, for example, through offtake records from harvested populations (e.g. commercial fisheries, Maunder et al. 2006; bushmeat, Rist et al. 2010). In conservation, encounter data derived from the reports of rangers patrolling protected areas (e.g. Leader-Williams, Albon & Berry 1990; Brashares & Sam 2005) or community based projects (e.g. Poulsen & Luanglath 2005; Stuart-Hill et al. 2005) are seen as an effective means of gathering a variety of data for natural resource management, including both poaching signs and encounters with species of interest (Arcese, Hando & Campbell 1995; Gray & Kalpers 2005).

The interpretation of encounter data requires assumptions to be made about how the number of encounters recorded relates to the true, underlying state of the study object, but these assumptions are easily violated. A large body of theoretical and empirical research has therefore examined processes leading to such violations, the extent to which models are robust to violations, and strategies for overcoming these issues (e.g. imperfect detection, MacKenzie et al. 2006; nonlinear relationships between the effort devoted to searching and the number of encounters, Maunder & Punt 2004; inter-observer differences and changing observer effectiveness over time, Sauer, Peterjohn & Link 1994). In ecological surveys, the risk of bias can be reduced through the use of rigorously designed sampling regimes (Williams, Nichols & Conroy 2002), but the avoidance of bias in opportunistically collected data relies on careful post hoc analysis.

In conservation, observations of rule-breaking collected opportunistically by enforcement agents in the course of their duties are seen as an attractive source of information about illegal activities because they are cheap and readily available. However, because the primary purpose of patrols is to deter rule-breaking, with data collection often a secondary concern, violations of common modelling assumptions are very likely and may be severe, for example, patterns of patrolling may be strongly nonrandom and rule-breakers are likely to alter their behaviour in response to the threat of patrols. Consequently, it is not always clear how such data can best be analysed, and whether they can be a useful source of information. These issues are clearly important for those wishing to use patrol data to inform management decisions. However, we believe that many biases present in patrol data also exist to a greater or lesser degree in encounter data used elsewhere in ecology and resource management. Thus, a better understanding of patrol data and its analysis has broader relevance to all those who use encounter data.

Here, we explore the use of patrol data as a source of information about rule-breaking, and use it to illustrate key features of encounter data. By analogy with the catch per unit effort (CPUE) approach used in the fisheries and bushmeat literatures, an obvious strategy for analysis is to treat the number of infractions detected per unit effort as an index of offences committed (Fig. 1). However, despite their intuitive appeal, the interpretation of such indices is not straightforward. Observed patterns in any type of encounter data result from the behaviours of two sets of actors (e.g. ranger patrols and rule-breakers; scientists and their target species) whose actions often violate analytical assumptions. Furthermore, the usefulness of encounter data depends on incentives for accurate reporting and the scale at which they are analysed. Analyses that neglect these considerations risk misinterpreting observed patterns. We conclude by proposing improvements to the collection and analysis of patrol data which address these difficulties, and suggesting how such insights could also improve the understanding and treatment of similar issues in other forms of encounter data commonly used in ecological and resource management settings.

Figure 1.

 An example of patrol data collected by the Cullman-Hurt Community Wildlife Project in Tanzania, showing (a) changes in the number of poachers arrested (shaded vertical bars) and patrol effort (solid line) between 1994 and 2003, and (b) the relationship between patrol effort (measured in patrol days) and the number of poachers arrested over the same period. The crude analysis in (b) provides no clear evidence that patrolling is a deterrent to poaching in this case, for example, the relationship could be reflecting changes in other factors over time.

Encounters per unit effort

Data collected by patrols share many similarities with those commonly used to study patterns of abundance in harvested populations (Table 1), and consequently, it is appealing to adapt methods from ecology and resource management to the study of rule-breaking. Measuring abundance directly is often difficult, so resource managers must base their decisions upon surrogate measures derived from changes in observed harvest levels over time (Milner-Gulland & Rowcliffe 2007). A common choice is CPUE. The use of CPUE as an index of abundance rests on the assumption that catch is proportional to both the abundance of the harvested population and the amount of effort invested in hunting (Hilborn & Walters 1992),

image(eqn 1)

where Ct,i is the observed catch, Et,i is the effort required to realise the catch and Nt,i is the size of the harvested population within area i at time t. q is a constant known as catchability. More generally, ‘catch’ could refer to encounters of any type, in which case Ct,i is the observed number of encounters and q is ‘detectability’ (because encounters do not necessarily result in capture), for example, in analyses of patrol data, the number of infractions encountered per unit of patrol effort has been used as an index of the number of infractions committed (Jachmann & Jeffery 1998).

Table 1.   A comparison between three common forms of encounter data and the extent to which assumptions are likely to be violated
CharacteristicEncounter-based ecological surveysFisheries and harvesting recordsEnforcement patrols
  1. +, violations are possible; ++, violations are likely; +++, violations are very likely; CPUE, catch per unit effort.

Data collectorScientists; research assistantsFishermen; hunters; harvestersEnforcement agents
Data generatorStudy speciesExploited speciesRule-breakers
Main aimAbundance estimation; occupancy modellingFishing; hunting; harvestingDetecting and punishing rule-breakers
Typical rationale for choice of routeMaximise statistical powerMaximise profitabilityMaximise encounters with rule-breakers
Encounters removed from population?NoGenerallyGenerally
Extent of violation of assumptionsEncounter-based ecological surveysFisheries and harvesting recordsEnforcement patrols
Accurate reporting by collector+++++
Generator does not respond to presence of collector+++++++
Perfect detection in sampled area+++++++++
Effort measures are appropriate+++++
Linear relationship between CPUE and abundance++++++++
Catchability does not vary+++++++++

Owing to the ready availability of catch data, CPUE has been among the most widely used indices of abundance in resource management, particularly in fisheries stock assessments (Hoggarth et al. 2006), but also in the bushmeat literature (e.g. Hill, McMillan & Fariña 2003). However, choosing an appropriate measure for each of the variables and for the catchability coefficient is not trivial.

What is the Appropriate Unit for Encounters?

In general, the definition of an encounter should be determined by a study’s objectives, for example, the number of elephants Loxodonta africana spotted along an aerial transect is an obvious unit for encounters if the aim is to estimate the size of that species’ population in an area (e.g. Jachmann 2002). In other cases, the unit of encounter might be an indirect sign of a species’ presence (e.g. the scats and footprints of American mink Mustela vision, Bonesi & Macdonald 2004; a burst of birdsong, Buckland 2006).

In patrol data, various indications of rule-breaking behaviour are commonly reported, including direct encounters with poachers or loggers and finding traps or snares, stumps of illegally felled trees or camp remains. Using patrol data to assess the effectiveness of enforcement measures as a conservation strategy (e.g. Hilborn et al. 2006) requires a decision about whether different types of infractions should be considered separately or analysed together. In general, analysing different types of infractions separately may be preferable if they are subject to different influences. However, related classes of infractions can act as substitutes for one another meaning that an observed reduction in one category of infraction might be accompanied by an unobserved increase in other categories of infraction, and need not reflect in an overall reduction in rule-breaking (Ehrlich 1996).

How Should Effort be Measured?

Data on numbers of encounters cannot be interpreted without a measure of the effort that produced them, but defining consistent and meaningful measures of effort is challenging (Bordalo-Machado 2006; Rist et al. 2008). Fishing effort, for example, is a compound measure which can include the area searched, the number of fishermen, the type of gear and other technologies used (Hovgård 1996; Sangster 1998) and the strategies employed to find fish (Hilborn 1985; Marchal et al. 2006). Similar effort measures are also used in studies of bushmeat hunting, including time spent hunting, distance travelled and number of hunters occupying an area (Rist et al. 2008). In both fisheries and bushmeat, the measurement of effort is further complicated if some of the apparent effort is directed towards areas or times that are unsuitable for the species or event being studied, for example, effective effort in longline fisheries for bigeye tuna Thunnus obesus is strongly dependent on the depth the gear reaches and the position of tuna in the water column (Bigelow, Hampton & Miyabe 2002).

Many measures of patrol effort have been used, including the number of times a grid cell is entered per unit time, distance patrolled or area surveyed per unit area per unit time (McShane & McShane-Caluzi 1984). Number of patrol days is widely used (e.g. Leader-Williams, Albon & Berry 1990; Jachmann & Billiouw 1997; de Merode et al. 2007), while other studies calculate ‘effective patrol man-days’ (Jachmann 2008), include time spent patrolling and number of scouts as separate predictors (Jachmann & Billiouw 1997; Holmern, Muya & Røskaft 2007), or employ cruder measures such as ranger patrols per day (Hilborn et al. 2006). Gaveau et al. (2009) simply distinguish ‘high’ and ‘low’ enforcement effort areas based on vegetation re-growth, interviews and models of accessibility.

What does the ‘Catchability’ Coefficient Represent?

Models of encounter data typically incorporate a coefficient that corresponds to the proportion of the target population that is encountered, on average, per unit of effort. In ecological surveys, this is the probability of detecting a species or event given that it is present in the area sampled. In distance sampling, detectability is generalised to be a function of the distance from the transect line to the sighting (Buckland 2001). In fisheries and bushmeat analyses, the analogous quantity is the catchability coefficient, generally described as the effectiveness with which a specific type of fishing gear or hunting equipment catches a particular species (Hilborn & Walters 1992). Similarly for patrol data, the detectability coefficient is related to the efficiency of patrolling and represents the probability of detecting an infraction per unit of patrol effort given that it is present within the area searched.

In each of these cases, the simplest models assume that catchability, or detectability, does not change over time or between areas. In reality, however, these coefficients incorporate the effects of many different influences on the probability of an encounter, including the characteristics of the data generator and data collector, their behavioural interactions and the temporal and spatial heterogeneity in environmental conditions (e.g. Arreguín-Sánchez 1996; Bibby, Burgess & Hill 2000); For example, patrol efficiency depends on the type of patrol (e.g. on foot, by vehicle and aerial surveys), expenditure on equipment, training and incentive payments, the morale and individual abilities of patrollers, and differences in terrain and weather conditions (Jachmann & Billiouw 1997; Jachmann 2008).

Interpreting patterns seen in encounter data

A key feature of all forms of encounter data is that they are the product of two sets of actors: a data generator and a data collector (Table 1). This is most obvious in the case of patrol data where the number of infractions detected depends upon the behaviour of both patrols and rule-breakers. In other forms of encounter data, this feature of the data has received less attention, perhaps because one set of actors is non-human, but it can still have important consequences. To successfully interpret encounter data, it is therefore necessary to understand how these two sets of behaviours interact.

How do Encounter-Based Indices Relate to the Size of the Sampled Population?

Referring to the use of CPUE measures in analyses of commercial fisheries, Hilborn & Walters (1992) note

‘The simplest assumption regarding the relationship between commercial catch and abundance is that the catch rate (CPUE) is directly proportional to abundance.... [This assumption] has been demonstrated to be wrong in almost every case where it has been possible to test – simply stated it is almost impossible for this relationship to be true.’

The causes of nonlinear relationships between encounter-based indices and the true size of the sampled population have primarily been studied in the context of fisheries, but many also apply to other types of encounter data (e.g. bird surveys, Bibby, Burgess & Hill 2000; bushmeat, Rist 2007). In the fisheries literature, the two general classes of nonlinear relationship between CPUE and true stock abundance are termed hyperstability and hyperdepletion (Fig. 2; Hilborn & Walters 1992), with hyperstability the more common (Harley, Myers & Dunn 2001; Lorenzen et al. 2006).

Figure 2.

 Two general classes of nonlinear, power-curve relationships between the infractions detected per unit effort and the total number of infractions committed (see eqn 2). Hyperstability describes relationships where the number of infractions detected per unit effort declines more slowly than the number of infractions, while hyperdepletion describes relationships where the number of infractions detected per unit effort declines more rapidly than the number of infractions.

Causes of nonlinearities are sometimes linked directly to changes in the size of the sampled population (e.g. when individuals differ in their detectability, those that are more easily caught will tend to be detected first and proportionally more effort will be needed to detect those that remain, resulting in hyperstability). A possible model for this is to incorporate a power curve into eqn 1 giving

image(eqn 2)

If β = 1, CPUE is proportional to abundance and the model simplifies to eqn 1. For 0 < β < 1, the model produces hyperstability and for β > 1 the model produces hyperdepletion (Fig. 2; Harley, Myers & Dunn 2001).

Nonlinear relationships can also be produced by changes in catchability over time, independent of the population size (e.g. improvements in rangers’ equipment or training can appear as hyperstability; spatial shifts in effort or learnt avoidance can appear as hyperdepletion; Walters 2003). In these cases, eqn 1 may be modified to allow q to vary over time and between areas:

image(eqn 3)

If not recognised, nonlinear relationships can have important consequences. In patrol data, hyperdepletion might encourage complacency, with managers overestimating the effectiveness of enforcement efforts, while hyperstability could result in overspending in the mistaken belief that rule-breaking is more prevalent than it really is. To date, we are aware of no studies of enforcement activities within conservation or resource management which have explicitly assessed the functional form of the relationship between the number of infractions detected per unit effort and the underlying number of infractions.

Do Encounters Affect the Behaviour of Data Generators?

All models of encounter data assume that the behaviour of data generators is not affected by the presence of data collectors, but violations of this assumption are common, for example, fish are known to move in response to fishing gear (e.g. Suuronen 1997; Handegard, Michalsen & Tjostheim 2003). Diana monkeys Cercopithecus diana have been shown to fall silent and retreat in response to the presence of humans (Zuberbühler 1997), and Adélie penguins Pygoscelis adeliae, red kangaroo Macropus giganteus and eastern grey kangaroos M. rufus all move in response to aerial surveys (Fewster et al. 2008). These types of behavioural responses could therefore be an important source of bias, but their consequences are rarely considered in analyses of encounter data.

The problem is particularly acute for patrol data because patrols are primarily a deterrent to rule-breaking. Therefore, an increase in patrol effort may produce both a decrease in the total number of infractions (because of deterrence or removal of rule-breakers through incarceration) and an increase in the proportion of those infractions that are detected (Fig. 3; Burton 1999). Deterrence is often treated as a single process, but changes in the recorded number of infractions actually reflect the aggregate effects of multiple behavioural responses to enforcement, for example, greater patrol effort might prompt poachers to make fewer hunting trips, or to cease to hunt altogether. Alternatively, they might substitute one type of infraction for another, trading off increased costs or lower efficiency for a reduction in the probability of being caught (e.g. switching to less conspicuous species or technologies, or changing their spatial and temporal patterns of hunting; Gibson 1995; Nyahongo et al. 2005).

Figure 3.

 Illustration of how a single relationship between catch per unit effort (CPUE) and effort could arise in different ways. In (a), increasing patrol effort increases detection and also produces a deterrent effect, leading to fewer infractions being committed. In (b), there is no deterrent effect of enforcement. However, both scenarios produce the same relationship between CPUE and effort. Hence, observed relationships between catch and effort may reveal little about underlying processes of interest (cf. Fig. 1).

Biases arising from the responses of data generators to data collectors cannot easily be detected and quantified using encounter data alone. To tackle this problem, CPUE measures in fisheries management are frequently validated against fishery-independent surveys (Hilborn & Walters 1992). However, alternative data sources must be carefully chosen if they are not to suffer from similar biases, for example, orange roughy Hoplostethus atlanticus move away from camera systems as well as fishing gear (Koslow, Kloser & Stanley 1995) and herring Clupea harengus actively avoid acoustic survey vessels (Vabø 2002).

Data Collectors’ Incentives

Another important, yet often neglected, consideration is the effect of incentives faced by data collectors. In the absence of appropriate incentives, data collectors in ecological surveys may be unwilling to invest more than the minimum required effort or, conversely, might try to anticipate what the ‘best’ results would be rather than accurately reporting their observations. Establishing the right balance of incentives for data collectors is therefore a critical part of survey design. Fishermen naturally have an incentive to fish as efficiently as possible because the profits from fishing relate directly to catch. However, the implementation of restrictions on the allowable catch can lead to misreporting of catches of both target species (Patterson 1998) and bycatch (Lewison et al. 2004). In conservation, the link between rangers’ rewards and the effort they invest, or number of infractions they detect, is not always clear. Indeed, rangers may face strong pressures to turn a blind eye to offences committed by friends, family or neighbours (Abbot & Mace 1999), face threats to their safety (Hart et al. 1997), or be involved in illegal activities themselves. All forms of encounter data may therefore be subject to accidental or deliberate omissions and falsification.

Well-designed management programmes can provide incentives for effective patrolling and accurate reporting, for example, the number of senior staff visits to ranger camps in Ghana’s National Parks was positively correlated with the amount of effort rangers expended on patrolling (Jachmann 2008). Similarly, Jachmann & Billiouw (1997) argue that increases in the payment of cash bonuses to scouts in the Luangwa Valley, Zambia, correlated with reductions in the numbers of elephants that were illegally killed.

Nonrandom Patterns of Sampling

Sampling regimes in ecological surveys are carefully designed to allow robust inferences to be drawn about the studied population. This generally requires samples to have been drawn at random, or at random from within defined strata. However, implementing statistically robust sampling designs is challenging (e.g. because of difficulties accessing randomly chosen sites), so effort is often directed towards areas where sampling is logistically feasible, or based on prior information about where encounters are likely. Nonrandom patterns of effort are also common among fishing vessels, hunters and ranger patrols, which all tend to concentrate their effort in areas where there is a high probability of encounters for the sake of ease or efficiency; For example, fishermen may invest in sophisticated technologies or use cues such as the presence of dolphins or seabirds to target areas that are suspected to contain more fish (Polacheck 1988).

Nonrandom sampling has several consequences for the interpretation of encounter data. When areas are chosen for ease of access, or based on prior information, the relationship between apparent effort and effective effort cannot readily be predicted, so appropriate measures of effort are difficult to define without extensive calibration (e.g. Burn & Underwood 2000). Nonrandom sampling can also mask important spatial patterns, for example, when the resources available to carry out ranger patrols are small relative to the area to be managed, difficult-to-reach areas may go unpatrolled for long periods, meaning that there is essentially no information about levels of rule-breaking in these areas (Fig. 4).

Figure 4.

 Distribution of patrol effort in Masoala National Park, Madagascar, between 2005 and 2007 (area of one grid cell = 9 km2). The shaded areas show an index of patrol effort, indicating that the patrol resources are concentrated around the periphery of the park, with darker grey shading representing more heavily patrolled cells. Large areas have not been patrolled at all during this period (white cells). This pattern may be efficient for producing deterrence but limits the value of the data for monitoring trends in rule-breaking because inferences cannot easily be made about unpatrolled areas.

If data are analysed at a sufficiently fine resolution to be able to distinguish between sampled and unsampled areas, potential biases can be avoided, but inferences cannot readily be made about areas that are not adequately represented (Walters 2003). If only aggregated data are analysed, however, differences in the unobserved patterns of sampling can confound their interpretation (cf. Fig. 1); for example, several studies have compared rule-breaking at the level of entire national parks over a number of years (e.g. Hilborn et al. 2006; Jachmann 2008). This approach is reasonable so long as the patrol effort is near-randomly distributed within the parks in question, but, if patrol coverage is patchy or inconsistent, apparent changes in the level of illegal activity might be caused by biases because of changing sampling or poaching patterns.

Spatial and temporal scale in analyses of encounter data

Encounter data are the result of complex, interacting processes occurring over a range of spatial and temporal scales, for example, forest clearance can affect whole parcels of land, and its effects remain detectable in patrol data for long periods of time. By contrast, individual poaching incidents are localised and once the hunter has left an area, little evidence may remain. Similarly, the spatio-temporal scales of rule-breakers’ behavioural responses to enforcement vary considerably, for example, lags of varying lengths can occur between increases in enforcement effort and any subsequent deterrent effect. If the higher level of effort is not maintained, deterrence may decay over time (Milner-Gulland & Clayton 2002). Small-scale avoidance behaviours, on the other hand, may change rapidly (e.g. hiding to evade detection by an active patrol). The scale at which encounter data are collected therefore has important consequences for how they can be used, and how they must be analysed.

As discussed, analyses of highly aggregated data risk drawing misleading conclusions and are also less likely to be informative for decision-making at relevant temporal and spatial scales for management. However, achieving high resolution requires greater sampling effort and interpreting data at finer scales may require sophisticated analytical techniques. Standard statistical techniques applied to encounter data, such as generalised linear models, assume that every data point is independent (McCullagh & Nelder 1989). However, at finer scales, this assumption is often violated, for example, if a hunter takes a route through a forest, setting snare traps as he goes, observations made close together in space will be more similar to one another than expected if they were independent. If not accounted for, autocorrelation can inflate the risk of false positive errors (e.g. Legendre 1993; Lichstein et al. 2002). In ecological surveys, careful design can minimise the potential effects of autocorrelation. When surveying clustered species or events, a common approach is to treat each cluster as a single encounter, then scale subsequent estimates according to the average cluster size (e.g. Rosenstock et al. 2002). Where data are collected opportunistically, autocorrelation cannot be reduced ‘by design’ but may be tackled through the incorporation of spatial covariates or explicit modelling of the autocorrelation (e.g. Nishida & Chen 2004).

Encounter data collected at finer spatial and temporal scales also often include large numbers of zero observations with no encounters, complicating statistical inference. In patrol data, large numbers of ‘true’ zeroes (e.g. areas in which there are no infractions) can occur if patterns of illegal activity are spatially autocorrelated (cf. Flores, Rossi & Mortier 2009). ‘False’ zeroes (e.g. areas where infractions were present, but remained undetected for some reason) can arise from imperfect detection (MacKenzie et al. 2006; Martin et al. 2005) or from inaccurate reporting (e.g. failure to report bycatch, Lewison et al. 2004). Explicitly modelling imperfect detection and autocorrelation can reduce the numbers of zeroes arising from these sources (e.g. MacKenzie et al. 2006; Flores, Rossi & Mortier 2009). In cases where the proportion of zero observations remains too large to be adequately represented by standard statistical distributions, ‘zero-inflated’ or ‘hurdle’ models are commonly employed (e.g. Minami et al. 2007; see Zuur et al. 2009 and Martin et al. 2005 for reviews).

How can the usefulness of patrol data be improved?

Improving the Recording of Patrol Data

The cheapest way to improve the usefulness of patrol data is to improve recording practices. The keeping of detailed, standardised records of ranger patrols has long been advocated (e.g. McShane & McShane-Caluzi 1984). Recording a greater variety of information about patrols would enhance our ability to distinguish between the many possible sources of variation within these data, for example, rangers naturally differ in their ability and motivation to uncover and report rule-breaking, and changes in the effectiveness of personnel over time can introduce bias into patrol records. Data collected by fishermen, fisheries observers and ‘citizen scientists’ suffer from similar problems (Sauer, Peterjohn & Link 1994; Thomas 1996). If properly recorded, many such sources of variability between patrols could be explicitly modelled as covariates in analyses of patrol data (cf. Punt 2000; Candy 2004). However, managers must also ensure that any requirement to collect additional data does not adversely affect patrol performance.

More accurate recording of patrol routes would help to determine whether there are spatial biases in sampling, and is essential for answering questions about fine-scale patterns of behaviour. Technological innovations such as GPS recorders and the CyberTracker system can help (Steventon 2002), but simple paper and pen recording systems can also be very effective if well designed and supported (e.g. the Event Book System; Stuart-Hill et al. 2005). In fisheries, the International Council for the Exploration of the Sea coordinates a large standardised data base containing information on commercial fisheries’ catches as well as trawl surveys and oceanographic information (http://www.ices.dk/). This and similar projects help to facilitate the use of fisheries data for stock management and research. Greater standardisation of the data collected by patrols, perhaps via the use of tailored data bases (e.g. BPAMP 2006), could enable large-scale comparisons of enforcement and illegal behaviour between different regions.

Understanding patterns and drivers of rule-breaking in conservation requires an understanding of how conservation measures affect resource users’ decision-making (Keane et al. 2008). Deterrence is a function of the probability of detection and punishment and the severity of punishment (Becker 1968), but many patrol records fail to track what happens to offenders once they are caught (Akella & Canon 2004). In practice, the actual punishment that a rule-breaker incurs may differ from the theoretical sanction and can vary considerably from case to case (Leader-Williams & Milner-Gulland 1993). As a result, it is difficult to infer the perceived risk involved in rule-breaking after the fact. To address this, there is a need for research into levels of knowledge and attitudes towards rules and enforcement measures to understand how such risks are perceived by potential rule-breakers (Keane et al. 2011).

Improving the Patrolling that is Done

Choosing an appropriate scale for the collection and analysis of patrol data involves a trade-off between the loss of relevant information at coarse spatial and temporal scales, and increased cost and analytical complexity at finer scales. With greater resources, considerable improvements in the usefulness of patrol data could be achieved by choosing sampling regimes to maximise the potential information that might be gained (e.g. stratifying patrol effort between different areas based on an understanding of human behaviour) or by adaptively managing patrolling patterns (cf. Thompson & Seber 1996). However, the benefits of this approach must be weighed against its costs (e.g. reductions in the deterrent effect of patrols). We are not aware of any studies that have asked whether ranger patrols can efficiently achieve multiple aims.

Improving the Analysis of Patrol Data

A number of techniques developed in the ecological literature could be readily adapted to the study of patrol data, for example, extensions of distance sampling and occupancy-based approaches are able to model spatially and temporally heterogeneous probabilities of detection (Buckland 2001; MacKenzie et al. 2006). Similarly, where multiple gear types or species are involved in fisheries, or catchability varies, CPUE measures are often ‘standardised’ by including relevant covariates in generalised linear models (Maunder & Punt 2004; Bordalo-Machado 2006).

Analytical approaches might also be improved by incorporating ideas from the social sciences. Traditional ecological analyses have been criticised for paying too little attention to whether they correctly identify causal processes (Ferraro 2005; Armsworth et al. 2009). Standard, regression-based approaches to patrol data assume that causality is strictly unidirectional, with the number of infractions committed being partially determined by the deterrent effects of patrolling, but not the other way around. In practice, however, this is rarely true. The principal aim of patrolling is to prevent rule-breaking efficiently, so managers commonly direct patrol effort towards areas where infractions are most likely. Consequently, at some scales of analysis, the level of patrol effort may be partly determined by the number of infractions committed in an area rather than being independently chosen.

Patrols and rule-breakers occupy heterogeneous landscapes, and factors such as ease of access may influence the decisions of both sets of actors about where they concentrate their effort. Consequently, areas that are easily accessed (e.g. near to paths or rivers) may be used more often by rule-breakers and also patrolled more often, potentially creating spurious correlations. These problems of endogeneity and selection bias are widely recognised in the social sciences (e.g. Maddala 1992; Kennedy 2001), so interdisciplinary collaborations in this area might prove to be particularly fruitful.

Validating Patrol Data With Alternative Sources of Information

Some of the difficulties of interpreting patrol data may only be overcome through calibration and validation against alternative sources of information on illegal behaviour. Hilborn et al.’s (2006) predictions about poaching based on changes in expenditure on enforcement correspond well with independent estimates of the buffalo Syncerus caffer, elephant L. africana and rhino Diceros bicornis populations in Serengeti National Park. A number of alternative approaches to studying rule-breaking are available to researchers, but comparisons between different sources of data on illegal behaviour remain rare (Gavin, Solomon & Blank 2010). There is also considerable scope for adapting approaches from other fields to develop novel ways of studying rule-breaking in conservation; for example, techniques from experimental economics offer opportunities to learn about individuals’ responses to threats of punishment or conditional rewards, the role of different institutional structures in legitimising rules and sanctions, the psychological effects of different enforcement regimes and the effectiveness of strategic dissemination of information about enforcement outcomes. So far, conservation has been slow to adopt these methodologies (but see Travers 2009).

Considering Rule-Breaking Behaviour in the Context of Wider Incentives

Ultimately, successfully interpreting data on how individuals respond to conservation measures such as ranger patrols requires approaches that treat enforcement as a part of a wider system, taking into account the myriad factors influencing individual choices (Ferraro & Pattanayak 2006; Keane et al. 2008). In Sumatra, for example, high international coffee prices increased rates of deforestation inside Bukit Barisan Selatan National Park, confounding the effects of law enforcement (O’Brien & Kinnaird 2003; Gaveau et al. 2009). It has also been shown that changes to the socio-political context of enforcement because of war or civil unrest can undermine its effectiveness (e.g. de Merode et al. 2007).

Conclusions

There is a strong desire within the conservation community to learn about and improve the effectiveness of our actions (Pullin & Knight 2001; Sutherland 2004). The enforcement of rules and agreements is widely recognised as being crucial to the success of conservation (Keane et al. 2008), and expenditure on enforcement consumes a large part of conservation budgets in many areas of the world (e.g. Jachmann 2008; Robinson, Kumar & Albers 2010). Patrol data sets are widely available and potentially valuable as sources of information about rule-breaking, but their complexity currently complicates their interpretation. With improvements to the collection and recording of patrol data and associated contextual information, and with the development of appropriate models to describe the underlying behavioural processes at work, this potential could be realised. However, if patrols are to become a practical source of information for the design of conservation interventions, the possible benefits of any changes must be traded-off against reductions in patrol effectiveness or increases in cost.

Here, we have argued that patrol data sets are just one form of a broader class of encounter data, and highlighted the considerable scope for improvements to their collection and analysis. As is so often the case, conservation can learn a great deal from the experiences of other disciplines. However, by highlighting the importance of understanding the behaviour of both data collectors and data generators, and other crucial sources of bias that might otherwise be neglected, we believe that a better understanding of patrol data also stands to benefit every field that relies on encounter data.

Acknowledgements

We are grateful to the Leverhulme Trust, Economic and Social Research Council (AMK) and a Royal Society Wolfson Research Merit Award (EJMG) for funding this research. We would also like to thank the Cullman-Hurt Community Wildlife Project, the Wildlife Conservation Society and the Association Nationale pour la Gestion des Aires Protégées, Madagascar, for kindly allowing the use of their data, and for their generous help and cooperation, and three reviewers for their comments on an earlier version of this manuscript.

Ancillary