Factors Influencing Farmers’ Participation in Contractual Biodiversity Conservation: A Choice Experiment with Northern Australian Pastoralists

Private landholders’ contributions to biodiversity conservation are critical in landscapes with insufficient formal conservation reserves, as is the case in Australia's tropical savannas. This study reports results from a discrete choice experiment conducted with pastoralists and graziers across northern Australia. The experiment was designed to explore the willingness of pastoralists and graziers to sign up to voluntary biodiversity conservation contracts. Understanding preferences for contractual attributes and preference heterogeneity were additional objectives. Such knowledge can increase effectiveness and efficiency of conservation programs by informing contract design, negotiation and administration. Random parameter logit modelling showed that of contract attributes, conservation requirement, stewardship payment, contract duration and flexibility in contract conditions significantly influenced choices. Land productivity was a significant factor as were attitudes. There was significant heterogeneity of preferences for all contract attributes. Models were run for best–worst scaling responses and the first preferences subset, with the latter model deemed superior. Latent class modelling distinguished four classes of decision‐makers and illustrated different decision heuristics. Conservation investment strategies, which offer farmers contract options that meet biodiversity requirements while accommodating heterogeneous attribute preferences, are likely to lead to increased participation rates. Complementary suasion efforts are also required which espouse the benefits that pastoralists derive from biodiversity and participation in voluntary conservation contracts.


Introduction
The tropical savannas of Australia are a vast area of grass-dominated landscapes that stretches across the north of the continent. The major land use is extensive beef cattle grazing on very large farms, which typically measure hundreds of square kilometres in size (Bortolussi et al., 2005). Tropical savannas also support an abundance of endemic plants and animals. The ecological condition and biodiversity value of Australia's tropical savannas has declined since European settlement due to a number of factors including grazing, changed fire regimes, and the introduction and spread of exotic plant and animal species (Lewis 2002;Woinarski and Ash 2002;Woinarski et al. 2007;Radford et al. 2014;Woinarski 2014). Tropical savannas are currently under-represented in the formal conservation estate, prompting calls for farmersthat is, pastoralists and graziersto be actively engaged in landscape-scale biodiversity conservation (Woinarski et al. 2007). This coincides with increasing interest by the north Australian pastoral industry to explore the provision of environmental services as one possible enterprise and income diversification strategy (Puig et al. 2011).
To date, participation of north Australian pastoralists and graziers in biodiversity conservation projects has been low (ABS 2011). This is likely due to a combination of paucity of such projects in northern Australia and ineffective targeting and design of past and existing projects (Hajkowicz 2009). Designing incentive programs that are both effective and efficient requires an understanding of, firstly, the financial resources required to incentivise a sufficient number of farmers to participate in on-farm conservation and, secondly, the way in which program and contract design and administrative features influence participation.
Choice experiments are a stated preference method and are ideally suited to help the design of new agricultural markets. A potential emerging market in northern Australia is for private biodiversity conservation services, whereby farmers receive a recurring stewardship payment in return for undertaking actions that support biodiversity. Policies and programs based on the notion of payments for environmental services (PES) have existed in other countries and parts of southern Australia for some time, but are a recent topic in a northern Australia context (Greiner et al. 2009a). Choice experiments have been used elsewhere to inform the design of agri-environmental and PES programs (Ruto and Garrod 2009;Espinosa-Goded et al. 2010;Christensen et al. 2011;Beharry-Borg et al. 2013;Broch et al. 2013;Kaczan et al. 2013). Very few and only locality-specific choice experimental studies with pastoralists and graziers have been conducted in northern Australia to date Adams et al. 2014) and consequently no comprehensive understanding exists of how the north Australian beef industry may engage with such programs to assist landscape-scale biodiversity conservation efforts. To fill this knowledge gap, a choice experiment was designed and implemented .
This study focuses on data analysis and presentation of results. In doing so, the paper pursues a number of objectives. It contributes to the body of empirical research using choice experiments to inform conservation policy. In particular, the research results generate the knowledge foundation for the design and tailoring of contractual biodiversity conservation to the distinct landscape and farm operating conditions of Australia's tropical savannas. The paper achieves this by exploring farmers' willingness to participate in contractual on-farm biodiversity conservation, preferences for and trade-offs between contract attributes, and the influence of socio-economic and Ó 2015 The Authors. psychological variables on stated participation. The paper also adds to the methodological discourse on choice experiments by comparing model results derived from the full best-worst scaling data set with results obtained from the first preference subset. The combination of mixed multinomial choice modelling and latent class modelling explores and explains the preference heterogeneity encountered among pastoralists and graziers.
The method section of the paper summarises the rationale for and key design features of the discrete choice experiment and choice modelling techniques employed. It also provides an overview of the study area and explains aspects of survey implementation. The results section provides relevant descriptive survey results and shows and interprets the choice experimental results. The paper concludes by discussing the results in the context of the international literature and offering conclusions as to the policy relevance of the findings.

Design of the discrete choice experiment
Choice experiments are based on the presumption that participants evaluate the alternatives presented to them in a choice task by choosing the alternative (among those presented in the task), which gives the greatest relative utility (Hensher et al. 2005). In the given research context, this means that a responding pastoralist will choose conservation contract A over B, if U (X A , Z) > U (X B , Z), where U represents his/her indirect utility function, X A the attributes of alternative A, X B the attributes of alternative B, and Z the personal (e.g. socio-demographic and attitudinal) and property characteristics (e.g. size, land productivity, farm profitability, ownership structure) that influence the pastoralist's utility. Choice analysis requires the inclusion of a stochastic error term e in the utility function to reflect the unobservable factors in the respondent's utility function. Thus, a pastoralist will choose alternative A over B, if V (X A , Z) + e A > V (X B , Z) + e B , where V is the measurable component of utility estimated empirically, and e A and e B reflect the unobservable factors in the pastoralist's utility function of alternative A and B, respectively.
There are a range of design decisions to be made, in particular relating to the expected make-up of the utility function, statistical properties of the experimental design, likely model to be used for data analysis and number of choice tasks, attributes and attribute levels defining an alternative, number of alternatives defining a choice, and response mechanism (Hoyos 2010;Bliemer and Rose 2011). All design decisions ultimately influence the results of choice experiments and resulting recommendations. A good design is able to explain more of the observed variance and minimises the stochastic element e. A detailed deliberation of all choice design considerations and decisions relating to the discrete choice experiment (DCE) for this study is given in Greiner et al. (2014). Table 1 offers a summary of other key choice design elements. Here, only the matters of response mechanism and model to be used for data analysis are revisited in detail as they bear direct relevance to the results. The response format was multiprofile 'best-worst scaling' (Marley and Flynn 2012). The more commonly used 'pick one' format only reveals the first preference among the alternatives while best-worst scaling is designed to deliver a full ranking of all alternatives contained in a choice set. By offering choice data in addition to the first preference, best-worst scaling offers a means of data augmentation and is particularly useful in situations where the sample size is expected to be lowas was the case hereor in situations where the number of choice tasks needs to be minimised (Potoglou et al. 2011;Lancsar et al. 2013). Best-worst scaling has also been found to be superior when dealing with qualitative data such as the different conservation requirements and different monitoring arrangements (Flynn et al. 2007).
Attributes and attribute levels were determined in a multistage process involving literature review, expert and industry consultation, and pilot testing. The final experiment is summarised in Table 2. The definition of conservation requirements was guided by (i) the idea of a multitenure reserve systems (Fitzsimons and Wescott 2008), which would see land taken out of cattle production and managed, by the pastoralist, exclusively for biodiversity conservation, and (ii) co-existence of grazing and biodiversity, based on the premise that conservation of many species of animals and plants is compatible with grazing, provided grazing land management respects the needs of these species (Woinarski and Ash 2002). Payment levels were guided by historical data about the land productivity of the tropical savannas, in particular the value of cattle sales per hectare during 1992-2011 as derived from farm survey data (ABARES, 2012) and industry comment.
The experiment adopted a Bayesian D-efficient design (S andor and Wedel 2001; Bliemer et al. 2009;Bliemer and Rose 2013). Priors were estimated from results of a pretest of the DCE. The design was optimised for random parameter logit (RPL) modelling of choice data. A final panel design was generated with 24 choice tasks being blocked into four versions of six choice tasks . Each choice task consisted of three contract alternatives and a 'none' option to mimic the voluntary nature of conservation contracts. Respondents were asked to pick their preferred option ('1st preference') and subsequently indicate the least preferred then second preferred option in any given choice task. This best-worst scaling format allowed for all contract alternatives and 'none' to be ranked. Figure 1 provides an example of one discrete choice task.
Prior to choice data analysis, protest respondents needed to be identified and removed from the analysis (e.g. Windle and Rolfe 2014). Protest respondents are respondents who had not engaged with the choice experiment, meaning their stated zero WTA was unlikely to be a reflection of their true WTA. A follow-up question was asked of those respondents who answered 'none' as their preferred alternative in all choice tasks to explore the underlying reason. If the reason was insufficient monetary incentive offered The contract stipulates and annual per-hectare conservation payment. The value shown is for year 2013 and the payment is indexed for the duration of the contract period, that is adjusted for inflation. The payment does not cover fixed costs: necessary infrastructure is paid for separately and upfront. Note: To enhance respondents' ability to assess the conservation payment in the context of their cattle enterprise, their business situation was established and an indicative value of per-hectare income from cattle production derived. Respondents were also prompted to consider the cost implications of each of the conservation Ó 2015 The Authors.
by any of the alternatives on offer, the respondent was retained in the choice data; if it was disagreement with the conceptual context of the valuation scenario, the respondent was removed from the choice data.

Survey of north Australian pastoralists and graziers
The objective of the research was to inform landscape-scale biodiversity conservation in Australia's tropical savannas by eliciting information from pastoralists and graziers. It was thus paramount to capture a geographically diverse and sufficiently large sample of properties in the empirical investigation. The research area ( Figure 2) included approximately one million sq km of tropical savannas (TS-CRC 2014), intersecting with the 'north live cattle export region' (Gleeson et al. 2012). The north live cattle industry is based on bos indicus type cattle, which are suitable for tropical conditions and resistant to ticks. Cattle graze on native pastures of low nutrient value, which typically carry between 2 and 10 head of cattle per sq km (PWC 2011). External monitoring: The administrating agency undertakes regular monitoring or contracts an independent provider for the task. Base level. Self: The pastoralist undertakes the monitoring but random spot-checks are conducted to validate results of self-monitoring. Each year the reports of 25% of program participants are validated.
The key market of cattle from this region is live export, predominantly to south-east Asia. The number of potential research participants was estimated to be <700 (Western Australia % 35, Northern Territory % 110, Queensland % 550) due to the generally large size of stations and consolidation of stations into larger pastoral enterprises. A recent survey of pastoral properties in the Northern  Figure 1 Illustration of a discrete choice task. Territory reported an average property size of approximately three thousand sq km across the Katherine and Barkly regions (DPIF 2010), which are also covered in this survey. To maximise opportunities for pastoralists to participate in the research and thereby maximise response rate and minimise participation bias of the sample (Wagner 2012), different response formats were offered to accommodate respondents' preferences and to capture decision-makers within business units. 1 A survey containing the choice experiment was conducted during April to July 2013, and 104 valid surveys were completed. Principally, the survey was administered face-to-face during visits by the lead researcher on pastoral properties or the company's head office. Visits were pre-arranged by telephone and all pastoralists and graziers who were prepared to participate in the survey and were available at a time matching the travel itinerary were interviewed. 2 In addition, research meetings with groups of pastoralists were conducted in association with industry and community events. Facilitation of meetings ensured integrity of the quantitative data in terms of independence of responses. In three instances, respondents completed the survey in their own time and provided the response by mail.
Responses captured approximately 15 per cent of the estimated sample frame with good coverage achieved in all three states/territories and a variety of property sizes and situations represented in the sample (Table 3). Total area coverage was approximately 250,000 km 2 , or about one quarter of the research area.

Choice data analysis
Prior to choice data analysis, four protest respondents (3.8 per cent, all from Queensland) were identified. As each survey contained multiple choice tasks , panel specification of the choice models was necessary. Panels were unbalanced because some respondents did not complete all six choice tasks and respondents who participated in the pretest and completed 12 choice tasks were included in the full sample. The design of survey and choice data sets are published (Greiner 2014). Data analysis was conducted in â NLOGIT 5 software (Econometric Software Inc 2012).
The increasing use of random parameter logit (RPL) and latent class (LC) models for the analysis of choice experiments in farming contexts has been underpinned by a recognition of the heterogeneity in farm conditions and farmer preferences, and the desire to make this heterogeneity relevant for 1 For corporation-owned properties, head office was initially contacted regarding participation in the survey. The CEO or GM decided who was best to complete the survey. In some instances, station managers were subsequently interviewed. In other instances, a relevant person at head office would complete the survey for either the entire property or agglomeration, or part thereof. 2 The lead researcher drove in excess of 25,000 km to visit pastoralists and conduct research meetings.
Ó 2015 The Authors. policy formulation (e.g. Jaeck and Lifran 2014;Schulz et al. 2014). RPL modelling was adopted as the principal data analytical method to establish the influence of contractual attributes on stated willingness to participate in contractual biodiversity conservation, trade-offs between them, and preference heterogeneity. RPL is a mixed multinomial logit model, which relaxes key assumptions constraining the interpretation of a multinomial logit (MNL) model, namely (i) IIDthat is, that unobserved effects are 'extreme value 1' distributed, independent and identically distributed, (ii) independence of observed choices and (iii) homogeneity of preferences (Hensher et al. 2005). RPL models take into account heterogeneity of the parameter values among respondents (Train 2000;Hensher et al. 2005;Marsh 2012;Mariel et al. 2013). Simulations are required in RPL models to provide parameter approximations. Parameters were estimated using 1000 Halton draws. In addition to the attributes, a number of covariates were included in both models as nonrandom parameters. Those found to be not significantly associated with choices in both models were deleted from the final models while significant covariates were retained. Table 4 details the variables contained in the final, parsimonious models.
To further explore heterogeneity of preferences among respondents, LC models were also estimated. LC models assign respondents into behavioural groups or latent classes, thus accounting for taste differences or different types of decision heuristics (Beck et al. 2011). Preferences are assumed to be homogenous within each latent class but differ between classes (Colombo et al. 2009). There is no common acceptance of the best criteria for determining the correct number of classes (Nylund et al. 2007) and parsimony and interpretability are key. Determining the final number of classes was therefore an iterative process, combining quantitative measures of model fit and meaning in a given context (Beck et al. 2011). RPL and LC models were run on the fully ranked choice data provided by the 'best-worst' scaling design. The same models were also run of a subset of the data which contained only the first preference choice made in each choice task. This is referred to as the '1st preference' data set, which mimics a 'pick one' choice design. Casewise deletion was used for missing covariate values, meaning that respondents for whom covariate values were not available were eliminated from the analysis. This resulted in 1643 and 598 observations being available in the fully ranked 'best-worst' choice data and '1st preference' subset, respectively.

RPL model results
The results for the RPL models for both best-worst data (i.e. full preference specification) and 1st preference data are summarised in Table 5. Both models delivered a good statistical fit as indicated by values of pseudo R 2 of 0.33 and 0.32, respectively. At a general level, considering coefficient direction and significance, the two models yielded very similar results. The ASC was positive and significant for both models meaning that there were preferences towards the 'none' option, which could not be explained by the variables contained in the model. Direction of attribute influence was consistent with economic theory. Higher conservation payments ('PAY') significantly increased likelihood of participation in conservation contracts. Longer contract terms ('YEARS') significantly reduced the likelihood of participation. The need for TOTAL EXCLUSION of cattle had a significant negative impact on likely participation, with the coefficient much larger compared to a requirement for a LONG SPELLING period each year. Both conservation levels were defined in terms of opportunity costs relative to short spelling of the contract area each year, which would not affect the cattle production from the contracted land ( Table 2). The 1st preference model indicated a significant disutility of LONG SPELLING while the disutility in the best-worst model was not significant. The inclusion of flexibility provisions into contracts, that is the right to negotiate contract suspension in exceptional circumstances, influenced stated uptake of conservation contracts significantly and favourably. External monitoring arrangements were found to be favoured by respondents over self-monitoring, but only the best-worst model found MONITOR to significantly influence preferences. From an interpretation perspective, the 1st preference model provided a better fit with the narrative that the lead researcher found emerging from the interviews, namely that opportunity costs from changes to grazing regimes mattered most for the 1st preference choice while monitoring arrangements played a minor role in 1st preference choices but were often considered in the subsequent choices of each choice task.
The lesser importance of monitoring compared to other attributes was also evident from stated attribute attendance. After completing the choice tasks, respondents were asked to quantify their attribute attendance: 'When making your choice decisions, how strongly did you consider each of the attributes?' The rating scale was 1 = never to 5 = always. Attribute attendance for MONITOR had a mean value of 3.07, with 15 per cent of respondents stating that they 'never' considered it and 23 per cent considering it only 'sometimes'. Mean attribute attendance of other attributes was as follows: CONSERVATION REQUIREMENT 4.54,PAY 4.37,YEARS 4.49,and FLEXIBILITY 4.18.
Of covariates defining the pastoral business, land productivity ('LANDPROD') was found to negatively and significantly influence participation probability, meaning that pastoralists grazing less productive land were significantly more likely to participate. Of covariates capturing social-psychological dimensions, respondents' intrinsic interest in biodiversity ('BIO-ATT') and favourable opinion of the concept of payments for environmental services as a policy mechanism ('PES-ATT') influenced participation positively and significantly. Covariates that were tested but not included in the parsimonious model because they were not statistically significant included respondent age, Ó 2015 The Authors. respondent education, property size, property ownership (corporate or family owned), previous participation in conservation programs and enterprise profitability.
The WTA estimates for the choice attributes were calculated as the negative of the ratio of each attribute coefficient to the price coefficient. Confidence intervals were estimated using the Krinsky and Robb (1986) procedure based on the unconditional parameter estimates and applying 1000 simulations to calculate the confidence intervals in this procedure. The results of the 1st preference model (Table 6) show that ceteris paribus, to get respondents to accept conservation requirements involving total exclusion of cattle from the contract area required an annual payment of approximately $11 per hectare per year while requiring them to spell land for long periods every year during times when biodiversity was sensitive to cattle grazing required a lower conservation payment of around $3.50 per hectare per year. Introducing flexibility provisions into the contract reduced the payment that pastoralists required by almost $6 per hectare per year. The inverse of this 'discount' in exchange for some level of contractual flexibility can be interpreted as a risk premium that respondents applied to contracts that they thought might limit the scope for farm-level responses to environmental variability.

Latent class model
A four-class latent class model supports the existence of preference heterogeneity for contract attributes and illustrates both the different   heuristics that respondents applied in the decision to participate in contractual biodiversity conservation and the diverse preference structures among respondents (Table 7). Common between all four classes was a significant positive influence of PAY at P < 0.001, indicating the central role of the level of stewardship payment in respondents' utility function and decision-making. Class 1, with a membership probability of approximately 32 per cent, had distinct preferences across all other contract attributes, which were very similar to the whole-of-sample for total exclusion, long spelling and flexibility, but had a higher time preference (average WTA for an additional contract year was $1.24/ha/year) and also derived a significant disutility from self-monitoring (WTA $4.85/ha/year). In contrast, class 4, with a membership probability of 33 per cent, did not distinguish between the different conservation requirements but focussed on the stewardship payment and contract duration. Probability of membership in classes 2 and 3 was <20 per cent each. Class 2 had a positive preference for long spelling arrangements (compared to short spelling), a strong dislike of total exclusion (equivalent to WTA $20.63/ha/year) but was insensitive to contract length. Class 3 associated a significant and high disutility with total exclusion and long spelling, and with longer contracts (mean marginal WTA of $0.30/ha/year). Model testing found no relationship between covariates and class membership probability, indicating that class probability was not related to sociodemographic or attitudinal factors.

Discussion
The research presented in this paper is based on a sample of 104 north Australian graziers and pastoralists. While the sample may appear small in absolute terms, it is large in relative terms as respondents collectively manage approximately 250,000 square kilometres of land, or approximately one quarter of the study area. The sample also succeeds in capturing the heterogeneity of business and socio-demographic conditions of pastoral enterprises across the tropical savannas. The sample size is sufficiently large to estimate all attribute coefficients at the level of P < 0.01 statistical significance because of the D-efficient choice experimental design, which requires a much smaller sample size than a random orthogonal design . A systematic review of discrete choice experiments based on design features and sample size by Bliemer and Rose (2011) supports the assertion. Limitations associated with small sample size arise, however, because the number of covariates that can be investigated in any given model is constrained because of loss of observations caused by casewise deletion of missing covariate values. This limitation was accepted in preference of imputation of missing values. RPL models were developed for the best-worst scaling data (i.e. full preference specification) as well as the subset of 1st preference data. The models yielded similar but not identical results and different WTA estimates for a number of attributes. Differences in estimates between the two elicitation methods have been reported in the literature before (Louviere and Islam 2008). In this case, there was virtually no difference for YEARS, but difference for some qualitative variables was up to 73 per cent although the direction of impact remained the same. The likely reason is inconsistency in the attribute trade-offs respondents made between the three choices required in each best-worst scaling task (Giergiczny et al. 2013). This assertion is corroborated by the lead researcher's observation that respondents found it increasingly hard to perform successive choices required by each choice task ( Figure 1). Consequently, the 1st preference data was deemed to produce more compelling models for an industry-wide analysis, which is the focus of this paper. The augmented choice data, based on full best-worst specification, may prove valuable for more in-depth analytical applications.
The model results about attribute preferences are consistent with expectations based on the literature on farmers' stated participation in environmental services programs elsewhere. Pastoralists and graziers require a greater monetary incentive to sign up to longer contract periods or alternatives causing higher opportunity costs, and they prefer flexibility Ruto and Garrod 2009;Espinosa-Goded et al. 2010;Christensen et al. 2011;Peterson et al. 2011;Yu and Belcher 2011;Broch and Vedel 2012;Jaeck and Lifran 2014).
Interestingly, the research does not find any statistically significant influence on the decision to participate in contractual biodiversity conservation contracts based on property size, whether properties are family-operated or corporation-owned, age and education level of pastoralists, and previous experience with conservation programs. This means that likely participation cannot be predicted on the basis of socio-demographic or business descriptors. The principal factors explaining participation choice across the northern pastoral industry are the contract attributes and foremost among them the conservation requirement and level of stewardship payment offered. However, it is important to note that pastoralists have heterogeneous preferences for contract attributes, as demonstrated by the RPL model results and illustrated by the latent class model results.
Across the industry, participation is distinctly positively influenced by favourable attitudes towards biodiversity and towards PES. This finding highlights the importance of complementing new PES-style programs with education and extension. Attitudes can be changed by explaining how PES work and illustrating the social co-benefits that voluntary conservation contracts can deliver to farmers (Greiner and Stanley 2013).
Introducing some level of contract flexibility significantly and positively influences contract adoption. Flexibility -defined in this context as possibly being able to give cattle access in exceptional circumstances to a grass resource that may have been built up through contractual biodiversity conservationgreatly reduces the premium that pastoralists seek for signing up to contractual biodiversity conservation. This finding needs to be seen in the context of the highly risky business environment that north Australian pastoralists operate in (Greiner et al. 2009b).

Conclusions
The research generates insights that can help enhance the emerging dialogue between the pastoral industry and potential investors in biodiversity conservation, including government, nongovernment conservation organisations and private businesses that wish to off-set their biodiversity impacts or appear conservation-minded.
This research establishes that the pastoral industry operating across Australia's tropical savannas would consider participating in contractual biodiversity conservation if suitable contracts were available. The choice experimental results support a narrative that decisions about participation in contractual biodiversity conservation are to a large extent driven by the financial merit offered by the contract options relative to costs-including opportunity and transaction costs, and costs associated with conservation action. These considerations are independent of property size, type of business ownership and operator age, but with clear consideration to risk aspects including contract duration and contractual flexibility.
Due to the very large size of enterprises, competitive allocation of contracts through mechanisms such as conservation tenders has limited applicability in the tropical savannas. Targeting pastoralists for participation in contractual biodiversity conservation needs to be based on identification of biodiversity assets. To maximise likelihood of participation of targeted pastoralists, a suite of contract options need to be offered. Contract conditions need to ensure that the requirements of target biodiversity are met while attribute options need to consider business-specific cost structures and the significant heterogeneity in preferences across the pastoral industry. Including some flexibility in contracts gives pastoralists more scope in responding to environmental variability and risk, and will increase program participation and reduce the conservation payments. Offering attribute choice, for example for monitoring arrangements, appeals to different personal preferences and thus also fosters participation.

Acknowledgements
Windle, John Rolfe, Matthew Beck and Michiel Bliemer for discussions on aspects of choice experimental design and analysis. Eddie Webber provided the map. I am grateful for the constructive feedback provided by two anonymous reviewers and the journal editor on the draft manuscript. An earlier version of this paper was presented at the 2014 conference of the Australian Agricultural and Resource Economics Society.