The CITES Programmes
Monitoring Illegal Killing of elephants (MIKE)
Central African Pilot Project
TECHNICAL REPORT No 1
RECOMMENDATIONS FOR A SURVEY DESIGN FOR THE CENTRAL AFRICAN FOREST REGION
Len Thomas, Rene Beyers, John Hart, Steve Buckland
Draft, 3 August 2001
1. SUMMARY and ACKNOWLEDGEMENTS
The objective of this report is to provide guidelines for the planning and execution of large-scale elephant surveys in the Central African forest. This report provides information to initiate extensive-scale inventories immediately.
We propose a multi-scale, design-unbiased, stratified survey in the forest zone using counts of elephant dung made from ground transects. The principal advantage of an unbiased survey design is that it guarantees an unbiased assessment of trends in elephant distribution and abundance across the entire survey region. At the same time the data are compatible with spatial modelling and other analytical approaches that can increase precision substantially, and provide a broader information base on elephants.
The lay out and results of the pilot project surveys, which covered three study areas, are presented and used to develop a proposed survey design. In the proposed design, transects will be systematically placed with sampling effort allocated using known distributions of elephants in the sub-region, and stratified by estimates of elephant abundance. Estimates of total required effort are based on estimates of variance of the pilot results. Areas where the occurrence and abundance of elephants are unknown would also be included in the survey sample.
Selected sites of significant elephant abundance that are also important for management, including designated MIKE sites, may be more intensively surveyed. These surveys will also utilize an unbiased design, and will be treated as separate strata in the overall analysis. The intensive sites will serve as bases for MIKE teams. In addition, we expect that monitoring programs at these sites will also include information on elephants from other types of surveys, which can be used in spatial modelling of elephant abundance, and provide data on illegal elephant killing (Law Enforcement Monitoring).
We illustrate this approach with a model design that covers the entire Central African forest zone. Guidelines for developing surveys in subsets of this zone (Study Areas) are provided, in relation to known constraints of access, available resources and limited field staff capacity.
The design we propose is a useful model for developing faunal monitoring programs over any landscape area where surveys will be stratified spatially, and where particular attention will be focused on a small and unrepresentative portion of the area which is of high faunal or management significance, but where knowledge of the status of the fauna over the total area is also desired.
We propose an incremental approach to the development of an elephant monitoring program in Central Africa. The proposed design can be modified in relation to constraints yet can serve as a foundation for further developments and improvements, including additions of new survey zones. The results will yield a representative assessment of elephant distribution and abundance at any spatial scale.
Many individuals assisted with the preparations, fieldwork and analyses presented in this report. In particular, we would like to thank and acknowledge the following: Fiona Underwood, Lee White, Peter Walsh, Richard Ruggiero and Andrew Plumptre who provided inputs for the design. Jean Marc Froment, Hilde VanLeeuwe, Leonard Mubalama, Esther Ntsame, Victor Mbolo, Yves Mihindou and Paulin Tshikaya all of whom made significant contributions to the realization of the fieldwork. The MIKE Pilot Project was funded by contributions of US Fish and Wildlife Service and Wildlife Conservation Society. Additional support was received from USAID/CARPE, IUCN, and IUCN/ Netherlands.2. INTRODUCTION
The general goal of MIKE is to monitor status and trends in elephant abundance, and illegal elephant killing over both space and time, and to determine what factors influence these, including CITES policy. This report and Technical Report 2 deal with monitoring elephant distribution and abundance. Monitoring of illegal killing, termed Law Enforcement Monitoring (LEM), will be treated in Technical Report 3. Site base maps, spatial modelling and reporting will be covered in Technical Report 4.
In proposing a survey design, we have assumed that the objective of MIKE is to develop a monitoring program that is representative of elephant populations across the entire range. This is to be achieved by a series of region-wide monitoring programmes throughout Africa and Asia.
Within the forested region of Central Africa (about 2,400,000 km2), it is not practical to survey elephants directly, so monitoring will be of elephant dung. Surveys on the ground involve standardized searches for elephant dung using line transects (Buckland et al. 1993) or combination reconnaissance ("recce") – transect methods (Walsh and White 1999). Despite problems (Plumptre, 2000), dung counts are the only practical alternative for large scale surveys of elephants in forest at present. Although conversions of dung counts to elephant numbers can yield highly variable estimates, the precision of estimates of dung density alone can be higher. Elephant population estimates derived from dung counts are correlated with independently derived estimates developed by other methods (R. Barnes, unpublished data).
2.1 Background to MIKE Survey Design in Central Africa
Global elephant range has been divided between two regions, Africa and Asia, and within each region, into sub-regional ranges. In Africa, sub-regional elephant ranges include western, southern, eastern and central ranges. Within Central Africa, DR Congo accounts for 57.7 percent of the total elephant range in the sub-region (Table 1)
Table 1. African Elephant Range in Central African Sub-Region (Data from African Elephant Data Base (Barnes et al. 1999) and expert opinion. See ANNEX 1).
Percent Sub-Regional Total
Central African Republic
In comparison with many other areas on the continent, where elephant populations have been highly fragmented for some time, and where elephants are now mainly restricted to protected areas and their vicinity, Central African forests still include a number of very large, mostly unobstructed, elephant landscapes. IN addition, significant elephant populations exist outside of protected areas. Nevertheless, the status of elephants in many areas of the sub-region is poorly known. Information on forest elephants is notably incomplete, and current elephant status in DR Congo, the largest elephant range state, and in a state of anarchy for the past five years, is in particular poorly known.
Essentially all of the originally proposed MIKE survey sites in the Central African sub-region are protected areas (See list in Annex 2). A monitoring program based only in these sites would be unlikely to provide a representative survey of elephant populations or of the factors affecting them.
Subjective placement of survey sites can lead to results with unknown biases, which would therefore be difficult to interpret. On the other hand, a completely random placement of monitoring sites can lead to an unnecessarily costly programme with wasted effort, relative to a more efficient scheme. A goal of the pilot project is to evaluate this problem and recommend possible solutions in a sub-regional scale survey design.
In this report, we propose a framework for the placement of sampling locations across the Central African forest zone. The justification and basic framework of this design are provided along with an outline of the formulae needed for optimising the sampling allocation. These inputs, along with information on survey costs, time and manpower needs, and the level of precision required in the final results, can be used by the MIKE programme to develop an optimised survey design. Methods of sampling at survey locations will be covered in a Technical Report 2.
2.2 Objectives and Overview of the Report
The Objectives of this report are to provide the following:An overview of survey design theory in relationship to MIKE objectives. An introduction to the overall design of the pilot project. The steps in setting up the large-scale survey. A model survey design for inventories of elephant populations in the Central African forest zone, using results of the pilot study. Guidelines to modify survey design where fieldwork is constrained by inaccessibility, low capacity and limited resources. Recommendations for research and additional inputs that will improve survey design. Recommendations for training and field team management to execute the surveys.
We begin this report with a review of design-based and model-based approaches to sampling. We then present a model design for the entire Central African forest region, based on available information on elephant range and stratified by estimates of relative abundance. We discuss how this design can be modified when access is constrained. We provide guidelines for defining sub-regional study areas to permit an adequate design when full sub-regional coverage is not feasible. We discuss how a strategy using site-based teams can be deployed to accomplish this in context of a nationally based elephant monitoring program. The paper concludes with recommendations for improving the efficiency of the current design.
3. DESIGN-BASED VS. MODEL-BASED APPROACHES TO SURVEY DESIGN AND ANALYSIS
It is impossible, and indeed ill advised, to attempt a complete census of most wildlife populations. Therefore, information on abundance and distribution will be developed by sampling the total population. In sampling, we select some part of a population to observe, and use our observations to make inferences about the whole population. For MIKE we will count elephant dung at a limited number of locations, and use the counts to estimate the abundance of dung (and by extension elephants) in the whole central African region. Two questions arise: firstly how best to determine what part of the population to sample – i.e. what sampling design to use; and secondly how to analyse the sample data once it is collected. These questions are linked, because decisions about the sampling design affect how the analysis can be done.
Broadly, there are two approaches to the design and analysis of sample surveys: design-based and model-based. These are summarized below, but are discussed in more detail in standard texts on sampling, such as Thompson (1992). Issues specific to forest elephant surveys are discussed by Walsh et al. (2000).
3.2 Design-based Surveys
In the design-based approach we select the part of the population to observe using a survey design that involves some element of randomness – for example completely random, or systematic random. Every part of the population must have some chance of being chosen. The different survey designs have known properties, and we use the properties of the design to tell us how to make inferences about the whole population. For example, if we used a completely random design to determine the locations for surveying elephant dung, then the mean dung density observed at our survey locations is an unbiased estimate of the mean dung density in the whole region.
The key advantage of the design-based approach is that we do not have to know anything about the underlying population or the factors affecting this. The validity of our estimates depends only on the design. The design-based estimates we commonly use can provide unbiased estimates, so long as the design has been correctly implemented. Another advantage is that the results of unbiased surveys can be analysed in standard, user-friendly software such as Distance 3.5. Thus data collected under this approach could potentially be analysed by MIKE staff and collaborators based in the region, providing they receive proper training and technical support.
One big disadvantage of a design based sampling programme is that we have to rigorously implement the survey design in order that the estimates we develop are valid. For example, if we run short of funds part-way through the survey and decide to abandon the sites that are difficult to access then our estimates are no longer valid.
The pilot project found that logistical difficulties, and physical impossibility of accessing some survey sites (due to warfare and insecurity) will likely constrain survey deployment at a sub-regional level. We provide guidelines on how to assemble survey blocks, or study areas, (essentially divisions of the sub-region) when a total sub-regional survey is not feasible. This exercise must be done before sampling starts, and should take into account available resources, time and accessibility in order to produce a valid design. While the design will ensure an unbiased sample within each study area, it will not be possible to extend the results to the entire sub-region. Nevertheless, study areas can be expanded as conditions and resources permit, and thus constitute the basis of a full sub-regional survey.
3.3 Model-based Surveys
In the model-based approach, we use a model of the population to make inferences from our sample data. For example, we may believe that elephant dung density is a function of habitat and is directly proportional to distance from nearest road. We fit this model to our survey data, and use the results to predict dung density.
The model-based approach has several advantages. The first is a sampling design issue: model-based approaches make few assumptions about how the data were collected. There is no need for a survey design based on randomisation (although randomly or systematically sampled, design-based data can be used). We would like to have information from sites that span the range of the variables in our model (e.g., from all major habitats and at various distances from roads), but these could be located in the most convenient places that fulfil these criteria. Such an approach is called "directed" or "purposive" sampling. Directed sampling is risky, however, because you can get very biased estimates if the model doesn’t contain all the important covariates.
Another advantage of the model-based approach comes from the modelling itself. By using variables of interest in the models, we can examine the data for effects of poaching and policy change. For example, if we find that dung density decreases close to roads everywhere, except in protected areas, we may infer that roads allow access for poaching (although strictly, such causation can only be inferred from a controlled study). In addition, the estimates from modelling may well be more precise than from a design-based estimate, because variation in dung density between locations is "explained" by the additional variables, rather than remaining in the error term.
The model based approach has several major disadvantages, however. The most important is that the properties of the estimate depend on the model alone. For example, if the model is correct, then the estimate is unbiased; if the model is not correct then the estimate may be badly biased and misleading. The problem here is that if the model provides biased results (and this is almost certain to be the case, given the dynamic nature of factors affecting elephants and how little is known of these), it will not be possible to analyse the same data post facto as if it were sampled in an unbiased design. (In contrast, data from design-based study can be used in a model design).
Another disadvantage is that the modelling of transect data is a major analytical exercise, and one for which relatively few people worldwide are qualified. Several aspects of the modelling are still subject of basic research, and there is currently no easy-to-use software available. Therefore, such analyses could not be done within the region without substantial technical support and investment in specialized training (to PhD statistics level).4. DESIGN AND RESULTS OF THE PILOT PROJECT SURVEYS
4.1 Design of the Pilot Surveys
The objectives of the pilot project with regard to elephant surveys were:
Table 2 gives a schematic overview of how the pilot surveys dealt with different spatial scales across the Central African region. On a regional scale three study areas were selected and within these areas different blocks or strata were selected representing different environmental conditions, e.g. protected versus unprotected, high versus low hunting pressure, close to human habitation versus remote, forest versus savannah.
It is important to note that the goal here was not to carry out a complete survey of these sites but rather to have representative blocks of the area with a high sampling density in order to study spatial variation at different spatial scales.
At the stratum level in the case of Lope (Figure 1 in ANNEX 3) and Odzala (Figure 2 in ANNEX 3) a grid was generated with grid cells 2"30’ by 2"30’ and sampling units were placed systematically on this grid. The location of the first sampling location was selected at random and subsequent sampling locations were placed such that the horizontal distances between the nearest segments of the sampling units were equal to the vertical distances between the nearest segments of the sampling units. In the case of Ituri (Figure 3 in ANNEX 3) transects that were laid out in 1994-5 were revisited to allow comparison between the two datasets.
In Lope and Odzala sampling units consisted of 5 transects of 200 m length each interspersed with 4 recces of 1 km length. Distance sampling was used on transects and dung was also recorded on recces without measuring perpendicular distances. Recce data were calibrated with transect data to provide estimates of encounter rates (see Technical Report 2). The original survey design proposed by Buckland, 2000 is listed in ANNEX 7.
TABLE 2. Schematic overview of Pilot Project survey designs at different spatial scales.
Regional study area
Study area: 14359 km2
Study area: 7858 km2
Study area: 14131 km2
Strata within study area
Sampling locations within Stratum
- systematic placement of sampling locations on grid with 2"30’ square cells
- 44 sampling locations total
- systematic placement of sampling locations on
grid with 2"30’ square cells
- 44 sampling locations total
14 locations surveyed in 2000 from 54 surveyed in 1994-1995. No grid
Transects within sampling location
1 km transect, 4 km recce at center of each sampled grid
1 km transect, 4 km recce at center of each sampled grid
- 1 km transect, 4 km recce at each sampling location
- 5 km linear transect at each sampling location
4.2 Results of the Pilot Surveys
Table 3 presents the results of the surveys using the 5 km recce transect design pooled across strata. Densities shown are mean dung densities per hectare. A more detailed analysis is presented in report 2.
Table 3. Dung densities and coefficient of variation for transects (pooled across all strata) at the three Pilot sites (long line transect data of Ituri not included).
95% CI (D)
K: number of samples, l: total transect length, n/l: encounter rate for transects per km, D: mean dung density per hectare, CI: 95% parametric confidence intervals, CV: coefficient of variation.
Estimates of dung disappearance and production rates necessary to convert dung densities to elephant densities are not well developed. The dung conversion factor is calculated as 1/mean defecation rate (in days) divided by 1/mean decay rate (in days).
Barnes et al. (1987) used a dung conversion factor of 0.0018 based on a mean defecation rate of 13 dung per day and a mean decay rate of 45 days. Hart found a dry season decay rate in Ituri of 55 days. Using the latter as a conservative estimate mean elephant densities for the three pilot sites can be estimated as follow:
0.9 elephants/ km2 for the Lope area
3.6 elephants/km2 for Odzala
1.9 elephants/km2 for Ituri
We consider these estimates tentative without better information on dung conversion factors.5. PROPOSAL FOR A MIKE SURVEY DESIGN IN CENTRAL AFRICA
We propose design-based surveys to estimate elephant abundance for the MIKE programme. This design will provide robust, unbiased estimates of elephant density and distribution, as well as trend information as the time series data are accumulated. The approach we suggest takes account of existing information on elephants to stratify sampling effort. We develop this survey design using the entire Central African Forest Zone as the study area. We then provide guidelines on applying this design to smaller blocks, and on the sampling of selected sites within these blocks that are of particular importance to elephants, including protected areas, and the originally designated MIKE sites.
We also recommend that additional data be collected to permit some preliminary modelling of factors affecting elephant numbers. This data would include Law Enforcement Monitoring (LEM) data, which will also be collected in the context of the MIKE programme. This data collection, however, should not be done at the expense of an unbiased survey of dung. These data for modelling would need to be of high quality, with established protocols and training of observers.
The use of a design-based sampling design has the additional advantage in that it offers the possibility of doing both design- and model-based analyses of the data that have been collected. This kind of strategy is applied in other situations where wildlife population estimation is difficult or where results may be controversial and critical for management, such as whale stocks (e.g., Hedley 1999). Statisticians commonly recommend the robust approach outlined above (Thompson 1992). Other relevant references include Manholland and Borkowski (1996), who consider good sample spread to be an essential component of biological surveys, and Hansen et al (1983), who present a compelling case for design-unbiased sampling designs.
5.2 Overview of Sampling Design
The study area for this design includes the entire Central African Forest Zone. We propose a systematic sampling design, based on a grid of sampling locations, with the grid randomly superimposed on the study area. We recommend the data be analysed as if they were collected from a random sample, although this will likely overestimate variance, producing a conservative analysis (Thompson 1992).
The design can be improved by stratifying sampling intensity of the survey region by expected abundance and to assign a greater density of samples in areas of high expected elephant abundance, thus producing an "uneven coverage design". Many studies have shown that variation in wildlife abundance also increases as density increases, and Walsh et al (2001) have shown this to be the case for elephant dung in central Africa. Therefore concentrating sampling effort in areas of greatest elephant abundance will increase overall precision.
Overall steps in establishing the sampling design are presented in Table 4, and described in further detail in the sections below.
Table 4. Overall steps in establishing the sampling design and allocation of sampling effort
Study Area: Defines the universe from which the sample is drawn. The study area must be identified in relation to overall objectives, but must take into account available resources and access. All sampling locations in the study area must be reachable by survey teams during the course of the study. The inferences we would make from the data collected apply to the study area. If we would take a subset of this area in which we would lay out the design then that subset would be our study area. Stratification of the study area is often done to decrease variance in density estimates or to have an unequal distribution of effort for other reasons (e.g. logistical). In this case we stratified our study area for relative elephant densities. We assigned very crude estimates of densities to strata. Ratios of estimates of strata densities rather than the absolute densities are important for allocating sampling effort. More effort is allocated to higher density strata. We stratified the area using a rational process that differed for each country. In this example we excluded MIKE sites from the extensive design since they may be sampled separately, and treated as separate strata (see intensive sites below). However they could easily be included in the extensive design if intensive sampling would not take place at these sites. Allocation of effort: a ratio of optimal sampling effort is calculated for each stratum based on relative density estimates of the strata (ANNEX 5). Optimal transect length and spatial configuration of transects at each sampling location is estimated using optimal allocation formulae and results from the pilot study (ANNEX 5 and Technical Report 2). Distribution and allocation of sampling locations in each stratum is determined using optimal allocation formulae (ANNEX 5). Sampling locations are placed on a grid in which the grid spacing is adjusted to accommodate optimal sampling density for each stratum. This is a function of available resources and tie for the survey, and the level of desired precision desired.
Note that several of the above steps in the sampling design are interdependent and that several iterations of this process may be necessary to achieve an optimal design.
5.3 Example of Proposed Extensive Design
Here, we present an example of the proposed extensive sampling design. This example is presented to demonstrate the design process, rather than as a suggested final design. All of the calculations presented below would be refined for the final design. Cartographic inputs for this design are presented in ANNEX 5 and summarized in Table 5
Table 5. Summary of steps performed in the producing a survey design for the Central African forest zone. Reference to cartographic inputs and outputs included in ANNEX 4.
1. Define study area
Determined elephant range and defined study area (Central African Rainforest)
Identified areas of high densities in the region: protected areas with probably high elephant abundance. Identified areas of definitely low abundance in Congo and DR Congo based on a buffer 25 km from major rivers and towns and some expert opinion. Identified strata in Gabon (documented expert opinion based on rapid assessment surveys), Cameroon, Equatorial Guinea and CAR (expert opinion).
Combined medium and high density strata because there was not a large enough high density area to support treatment as separate strata (i.e., number of sites in that stratum would be lower than the minimum 15-20). Included unknown but definitely low density areas in the low density stratum. Included unknown but possibly high density areas in the medium/high density stratum. Classified original MIKE sites in separate strata as Intensive Sites.
3. Allocate proportion of effort within strata
Estimated density for each final stratum and used optimal allocation formulae to determine relative sampling effort for each stratum (ANNEX 5)
4. Determine sampling effort at sampling sites
Defined total transect length at each sampling location using optimal allocation formulae (ANNEX 5)
5. Determine distribution and allocation of sampling units
Placed sampling locations on a grid with distance between locations (grid spacing) proportional to optimal sampling effort for each stratum.
The proposed extensive survey program consists of a systematic grid of sample locations, located throughout the survey region. The region is divided into discontinuous strata, of anticipated elephant abundance with a different density of sample points in each stratum. The strata where samples will be located are a medium/high elephant density (hereafter called high density) stratum and low elephant density stratum.
Formulae for optimal allocation of survey effort are given in the ANNEX 5, and these formulae are applied to the pilot data.
The predicted optimum ratio of relative density of points in the high:low density strata is 1:0.44. This implies, for example, that if the grid spacing of sample locations in the high density stratum is 80 km, then the grid spacing in the low density stratum should be (802 / 0.44) 0.5 = 119.6 km. Because the strata are intermingled, it is useful to place the sample locations for all strata on the same grid, varying the spacing to achieve close-to-optimal sampling intensity. In the above example, it seems reasonable to place the low density sample locations on a 120x120km grid. This means that the minimum distance two samples from different strata is 40km.
In ANNEX 5, we discuss the assumptions required to derive the optimization formula. If the exact optimal spacing cannot be achieved in the low density stratum, it is better to round the grid spacing for this stratum up, so that the sampling intensity in the low density stratum is less than that suggested by equation (3), and slightly closer to that of equation (7).
A purpose-built software tool has been written to produce surveys, based on specified grid spacing for each stratum. The output from one run is shown in ANNEX 5. At the above sampling intensity, the expected number of sampling locations is 77 in the high density stratum and 63 in the low density stratum (140 in total). The number of samples in each instance of the design will vary slightly, as the grid is given a random starting point. In the example, there are 137 sample locations, 75 within the high density stratum and 62 within the low density stratum.
The model survey plan demonstrates a number of important points. Firstly, there is one isolated point in the extreme south of the survey area, in a small patch of high-density stratum embedded in a large area where there are no known elephants. Small areas like this contribute almost nothing to the overall estimate, but isolated points are likely to be different in character to other points in their stratum and so contribute to the variance of the estimates. Consideration should therefore be made to rationalizing them. Secondly, the density of points is high enough that portions of the study area can be missed in some years and there will still be enough left to allow robust estimates of density in the remaining areas. For example, if it were not possible to survey in the DRC in one year, there are still 14 points left in the high density stratum and 35 in the low density stratum. Note that it would then not be possible to compare trends over time across the whole study area, but only in the subset of area that was covered in all years. Ideally, the minimum sample size within a stratum is 15-20 points. See further discussion of creation of survey blocks, above.
The estimates of the potential precision of the density estimates under this design are approximate as that part of the total variance contributed by dung encounter rate is based on a sample size of 3 sites, none of which were selected at random from the entire survey area (see ANNEX 5 for details of estimation). Secondly, we have assumed that each sample location comprises just one 5km recce-transect, as in the pilot survey. Lastly, precision can be markedly improved by spatial modeling (e.g., Thomas and Buckland, unpublished)..
Nevertheless, using the methods outlined in the appendix, the expected estimate of density pooled across all three sites is 0.93 dung piles/ha. The predicted CV using only the transect part of the recce-transect is 32.7%, and using both recce and transect data is 30.6%. From equation (8), the approximate percent change in abundance detectable is 90.6% for transect only data and 84.7% for recce-transects.
These CVs could be reduced by increasing the sample effort at each sampling location. Additional effort will have the greatest effect there if it is spaced beyond the range of short-distance autocorrelation in encounter rates (see Thomas and Buckland unpublished aqnd Report 2). For example, four, 5 km recce-transects could be situated in a star pattern centered around the sampling point, but starting 1.5km from that point, so that they are spaced 3km apart. Assuming that there is then no correlation among these counts, the predicted CV becomes 16.3% for transect only data and 13.3% for recce-transect data. Trend sensitivity is then 45.2% for transect only data and 36.8% for recce-transect data.
Several configurations of transects or recce-transects are possible. Based on analysis of pilot data (Technical Report 2), we recommend a design consisting of 4 one kilometer transects separated by three kilometers each. The orientation of these can vary. One preferred example is given in figure 1.
FIGURE 1: example of transect placement at a sampling location
We stress that the optimization algorithm is designed to produce an allocation of survey effort to obtain the best precision of the design-based estimate. If the assumptions used are incorrect, then the allocation will not be optimal (in that the variance will be higher), but the estimates will not be biased. For example, a design which allocates 120 of 140 samples to the low density stratum and only 20 to the high density stratum is far from optimal. However, the expected estimate of density is unchanged. The predicted CV for transect only data is increased from 32.7% with optimal allocation to 50.3%, and for recce-transect data from 30.6% to 47.1%. So, it is a good idea to get as close as possible to the optimal allocation, but not essential for unbiased estimation.
5.4 Integration of Existing MIKE Sites in the Survey Design.
Originally, 16 MIKE survey sites were designated for Central Africa. Others have been proposed or anticipated. Twelve sites are located in the Forest Zone (See ANNEX 2). We recommend that current MIKE sites be retained as intensive sites only if they have high elephant abundance, or are of interest for other reasons. Also, intensive site surveys should not divert resources from the extensive program. If an intensive site is not surveyed in any particular year, the grid of extensive points can be extended to cover the intensive site, and these extensive sample locations can then be included in the extensive survey.
The geographic limits of the intensive sites can be defined based on the interest of site-based partners, or the mandate for the site. Once geographically defined, this area would be removed from the zone to be sampled by the extensive surveys.
Inventories at intensive sites will be design-unbiased, as in the extensive surveys. The design for this would be a sampling program similar to that developed and tested in the MIKE pilot sites (systematic placement of transects or recce-transects). Spacing of survey locations (grid size) can be determined based on resources available and objectives of site-based partners. These designs could be stratified based on the individual characteristics of each site (e.g. for estimated relative elephant densities or access). The design at each site need not be the same, and the same formulae as for the extensive program can be used to optimize the allocation if strata are employed. This could be produced by the local site officer. With adequate sampling density, design-based estimates of abundance of elephants can be provided for each site. Results from each site would constitute a separate stratum in the overall analysis.
The intensive sites also offer opportunities for model-based estimates of abundance. The effects of covariates related to poaching could be investigated at the local level. In addition, intensive sites would offer opportunities to develop a broad range of additional information concerning elephants at these local levels that will aid in interpreting results of the extensive-scale surveys.
We emphasize that while these intensive MIKE sites would be extremely useful in the MIKE program, they are no substitute for the extensive set of locations outlined above. If potential intensive sites were not able to participate in the programme then they should be included in the extensive programme instead.6. DESIGNING SURVEYS UNDER CONSTRAINING CONDITIONS
During the pilot project survey teams encountered numerous problems in gaining access to designated survey locations. These difficulties were due to logistical constraints (lack of transportation or supply lines), but they were also due to security concerns in conflict zones. It is clear that these same constraints will prevail in the next phase of MIKE. Often these constraints are not possible to predict in advance, and may change (sometimes rapidly) over time. A rational process to allocate inventories in a subset of the entire survey region (all of central Africa) is needed to allow constraints to be met, at the same time maintaining the analytical integrity of the design.
We recommend creation of what we term "study areas". These would be large contiguous areas (at least 50,000 km 2) where accessibility is not precluded from any area of the survey block. These areas would also contain a wide range of conditions within the range of elephant occurrence, including projected high and low density zones, and both protected and unprotected areas. It is important to emphasize that inventories must be conducted at all sampling locations in the study area, as described in the survey design section above. Study areas could be designated to correspond with regional geography (national or provincial borders), and this may correspond with the national elephant programs that will execute MIKE in the next phase.
The allocation of sampling locations within a study area follows the same process as outlined above for the whole region. In each case grid spacing should be developed to permit a minimum 15-20 sampling locations per stratum. More locations will lead to better precision from the design-based estimates, and more reliable spatial modelling.
Finally it is important to recognize that results of surveys in a truncated study area, can only apply to the area from which the sampling points were developed. Thus, if certain provinces or areas of a country are excluded from the survey due to insecurity or some other constraint, then the results developed from the accessible survey zone cannot be applied to the inaccessible area.
A map showing information on access in the Central African forest zone at the time of the Pilot Project is provided in ANNEX 6.7. RECOMMENDATIONS FOR RESEARCH AND ADDITIONAL INPUTS THAT WILL IMPROVE SURVEYS AND SURVEY DESIGN.
7.1 African Elephant Database Update
It is clear that the survey area should be restricted to the elephant range to avoid waste of survey effort. Preliminary evidence indicates that the current elephant range in Central Africa may vary significantly from that shown in the African Elephant Database. In some countries (DR Congo, Cameroon) the range appears to be undergoing rapid reduction and fragmentation. An update of the current distribution of elephants in Africa would be very useful for the MIKE programme. We strongly recommend that this update be accomplished through an update of the African Elephant Database.
7.2 Dung Deposition and Disappearance Rates
The current elephant inventories use dung density as an index of elephant abundance. It is desirable that this be convertible to elephant numbers. The conversion is based on estimates of dung disappearance and defecation rates, which are variable.
Ideally, for unbiased estimation of dung disappearance rates, dung decay should be studied at the extensive sampling locations. However, this is almost certainly impossible, therefore we recommend that studies of disappearance rates of dung be done at all MIKE sites. It is highly recommended to start these studies one to two months before the surveys in nearby blocks and continue them up to two months after the surveys have ended. There is a relationship between rainfall, leaf-fall and other factors and dung disappearance rates but the nature of this relationship may differ from area to area. With rainfall data a model can be constructed to predict dung survival (Barnes 1997). It is recommended that rainfall data be taken at all sites where dung decay is being monitored.
We recommend that efforts also be made to study defecation rates. A possible site for this is Garamba National Park where domestic elephants are kept. The park and the surrounding hunting reserves include both forest and savannah habitats. It will be essential to ensure representative diets, habitats etc.
7.3 Use of Data Collected on Travel between Sampling Locations
Under the suggested design, the sampling locations will be widely spaced (likely 50-200 km). When travelling between locations, maximum use will be made of the existing road network; nevertheless, substantial travel on foot can be anticipated to reach the locations.
It is tempting to view this travelling as time when additional survey information could be collected. This information could not form part of the design-based estimates, but it could potentially be used in spatial modelling estimation. However, it is important to remember that the priority while travelling is to reach the next survey location as fast as possible. Therefore the field teams should not be asked to do anything that slows them down to any significant extent.
It would seem very useful to ask teams to record any elephants, carcasses and direct evidence of poaching activity that they observe, and the locations of these. They should also plan to gather information on local elephant distributions from villages. At the other extreme, it is probably not worth asking the teams to record all elephant dung as they travel, because this is likely to slow them down significantly. Nor will they be able to do a comprehensive village survey, though some effort to come up with a systematic sampling design for the village questionnaires is worthwhile (See Report 3 for a possible model in relation to data collection on illegal killing). In the moving between locations in the forest, it may be worth asking them to record dung using a "recce" protocol for perhaps one hour a day, or for a fixed distance. The exact protocol remains to be discussed, but this should not hold up progress in designing and implementing the survey.
7.4 Possible Refinements of the Design
The initial design of MIKE inventories can be refined. Possible refinements include:Re-stratification following first surveys Modelling of covariates to improve efficiency of survey design
On balance, given the current lack of information about elephant abundance, we think that the stratified systematic design we have proposed is the best option at present. Once the survey has been running for one or two survey cycles, and the first maps of estimated abundance have been produced, it would then be worth re-evaluating whether other designs would be more efficient. There is no reason why the design cannot be changed over time.
7.5 Guidelines for Management and Training of Field Teams
Logistical difficulties will continue to constrain MIKE inventories in Central Africa in the foreseeable future. Based on the Pilot experience we recommend the following for these large-scale surveys:Develop mobile survey teams Invest in selected sites as logistical and training bases for mobile teams. An initial selection could include designated MIKE sites, or some subset of these. Ensure adequate quality control of data and supervision of field work. Hold regular evaluation and training sessions.
Further details on these can be found in the technical reports on training.8. SUMMARY RECOMMENDATIONS FOR SURVEY DESIGN Define objectives carefully and be sure they are well understood by all participants in the surveys. Agree the level of precision needed in the results (optimal allocation). Determine resources needed and their availability. Define study area in relationship to constraints affecting field operations. Ensure resources for training of field teams. Ensure appropriate input of statistical advice. Use design-unbiased systematic placement of sampling locations. Stratify sampling effort. Use pilot studies to improve design.
Barnes, R.F.W. and Jensen, K.L. 1987. How to count elephants in forests. IUCN African Elephant & Rhino Specialist Group Technical Bulletin 1, 1-6
Barnes, R.F.W., Asamoah-Boateng, B., Naada Majam, J. andAgyei-Ohemeng, J. 1997. Rainfall and the population dynamics of elephant dung-piles in the forests of southern Ghana. African Journal of Ecology, 35, 39-52
Barnes, R.F.W., Craig, G.C., Dublin, H.T., Overton, G., Simons, W. and Thouless, C.R. 1999. African Elephant Database 1998. IUCN/SSC African Elephant Specialist Group. IUCN, Gland, Switzerland and Cambridge, UK. vi + 249pp.
Buckland, S.T., Anderson, D.R., Burnham, K.P., Laake, J.L., 1993. Distance sampling. Estimating abundance of biological populations. Chapman & Hall
Hansen, M.H., Madow, W.G. and Tepping, B.J. 1983. An Evaluation of Model-Dependent and Probability-Sampling Inferences in Sample Surveys. J. Amer. Statist. Assoc,, 78, 776-793.
Hedley, S.L., Buckland, S.T. and Borchers, D.L. 1999. Spatial modelling from line transect data. Journal of Cetacean Research and Management 1, 255-64.
Munholland, P.L. & Borkowski, J.J. 1996. Simple Latin Square Sampling+1: A Spatial Design Using Quadrats. Biometrics, 52, 125-136.
Plumptre, A.J. 2000. Monitoring mammal populations with line transect techniques in African forests. Journal of Applied Ecology 37 (2), 356
Thompson S.D. (1992). Sampling. New York. Wiley.
Walsh, P.D., White, L.J.T. 1999. What will it take to monitor forest elephants. Conservation Biology, 13 (5), 1194-1202
Walsh, P.D., White, L.J.T., Mbina, C., Idiata, D., Mihindou, Y., Maisels, F., Thibault, M. 2001. Estimates of forest elephant abundance: projecting the relationship between precision and effort. Journal of Applied Ecology, 38 (1), 217-
- Central African Study Area, MIKE Sites and Pilot Project Sites
- MIKE Survey Sites provided to the Pilot Project (April 1999) and treated as intensive sites in the proposed survey design
- Sampling plans and sampled locations in
- Cartographic inputs/outputs for sampling effort allocation
a. Study area
b. Stratification process
c. Results of stratification
d. Allocation of sampling points
- Formulae for calculating optimal allocation of survey effort among strata
- Access map for the Central African forest zone
- List of technical consultants reports