How predictive is traditional ecological knowledge? the case of the Lacandon Maya Fallow Enrichment System

 

Samuel Israel Levy Tacher and John Duncan Golicher

Samuel Israel Levy Tacher. Agronomist. Ph.D in Ethnobotany, Colegio de Postgraduados, Chapingo, México. Researcher, Colegio de la Frontera Sur, Unidad San Cristobal de las Casas, Chiapas, México. e-mail: slevy@sclc.ecosur.mx.

John Duncan Golicher. PhD in Ecology, University of Edinburgh, Scotland. Researcher, Colegio de la Frontera Sur, Unidad San Cristobal de las Casas, Chiapas. México. e-mail: dgoliche@sclc.ecosur.mx.

Resumen

La habilidad de predecir mediante una asociación estadística o conocimiento causal es la clave característica del conocimiento científico. Sin embargo, el poder predictivo del conocimiento ecológico tradicional basado en datos obtenidos de observaciones pocas veces ha sido evaluado o criticado. Analizamos el poder predictivo de los resultados de un estudio de caso en el manejo de acahuales (vegetación secundaria). Para este propósito aplicamos tres análisis estadísticos contrastantes; análisis de covarianza, modelación espacial descriptiva y modelo gráfico. Nuestros resultados sugieren la necesidad de contar con cierto grado de abducción o explicación a posteriori para el análisis de las observaciones derivadas de las estrategias realizadas por los indígenas. A pesar de ello pudimos reconocer el poder predictivo del conocimiento indígena a partir de una combinación de experiencia, deducción e inducción probabilística. El presente estudio de caso sugiere que el conocimiento ecológico tradicional lacandón es de utilidad para guiar los esfuerzos hacia la restauración del ecosistema selvático, particularmente con el uso de Ochroma pyramidale como agente de enriquecimiento del suelo.

Summary

The ability to predict through causal understanding or statistical association is a key feature of scientific knowledge. However, the predictive power of traditional ecological knowledge in an observational setting is rarely critically evaluated. We analyze the results of an applied case study of indigenous fallow management in the context of its predictive power. In order to do this we applied three contrasting statistical analyses; visual pattern matching, general linear modelling and graphical models. Our results suggest that a degree of abduction or a posteriori explanation is typically needed in the analysis of the observations of natural phenomena that indigenous people make. Despite this, contextually useful predictive power may be extracted from such information through a combination of experience, deduction and probabilistic induction. The present case study suggested that the traditional ecological knowledge of the Lacandon people could provide useful guidance for the restoration of ecosystem function. In particular, the results support the use of Ochroma pyramidale as a soil enriching agent.

Resumo

A habilidade de predizer mediante uma associação estatística ou conhecimento causal é a clave característica do conhecimento científico. No entanto, o poder preditivo do conhecimento ecológico tradicional baseado em dados obtidos de observações poucas vezes tem sido avaliado ou criticado. Analisamos o poder preditivo dos resultados de um estudo de caso no manejo de "acahuales" (vegetação secundária). Para este propósito aplicamos três analises estatísticas contrastantes; analise de co variação, modelação espacial descritiva e modelo gráfico. Nossos resultados sugerem a necessidade de contar com certo grau de abdução ou explicação a posteriori para a análise das observações derivadas das estratégias realizadas pelos indígenas. A pesar disto podemos reconhecer o poder preditivo do conhecimento indígena a partir de uma combinação de experiência, dedução e indução probabilística. O presente estudo de caso sugere que o conhecimento ecológico tradicional lacandon é de utilidade para guiar os esforços para a restauração do ecossistema selvático, particularmente com o uso de Ochroma pyramidale como agente de enriquecimento do solo.

KEYWORDS / Epistemology/ Lacandon / Ochroma pyramidale / Restoration / Soil Organic Matter /

Received: 22/04/2004. Accepted: 06/08/2004.

Traditional ecological knowledge (TEK) is being increasingly respected as a source of information which can be used to guide conservation, management and restoration of natural resources (Gadgil et al., 1993; Ford and Martínez, 2000). However, in order to draw on this reservoir of information ecologists must learn to respect its cultural significance while evaluating its predictive power (Huntington, 2000; Alhamidi et al., 2003). Predictive traditional ecological knowledge (PTEK) has to be sifted from a pool of tradition knowledge and wisdom (TEKW) that mixes empirical knowledge of the natural world with moral, ethical and spiritual values (Doubleday, 1993; Inglis, 1993). This is a complex task made yet more challenging by the need to accept and combine epistemologies drawn from both ecological and social science (Berkes et al., 2000). The validity of the epistemology used by researchers aiming to produce predictive ecology is formally evaluated through the process of peer review (Ford, 2000). Yet the way in which TEK tests its own predictive power has received much less attention. This paper aims to critically evaluate the predictive value of TEK in the context of an applied case study.

It is common to draw a rather sharp divide between TEK and conventional scientific understanding (e.g., Berkes, 1993; Deloria, 1996; Stevenson, 1996). Yet, far from being incompatible world views, the two strands of knowledge have to confront very similar problems. Ecologists naturally place great emphasis on experimentation as the ultimate arbiter of cause and effect (Underwood, 1997). Experiments are typically designed with the explicit intention of negating the hypothesis of no effect. However, knowledge of ecosystem structure and function is more naturally derived from observational data. This leads to an interesting dilemma. Inference drawn from observation alone is often criticized by trained scientists who emphasize the weakness in the underlying epistemology (Peters, 1991). Confounding, explaining away and the problem of inadequate or «pseudo-replication» reduce predictive power (Hurlbert, 1984). These are not new problems, neither are they necessarily resolved through the use of more sophisticated survey design and statistical analysis (Hawkins, 1986).

The ability to predict is often linked to the ability to understand causal mechanisms. However prediction and understanding are rather different matters. Causality may sometimes be deduced from observations, yet a test of predictive power is usually simplest in an experimental setting (Holland, 1986). Despite this, experimental approaches do not provide all the information that is relevant at the temporal and spatial scale at which ecological phenomena occur (Vepsalainen and Spence 2000). Because intuitive understanding based on natural history places greater emphasis on spatial and temporal contiguity, it remains a commonly used component of natural resource management (Schrader-Frechette and McCoy, 1993). PTEK itself must also be the product of a successful natural history based inference. This could consist of an incremental process in which predictive hypotheses are tested and strengthened through a mixture of deduction and induction. One approach that can achieve this is Bayesian inference (Ellison, 1996). Bayesian methods are now becoming widely accepted in the field of ecology and resource management. One of the aims of the present work was to begin to look for parallels between the ideas that underlie Bayesian methods and TEK.

An evaluation of the predictive values of TEK has to take place within the context of a case study. The underlying motivation for the research was practical and applied. We aimed to evaluate the potential of PTEK to guide ecosystem restoration in the Lacandon area of southern Mexico. A central question asked is to what extent can the traditional knowledge involved in the management of the fallow vegetation under rotational slash and burn farming be considered predictive. In order to address this question data was collected that corresponded closely to the type of evidence that would be available to the indigenous farmers themselves. This provided a good opportunity to begin to evaluate the epistemology of TEK in a wider sense. The statistical methods used were directed at critically evaluating the information base from which indigenous knowledge could be derived. In particular, the extent to which inductive, deductive and abductive reasoning are used in the process that generates PTEK was investigated.

Study area

The study took place Jan-Jul 2003, within the communally owned land of the Lacandon community of Lacanhá Chansayab in the municipal region of Ocosingo, Chiapas, 16º46'08'' N, 91º08'12''W and »250m altitude. The area is part of the Lacanhá river watershed, which drains into the Usumacinta (Muench, 1982). Soils vary markedly due to the occurrence of parallel series of calcareous outcroppings and alluvial deposits (Anonymous, 1974). Humic acrisols are associated with rendzinas on calcareous outcroppings which are interspersed with eutric regosols. The climate is warm and humid with a short dry season in Mar-May. Mean annual temperature is 25ºC with mean annual rainfall between 2300 and 2500mm (García, 1973; Anonymous, 1974, 1988; Muench, 1978).

In its undisturbed state the vegetation forms a tall perennial rain forest (Miranda and Hernández, 1963; Miranda, 1998; Pennington and Sarukhán, 1998) or a lower montane rain forest (Breedlove, 1973). The Lacandon themselves recognize two types of mature forest, which they call monte alto and chaparral. The latter consists of lower vegetation, below 20m in height, occurring on temporally flooded areas within the high forest. Completely undisturbed natural vegetation is uncommon. The Lacandon have subsisted in the area for centuries using a rotational slash and burn system (milpa) to produce maize. The secondary vegetation that derives from this is known as acahuales in spanish or jurupchés in Lacandon (Levy et al., 2002).

Older members of the Lacandon communities have a rich understanding of vegetation properties. This knowledge includes a complex classification system based both on patterns and processes that occur as the vegetation cover develops following disturbance through slash and burn. However, many younger members of the community have adopted simplified but economically more profitable farming methods in which soil fertility and weed control are obtained through the use of chemical inputs. Cattle ranching is also spreading through the region. The consequences of these changes are both loss of forest cover and loss of TEK (Levy, 2000; Levy et al., 2002).

A key element of the traditional fallow management system is the deliberate planting and encouragement of the fast growing tree Ochroma pyramidale (Cav. ex Lam.) Urban, known to the Lacandon as Chujum in the fallow vegetation. This tree is believed by the Lacandon to produce larger amounts of soil organic matter than any of the other components of the fallow vegetation. The tree may also help to control weeds by quickly producing a dense continuous vegetation cover (Levy, 2000). The traditional Lacandon system of rotational slash and burn (milpa) farming involves a cropping period of 4 or 6 years followed by an extended tree fallow (acahual) of between 7 and 12 years. Following the last cropping period Ochroma seeds are deliberately spread within the fields with the intention of improving the soil in preparation for the next cycle of cultivation (Levy, 2000). This practice clearly has important implications for sustainable management in the area. The predictive power of the Lacandon’s beliefs deserves being put to the test. We investigated the Lacandon’s postulate that soil organic matter (SOM) accumulates more quickly under the cover of Ochroma than any other type of vegetation. Like the Lacandon themselves, the only evidence available at this stage of the investigation was observational rather than experimental.

Methods

Site selection

A subset of areas of fallow under traditional cultivation was selected from those available, using the following criteria: 1) they contained some small but notably homogeneous pure stands of Ochroma that were thought to have established as a result of deliberate management action; 2) the stands of Ochroma were embedded in a heterogeneous mosaic of other vegetation types; and 3) the whole area shared a common history of use in terms of time since the last disturbance and previous cycles of cropping.

A total of fifteen fallows were found which partially met these criteria. Unfortunately, reliable histories of use and management could be obtained for only seven of them. This stresses the urgency of the task of documenting the Lacandon’s TEK before it is lost. The time since the last maize crop for these areas ranged between 5 and 22 years. An additional plot was added for comparative purposes. This consisted of a pasture of approximately 0.5ha and had been chosen for a pilot project in 1996, in which half the area was sown with pure Ochroma. Half of this area now consists of a stand of Ochroma around 13m in height with trees with diameters of 20 to 25cm at breast height. This was the only plot in which the alternative to Ochroma cover was a non woody vegetation type.

Survey method

The areas were geo-referenced using a GPS (Garmin 12XL, accuracy 3m). In each fallow the four corners of a rectangular area were marked. The initial aim was to use 100m long transects in all plots. However, in some of the suitable plots this would have led to the inclusion of areas with differing use history. Therefore, two sizes of survey site were included in the survey depending on the size of the acahual, 30x100m and 30x50m. These were marked out on the ground using tape and compasses. Openings were cut in the vegetation with a machete. Once the rectangular areas were established, four parallel transect lines 10m apart were cut through the vegetation. Ten equidistant points were selected along the transect lines leading to a total of 40 points for each plot. At each point a circular subplot was marked. The radius of this plot was approximately proportional to the estimated height of the vegetation. The height classes were 0-5, 5-10 and >10m. The radius of the corresponding subplots were 1.5, 2 and 2.5m.

At each point the cover of Ochroma was estimated. A 100% cover was assigned to areas completely covered by Ochroma and 0% to areas with any other type of vegetation. The % cover of Ochroma in intermediate situations, when branches of Ochroma intermingled with other elements of the vegetation were estimated visually. Such situations were comparatively uncommon and adopting more sophisticated methods of measurement would have added little useful precision. The identity of the woody species in the non-Ochroma fallow was noted but the information was too complex to be used directly in the statistical analysis and was reserved for descriptive purposes. Litter depth was recorded at a total of 20 randomly sited measurements within each marked circle. Distance from the mineral soil surface to a point corresponding to the highest point of the litter after lightly compressing the surface by hand was measured.

In each circle soil samples were collected for analysis. Litter was cleared from the mineral soil surface and a 20x30cm sample of the first 5cm of soil depth was taken. Four such samples were taken in each circle, distributed in the centers of the four quadrants of the circular plots and mixed to form a compound sample. The soil samples were used for three purposes: 1) evaluation of the soil organic matter (SOM) and content in each of the 40 samples for each plot; 2) a complete quantitative analysis on two compound samples for each of the 8 plots; and 3 seed bank analysis. The results of the seed bank analysis are not included here. For the complete soil analysis 14 samples were chosen from each plot. Of these, 7 corresponded to points with 100% cover of Ochroma and the other 7 to areas with 0% cover of Ochroma. These were mixed to form two compound samples.

Determinations of SOM and total N2 were carried out using the procedures in the Mexican normative framework NOM (Anonymous, 2000). The Walkley-Black oxidation method was used for the determination of SOM in all the samples. This method typically underestimates SOM, and a correction factor of 1.3 was therefore applied, as suggested by the NOM.

Acahual use history

The history of usage of the acahuales was obtained through semi-structured interviews with the owners of the plots. Questions were aimed at finding out 1) the length of the current and previous fallow period; 2) the number of previous milpa cycles that the plot had been subject to; and 3) the moment in which Ochroma was sown or developed spontaneously within the vegetation.

Statistical analysis

The initial assumption was that complex saturated statistical models might be a default for representing natural systems. This is not necessarily because the physical processes involved interact in a truly complex manner. Rather, it is that historical and spatial contingencies are inevitable components of observational data. The inductive element in PTEK epistemology would seem to be based on intuitive model building that seeks to simplify these complex patterns in order to produce useful syntheses. Our analysis attempted to replicate this process using various statistical tools. Rather than restrict our attention to one model we compared a range of models. We separated the analysis into three parts:

1- Spatial pattern matching. Recognition of spatial pattern is an essential component of natural history (Legendre and Legendre, 1999). Spatial autocorrelation is the norm, not the exception, in observational contexts, even though classically trained ecologists may occasionally overlook the statistical consequences (Legendre, 1993). In order to investigate spatial patterns we constructed models using universal kriging (Ripley, 1981; Cressie, 1993). This was implemented using the "spatial" package (Venables and Ripley 1999) within the statistical environment R (Ihaka and Gentleman, 1996). The overall distribution of Ochroma cover, litter depth and SOM in each plot were first modelled as 3rd order polynomial trend surfaces. The local spatial autocorrelation was then included in the model by fitting an exponential covariance function to the data in order to adjust the model to local effects with reference to the variogram. Geostatistical models of this type are powerful tools for making spatial predictions for unknown quantities. In this situation they were used in a simpler capacity in order to produce visual representations of the patterns observed in the field. It is recognized that, because cover was measured as a proportion, the assumption is that the distribution of the regionalized variable is Gaussian. In these circumstances, indicator kriging is the preferred technique for inferential spatial modelling. However, the interpretation of maps produced by indicator kriging is less straigthforward. The use of universal kriging was chosen because it produced simple and easily interpretable maps of all three variables for each plot that matched our perceptions from field experience. These sort of visual patterns correspond closely to the type of evidence that is available to the farmers themselves.

2- Parametric analysis. The second analysis ignored the spatial component and concentrated simply on the relationships between Ochroma cover, litter depth and SOM. These models were interpreted in a semi-inferential capacity in order to find numerical patterns. Care should be taken in the interpretation of models based on observations that are known to be non-independent. Experimentation is needed in order to fully neutralize unwanted dependencies. General linear models were fitted to the data taking SOM and litter as response variables, and Ochroma cover and plot as explanatory variables. The plot identity was included in the models as a fixed factor. The justification for modelling the effects of plot as fixed rather than random, which would be more natural in an experimental context involving blocking, was that we considered that the effects of each plot were at least partly explainable in the context of our additional observations. Each plot had a different history of use and the nature of the non-Ochroma vegetation varied between plots. It was thus inappropriate to ascribe variability to a set of unknown confounding factors, as would be possible in an experimental setting.

The model formulae used in R took the form

lm (OrganicMatter~Cover*as.factor(Plot)) for Model 1, and

lm (Litter~Cover*as.factor(Plot)) for Model 2.

These models can also be described as analyses of covariance (ANCOVA). A key element was the inclusion of the interaction between plot and Ochroma cover. In other words, it was supposed that the slope of the regression line for each plot could be different (Faraway, 2002). From the perspective of prediction such a model is overly complex. If the effects of Ochroma cover were found to be sufficiently uniform over all plots the interaction terms could be dropped from the model, leaving a simpler predictive tool. Thus, the level of model complexity required to find an adequate fit to the data was evaluated using a stepwise procedure that tested whether simpler model formulae would be accepted by An Information Criteria (AIC; Akaike, 1974; Burnham and Anderson, 2002).

Confidence intervals were computed for the parameters of the model using treatment contrasts. Under treatment contrasts the intercept of the model represents the mean value for SOM or litter depth in plot 1 with no Ochroma cover. The second parameter represents the slope of a regression line of SOM or litter depth on Ochroma cover for plot 1. Further parameters for each plot and the PlotCover interactions represent differences in the intercept and slope for a regression model for each subsequent plot.

3- Categorical analysis using log linear models and Bayesian networks. The Lacandon clearly do not have access to the numerical data on SOM produced. In order to find statistical techniques that more closely matched the process of generating PTEK the observations were converted into classes. The frequency of observations in each combination of classes was modelled. Each of the four data vectors representing the age of the fallow, percent cover of Ochroma, litter depth and SOM were simplified into two classes, high and low, dividing the data around the median value. This categorical data was first investigated by fitting a series of log linear models in order, again, to investigate necessary model complexity using AIC. However, the conditional dependencies identified by log linear modelling are more intuitively depicted using graphical models. Bayesian networks are particularly powerful graphical models that can learn their structure and parameters from the data. They are particularly useful tools for investigating conditional dependencies and can be used to represent either inferred or known causal relationships between variables (Pearl, 1995). The Hugin PC algorithm (Spirtes et al., 2000) was used to fit Bayesian networks to the classified data.

Results

Description of plot characteristics

Table I provides a breakdown of the characteristics of each of the plots. The current fallow period for the plots varies from five to 22 years, with a mean of around 12 years. In five of the eight plots, seeds of Ochroma were scattered at the end of the 1st cycle of maize cultivation; in two of them the decision to sow Ochroma was taken after the 4th cycle of cultivation. The motivation given by the Lacandon for sowing Ochroma was not only to enrich the soil for future use but also to prevent invasion of the fallow by Piper spp. and the fern Pteridium aquilinum, which are aggressive weeds. More than half the areas had a comparatively short (5-12 years) rest period with a disturbance index of over 1, showing that they are under cultivation for longer periods than they are left fallow. A major source of differences between plots was the variability in microtopography and in the surrounding vegetation.

Spatial analysis

Figure 1 shows the spatial patterns observed in each of the eight plots for cover of Ochroma, mean litter depth and percent SOM. Many factors have to be taken into account in order to explain these patterns. It is worth looking at these in more depth in order to understand the sort of inferential task the Lacandon farmers would have to perform in order to derive quantitative predictions of the effect of Ochroma on SOM.

Plots 1 and 2 show a clear cut relationship between Ochroma cover and SOM. The spatial pattern agrees well with the postulate that Ochroma cover leads to soil enrichment through the increase of the amount of litter that is incorporated into the soil. As would be expected, areas with the highest percentage of SOM are at the centers of each patch of Ochroma. Plot 3 provided the worst visual match to the expected pattern, with SOM being concentrated in areas away from the cover of Ochroma. This may have been due to the presence of seasonal stream beds in the plot that led to the SOM being washed from its source of origin. In this context the microtopography of the plot may be particularly important. It must be mentioned that this plot had sloping areas with gradients of over 8%. Organic matter here tended to accumulate in hollows and channels. In plot 4 two extreme measurements contributed greatly to ensuring that the pattern strongly matched expectation. These measurements were almost certainly due to the erroneous inclusion of litter in the samples. This mechanism cannot be ruled out in the other plots, although care was taken to ensure that only mineral soil was collected throughout. Plot 5 also had an unusual SOM pattern. In this case it may again have been attributable to the effects of slope within the plot compounded by the presence of a Mayan ruin beneath the soil. In plot 6 the cover of Ochroma also had little clear connection with SOM distribution. The history of use of this plot was atypical. It had remained partially open and was still being used to produce pineapple. The pseudoexperimental plot 7 matched expectations, but in this case the non-Ochroma area was a mixture of pasture and small shrubs rather than tree fallow. It was therefore expected to have much lower SOM than a woody fallow with any composition. Finally, plot 8, another mainly flat area surrounded by mature vegetation showed the expected pattern.

Analysis of covariance

Table II shows the analysis of variance for the two parametric linear models. Note that all terms are significant, while Table III gives a breakdown of the model parameters and confidence intervals. Backwards stepwise analysis of the statistical models linking Ochroma cover to SOM (Model 1) and litter depth (Model 2) suggested that the factor plot and the interactions between plot and Ochroma cover should be retained in both cases. The initial AIC for Model 1 was 739. When the interaction term was removed AIC rose to a value of 756, showing that the complexity represented by the interaction term was needed to represent the data. In the case of Model 2, AIC for the complex model was 500 but only rose to 505 when the interaction term was removed. The complex model is thus only marginally superior on information theoretic grounds and the simpler model for litter depth may perhaps be preferred. The conclusions are simple. Litter depth is quite predictable from knowledge of the percent Ochroma cover alone, while the relationship between Ochroma cover and soil organic matter is much more specific to each plot. Figure 2 shows the variation in the slopes of the regression lines between plots. The possible explanations for this variation have already been pointed out.

 

Categorical modelling

The first two approaches to the analysis already suggest that complex models seem to be needed to fully represent the data. However, in this analysis it was attempted to find predictive power through deliberate simplification. A saturated log linear model was fitted to the categorical data including the age of the plot, litter depth, SOM and Ochroma cover. Stepwise backwards model selection was then used to find simpler models than using AIC as a criteria. Again, the simpler models were not accepted by this criteria. High level interactions terms were also apparently needed to represent the structure in the categorical data. However, when age of the plot was dropped from the saturated model, stepwise backwards model selection did suggest the removal of the interaction between organic matter and litter, while keeping the rest of the terms. This model was similar to the Bayesian network that the PC algorithm fitted to the data. Interestingly, the algorithm almost found the parsimonious causal model that would be expected. However, an additional arc between cover and SOM was added that had no causal interpretation. Figure 3 shows the Bayesian network predictions after evidence was entered at the node for Ochroma cover. If Ochroma cover is high, SOM is higher than it would be if Ochroma cover is low. Information on the age of the plot can also be used to refine this prediction. This Bayesian network provides a much simpler predictive model than that suggested by analysis of covariance, but it should not be forgotten that this is in large part due to the deliberate exclusion of the interactions and dependencies that we have already observed when plot was included as a factor. Simplification can be justified in terms of the utility of small models for prediction. Updating the model by allowing it to learn from subsequent experience is required in order to prevent over-generalization. Bayesian networks should not be seen as static means of inference. Rather, they are tools for turning available data into models that can learn and adapt their structure and parameters as more information becomes available.

Discussion

The statistical analysis showed Occam’s razor to be rather blunt. A simple model that could be accepted under Akaike’s information criterion was not available. Once it became clear that the best statistical models involved complex interactions, detailed explanation of individual observations were obtained through a form of abduction. Abductive reasoning is often used in diagnosis and case centred explanation, usually in an informal way (Pierce, 1955). Abduction is a task that can be described in the following terms. Given some proposition find a set of additional propositions that together with background knowledge account for some new observation. Abductive reasoning is discouraged in formal scientific research that seeks to establish predictive power. The reason for this distrust is particularly well understood by the artificial intelligence community that has attempted to model such reasoning. Abduction is closely associated with so called "commonsense", and is an essential element of natural history. However, it is very difficult to formalize and it is not predictive (Shanahan, 1989). Abduction assumes that given any effect a set of causes can be found, but does not permit the prediction of effect size from knowledge of a single cause. This is because under abduction additional causes may always be added to produce "better" explanations. Abduction alone does not allow general predictions to be made regarding the effect of enriching a fallow with Ochroma.

If abduction is not predictive, could predictions concerning the properties of Ochroma be derived from induction? There is a difficulty with purely inductive inference that Hume (1748) pointed out. Classical induction is based on a circular argument that invokes the principle of uniformity of nature. In other words, if the Lacandon observe that SOM is higher (the soil is darker) under Ochroma in one plot, it may be assumed that it will be higher in others. The only way to know that nature is uniform is by observing it to be uniform. This was the cause of Hume’s scepticism regarding induction. However, it is not the tautology of this argument that causes problems when trying to predict in real life. Rather, it is the simple lack of uniformity in real patterns of nature that was drawn out by our analyses. In one sense, lack of uniformity is trivial and can be expected a priori. We knew before taking any measurements that no two plots were alike. It is this lack of uniformity in natural systems that challenges ecologists, farmers and hunters trying to derive predictive rules from observations (Schrader-Frechette, 1994). Inferential techniques that assume uniformity as an a priori condition can be pointless in an observational context, although this has not prevented them from being inappropriately used (Berger and Selke, 1987; Johnson, 1999).

Observational data such as the present one is inevitably noisy, fuzzy, uncertain or probabilistic. Ecologists and resource managers alike have sought to develop suitable paradigms to deal with this (Simberloff, 1980). Induction under uncertainty requires a probabilistic framework. The adoption of a Bayesian approach to the inference problem may be a natural consequence (Hillborn and Mangel, 1997). There is still a need to invoke the uniformity of nature in order to predict, but in a non intuitive manner. Instead of stating that ‘SOM is universally higher under Ochroma than other types of vegetation’, a better statement may be that ‘it is believed that SOM is probably higher under Ochroma’ after some simplifying assumptions have been made. The inherent difficulty with this sort of statement was summed up by Bertrand Russell (Jeffreys, 1933): "Induction appears to me to be either disguised deduction or a method of making plausible guesses". Bayesian inference may be a step towards making useful predictions, but the criticism that such predictions may be guesswork remains, even when inference is informed by experience. Our small Bayesian network learnt both its structure and parameters from the data in order to produce probabilistic guesses. From one viewpoint it was an oversimplification of a complex reality. However the strength of this sort of model is that as more data becomes available probabilities may be strengthened, revised or even reversed. Bayesian methods may thus reflect the route through which real PTEK is generated over time in noisy systems, although it is far from clear whether the counter intuitive calculus of inverse probabilities is actually reproduced by the human cognitive system (Eddy, 1982; Efron, 1986; Gigerenzer and Murray, 1987; Anderson, 1998).

The alternative method through which the Lacandon could develop a generally applicable pool of PTEK is of course through deduction. In this case the deductive argument is simple. Ochroma produces very large quantities of litter. Litter decomposition is responsible for producing SOM. It therefore would follow that SOM should be higher under Ochroma than under other types of vegetation. This argument, if valid, is predictive. Unfortunately, the predictions derived from deductions, even when they are based on such apparently sound premises, should still be tested at some stage through experience. For example, from the perspective of TEK, the doctrine of signatures used in traditional medicine may seem to be based on sound deduction. It is well known not to be predictive. As we have already noted, observational data alone does not provide clear cut tests of deductions. A single deduction is not enough to explain the sort of results that nature throws up. There is always noise around the signal, providing a temptation to return to the special pleading of abductive reasoning in order to find subsidiary explanations for changes in effect size.

Conclusion

The challenges faced in interpreting the results are not unusual in observational studies. Practical conclusions can be derived from our results. There is a sufficiently strong suggestion that Ochroma has useful properties for soil enrichment to justify further investigation. Long term controlled experimentation is needed in order to fully assess the reliability of the Lacandon’s knowledge regarding the properties of Ochroma. These studies need to be carried out within the complex ecological setting of the region in which the TEK was generated. Given the urgency of finding viable solutions to the problems of ecological degradation in the Lacandon area, such work would be fully justified by the present results.

ACKNOWLEDGEMENTS

The authors thank the people of Lacanhá Chansayab: Manuel Castellanos Chankin and Kin Bor, for sharing their knowledge and experience; and Irene Pisanty, Eduardo Peters and Gloria Luz Portales for their valuable assistance and support. Financial support was provided by Fondo Sectorial CONACYT-CONAFOR (2003-C03-9950), National Institute of Eclogy of México (INE), International Conservation México A.C. and Etnobiología para la Conservación A.C.

REFERENCES

1. Alhamidi SK, Gustafsson M, Larsson H, Hillbur P (2003) The cultural background of the sustainability of the traditional farming system in the Ghouta, the oasis of Damascus, Syria. Agriculture and Human Values 20: 231-240.        [ Links ]

2. Akaike H (1974) A new look at the statistical model identification. IEEE Trans. Aut. Cont. 19: 716-723.        [ Links ]

3. Anderson JL (1998) Embracing uncertainty: The interface of Bayesian Statistics and Cognitive Psychology. Cons. Ecol. 2: 1-28.        [ Links ]

4. Anonymous (1974) Estudio de gran visión de la zona Lacandona, Chiapas. Comisión de Estudios del Territorio Nacional. Secretaría de la Presidencia. México DF, México. 99 pp.        [ Links ]

5. Anonymous (1988) Las Margaritas (E15-12, D15-3). Carta climática. Esc. 1:250,000. Instituto Nacional de Estadística Geografía e Informática. México DF, México.        [ Links ]

6. Anoymous (2000) Norma Oficial Mexicana NOM-021-RECNAT-2000. Secretaría de Medio Ambiente y Recursos Naturales. México DF, México. 75 pp.        [ Links ]

7. Berger JO, Sellke T (1987) Testing a point null hypothesis: the irreconcilability of P values and evidence. J. Am. Stat. Assoc. 82: 112-122.        [ Links ]

8. Berkes F (1993) Tradicional ecological knowledge in perspective. In Inglis TJ (Ed.) Traditional ecological knowledge: concepts and cases. International Program on Traditional Ecological Knowledge. IDRC. Ottawa, Ontario, Canada. pp. 1-9.        [ Links ]

9. Berkes F, Colding J, Folke C (2000) Rediscovery of traditional ecological knowledge as adaptative management. Ecol. Applic. 10: 1251-1262.        [ Links ]

10. Breedlove DE (1973) The phytogeography and vegetation of Chiapas (México). In Graham A (Ed.) Vegetation and vegetational history of northern Latin America. Elsevier. Amsterdam, Netherlands. pp. 149-165        [ Links ]

11. Burnham KP, Anderson DR (2002) Model selection and multimodel inference: A practical information-theoretic approach. 2nd ed. Springer. New York, NY, USA. 496 pp.         [ Links ]

12. Cressie NAC (1993) Statistics for Spatial Data (Revised Edition). Wiley, New York, NY, USA. 900 pp.        [ Links ]

13. Deloria V (1996) If you think about it, you will see that it is true. ReVision 18: 37-44.        [ Links ]

14. Doubleday NC (1993) Finding common ground: natural law collective wisdom. In Inglis TJ (Ed.) Traditional ecological knowledge: concepts and cases. International Program on Traditional Ecological Knowledge. IDRC. Ottawa, Ontario, Canada. pp. 41-53        [ Links ]

15. Eddy DM (1982) Probabilistic reasoning in clinical medicine: Problems and opportunities. In Kahneman D, Slovic P, Tversky A (Eds.) Judgement under uncertainty: Heuristics and biases. Cambridge University Press. Cambridge, UK. 237 pp.        [ Links ]

16. Efron B (1986) Why isn’t everyone a Bayesian? Am. Stat. 40: 1-11.        [ Links ]

17. Ellison AM (1996) An introduction to Bayesian inference for ecological research and environmental decision-making. Ecol. Applic. 6: 1036-1046.        [ Links ]

18. Faraway JJ (2002) Practical Regression and Anova using R. www.stat.lsa.umich.edu/~faraway/book/        [ Links ]

19. Ford D (2000) Scientific method for ecological research. Cambridge University Press, Cambridge, UK. 564 pp.        [ Links ]

20. Gadgil M, Berkes F, Folke C (1993) Indigenous knowledge for biodiversity conservation. Ambio 22: 151-156.         [ Links ]

21. García E (1973) Modificaciones al sistema de clasificación climática de Koppen. 2ª ed. Instituto de Geografía, Universidad Nacional Autónoma de México. México DF, México. 246 pp.        [ Links ]

22. Gigerenzer G, Murray DJ (1987) Cognition as intuitive statistics. Erlbaum. Hillsdale, NJ, USA. 430 pp.        [ Links ]

23. Hawkins CP (1986) Pseudo-understanding of pseudoreplication: a cautionary note. Bull. Ecol. Soc. Am. 67: 184-185.        [ Links ]

24. Hilborn R, Mangel M (1997) The ecological detective: confronting models with data. Princeton University Press. Princeton, NJ, USA. 320 pp.        [ Links ]

25. Holland PW (1986) Statistics and causal inference. J. Am. Stat. Assoc. 81: 945-960.        [ Links ]

26. Hume D (1748) An inquiry Concerning the Principles of Morals Human. Beauchamp T (Ed.). Oxford Press. Oxford, UK. 304 pp.        [ Links ]

27. Huntington PH (2000) Using traditional ecological knowledge in science: Methods and applications. Ecol. Applic. 10: 270-274.        [ Links ]

28. Hurlbert SH (1984) Pseudoreplication and the design of ecological field experiments. Ecol. Monogr. 54: 197-221.        [ Links ]

29. Ihaka R, Gentleman R (1996) R: a language for data analysis and graphics. J. Comp. Graph. Stat. 5: 299-314.        [ Links ]

30. Inglis JT (Ed.; 1993) Traditional ecological knowledge: concepts and cases. International Program on Traditional Ecological Knowledge. IDRC. Ottawa, Ontario, Canada. 214 pp.        [ Links ]

31. Jeffreys H (1933) Probability statistics and the theory of errors. Proc. Roy. Soc. Lond. Series A 140: 523-535.        [ Links ]

32. Johnson DH (1999) The insignificance of statistical significance testing. J. Wildlife Manag. 63: 763-772.        [ Links ]

33. Legendre P, Legendre L (1999) Numerical Ecology. 2nd ed. Elsevier. New York, NY, USA. 553 pp.        [ Links ]

34. Legendre P (1993) Spatial autocorrelation: trouble or new paradigm? Ecology 74: 1659-1673.        [ Links ]

35. Levy SI, Aguirre RJR, Martínez RMM, Durán FA (2002) Caracterización del uso tradicional de la flora espontánea en la comunidad lacandona de Lacanhá, Chiapas, México. Interciencia 27: 512-520.        [ Links ]

36. Levy SI (2000) Sucesión causada por roza-tumba-quema en las selvas de Lacanhá Chansayab, Chiapas. Thesis. Colegio de Postgraduados. Montecillos, Edo. de México, México. 165 pp.        [ Links ]

37. Miranda F, Hernández E (1963) Los tipos de vegetación de México y su clasificación. Bol. Soc. Bot. Méx. 28: 29-179.        [ Links ]

38. Miranda F (1998) La vegetación de Chiapas. Gobierno del Estado / Consejo Estatal para la Cultura y las Artes. Tuxtla Gutiérrez, Chiapas, México. 426 pp.        [ Links ]

39. Muench P (1978) Los sistemas de producción agrícola en la región Lacandona (Estudio agronómico preliminar). Tesis. Universidad Autónoma de Chapingo. Chapingo, México. 267 pp.        [ Links ]

40. Muench P (1982) Las regiones agrícolas de Chiapas. Geografía Agrícola 2: 33-44.        [ Links ]

41. Pennington TD, Sarukhán J (1998) Árboles tropicales de México: manual para la identificación de las principales especies. 2ª ed. Fondo de Cultura Económica. México DF, México. 413 pp.        [ Links ]

42. Pearl J (1995) Causal diagrams for Empirical research. Biometrika 82: 669-688.        [ Links ]

43. Peters RH (1991) A Critique for Ecology. Cambridge University Press. Cambridge, UK. 352 pp.        [ Links ]

44. Pierce C (1955) Abduction and induction. Dover. New York, NY, USA. 309 pp.        [ Links ]

45. Ripley BD (1981) Spatial Analysis. Wiley. New York, NY, USA. 340 pp.        [ Links ]

46. Schrader-Frechette KS (1994) Ethics of scientific research. Lanham MD (Ed.) Rowman and Littlefield. Boston, MA, USA. 256 pp.        [ Links ]

47. Schrader-Frechette KS, McCoy E (1993) Method in Ecology. Cambridge University Press. Cambridge, UK. 338 pp        [ Links ]

48. Shanahan MP (1989) Prediction Is Deduction but Explanation Is Abduction. Proc. IJCAI 89: 1055-1060.        [ Links ]

49. Simberloff D (1980) A succession of paradigms in ecology: essentialism to materialism to probabilism. Synthese 43: 3-39.        [ Links ]

50. Spirtes P, Glymour C, Scheines R (2000) Causation, Prediction, and Search. 2nd ed. Adaptive Computation and Machine Learning. MIT Press. Cambridge, MA, USA. 526 pp.        [ Links ]

51. Stevenson MG (1996) Indigenous knowledge in environmental assessment. Arctic 49: 278-291.        [ Links ]

52. Underwood AJ (1997) Experiments in Ecology: their logical design and interpretation using analysis of variance. Cambridge University Press. Cambridge, UK. 552 pp.        [ Links ]

53. Venables WN, Ripley BD (1999) Modern Applied Statistics with S-Plus. 3^rd ed. Springer. New York, NY, USA. 510 pp.        [ Links ]

54. Vepsalainen K, Spence JR (2000) Generalization in ecology and evolutionary biology: From hypothesis to paradigm. Biol. Phil. 15: 211-238.                                                           [ Links ]