a process-led approach to modeling land change in agricultural landscapes: a case study from...

19

Click here to load reader

Upload: rheyna-m-laney

Post on 25-Aug-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

Agriculture, Ecosystems and Environment 101 (2004) 135–153

A process-led approach to modeling land change in agriculturallandscapes: a case study from Madagascar

Rheyna M. Laney∗Sonoma State University, 1801 E. Cotati Avenue, Rohnert Park, CA 94928, USA

Abstract

This study develops a model to predict pattern from process, providing a counterpoint to land-change studies that interpretprocess from pattern. This less-represented approach to land-change analysis uses theory to guide the characterization ofland-cover patterns and to shape more directly the process–pattern relationships explored. The study presents a land-changemodel explicitly structured by a human–environment theory—the induced-intensification thesis of agricultural change. It usesthe thesis to characterize the range of management strategies adopted by farmers in the Andapa region of Madagascar as theyrespond to rising land pressure. The study then identifies the particular land-cover consequences of each of these strategies.These farmer-level process–pattern linkages form the basis of an aggregate village-level land-change model, which predictsthe land-cover outcomes of different dominant trajectories of agricultural change. The process-led model is found to be robust,successfully reproducing village-level landscape outcomes for most land-cover classes. Yet, the model also reveals that thedominant village-level process (the strategy followed by most farmers) does not necessarily create the spatially dominantland-cover outcome. This breakdown in the village-level process–pattern linkage occurs because of the well known, but oftenover-looked, methodological problem of the ecological fallacy. Aggregate variables reveal relationships at broader levelsthat do not necessarily parallel relationships found at the individual level. This study considers different types of aggregationeffects, and clarifies why this case study is particularly susceptible to the adverse implications of the ecological fallacy. Despitethese methodological issues, a process-led approach proves to be a valuable means to incorporate theory more explicitly intoland-change models than achieved in pattern-led studies.© 2003 Elsevier B.V. All rights reserved.

Keywords:Land-change modeling; Agricultural change; Scale dynamics; Madagascar

1. Introduction

Land-use/cover change (LUCC) research seeks toimprove understanding of land changes under diversecultural, economic, political, institutional and envi-ronmental situations (Lambin et al., 1999). Emergingas “integrated land science” (Klepeis and Turner,2001), this agenda has stimulated a range of stud-ies, including two broad approaches to land-change

∗ Tel.: +1-707-664-2183; fax:+1-707-664-3332.E-mail address:[email protected] (R.M. Laney).

modeling (Geoghegan et al., 1998; Veldkamp andLambin, 2001). “Empirical” and largely inductiveland-change models are predicated on statistical cor-relations between land change and a suite of explana-tory variables that provide insight into this change(e.g.Reis and Margulis, 1991). These models employa range of techniques, including multivariate statistics,transition probabilities and simulations to identifythe variables and relationships in question (Lambin,1997). “Theoretical” and largely deductive mod-els draw on general concepts and theories from thehuman and environmental sciences to guide which

0167-8809/$ – see front matter © 2003 Elsevier B.V. All rights reserved.doi:10.1016/j.agee.2003.09.004

Page 2: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

136 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

explanatory variables to “test” in regard to land change(e.g.Serneels and Lambin, 2001; Turner et al., 2001;Walker et al., 2000). This latter approach signalsgrowing interest in moving from statistical relation-ships per se to a conceptual base for understandingwhy land change occurs (Lambin et al., 2000).

Both above approaches are pattern-led. They sharea predisposition to identify land-cover patterns first,be they spatial, temporal or both, and then developmodels that explain or replicate those patterns. Theprocesses in question do not greatly influence how theland-cover patterns (dependent variables) are charac-terized. A process-led study, on the other hand, usestheory to guide the characterization of land-cover pat-terns and to shape more directly the process–patternrelationships explored in the land-change model (e.g.agent-based models;Parker et al., 2002). This studytakes this latter approach.

Leading with process is an important approach toexplore for several reasons. It structures the modelaround the critical human–environment relationshipsidentified within the theory, and focuses attention onthe data required to explore those relationships. Incontrast, a pattern-led approach too frequently drawson readily available data (often proxy variables) thatmay not fully represent the processes at work, forcingthe analysis to sidestep the role of these processes(Irwin and Geoghegan, 2001). Pattern-led studies alsotend to explore only those processes and conceptsthat are likely to explain the observed cover pat-terns. These cover patterns are often conspicuous,such as those observed from space-based platforms,and they are heavily influenced by the techniques ofspatial-pattern analysis employed. Thus, these studiespotentially miss cover patterns that signal importantprocesses of change, but are not easily discerniblefrom space or are not captured by the particular patternanalysis applied. Finally, with a process-led approach,theory more directly guides the level of analysisand the grain of cover data chosen. Pattern-led stud-ies, on the other hand, are likely to search for theprocesses that best explain cover patterns at the grainand level of their cover data. Process-led studies can-not dissociate the model design from the influence ofland-change data parameters and spatial-pattern anal-ysis either, but theory can more purposefully structurehow the patterns of land change are characterized andanalyzed. On the downside, a process-led approach

may “blind” the analyst to alternative processes atwork, but in principal, the test results should signalweaknesses and demand a re-evaluation of the dataand of the mode of analysis.

The process-led approach used here employs theinduced-intensification thesis, which characterizesagricultural systems and their change in the lessdeveloped world (Chayanov, 1966; Boserup, 1965;Turner and Ali, 1996). The study applies this thesisto a case study in the Andapa region of northeastMadagascar (Fig. 1). In Andapa, recently enforcedconservation laws have abruptly ended the agricul-tural frontier and smallholders now respond to rapidlyrising land pressure. The study uses the thesis’ typol-ogy of agricultural change trajectories to differentiatevarious farmer “strategies” in response to that pres-sure. It also identifies the land-change “profiles,” orthe kinds, rates and quantities of cover change (butnot their spatial pattern), produced by those strategies.The study statistically characterizes these farmer-level“strategy-profile” linkages, and then uses them to cre-ate a village-level model of landscape change underdifferent dominant trajectories of agricultural changefound in the Andapa region.

The development of the model is predicated onestablishing two fundamental hypotheses. The firsthypothesis maintains that each type of farmer-levelmanagement strategy creates a characteristic farm-levelland-cover profile. If these strategy-profile linkagesarenot characteristic, then a process-led model wouldnot be able to predict characteristic land-cover out-comes. To test the hypothesis, the study uses discrim-inant analysis to evaluate cover profiles for varianceby strategy, separability between strategies, and tem-poral consistency. The second hypothesis assertsthat a village-level dominant management strategyproduces a characteristic village-level land-cover out-come. This hypothesis is based on the assumptionthat farmer-level strategies and cover profiles can beusefully aggregated to the village level. To test thissecond hypothesis, the study develops a model that ag-gregates the cumulative cover consequences of manyfarmers operating in the same landscape, and thenevaluates, through descriptive statistics, whether thelandscapes produced by particular dominant strategiesare characteristic.

The study carefully tracks process–pattern linkagesacross levels of analysis in the aggregation model,

Page 3: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 137

Fig. 1. The Andapa region.

monitoring how the aggregation process itself affectsthe portrayal of those linkages. Aggregation effectshave long been appreciated in spatial geography(Openshaw and Taylor, 1981), remote sensing andGIS (Bian and Butler, 1999), and ecology (Rastetteret al., 1992), but they are only just beginning to besystematically investigated in the land-change litera-ture (e.g.Walsh et al., 1999). This study differentiatesbetween two aggregation effects, showing conditionsin which aggregation’s “smoothing effect” exposesunderlying broad-level, large-grain signals and un-covers broad-level process–pattern relationships, andthose in which aggregation’s “canceling and ampli-fying effects” obfuscate low-level, fine-grain signalsand reveal false broad-level relationships. The resultsof this study confirm the long-established spatial dy-namic that the level of the analysis affects the resultsobtained (Haggett, 1964) by showing that changingthe level of analysis changes the process–pattern re-lationships identified. It goes a step further, however,to reveal conditions in which changing spatial scalecan revealfalseprocess–pattern relationships.

2. The induced-intensification thesis

The induced-intensification thesis, so coined byTurner and Ali (1996), captures an interdisciplinaryarray of similar agricultural change theses in whichsmallholders respond to shifts in demand for produc-tion, leading to shifts in the intensity of cultivation.Demand pressures may arise from many sources, in-cluding population growth, aspirations and markets(Brush and Turner, 1987). Farmers choose to inten-sify (or to follow other strategies) based on thesepressures, as well as on the rationale inherent to theirsubsistence, market, or mixed production system.Demand pressures, and responses to it, are furthermediated by environmental conditions, access to tech-nology, land tenure and other resource institutions,and land use policies.

Agricultural change under rising pressure is notlinear, but follows a stair-step trajectory of intensifica-tion slopes (Brookfield, 1984; Turner and Ali, 1996).Within any techno-managerial strategy, changes inpressures are typically met by concomitant changes in

Page 4: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

138 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

sh

ort fallow

irrigated

long fallow

no change

[b]

[a]

outp

ut/la

nd a

rea/

time

inputs/land area/time

outp

ut expansion

total land area under production

disintensification

innovativeintensification

excessive cropping frequency

non-innovativeintensification

Fig. 2. The induced-intensification thesis’ model and typology ofsix trajectories of agricultural change, showing a stair-step growthpattern in agricultural development. The concave curvature of eachintensification slope represents declining marginal returns to in-puts within a techno-managerial level. Points of intersection be-tween intensification slopes represent thresholds to intensificationbetween techno-managerial levels.

input intensity, but as declining marginal returns setin for a particular strategy, qualitative changes are re-quired to shift the techno-managerial system to a newstate (e.g. rain-fed to irrigated cultivation), calledinno-vative intensification(Fig. 2). In some cases, however,

Table 1Definitions of management strategies

Management strategy Definition

Hill-rice sectorExcessive cropping frequency Increasing the amount of land area under production each year without clearing new land,

inducing a decline in fallow cycles.Expansion into claimed forest Cutting family-claimed forest to put new land into productionExpansion into the reserve Cutting protected reserve land to put new land into productionFalse expansion Borrowing or renting another farmer’s land, thereby stabilizing fallow cycles on the borrower’s

own land. At the village level, this strategy can only dominate if many farmers borrow landfrom other villages.

False intensification Lending or renting land to others. The lender does not respond to its own pressures, but to asocial obligation to those in need.

No change Maintaining the same amount of land under production each year, stabilizing cropping intensity

Market-crop sectorExpansion Adding enough new vanilla and coffee trees to exceed LMUs demographic growthNo change Not adding new plants, or not planting enough to exceed LMUs demographic growth

barriers to this shift may be severe, and absent alterna-tives, may lead toexcessive cropping frequency, wherefarmers seek increases in output by increasing landarea in cultivation, inducing a decline in the fallowcycles. The induced-intensification thesis has codifiedthese strategies, and others, into a typology, and thiscase study uses this typology, with some modifica-tions, to structure its land-change model.Table 1iden-tifies and defines those strategies relevant to this paper.

3. Study area

The rugged tropical forest of the Andapa region wassparsely populated until the 1920s when French andReunionais coffee and vanilla plantation owners dis-covered the valley’s rich, alluvial soils (Neuvy, 1989).Immigrants poured into the region to work as hiredhands on the new plantations, but quickly establishedsmall plantations of their own. Smallholders provedresilient to fluctuations in the vanilla market, and by1925 small farmers produced 90% of all vanilla fromthe region (Portais, 1972). Madagascar soon becamethe largest producer of vanilla in the world, with nearlyall of this production located in the “vanilla triangle”between Antalaha, Sambava and Andapa (Fig. 1, in-set). Coffee became an important crop as well, dom-inating in the region’s drier areas. Andapa farmersnever shifted entirely to market production, however,always maintaining a mixed cropping system that in-cluded subsistence production as well.

Page 5: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 139

Hill rice and irrigated rice are the main subsistencecrops. Hill rice is managed in a short-fallow rotationsystem, called “tavy”, where a field is cleared, burned,and planted for a single year, and then left in an unman-aged fallow for several years before being cultivatedagain. Irrigated fields are ubiquitous in the Andapavalley proper (Fig. 1), but in the foothills surround-ing the valley and adjacent to the reserves, which arethe focus of this study, less than 20% of householdsmanage irrigated fields. Moreover, these fields averageless than a quarter hectare per household, leaving hillrice the principal source of subsistence in the foothillvillages.

In the early 1950s, French colonial authorities es-tablished two nature reserves, each lying fairly closeto established villages (Fig. 1). The population grewrapidly, and when land pressure became especiallyhigh in the late 1980s, farmers began to cultivate inthese reserves. Authorities enforced the reserve bound-aries in the early 1990s, thereby arresting agriculturalexpansion. As a result, farmers have had to adopt newstrategies to cope with their increasing consumptionneeds. Very few have chosen to leave, and off-farmemployment opportunities are rare in this remote re-gion. Most have, instead, mobilized strategies withinthe existing production system.

Farmers in the Andapa region organize land accessin a mixed collective/private system. Extended fam-ilies control most fallow land, while individuals andhouseholds own the market fields. The principal unitof analysis in this study is the land management unit(LMU), which, depending on the lifecycle of the fam-ily, may equal a nuclear household, or it may encom-pass multiple generations headed by a family elder.The study focuses on four villages, Mandena, Betso-manga, Befingotra and Andilandrano (Fig. 1), repre-senting a range of agroecological zones and levels ofdemographic pressure found in the region.

4. Statistically characterizing land-cover profiles

Establishing a linkage between agricultural pro-cesses and land-cover outcomes rests upon the fun-damental hypothesis that farmer-level managementstrategies produce characteristic land-cover outcomes.If cover outcomes are highly variable by strategy,or quite similar between strategies, or inconsistent

over time, then there is little hope for predictingdistinct village-level land-cover outcomes under dif-ferent dominant strategies. The first challenge, there-fore, is to identify the kinds, rates, and quantities ofland-cover change, or the land-cover profile, producedby each management strategy, and to evaluate thesecover profiles for variance by strategy, separabilitybetween strategies, and temporal consistency.

4.1. Farmer-level management strategies

About 180 LMUs were interviewed in the four studyvillages, revealing crop choices, land access strategies,field sizes, cropping intensities, outputs, yields, andconsumption levels between 1987 and 1998. Thesedata were used to identify the three managementstrategies that each LMU followed in its hill-rice,irrigated-rice and market-crop sectors (Laney, 2002).The irrigated-rice sector is not discussed in this paper,however, because relatively few LMUs had irrigatedfields in the study area, and the study could not de-velop statistically robust cover profiles. The hill-riceexpansion strategies, while presented inTable 1, arenot included in the land-cover model either because,with enforcement of anti-deforestation laws in 1994,they were no longer viable strategies in the 1995–1998time period.Table 2reports the number of LMUs thatfollowed each strategy in each village over two timeperiods, 1991–1994 and 1995–1998.

4.2. Land-cover profiles

The study develops an individual land-cover profilefor each strategy within the hill-rice and market-cropsectors. Profiles describe both the land-cover changesand end states produced by each strategy. Coverchanges specify the amount of land area that increasedor decreased in each cover class over the 4-year studyperiods. End states indicate the amount of land areafound in each cover class at the end of the 4-yearperiods. Because LMUs are heterogeneous in termsof farm size, plot size, and historical fallow cycles,those that follow the same management strategy donot produce the exact same cover changes and endstates. Consequently, a profile statistically character-izes the average and range of cover changes and endstates associated with a particular strategy.

Page 6: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

140 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

Table 2Number of land management units following each management strategy

Management Strategy Total Mandenaa Betsomangab Befingotraa Andilandranoa

1991–1994

1995–1998

1991–1994

1995–1998

1991–1994

1995–1998

1991–1994

1995–1998

1991–1994

1995–1998

Hill-rice sectorExcessive crop frequency 16 35 9 4 5 10 5 12 6 9Expansion, forest 13 – – – 7 – – – 6 –Expansion, reserve 32 – 5 – 14 – 13 – – –False expansion 33 44 15 18 8 11 6 9 4 6False intensification 5 28 4 9 1 10 – 6 – 3No change 51 53 19 17 10 18 7 6 15 12

Market-crop sectorExpansion 18 26 11 19 6 18 19 8No change 32 24 39 31 27 15 14 25

a Dominant market crop is vanilla.b Dominant market crop is coffee.

Table 3Farmer-level cover profiles in the hill-rice sectora

a Cover changes and end states give the average, in hectares. Variability is shown as one standard deviation. Shaded cells identify thecover changes and end states that differentiate between the strategies in the discriminant analysis.

Land-cover classes include market crops (MRKT),hill rice in cultivation (HILL), young fallow land un-der 5 years of age (FALLOW 1–4), and older fallowland 5 years of age or more (FALLOW 5+).1 Theland-cover data were developed through cover-historymaps reconstructing the 12-year cover history of everyparcel in the study area. Each parcel was sketched ona 1990 SPOT panchromatic image (10 m resolution),

1 Identifying a threshold for fallow-age classes took both man-agement issues and remote sensing technology into account be-cause, in the full study, satellite imagery is used to verify theland-cover model (Laney, 1999). Farmers indicate that when fal-low cycles drop below 4–5 years they are operating under ex-treme land pressure, making this age break a potential indicator ofchanging management strategies. Using the remote sensing anal-ysis technique, divergence, this same age break yields the mostaccurate cover classification.

digitized, and given an identification code. Land-usehistories (based on farmer recollections of land use,SPOT and Landsat TM images of 1990 and 1994,aerial photographs of 1992, and site verifications be-tween 1996 and 1998 during the interview period),were recorded in a database and linked to the digitizedmap with a shared identification code (Fig. 3). Thisdatabase includes over 2500 parcels in the four studyvillages, revealing the amount of land area that eachLMU had under each cover class each year.

To create a land-cover profile for a particular man-agement strategy, the study filtered and selected allLMUs that followed that strategy from within thislarger database. From this subset, the study identifiedthe average and range of cover changes and end statesthat occurred on all of those LMUs’ land.Table 3shows each strategy’s mean cover profile over the

Page 7: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 141

CODE SIZE 87 88 89 90 91 92 93 94 95 96 97 98101 2 3 4 HR 1 2 3 4 5 HR mkt mkt102 12 13 14 15 16 17 18 19 HR 1 2 3103 1 2 3 4 5 6 7 8 9 10 11 HR104 2 3 HR 1 2 3 4 HR 1 2 3 4105 mkt mkt mkt mkt mkt mkt mkt mkt mkt mkt mkt mkt 114 0.9 2 3 4 5 6 7 8 9 HR 1 2 3106 1.0 5 6 7 HR 1 2 3 4 5 HR 1 mkt107 1.1 20 HR 1 2 3 4 5 HR 1 2 3 4

101 102 103

104106107

114108 109

1.01.31.50.90.3

COVER-HISTORY DATABASE

SPATIAL MAP

1 mile

105

Fig. 3. Cover-history map of Betsomanga, illustrating how the cover-history database is linked to the spatial map. In the fields, “code” isthe number matching the spatial map to the database, “size” is the size of field in hectares, and “87–98” identify the years covered inthe database. Data encoding includes the numbers “1–20” to represent fallow ages, “HR” marks when hill rice was cultivated, and “mkt”represents fields under market crops.

1991–1994 time period as well as the variability asso-ciated with each cover variable, represented as a stan-dard deviation. Each mean profile is also illustratedin Figs. 4 and 5, in this case shown as proportions oftotal land area owned.

For the market sector, separate profiles are devel-oped for coffee-dominant and vanilla-dominant mar-ket sectors (Table 4; Fig. 5). Vanilla fields have a

Table 4Farmer-level cover profiles in the market-crop sectora

Strategy Cover change(CHNG MRKT)

End state(MRKT)

Expansion—vanilla 0.2± 0.25 0.8± 0.2No change—vanilla 0.05+ 0.05 0.8± 0.1Expansion—coffee 0.3± 0.2 1.3± 0.5No change—coffee 0.1+ 0.1 0.8± 0.15

a Cover changes and end states show the average, in hectares.Variability is shown as one standard deviation.

planting density of 3300 plants/ha, while coffee fieldshave a planting density of 1100 plants/ha. Thus, in thecoffee-dominant areas, the expansion strategy createsa much larger signal in the CHNG MRKT variable.Note that the CHNG MRKT variable increases a bitwith theno changestrategy as well (Fig. 5). The studyclassified LMU as experiencingexpansiononly if itplanted enough new plants to increase the number oftrees per capita in the family. This avoided classify-ing LMUs asexpandingwhen they only planted a fewtrees.

4.3. Statistical analysis of land-cover profiles in thehill-rice sector

To test the hypothesis that different processes ofagricultural change produce distinct land-cover out-comes, the study evaluates whether these cover profiles

Page 8: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

142 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

CHNG HIL

L

CHNG FALL

1-4

CHNG FALL

5+

HILL

FALL

1-4

FALL

5+

COVER CHANGES END STATES

CHNG HIL

L

CHNG FALL

1-4

CHNG FALL

5+

HILL

FALL

1-4

FALL

5+

excessive cropfrequency

false intensification

false expansion

no change

COVER CHANGES END STATES

1086420

-2-4-6-8

-10

40

20

1086420

-2-4-6-8

-10

40

20

1086420

-2-4-6-8

-10

1086420

-2-4-6-8

-10

40

20

40

20

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent

perc

ent

perc

ent

perc

ent

Fig. 4. Farmer-level cover profiles for hill-rice management strategies.

are distinct from each other. If each profile is dis-tinct, the hypothesis is supported. Profiles are distinctif the distribution of cover outcomes surrounding themean changes and end states for one strategy do not

CHNG MRKT

MRKT

CHNG MRKT

MRKT6

4

2

0

20

10

20

10

6

4

2

0

6

4

2

0

6

4

2

0

20

10

20

10

expansion, vanilla

no change, vanilla

expansion, coffee

no change, coffee

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent

perc

ent

perc

ent

perc

ent

COVER CHANGE

END STATE

COVER CHANGE

END STATE

Fig. 5. Farmer-level cover profiles for market-crop managementstrategies.

overlap significantly with another strategy’s outcomes.When the mean outcomes are close and the standarddeviation of possible outcomes around those meansoverlap significantly, as appears in the case of theex-cessive cropping frequency and false intensificationstrategies(Table 3), then this suggests that the twodifferent strategies are creating the same land-coveroutcome.

The study uses discriminant analysis to establish ifcover profiles are, in fact, statistically distinct (Klecka,1980). It applies discriminant analysis, as opposedto other regression-based methods, because thereare more than two dependent variables (manage-ment strategies).2 The discriminant technique de-velops a set of classification rules that describe thecover changes and end states (independent variables)

2 Discriminant analysis is based on several assumptions (Klecka,1980). Laney (1999)provides a full accounting of how eachof these assumptions was tested. Only one assumption, normalindependent variables, was violated in a few cases. In these cases,the variables were log transformed in order to achieve normalitybefore being included in the models.

Page 9: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 143

associated with each management strategy. Theserules are derived from a sample of cases, and thenapplied to the remaining cases for verification. Asuccessful discriminant model is able to classify eachremaining LMU to the management strategy that itactually followed, based on the cover changes andend states that occurred on its land. A high rate ofmisclassification between two management strategiessuggests that the two strategies share very similarcover profiles. This study conducted discriminantanalysis on the hill-rice sector strategies only. For themarket sector, theexpansionandno changestrategiesare so distinct (in the CHNG MRKT variable) thata statistical assessment of their separability is notnecessary (Table 4).

In the study’s first discriminant test, the rules weredeveloped from the cover profiles of half of all LMUsin the 1991–1994 time period (Table 2; TOTAL).These rules were then applied to the other half of theLMUs in the same time period. High rates of mis-classification of LMUs to the wrong strategy wouldsuggest that the strategies have similar land-coverconsequences. In a second test, the rules were devel-oped from the cover profiles of all LMUs from the1991–1994 time period. These rules were then appliedto the 1995–1998 time period. A successful discrimi-nant model classifies each LMU to the managementstrategy that it actually followed in the 1995–1998time period. High rates of misclassification suggestthat either the strategies have similar cover conse-quences, or their cover profiles were not consistentover time.

Discriminant analysis produces many kinds of out-put describing the classification rules, the independentvariables that discriminate the dependent variables,and the statistical significance of the results.Table 3highlights (with shading) the cover variables that dis-criminate strategies well enough to be statistically sig-nificant, and which strategies are shown to be distinctfrom each other by those variables.

4.4. The distinctiveness of land-cover profiles

The two intensification strategies,excessive crop-ping frequencyandfalse intensification, considered asa pair, produce a land-cover profile that is decidedlydifferent from the cover profile produced by thefalseexpansionandno changestrategies, again considered

as a pair (Table 3). The CHNG HILL variable dif-ferentiates these two sets of strategies. The latter twostrategies produced outcomes distinct from each otheras well. The HILL and FALLOW 1–4 end-state vari-ables discriminated between the two strategies. Theformer two intensification strategies, however, werenot statistically different from each other.

The classification matrix for the first test, restrictedto the 1991–1994 time period, verifies this interpre-tation (Table 5). The matrix shows the number ofLMUs correctly assigned to the management strategythat it actually followed, and the number of LMUsmiss-classified to another strategy. Thefalse expan-sion andno changestrategies experienced little mis-classification with other strategies. As well, their levelsof classification accuracy (percent correct) are greaterthan would be expected by chance (reported as maxi-mum chance criteria,Cmax in Table 5).3 Theexcessivecropping frequencyandfalse intensificationstrategies,however, experienced very high rates of misclassifi-cation between each other. The land-cover outcomesproduced by these two different processes are statisti-cally inseparable.

The results of the second discriminant test estab-lish temporal consistency. Only a slight degradationof overall classification accuracy (74–67%) occurswhen classifying all 1995–1998 cases, based on the1991–1994 classification rules (Table 6). Otherwise,the patterns of misclassification between strategies arethe same, suggesting that strategies create essentiallythe same landscape profiles over time.

These moderately successful results, with threecharacteristic profiles (considering the two intensifi-cation strategies as a set), and temporal consistency,provide enough evidence to suggest that a land-changemodel based on these strategy-profile linkages maybe successful in identifying characteristic land-coveroutcomes in landscapes dominated by differentstrategies.

3 The maximum chance criteria,Cmax, is the percentage ofcorrectly classified cases that would be achieved if all cases wereassigned to the group with the highest probability. Proportionalchance criteria,Cpro, accounts for the relative proportion of casesin each group, and is calculated by squaring the proportions ofeach group. Of the two tests, whichever has the highest figure isthe measure that the model should outperform (Hair et al., 1995),Cmax is the higher figure, and is the only one reported in bothTables 5 and 6.

Page 10: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

144 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

Table 5Classification results of (withheld) 1991–1994 casesa

True strategy Classified strategy

False expansion Excessive cropfrequency

No change Falseintensification

Percentagecorrect

False expansion 14 1 5 0 70.0Excessive crop frequency 0 12 0 0 100.0No change 5 0 15 0 75.0False intensification 0 3 0 0 0.0

Total 19 16 20 0 74.5

a Cmax = 0.36 (see Footnote 4).

Table 6Classification results of 1995–1998 casesa

True strategy Classified strategy

False expansion Excessive cropfrequency

No change Falseintensification

Percentagecorrect

False expansion 35 2 5 0 83.3Excessive crop frequency 5 23 0 6 67.6No change 15 3 36 0 66.7False intensification 5 10 1 10 38.5

Total 60 38 42 16 66.7

a Cmax = 0.34 (see Footnote 4).

5. Modeling village-level land-cover outcomes

Modeling characteristic village-level cover out-comes produced by different trajectories of changerests on a second critical hypothesis that dominantstrategies produce distinct village-level cover out-comes. While strategies may produce distinct coveroutcomes at the farmer level, the cumulative andcontradictory effects of many farmers acting within alandscape may obscure these process–pattern linkagesat the broader level.

A dominant village-level trajectory is defined asthe set of strategies adopted by most LMUs in a vil-lage. Basing dominance on the number of LMUs andnot on a spatial factor, such as the strategy affectingthe greatest land area, reflects the study’s process-led,and not pattern-led, approach. The study’s overrid-ing goal is to connect the driving forces of changeto management strategies to landscape effects. The-oretically, or at least by convention within theinduced-intensification thesis, a dominant force in-ducesmostLMUs to adopt a particular strategy. Yet,

this definition leaves open the possibility that a fewland-rich LMUs following other strategies may pro-duce cover outcomes that overwhelm the outcomes ofthe dominant strategy. If this occurs, then the hypo-thetical linkage between dominant forces, dominantstrategies and dominant land-cover outcomes is notverified. This potential problem is evident in one ofthe study villages, Befingotra, where most LMUsfollow excessive crop frequency, but false intensifi-cation affects the greatest land area in the hill-ricesector. Nevertheless, the study maintains a definitionof dominance that reflects its process-led approach.

Since LMUs follow many different managementstrategies within a village, dominance does not nec-essarily describe a majority of LMUs and may onlyreflect the statistical mode (Table 2). For this reason,the study develops a land-cover model that incorpo-rates farmer variability in land use. It does not simplymodel the cover outcomes associated with the dom-inant process. Rather, it aggregates the cumulativeconsequences of the non-dominant strategies occur-ring along with the dominant strategy. Thus, just as

Page 11: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 145

farmer-level cover profiles represent a range of possi-ble cover outcomes, village-level land-cover outcomesinclude a range of potential outcomes, given bothfarmer variability and the presence of non-dominantstrategies in the village.

To test the hypothesis that farmer-level process–pattern linkages manifest at broad levels, the studyevaluates two criteria of success in the model.

1. Different village-level dominant strategies producestatistically different cover outcomes.

2. A single dominant strategy produces similar coveroutcomes at the farmer and village levels.

Criteria 1 verifies that two villages, one experienc-ing dominant strategy “A” and the other strategy “B”have ranges of potential cover outcomes that do notoverlap. Criteria 2 confirms that the cover outcome invillage “A” is similar to the farmer-level cover pro-file associated with strategy “A”. In other words, theunique cover patterns identified at the village level canbe attributed to the correct processes that created them.

5.1. The model

The model aggregates the farmer-level cover pro-files of all LMUs in a village. To predict a future land-scape outcome, the model projects the cover profilesdeveloped from the 1991–1994 time period (Table 3)into the 1995–1998 time period, accounting for thechange in number of LMUs following each strategy(Table 2).

The model is represented inEq. (1). To produce anend-state outcome (K) for a cover type (1) in 1998(K98

1 ), the model multiplies the end state (COV) ofcover type 1 produced by a single LMU following amanagement strategy (ms1) over the 1991–1994 timeperiod (COV1 ms94

1 ) times the number of LMUs fol-lowing that strategy in a village in the 1995–1998 timeperiod (N98). It then adds these subtotals across allstrategies (ms1, ms2,. . . ) for cover type (1). The sameequation is used to derive all village-level change andend-state outcomes:

K981 = (COV1 ms94

1 N98) + (COV1 ms942 N98) + · · ·

(1)

The aggregation model must be able to handle the factthat the cover change and end-state variables (COV)

have both means and distributions, accounting for therange of land-cover outcomes associated with eachstrategy. Moreover, the distributions are not necessar-ily normal; some, in fact, are quite skewed. Assum-ing normality would distort the model’s predictions intwo ways: (i) averages would suggest false central ten-dencies, causing error in predicted average outcomes;(ii) standard deviations might be “overconfident”, sug-gesting narrower distributions than are representativeof true variability. The model, therefore, must be ableto aggregate non-normal distributions.

The study uses Monte Carlo simulation to performthe aggregation. Monte Carlo simulation provides ameans to propagate quantitative uncertainty throughnumerical models (Dakins et al., 1996). The techniquesimulates the actual (normal or non-normal) distribu-tions of each input parameter (COV), and creates sce-narios that reflect the distribution of possible outcomes(K). A separate set of simulations is run for each of theeight cover change and end-state variables. Output foreach variable is an empirical frequency histogram ofthe results of 500 simulations. The histogram revealsan average and a distribution of potential outcomes forthe cover variable.4 This study ran the Monte Carlosimulations within the risk-analysis software CrystalBall (Decisioneering, 2000).

Sets of simulations were run for each village sepa-rately. In each simulation,N represented the numberof LMUs following each management strategy in thevillage over the 1995–1998 time period. Finally, inorder to facilitate comparative analysis between vil-lages of different sizes, the model’s output is con-verted into proportions of total land area within eachvillage.

5.2. Landscape outcomes at the village level

Overall, the aggregation model performs moder-ately well in predicting village-level land-cover out-comes. Actual land covers (based on the land-coverdatabase (Fig. 3)), fall within the model’s distributionof predicted covers (uncertainty bars with two stan-dard deviations) in 75% of cases in the four villages

4 Since the output is a frequency histogram showing the resultsof empirical simulations, no statistical test of accuracy is made.The accuracy of the model is tested empirically against satelliteimagery inLaney (1999).

Page 12: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

146 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

CHNG HIL

L

CHNG FALL

1-4

CHNG FALL

5+

HILL

FALL

1-4

FALL

5+

COVER CHANGES END STATES

Betsomanga

Mandena Andilandrano

Befingotra

1086420

-2-4-6-8

-10

40

20

1086420

-2-4-6-8

-10

COVER CHANGES END STATES

CHNG HIL

L

CHNG FALL

1-4

CHNG FALL

5+

HILL

FALL

1-4

FALL

5+

40

20

40

20

40

20

1086420

-2-4-6-8

-10

1086420

-2-4-6-8

-10

perc

ent c

hang

e

perc

ent

perc

ent c

hang

e

perc

ent

perc

ent c

hang

e

perc

ent

perc

ent c

hang

e

perc

ent

= Model's predicted outcomes = Actual cover outcomes | = Uncertainty bar

Fig. 6. Actual and predicted village-level land-cover outcomes for the hill-rice sector. Predicted land covers show the output of theland-change model. The uncertainty bars show two standard deviations around the predicted mean, and represent variability given both therange of cover outcomes created by farmers following the same strategy and the presence of non-dominant strategies in the village.

(Figs. 6 and 7).5 The model was best able to predictthe CHNG HILL, HILL, CHNG MRKT and MRKTvariables, with actual and potential covers overlap-ping in all but one case. The model’s uncertainty barsaround these four variables are also fairly narrow. Themodel performs most poorly with the FALLOW 1–4and FALLOW 5+ change and end-state variables, pro-ducing very wide uncertainty bars and failing to pre-dict actual cover outcomes in many cases (Fig. 6).

In the hill-rice sector, the model predicts distinctlandscape outcomes between villages experiencing

5 The model fits some villages more poorly than others becausethe profiles are derived from a regional average, and the modelcannot account for certain village-specific patterns. For example,Betsomanga had a much higher than average amount of landin old fallow at the beginning of the study, so the model doesnot anticipate this village’s exceptionally large decline and highremaining end state in old fallow.

CHNG MRKT

MRKT

Betsomanga

Mandena Andilandrano

Befingotra

CHNG MRKT

MRKT

= Model's predicted outcomes

= Actual cover outcomes

| = Uncertainty bar

perc

ent c

hang

e

6

4

2

0

40

20

40

20

6

4

2

0

6

4

2

0

6

4

2

0

40

20

40

20

perc

ent

perc

ent

perc

ent

perc

ent

perc

ent c

hang

e

perc

ent c

hang

e

perc

ent c

hang

e

Fig. 7. Actual and predicted village-level land-cover outcomes forthe market-crop sector.

Page 13: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 147

Village Dominant Strategy Patterns[a] Befingotra excessive crop frequ. excessive crop frequ.

[b] Betsomanga no change excessive crop frequ.

[c] Andilandrano no change excessive crop frequ.

[d] Mandena false exp./ no chg unclear

[e] Befingotra exp. / no change - vlla expansion - vlla

[f] Betsomanga no change - coffee expansion - vlla

[g] Andilandrano no change- vlla no change - vlla

[h] Mandena exp. / no change - vlla expansion - vlla

Hill-riceSector

Market-cropSector

Fig. 8. The dominant strategy that actually occurred in each village, and the strategy that one would interpret occurred in each villagefrom its landscape patterns alone, 1994–1998.

different dominant strategies. For example, the areaestimates and uncertainty bars for the CHNG HILLand HILL variables for Andilandrano (no change)and Befingotra (excessive cropping frequency) do notoverlap (Fig. 6). These results support Criteria 1,showing that different dominant strategies producedistinct land-cover outcomes. But, these two variablesare the only ones that consistently diverge betweenvillages experiencing different dominant strategies.The fallow variables, on the other hand, display suchhigh variability that their uncertainty bars overlapbetween villages in nearly all cases (Fig. 6).

It is not surprising that the fallow variables areso indistinctive. Their variability in the farmer-levelcover profiles is high (large standard deviations,Table 3), and this variability propagates in the ag-gregation model. Nevertheless, their failure to lenddiscriminating power to landscapes experiencing dif-ferent dominant hill-rice strategies leaves only theCHNG HILL and HILL indicator variables to differ-entiate landscapes. Referring back to the farmer-leveldiscriminant model (Table 3), both the FALLOW1–4and the HILL end-state variables differentiate land-scapes experiencingfalse expansionand no change.The aggregate model indicates, however, that at thevillage level, variability is too high in the fallow vari-ables and only the HILL variable is available to dif-ferentiate landscapes experiencing these two differentstrategies.

Although the model reveals distinct village-levelcover outcomes (Criteria 1), to fully support the sec-ond hypothesis, the characteristic cover outcomes

must be attributable to the dominant strategy occur-ring in the villages (Criteria 2). Befingotra, whichexperiencedexcessive cropping frequency, illustratesa case that fulfills both criteria. First, its landscapeis distinct. The CHNG HILL and HILL variables areboth quite high, and their uncertainty bars do notoverlap with any of the other three villages (Fig. 6).Second, a high reading in these variables correspondswith the farmer-level cover profile forexcessive crop-ping frequency(Fig. 4). Befingotra’s cover profile isalso consistent with thefalse intensificationstrategy(Fig. 4), which is expected since these strategies’cover profiles are not statistically different.6 The dom-inant strategy clearly produced Befingotra’s distinctvillage-level landscape pattern, revealing a consistentprocess–pattern relationship (Fig. 8(a)).

Betsomanga and Andilandrano represent cases inwhich the village-level cover outcomes do not reflectthe dominant strategy, and Criteria 2 does not hold.Both villages experience the same dominant strategy,no change, and both share cover outcomes that aredistinct from the other two villages in the CHNGHILL and HILL variables (Fig. 6)—upholding Cri-teria 1. Yet, their village-level cover outcomes arenot consistent with the farmer-level cover profileof the no changestrategy. Theno changestrategyshould decrease the CHNG HILL variable (Fig. 4),yet both villages show a moderate increase in CHNG

6 Befingotra’s fallow variables are consistent with theexcessivecropping frequencyprofile as well, but the fallow variables are notstatistically significant, and thus cannot be used reliably.

Page 14: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

148 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

HILL ( Fig. 6). The increase is attributable to the twonon-dominant strategies,excessive cropping frequencyand false intensification, that are also occurring inthese villages. Both of these non-dominant strategiesincrease the CHNG HILL variable, and together, theyimpact a greater proportion of land area than thenochangestrategy. These “canceling effects” also createspurious signals in the HILL variable. At the farmerlevel, theno changestrategy creates a relatively lowHILL end state (Fig. 4), while the non-dominantex-cessive crop frequencystrategy produces a high HILLend state. With significant numbers of LMUs follow-ing the latter non-dominant strategy (Table 2), theseLMUs produce a moderately high HILL end-statevariable at the village level. This landscape patterncorresponds more closely withexcessive croppingfrequencythan no change. As a result, the domi-nant strategy does not create the dominant landscapepattern, and the process–pattern relationship breaksdown (Fig. 8(b) and (c)).

Mandena demonstrates confusing results, withstrong but inconsistent signals from non-dominantstrategies in its landscape outcome. The village’s twodominant strategies,false expansionand no change,should leave less than 10% of land area in hill riceproduction (Fig. 4), but the village actually has nearly20% land area in production (Fig. 6), which stronglysuggests theexcessive cropping frequencystrategy(Fig. 4). As well, Mandena should experience a de-crease in the CHNG HILL variable (Fig. 4). But,the village has a zero net change in this variable(Fig. 6). In this case, the two non-dominant strategies,excessive cropping frequencyand false intensifica-tion, increase the variable, thereby “canceling” thedominant signal. In composite, Mandena’s villagelevel outcome does not clearly reflect any strategy(Fig. 8(d)).

Turning to the market sector, three of the fourcases, Andilandrano, Befingotra and Mandena sup-port the second hypothesis. Andilandrano has a coverpattern quite distinct from those in Befingotra andMandena, where different dual strategies dominate(Criteria 1). As well, their cover patterns closelymirror the farmer-level profiles of their dominantstrategies (Criteria 2) (Fig. 8(e), (g) and (h)). Bet-somanga does not support the hypothesis, however.It fails Criteria 1 because its cover pattern is notdistinct from patterns in Befingotra and Mandena.

It fails Criteria 2 because its cover pattern lookslike an expansion—vanillaprofile, whereas its dom-inant strategy isno change—coffee(Fig. 8(f)). Inthis case, the non-dominantexpansion—coffeestrat-egy is “canceling” the dominantno change—coffeestrategy’s signal, and outcome is a spurious signalthat resemblesexpansion—vanilla(Fig. 8(f)).

6. Discussion

The characteristic farmer-level strategy-profilelinkages found in this study demonstrate the capac-ity for a process-led modeling approach to establishprocess–pattern linkages. The fact that two differ-ent management strategies can produce statisticallysimilar land-cover outcomes (e.g.excessive croppingfrequencyandfalse intensificationin the hill-rice sec-tor) implies, however, that management strategies donot necessarily create “thumbprints” on the landscapethat clearly reveal the processes that created them.This result parallels landscape studies in the 1960sand 1970s, which observed that a particular spa-tial pattern can result from very different processes(Getis and Boots, 1978; Rediscovering GeographyCommittee et al., 1997).

This study also shows that a land-cover changemodel based on aggregating the cumulative conse-quences of farmer-level land-cover outcomes can suc-cessfully reproduce village-level landscape outcomes.The model correctly predicted actual land-cover out-comes for many critical cover variables, although notfor the fallow variables. Nevertheless, results refutethe second hypothesis thatdominantstrategies cre-ate thedominantvillage-level landscape outcomes. Infour out of eight cases, the village-level dominant pro-cess did not create the village-level dominant patternFig. 8(b),(c),(d) and (f)).

Two dynamics, spatial and scalar, may cause thebreakdown in the village-level process–pattern link-age. First, the linkage may be in peril because adominant process in this study is not necessarily spa-tially dominant. Following a process-led approach,this study defines dominance based on the strategychosen by most LMUs (statistical mode). But, a fewland-rich LMUs following a non-dominant strategymay produce land-cover outcomes that overwhelmthe dominant strategy’s cover effects. This spatial

Page 15: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 149

dynamic is not, however, the cause of the break-down in the process–pattern linkage in this study.For example, in Betsomanga and Andilandrano, mostLMUs follow the no changestrategy, and most landwas managed under theno changeas well. Yet, theland-cover outcomes in these villages strongly reflectthe non-dominantexcessive crop frequencystrategy(Figs. 6 and 8(b) and (c)). Spatially predominant“non-dominant” strategies did not cause the discon-nect between dominant processes and patterns.

A second scale dynamic, aggregation, maybe the most important causal factor weakeningprocess–pattern linkages in this study. The study’ssecond hypothesis rests on the assumption that ag-gregate variables reveal broad-level process–patternlinkages. Aggregate process variables should revealthe dominant strategy, and aggregate cover variablesshould reflect this same strategy’s cover outcomes.This study’s results clearly show, however, that thisassumption does not hold.

6.1. The ecological fallacy

The effects of data aggregation on objects of studyis recognized in many disciplinary contexts, rang-ing from aggregating pixels in remote sensing andraster GIS (e.g.Bian and Butler, 1999), to landscapepatterns in landscape ecology (e.g.Turner, 1987), tocensus tracks in political and spatial geography (e.g.Openshaw, 1984), and people in sociology (Robinson,1950). Addressing both spatial and non-spatial ag-gregation, they all face the more generally under-stood problem of the ecological fallacy, which assertsthat aggregated variables are not equal to individualvariables, and that aggregation reveals relationshipsbetween variables at higher levels of abstraction thatdo not hold at the individual level (Gehlk and Biehl,1934; Robinson, 1950).7 Correlations can even re-verse from positive to negative, or vice versa, as dataare made coarser (Openshaw, 1978).

7 When two variables describing individuals are aggregated toa higher level, the statistical object becomes the group, and therelationship between the two groups is anecological correlation(Robinson, 1950). The different underlying distributions of thetwo variables will preclude the grouped variables from having thesame relationship that they had at the individual level because,with each aggregation, the variables change independently, relativeto their respective underlying distributions.

The ecological fallacy does not suggest that, whenbroad-level correlation coefficients are different fromlower-level coefficients, the broad-level relationshipsare necessarily false. Different forces may operateat different levels, generating different relationships(Levin, 1992). Aggregation may simply parallel thisscale dynamic and expose these relationships.Fig. 10,scenario (b) illustrates this dynamic in the context ofa land-change study. In this scenario, the aggregationof each set of process and pattern variables reveals anew set of linkages at the broad level, but the integrityof the linkage holds within levels.

Walsh et al. (1999)experience this dynamic intheir study of land change in Thailand. They identifystrong relationships between social variables (e.g.population density) and certain landscape patterns(e.g. forest patches) in fine-grain data, and betweenbiophysical variables (e.g. topography) and broaderlandscape patterns in the coarse-grain data. They notethat, in the land-cover data, forest patches attributableto social factors are lost as spatial aggregation ab-sorbs the fine-grain covers into the more dominantsurrounding land covers. Similarly, in the explanatory(process) data, variation and patterns in the socialvariables become less pronounced at coarser grains,while patterns in the broad-level biophysical variablesbecome more prominent.8 Aggregation, in theWalshet al. (1999)case, has a “smoothing effect” on bothsides of the process–pattern relationship, revealing abroad-level relationship between biophysical processvariables and landscape patterns that are otherwiseobscured by low-level, fine-grain, high-frequencyvariability. Fig. 9(a) illustrates this “smoothing effect”abstractly.

Notably, the Walsh study aggregates its data vari-ables from fine-grain to large-grain through spatialaggregation (pixels). This study, on the other hand, ag-gregates through summations and averages of the in-dividual (LMU-level) variables, and the groupings (bystrategy) are not spatial. Nevertheless, both spatial andnon-spatial aggregations contend with the same issue

8 While Walsh et al. (1999)identify biophysical factors atbroader levels, others have argued that other social variables mayalso only emerge at these broader levels.Meyer and Turner (1992),for example, suggest that relationships such asI = PAT (impactis a factor of population, affluence and technology) are lost underhigh frequency low-level variability and can only be detected inbroad level (aggregated) studies.

Page 16: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

150 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

[a] Smoothing [b] Canceling and effect amplifying effects

Broad Level

Low Level

DominantNon - Dominant

Fig. 9. Aggregation’s smoothing vs. its canceling and amplifyingeffects operating on both sides of the process–pattern linkage.

of the ecological fallacy, where broad-level relation-ships do not equal relationships found at lower levels.

6.2. Aggregation’s “canceling” or “amplifying”effects

There is one critical difference between this studyand the Walsh study that make this study perhapsmore susceptible to identifying false broad-level re-lationships after aggregation rather than finding justdifferent, but valid, broad-level relationships. In theWalsh study, the two social and biophysical drivers ofland-use and land-cover change each operate at dif-ferent levels and grains. The aggregation process par-allels this scalar dynamic. In this study, the dominantand non-dominant proximate causes of land-cover out-comes all operate at the same low level and fine grain.There is no broad-level, large-grain pattern for thesmoothing effect of aggregation to recover.

When the signals of dominant and non-dominantstrategies manifest at the same fine grain, thenon-dominant signals interfere with the dominantsignal. When averaged or summed, the non-dominantsignals either cancel or heighten the amplitude of thedominant signal.Fig. 9(b) illustrates this effect. Theaggregate variables are sensitive to the variability cre-ated by the non-dominant strategies because there isno low-frequency dominant pattern to overpower thehigh-frequency variability. For aggregation to exposethe dominant signal, the dominant strategy’s fine-grainsignal must be considerably “stronger” (of higheramplitude and/or of an overwhelming majority). Oth-erwise, the aggregate variables may reveal the signalsof non-dominant strategies, or perhaps even spurioussignals. This “canceling” effect caused four of eight

Fig. 10. Changing process–pattern relationships across levels ofanalysis. “A” and “B” each represent different sets of drivingforces, management strategies and landscape patterns expected tooccur together. Case (a) represents the situation in which aggregatevariables properly translate to the broader level what is occurringat the lower level. Case (b) suggests that different processes andpatterns emerge at the broader levels. Cases (c) and (d) representcases where, because of the ecological fallacy, aggregate variablesmisrepresent dominant processes or patterns at the broader level.

cases in this study to show inconsistency betweendominant processes and dominant landscape outcomesFig. 8(b),(c),(d) and (f)). Fig. 10, scenario (c) illus-trates this worst-case scenario of the ecological fallacy.

Like (Walsh et al., 1999) and others (Haggett, 1964;McCarty et al., 19569) have long noted, this studyshows that research teams operating at different lev-els of analysis will likely reach different conclusionsabout those relationships. But even further, this studyshows that, where dominant processes only operate atthe low level and fine grain, broad-level, large-grainstudies may reach inaccurate conclusions.

6.3. Effect of different methods of aggregation

This paper has focused attention on the patternvariables, highlighting the ecological fallacy on thepattern side of the relationship (Fig. 10(c)). Processvariables, however, are equally subject to the eco-logical fallacy. Laney (2002)explores the effect of

9 Cited in Meyer, W.B., Gregory, D., Turner II, B.L., McDowell,P.F., 1992. The Local Global Continuum. In: Abler, R.F., Marcus,M.G., Olsen, J.M. (Eds.), Geography’s Inner Worlds. RutgersUniversity Press, New Jersey.

Page 17: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 151

different methods of aggregation on the process vari-ables used to identify dominant strategies, and findsthat different forms of aggregation reveal differentvillage-level dominant strategies.10 The dominantstrategies identified in this paper may be a product ofthe ecological fallacy (Fig. 10(d)).

In fact, there are many potential scenarios wherebyprocess–pattern linkages may fail to transcend lev-els of analysis (Fig. 10). Two distinct linkagesmediate relationships between driving forces, man-agement strategies and land-cover outcomes in theinduced-intensification thesis’ model of agriculturalchange, and each set of variables describing thesecomponents are subject to independent, unparallelmodifications through the aggregation process.

7. Conclusions

This case study establishes that a process-led ap-proach, structured within the induced-intensificationthesis, can be used to identify process–pattern relation-ships. Farmer management strategies, as defined bythe thesis, produce characteristic land-cover outcomesin most cases, and these outcomes are consistent overtime. This farmer-level process–pattern linkage pro-vides the basis for developing a village-level model ofland change. The model, which aggregates the cumu-lative cover consequences of farmers following manystrategies in the same landscape, predicts village-levellandscape outcomes and identifies characteristic land-scapes in villages supporting different dominantstrategies.

Results reveal, however, that village-level land-cover outcomes are not necessarily indicative of thedominantstrategy. A breakdown in the village-levelprocess–pattern linkage occurs because of thewell-known methodological problem of the ecologi-cal fallacy—aggregate variables reveal relationships

10 The method used to identify the dominant processes reportedin this paper (Fig. 8) identifies the dominant village-level strat-egy by interpreting each LMU’s production variables (e.g. crop-ping intensity), classifying each LMU to a strategy, and thenidentifying the strategy followed by most LMUs in the vil-lage (an LMU-aggregate). A second method, also common ininduced-intensification studies, aggregates the production variablesto the village-level first, and then classifies the village as a wholeto a single management strategy (a village-aggregate).

at broader levels to do not necessarily parallel re-lationships found at the individual level. This studyconsiders different types of aggregation effects, andclarifies why this case study is particularly suscep-tible to the adverse implications of the ecologicalfallacy. In the Andapa region, the dominant man-agement strategy does not produce an underlyingbroad-level, large-grain signal for the aggregationprocess to recover. Instead, farmers following domi-nant and non-dominant strategies all produce signalsat the same low level and fine grain. Aggregation,in this case, is particularly sensitive to the con-founding signals of the non-dominant strategies.Aggregate variables reveal the land-cover patternsof strategies that were not, in fact, chosen by mostfarmers. This study’s bottom up modeling tech-nique, which traces process–pattern linkages throughthe aggregation process, exposes these problems.By monitoring this scalar dynamic very closely,the study avoids identifying false process–patternrelationships.

Previous research has shown that broad-levelLUCC studies may identify different process–patternrelationships than low-level LUCC studies (Walshet al., 1999). This research identifies situations inwhich broad-level studies are susceptible to identi-fying false process–pattern relationships. While notoffering a solution to this problem, the study high-lights the importance of carefully monitoring howprocess–pattern linkages transfer across levels ofanalysis. Aggregation needs to be added to the suiteof scale dynamics confounding cause–effect relation-ships in human-dimensions of global environmentalchange research agenda (Gibson et al., 2000).

Acknowledgements

This research was funded by a National Sci-ence Foundation Dissertation Support Grant and aFulbright Scholarship. The author wishes to thankWilliam McConnell for his substantial contributionto this manuscript, as well as B.L. Turner, II, andthe anonymous reviewers who provided valuablecomments on earlier drafts. The author also thanksPascaline Lahady Charlotte for her untiring help inthe field, and the Andapa farmers for sharing theirpersonal information.

Page 18: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

152 R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153

References

Bian, L., Butler, R., 1999. Comparing effects of aggregationmethods on statistical and spatial properties of simulated spatialdata. Photogramm. Eng. Remote Sens. 65, 73–84.

Boserup, E., 1965. The Conditions of Agricultural Growth. Aldine,Chicago.

Brookfield, H., 1984. Intensification Revisited. Pacific Viewpoint25, 15–44.

Brush, S., Turner II, B.L. (Eds.), 1987. Comparative FarmingSystems. Guildford Press, New York.

Chayanov, A.V., 1966. Peasant farm organization. In: Thorner, D.,Kerblay, K., Smith, R.E.F. (Eds.), Proceedings of the Theoryof Peasant Economy. Irwin, Homewood, IL.

Dakins, M.E., Toll, J.E., Small, M.J., Brand, K.P., 1996.Risk-Based Environmental Remediation: Bayesian Monte CarloAnalysis and the Expected Value of Sample Information. RiskAnalysis 16 (1), 67–69.

Decisioneering Inc., 2000. Crystal Ball 4.0. Denver, CO.Gehlk, C.E., Biehl, K., 1934. Certain effects of grouping upon the

size of the correlation coefficient in census tract material. J.Am. Stat. Assoc. 29, 169–170 (Suppl.: Proc. Am. Statist. J.).

Geoghegan, J., Pritchard Jr., L., Ogneva-Himmelberger, Y., RoyChowdhury, R., Sanderson, S., Turner II, B.L., 1998. Socializingthe pixel and pixelizing the social in land-use and land-coverchange. In: Liverman, D., Moran, E.F., Rindfuss, R.R., Stern, P.(Eds.), People and Pixels: Linking Remote Sensing and SocialScience. National Academy Press, Washington, DC.

Getis, A., Boots, B.N., 1978. Models of Spatial Processes: AnApproach to the Study of Point, Line, and Area Patterns.Cambridge University Press, New York.

Gibson, C.C., Ostrom, E., Ahn, T.K., 2000. The concept of scaleand the human dimensions of global change: a survey. Ecol.Econ. 32, 217–239.

Haggett, P., 1964. Regional and local components in thedistribution of forested areas in southeast Brazil: a multivariateapproach. Geogr. J. 130, 365–380.

Hair, J.F., Anderson, R.E., Tathan, R.L., Black, W.C., 1995.Multivariate Data Analysis, 4th Edition. Prentice Hall, NewJersey.

Irwin, E.G., Geoghegan, J., 2001. Theory, data, methods:developing spatially explicit economic models of land usechange agriculture. Ecosyst. Environ. 85 (1–3), 7–23.

Klecka, W.R., 1980. Discriminant Analysis. Series: QuantitativeApplications in the Social Sciences No. 07-019. Sage, NewburyPark.

Klepeis, P., Turner II, B.L., 2001. Integrated land history andglobal change science: the example of the southern Yucatánpeninsular region project. Land Use Pol. 18, 239–272.

Lambin, E.F., 1997. Modelling and monitoring land-cover changeprocesses in tropical regions. Prog. Phys. Geogr. 32, 375–393.

Lambin, E.F., Baulies, X., Bockstael, N., Fischer, G., Krug, T.,Leemans, R., Moran, E.F., Rindfuss, R.R., Sato, Y., Skole, D.,Turner II, B.L., Vogel, C., 1999. Land-Use and Land-CoverChange: Implementation Strategy. IGBP Secretariat, Stockholm,Sweden.

Lambin, E.F., Rounsevell, M.D.A., Geist, H.J., 2000. Areagricultural land-use models able to predict changes in land-useintensity? Agric. Ecosyst. Environ. 82, 321–331.

Laney, R., 1999. Agricultural change and landscape trans-formations in the Andapa region of Madagascar. Ph.D.dissertation. Clark University, Worcester.

Laney, R., 2002. Disaggregating induced intensification forland-change analysis: a case study from Madagascar. Ann.Assoc. Am. Geogr. 92, 702–726.

Levin, S.A., 1992. The problem of pattern and scale in ecology.Ecology 73, 1943–1967.

McCarty, H.H., Hook, J.C., Knos, D.S., 1956. The measurementof association in industrial geography. University of IowaDepartment of Geography Reports.

Meyer, W.B., Turner II, B.L., 1992. Human population growthand global land-use/cover change. Ann. Rev. Ecol. Syst. 23, 39–61.

Neuvy, G., 1989. Le développement agricole du bassin supérieurde la lokoho, à Madagascar. Les Cahiers d’Outre-Mer 42, 135–154.

Openshaw, S., 1978. An empirical study of some zone-designcriteria. Environ. Plann. A 10, 781–794.

Openshaw, S., 1984. Ecological fallacies and the analysis of arealcensus data. Environ. Plann. A 16, 17–31.

Openshaw, S., Taylor, P.J., 1981. The modifiable areal unitproblem. In: Wrigley, N., Bennett, R.J. (Eds.), QuantitativeGeography: A British View. Routledge & Kegan Paul, London.

Parker, D.C., Berger, T., Manson, S.M. (Eds.), 2002. Agent-BasedModels of Land-Use and Land-Cover Change. LUCC ReportSeries No. 6. IGBP/IHDP-LUCC Project. LUCC Focus I Office,Bloomington, IN.

Portais, M., 1972. Les Cultures Commerciales dans un MilieuGeographique Original: La Cuvette d’Andapa. Centre deTananarive, Office de la Recherche Scientifique et TechniqueOutre-Mer.

Rastetter, E.B., King, A.W., Crosby, B.J., Hornberger, G.M.,O’Neill, R.V., Hobbie, J.E., 1992. Aggregating fine-scaleecological knowledge to model courser-scale attributes ofecosystems. Ecol. Appl. 21, 55–70.

Rediscovering Geography Committee, Board on Earth Sciencesand Resources, Commission on Geosciences, Environment, andResources, National Research Council, 1997. RediscoveringGeography: New Relevance for Science and Society. NationalAcademy Press, Washington, DC.

Reis, E.J., Margulis, S., 1991. Options for slowing Amazonjungle clearing. In: Dornbusch, R., Poterba, J.M. (Eds.), GlobalWarming: Economic Policy Responses. MIT Press, Cambridge,MA, pp. 335–375.

Robinson, W.S., 1950. Ecological correlations and the behavior ofindividuals. Am. Soc. Rev. 15, 351–357.

Serneels, S., Lambin, E.F., 2001. Proximate causes of land-usechange in Narok district, Kenya: a spatial statistical model.Agric. Ecosyst. Environ. 85, 65–81.

Turner, M.G., 1987. Landscape ecology: the effect of pattern onprocess. Ann. Rev. Ecol. Syst. 20, 171–197.

Turner II, B.L., Ali, A.M.S., 1996. Induced intensification:agricultural change in Bangladesh with implications for Malthus

Page 19: A process-led approach to modeling land change in agricultural landscapes: a case study from Madagascar

R.M. Laney / Agriculture, Ecosystems and Environment 101 (2004) 135–153 153

and Boserup. In: Proceedings of the National Academy ofSciences’93, pp. 14984–14991.

Turner II, B.L., Villar, S.C., Foster, D., Geoghegan, J., Keys,E., Klepeis, P., Lawrence, D., Mendoza, P.M., Manson,S., Ogneva-Himmelberger, Y., Plotkin, A.B., Salicrup, D.P.,Roy Chowdhury, R., Savitsky, B., Schneider, L., Schmook,B., Vance, C., 2001. Deforestation in the southern Yucatánpeninsular region: an integrative approach. For. Ecol. Manage.154, 353–370.

Veldkamp, A., Lambin, E.F., 2001. Predicting land-use change.Agric. Ecosyst. Environ. 85, 1–6.

Walker, R., Moran, E.F., Anselin, L., 2000. Deforestation andcattle ranching in the Brazilian Amazon: external capital andhousehold processes. World Dev. 28, 683–699.

Walsh, S.J., Evans, T.P., Welsh, W.F., Entwisle, B., Rindfuss,R.R., 1999. Scale-dependent relationships between populationand environment in northeastern Thailand. Photogramm. Eng.Remote Sens. 65, 95–105.