Modeling land use and land cover change in an Amazonian frontier settlement: strategies for addressing population change and panel attrition

Abstract

Research on tropical deforestation has been prolific, yet few studies have assessed the long-term dynamics of frontier migration and the resulting impacts on deforestation. These lacunae arise from the difficulty of obtaining the panel data required to evaluate the dynamic socioeconomic and land use processes of the advancing and aging frontier. Furthermore, the quality and design of household surveys reported in the land use literature are often not transparent, limiting possibilities for comparing results. This article first describes a three-round spatial panel survey of households in a settled and heavily deforested Amazon frontier region. We detail several methods that are employed to ensure and assess data quality. Second, we estimate forest clearing at the agent (household) level, using several sets of explanatory variables and sub-samples that would be generated by applying different field methodologies. We find the definition of the panel agent and the sampling frame to influence our estimations.

Keywords:

• Brazilian Amazon
• land use change
• farm households
• remote sensing
• household survey
• panel attrition

1. Introduction

Tropical deforestation is a striking form of land cover transformation attracting the attention of researchers across multiple disciplines who seek to describe, explain, and predict the progression of the deforestation frontier by examining agents, proximate drivers, and underlying causes. Brazil contains the largest area of dense tropical forest in the world and despite numerous policy initiatives to slow deforestation, forest removal continues at the rate of 18,000 km² per year (INPE 2007). Thus, the country is said to house the most active land use frontier in the world (Morton et al. 2006). The Brazilian Amazon comprises almost 70% of the tropical forests within South America and nearly 60% of the territory within the nation, yet is home to only 11% of the country's population (Kirby et al. 2006). This population is concentrated in urban centers and along the deforestation frontier, including most of the 218 municipalities that are now classified as ‘deforested’, with an average level of deforestation equaling 65% (Celentano and Veríssimo 2007; based on study of 408 municipalities with original land cover more than 50% forest). The low population density and governmental history of using land settlement to address socioeconomic issues leaves the vast remaining area of forest at risk of conversion in the future.

Research examining the underlying causes of tropical forest conversion has been prolific, yet few studies have assessed the long-term dynamics of frontier migration and the resulting household impacts on deforestation. These lacunae are directly related to the difficulty of obtaining the panel data required to track and evaluate the underlying dynamic processes related to forest cover change. Moreover, the maintenance of quality panel data is complicated by population drift and attrition. Given these difficulties, researchers typically work with macro-scale data (e.g., municipality or county level) and extend these results to infer individual decision making (Miller and Plantinga 1999; Caldas et al. 2007). Although micro-level studies exist, most rely on cross-sectional surveys, which often include retrospective questions to infer dynamic decision making (Pedlowski and Dale 1992; Godoy et al. 1997; Pichon 1997a,b; Faminow 1998; Shively 2001; Coxhead, Shively, and Shuai 2002; Walker, Perz, Caldas, and Silva 2002; Browder, Pedlowski, and Summers 2004). Notable exceptions include the work conducted at the Anthropological Center for Training and Research on Global Environmental Change at Indiana University (VanWey, D'Antona, and Brondízio 2007; Sirén and Brondizio 2009).

Studies at the agent level suggest that wealth accumulation is closely linked to higher levels of deforestation (Jones, Dale, Beauchamp, Pedlowski, and O'Neill 1995; de Almeida and Campari 1995; Caldas et al. 2007; Zwane 2007). In addition, other studies indicate the biophysical conditions of the property to be important drivers. For example, large land holdings are shown to permit the retention of forest tracts and provide enough land to allow for substantial fallow periods (D'Antona, VanWey, and Hayashi 2006). On the other hand, the proximate causes often identified at the macro-level continue to be road creation and improvement (i.e., paving and other public support for colonization projects that continue to be approved by the government) (Pfaff et al. 2007; also see Kirby et al. (2006 for a review). Finally, the household life cycle posits that the demographic composition of the household shape land use and land use change (Walker 2004, 2008; VanWey et al. 2007).

Given the importance of population dynamics at both the micro- and the macro-level, we argue that it is critical to track both changes in land use and the agents of those changes (family farmers in the case of our study area) over time. We describe various aspects of our survey methodology that allowed this tracking along with the construction of a survey and geospatial database with multiple temporal and spatial linkages. We show that descriptive statistics and models of land use and land cover (LULC) will vary with different tracking methods because of the resulting differences in samples, supporting the common call for greater transparency in methodology (Parker et al. 2008). We describe our survey methodology in detail and assess the convergent validity of the survey, remote sensing, and census data. Our experiences offer lessons for other researchers collecting data, as well as providing a foundation for other researchers to use the resulting publicly available data to advance land use science. ¹

2. LULC change in the Amazon and the need for high-quality panel data

Research on the proximate and underlying causes of deforestation has been undertaken at the national, state, municipal, and household scales (see Barbier and Burgess 1997; Wibowo and Byron 1999; Geist and Lambin 2001 for reviews). Many early studies attributed deforestation in the tropics, and particularly the Brazilian Amazon, to single factors such as population growth, migration, government policy, or road construction; however, it has since been recognized that a complex system of variables, contributing at different spatial and temporal levels, is at work (Bilsborrow 2002; Wood and Porro 2002; Sills and Pattanayak 2006; Rindfuss et al. 2008). Household level panel data play a key role in identifying the impacts of socio-demographic factors, government policy, and economic shocks that are vital to understanding landscape change (Sunderlin, Angelsen, Resosudarmo, Dermawan, and Rianto 2001; Andersen 2002; Pan and Bilsborrow 2005; Pan, Carr, Barbieri, Bilsborrow, and Suchindran 2007). In particular, agent-based models rely on such micro-level data to formalize representations of behavior. This ‘bottom up’ approach uses micro-level data to explain macro-level phenomena. Data used in these models are generally derived from surveys, participant observations, field and laboratory experiments, companion modeling, and GIS and remotely sensed spatial data. With the exception of the latter, the cross-sectional nature of many of these data sources has been a major drawback to modeling alignment and success (Robinson et al. 2007).

Whether data are used to inform simulation models or to analyze governmental policy impacts, the combined use of survey, Geographic Information Systems (GIS) and remote sensing data can make significant contributions. The recent surge of studies that combine satellite and survey data can be attributed to the greater availability of remote sensing data (and other spatial data) along with the increased understanding that can be gained from such analysis. Recent approaches include the use of GIS to improve distance measurements (Staal, Baltenweck, Waithaka, de Wolff, and Njoroge 2002), the use of satellite remote sensing data to improve survey sampling (Binford, Lee, and Townsend 2004), the use of multi-level models with data at the household village and pixel levels (Pan and Bilsborrow 2005; Vance and Iovanna 2005), and the analysis of land cover change (McCracken et al. 1999; Pfaff 1999; Geoghegan et al. 2001; McCracken, Siqueira, Moran, and Brondizio 2002; Rudel, Bates, and Machinguiashi 2002; Staal et al. 2002; Munroe, Southworth, and Tucker 2004); also see Nelson and Geoghegan (2002) for a review.

Robinson et al. (2007) identify key areas in land use science that continue to require attention including the identification of the agents of land use change, the analysis of agent behavior, as well as temporal aspects of LULC change. Although the call for high-quality panel data for developing regions has been made by the land-use community with respect to the issues above (Parker et al. 2008), this community of researchers is not alone. From meta-analysis (Doss 2006) to program evaluation (Anderson and Feder 2007), researchers across the spectrum would greatly benefit from panel data that encompass a wide range of sources that are clearly defined both qualitatively and quantitatively.

3. Panel attrition and conditioning

Panel data are essential to understanding dynamic decision making, yet the attrition that occurs in panel surveys can reduce the benefits of use (Lillard and Panis 1998; Ziliak and Kniesner 1998; Glewwe and Jacoby 2000). The most significant issue affecting panel surveys is non-random attrition that is systematically related to the outcome of interest in a way that causes bias in estimation (Fitzgerald, Gottschalk, and Moffitt 1998; Olsen 2005; Burton, Laurie, and Lynn 2006). In addition to introducing bias, attrition reduces analytical power by decreasing the number of observations (Ferland, Tremblay, and Simard 2007). Because attrition bias has been found to be model specific, and because it is difficult to predict which variables will impact attrition and in which ways, methods to limit attrition remain important (Hawkes and Plewis 2006).

There are several pre- and post-survey approaches that can be used to reduce attrition and/or attrition bias. Pre-survey methods include attention to survey design, expanding the sample size, tracking individuals, and collecting independent comparison data, whereas post-survey methods include the use of models to identify and correct attrition bias. Panel attrition and participant migration are often linked in developing countries, where the major reason for non-response is the high degree of mobility of the population rather than refusal to answer the questionnaire (Thomas, Frankenberg, and Smith 2001). In these cases, tracking has been found to reduce attrition by up to 45% (Hill 2004). Studies also suggest that the quality of the interviewers plays a key role in reducing attrition that is attributable to refusal (Hawkes and Plewis 2006). Similarly, Olsen (2005) argues that continued participation in a panel depends heavily on whether the participants feel that the study is important, and that the interviewers play a large role in ‘selling’ the survey through their enthusiasm and level of experience. The pre-survey approaches undertaken in our study include a participant registry to aid in tracking individuals who moved, gifting respondents with calendars including a map of the region and results from previous rounds, and the use of Global Positioning Systems (GPS) to locate household lots.

Regardless of the measures taken to reduce attrition, some panel conditioning is inevitable in panels of considerable length. Given the dramatic changes occurring in the survey region with the creation of new settlements and continuing in-migration, a panel that retains the same observational unit is expected to become less representative of current land use patterns and population dynamics, a problem that is most likely exacerbated by attrition. In an effort to examine and reduce the extent of panel conditioning in our project, we expanded the size of our sample to collect comparative data and updated our survey questions to obtain information on new trends and activities. The wider sampling (i.e., increase in sample size within the original survey region) helps to address the problem of conditioning by providing a better overall view of welfare and land use. We use this expanded sample in conjunction with external census data to assess the representativeness of our panel.

Post-survey data can be tested and corrected for attrition with independent data, observations from the expanded survey, and information on households who attrit using instrumental variables, probability, and hurdle models. Methods for detecting the presence of nonrandom bias include estimating attrition with variables measured in the previous wave (Hawkes and Plewis 2006) and using attrition indicator variables to form interaction terms that are included as explanatory variables (Maluccio 2004). The most common correction models for significant attrition bias include weighted least squares for selection on observables (Fitzgerald et al. 1998); the use of sample weights as determined by the inverted response rate of the group (Ferland et al. 2007), ² and the Heckman selection approach for selection on unobservables (Maluccio 2004). We employ the latter in our estimations of land use to both test and correct for attrition using different survey samples applicable to our varied sampling approaches.

4. The study region

The Ouro Preto do Oeste (OPO) region (comprising six municipalities) is located in central Rondônia, an Amazonian state in southwestern Brazil near the border with Bolivia (Figure 1). This region is ideal for the analysis of LULC change on a typical ‘old frontier’, as it is representative of the ‘arc of deforestation’ across the southern Brazilian Amazon (Lele et al. 2000; Alves 2002a), is a priority area for monitoring and managing development pressures (Ministério do Meio Ambiente 2001), and encompasses a number of government-sponsored settlements established at different times. In addition, the state of Rondônia is the Brazilian state that has experienced the most extensive and rapid land transformation (from forest to farmland) within the last 20 years (Alves 2002b). Furthermore, given large average lot sizes (approximately 71 ha) survey and Landsat data can be matched at the household level (See Brondizio, Moran, Mausel, and Wu 1996; Brondizio et al. 2002 for such methods).

Figure 1. Study area.

The climate of OPO is classified as humid tropical, or Awi in the Köppen classification system, and experiences a distinct dry season in the months of July and August (RADAMBRAZIL 1978). Temperatures in the region average 24°C with precipitation totals near 2300 mm, resulting in both dense and open tropical forests (INPE 2000). However, most of the land in the survey region has been converted to pasture with small patches of perennial (e.g., coffee and cacao) and annual crops (e.g., corn and rice) as well as small tracts of forest (Pedlowski 1997). Topography is a mix of rolling hills and flat valleys surrounding several steep and rocky inselbergs (Numata et al. 2003). Soils vary throughout the region based on underlying geology, slope, and climate, but are dominated by Podzólico Vermelho Amarelo and Podzólico Vermelho Escuro, roughly equivalent to oxisols and ultisols in the US soil classification system.

The state of Rondônia experienced significant in-migration with the construction of two federally funded highways in the early 1960s (Paraguassu-Chaves 2001), representing a guided effort by the Brazilian government to demonstrate control over a greater area within the Amazon. From 1964 to 2005, the national land reform agency settled 84,434 families in the state (Imazon 2007) including many in new settlements recognized and regularized by INCRA (Instituto Nacional de Colonização e Reforma Agrária – National Institute for Colonization and Agrarian Reform) over the past 10 years (Sparovek 2003). Deforestation increased within the state of Rondônia from approximately 2% in 1977, to 20% in 1996, to over 60% by 2005 (Alves 2002b; INPE 2007). Moreover, researchers have found approximately 80% of the state's deforestation to occur within 12.5 km of the major highway, BR-364 (Alves 2002b), running from the southwest through the study region to the northern capital, Porto Velho.

OPO was the first Integrated Colonization Project to be launched in Rondônia in 1971, with an initial goal of settling 500 families; however by 1974 approximately 4000 lots had been distributed to immigrant families (Martine 1980; Pedlowski 1997; Oliveira 2002). This rapid influx of migrants has been attributed to a combination of factors including the abandonment of colonization schemes along the Transamazon Highway and the relatively fertile soils of central Rondônia (Leite and Furley 1981; Coy 1987; Martine 1990; Browder 2002).

5. Survey methodology for panel data quality

In addition to meticulous oversight of data entry (e.g., double entry by oral verification), our survey methodology employs several strategies to assure and evaluate data quality. To address attrition and panel conditioning in the third round of the survey in 2005, we implemented a pre-survey ‘registry’ of households in the panel, expanded the sample size, tracked households and individuals who moved, and updated survey variables. Furthermore, we assess the convergent validity of household responses with GIS data collected at the same scale (Caviglia-Harris and Harris 2005; Cohen 2005) and compare average survey values to average values from the agricultural and population census at the municipality level.

5.1. Expanded sampling and household tracking

The full set of survey data consists of three rounds of data collected in 1996, 2000, and 2005. In the 1996 survey round, data were collected from a stratified random sample of households that defined municipality as the strata with a random draw of rural lots based on a fixed proportion of the rural properties. By selecting a random starting point and interviewing households on lots at intervals required to obtain the desired sample size in each strata, the sampling process ensures variation in topography, soil type, distance to markets, and distance to the central city (Casley and Kumar 1988). Additionally, we interviewed a convenience (or intercept) sample of households involved in the Association of Alternative Producers (APA) – a local non-governmental organization that promotes sustainable agricultural and forestry practices – to investigate the adoption diffusion of these practices. For completeness, these association members are included in our report of sample size; however, these observations are not included in any further analyses, as the convenience method of identification would bias the sample. The 1996 survey round yielded 196 household interviews on 196 lots: 171 in the stratified random sample and 25 in the intercept sample of APA members. Revisiting the same 196 lots, we obtained 193 interviews in the 2000 survey year, losing 1 lot each from the stratified random sample and the convenience sample (Table 1).

Table 1. Household surveys collected by survey year

	Year
	1996	2000	2005
Stratified random sample	171	170	361
Original sample			177
Tracked households			67
Additions to stratified sample			117
Treatment group – APA sample collected by convenience sampling	25	24	38
Original sample			17
Tracked households			9
Additions to convenience sample			6
Total	196	193	399

In the 2005 survey year we expanded the target sample size. Understanding that the correct sample size is not a percentage of the population but rather a function of the variability of the characteristic measured and the degree of precision required, we followed the framework outlined in Casley and Lury (1982) to estimate the minimal sample required for key variables of interest. ³ Depending on the variable used to proxy land cover or welfare, we estimate the minimum required sample size to be between 2 and almost 9000 households (Table 2). However, nearly 60% of our key variables can be adequately represented with 202 or fewer observations. Given that we did not have the resources to interview thousands of households, we used this more conservative estimate of the required sample size, increasing the target size of the control sample from 171 to a minimum of 200 lots. In doing so, we also expanded the control sample to include lots from new settlements established since 1996, increasing the sample to 3–7% of the municipality population (Table 3). We further adjusted the sample to obtain information on migration, by tracking individuals and entire households that moved from the original surveyed lots.

Table 2. Estimation of minimal required sample size for different land use and welfare indicators

Variable	Definition	Mean	SD	Estimated sample size
1996
Cattle	Number of cattle owned on the lot	71.44	83.73	7010
Vehicles	Number of vehicles (including motorcycles, cars, trucks, and tractors) owned by the household	0.43	0.67	0
Lot size	Lot size, in hectares, as reported by household	70.64	46.12	2127
Agriculture	Number of hectares in agriculture reported by household	7.42	6.54	43
Pasture	Number of hectares in pasture reported by household	46.45	37.69	1420
Forest	Number of hectares in primary forest reported by household	16.67	18.88	356
Agroforestry	Number of hectares in agroforestry reported by household	0.11	0.46	0
2000
Cattle	Number of cattle owned on the lot	96.64	94.59	8946
Vehicles	Number of vehicles (including motorcycles, cars, trucks, and tractors) owned by the household	0.80	0.91	1
Lot size	Lot size, in hectares, as reported by household	63.49	34.97	1223
Agriculture	Number of hectares in agriculture reported by household	6.16	7.06	50
Pasture	Number of hectares in pasture reported by household	45.27	31.36	984
Forest	Number of hectares in primary forest reported by household	11.86	14.20	202
Agroforestry	Number of hectares in agroforestry reported by household	0.19	1.33	2

Table 3. Percentage of households interviewed by municipality in each survey year

	1996			2000			2005
	Number of lots	Number interviewed	Percent interviewed	Number of lots	Number interviewed	Percent interviewed	Number of lots	Number interviewed	Percent interviewed
Ouro Preto do Oeste	2007	45	2.24	2087	44	2.11	2085	80	3.84
Vale do Paraíso	935	20	2.14	937	20	2.13	937	36	3.84
Nova União	592	28	4.73	1217	28	2.30	1212	79	6.52
Teixeirópolis	387	11	2.84	387	11	2.84	387	16	4.13
Urupá	1561	35	2.24	1782	35	1.96	2717	75	2.76
Mirante da Serra	1140	32	2.81	1141	32	2.80	1376	72	5.23
Total	6622	171	2.58	7551	170	2.25	8714	361	4.14

In total, we increased the sample for the 2005 survey round to 399, including 177 lots from the original stratified random sample (increasing because of lot subdivisions), 67 lots corresponding to individuals (with information for their complete households) that moved from the original stratified random sample and were tracked to their current locations, 38 lots corresponding to APA members, and 117 lots selected by the original stratified random sampling methodology. Of these 117 lots in the expanded random sample, 60% were drawn from new settlements established by the land reform agency INCRA since 1996. The remainder were drawn from within the original settlements (Table 1).

5.2. Updated survey variables

The data collection efforts in 1996 and 2000 provided (i) information on outputs and inputs for farm production; (ii) hectares reported in different land uses, including forest, pasture, and crops; (iii) measures of wealth, including consumer durables, farm equipment, livestock, and self-reported value of parcels; and (iv) a standard set of socioeconomic characteristics, including some ‘pre-sample’ characteristics such as state of birth, number of years in Rondônia, and how the lot was acquired. At least one member of each household was interviewed to collect socioeconomic information on all members residing on the lot.

In 2005 we maintained the same core set of questions as in the earlier rounds. In addition, a pre-survey or ‘registry’ focused on confirming residents on the lot and open-ended questions to elicit information on important changes and current trends in the region. Based on insights from the registry, interviews with key informants, and analysis of prior rounds, the survey instrument was expanded to include (i) updated measures of wealth to reflect new trends, (ii) input and output quantities for any new farm activities, (iii) expanded measures of human capital, and (iv) indicators of current and past shocks that are not correlated across the entire region. Furthermore, we expanded spatial data collection by geo-referencing the lots and relevant regional infrastructure. This improved upon both the accuracy and the precision of the spatial data by mapping the road network and identifying global positioning system (GPS) points for the individual lots, urban centers, agricultural markets, and dairy-processing plants. The collected geospatial data include classified Landsat Thematic Mapper (TM) satellite images, shuttle-derived digital elevation models, lot boundaries digitized from Brazilian settlement maps and GPS data collected to identify surveyed lots, markets, and infrastructure. All lots surveyed in any of the 1996, 2000, or 2005 rounds were located in a GIS to match them with these geospatial data, including current and historical land cover as established by remote sensing, biophysical indicators (such as slope and soil type) from secondary sources, and distance measurements (to market, city center, etc.) based on road networks mapped with GPS.

5.3. The survey registry

Before the survey in 2005, we conducted a ‘registry’ of lots in the panel, with the goals of determining how many households had moved, identifying and tracking both households and individuals who had moved off lots, assisting survey teams in locating lots, and scoping out important trends and changes that should be addressed in the survey instrument. Key objectives were to enhance the efficiency of fieldwork and reduce attrition. A trained interviewer with extensive experience in the region visited all lots in the original random stratified sample and identified the individuals and households that had moved since the 2000 survey. Calendars with the survey weeks were highlighted and maps of the region were distributed to help build a sense of reciprocity and commitment to participate in the survey. In cases in which household members or entire households had moved, the interviewer elicited information on why they moved, where they moved, their current occupation(s), and their contact information. The registry information helped us establish a feasible sampling plan for households and individuals who had left their lots, reduced the amount of information that had to be collected during the survey process, and helped interviewers locate households in the sample when we returned three months later to conduct the survey. ⁴ In addition, we were able to identify and untangle the relatively few complicated situations involving sub-divided lots, family feuds, and multiple moves to and from lots. The registry data (including observations and directions to the lots) and photos taken of each family during the registry were embedded in the questionnaires to assist with lot identification and the interview. The interviewers presented copies of the photos as gifts to the family.

5.4. Interview efficiency

To evaluate the impact of our survey methodology on interviewer efficiency, costs per interview and number of surveys completed per day are investigated. Our survey field budget can be broken down into three categories: (i) registry expenses (15%); (ii) household interview expenses (35%); and (iii) fixed costs (50%), including equipment, housing, travel, and consulting fees. The ‘fixed costs’ are independent of the number of surveys completed. The costs per completed interview (net of these fixed costs) are primarily a function of the time required to locate and arrive at a household and the time required to obtain consent for and conduct the interview. We calculate these costs based on number of interviews completed per day and two components of the survey budget: (1) car rental and fuel and (2) enumerator pay. ⁵ The cost per interview ranges from $19 to $762. The average cost per household in the tracked sample (households and individuals who moved to new locations) was $61. The average cost per household in the new random sample (from original and new settlements) was $55. The average cost of interviewing the original sample was only slightly less, at $48 per household. We believe that the registry helped contain the cost of both the tracked and the original samples. In comparison, the cost of registry data is estimated to be approximately $55 per household visited. These costs include car and fuel costs, enumerator pay, in addition to training costs. Dividing these costs over the total number of surveys as part of the original and tracked sample (because the registry was applied to both these samples, but not the ‘new’ sample) adds an additional $24 per survey. In other words, the ‘new’ sample – that with the least restrictions on the household or lot to be interviewed ⁶ – was the most cost effective at $55 per survey, compared with a total ⁷ cost of $85 per survey for the tracked sample and $72 per survey for the original sample.

We expanded our study area from 6000 km² to approximately 20,000 km² (including neighboring municipalities where households had moved) in order to track households. Nearly three-quarters of individuals who had left lots remained within this expanded study area (72% according to the registry). The registry information allowed us to assign original and tracked households to survey teams in an efficient manner, so that they could interview tracked households interspersed with the original households when they first visited an area. This field plan – and specifically the limit on distance traveled by our interviewers – did clearly affect the sample by excluding more than one-quarter of individuals who moved further away. Thus, our relatively low tracking costs stem from three factors: (1) a stable population (i.e., the majority of moves were within the survey region), (2) the use of a participant registry, and (3) pre-defined limits on how far we would track a household. Our survey costs for the tracked households would likely be considerably higher if we followed the remaining 28% that migrated to more distant locations.

To gain further insight on efficiency, we examine patterns in the number of interviews completed per day as the survey proceeded and by distance from our headquarters (the city of OPO). As expected, we find that interviews per day increased over the first couple of weeks of the survey (Figure 2) but declined toward the end of the survey in a statistically significant manner (as confirmed with a regression below). We attribute the initial increasing efficiency to a learning effect and the later decline to the increased difficulty of locating households, because of a combination of the spatial organization of the settlements, the sampling methodology, and panel tracking. The original random sampling by lots minimizes clustering of households in the sample. This meant that interview teams would proceed down roads, stopping periodically at households in the sample and thereby interviewing a relatively large number of households even though the distance between properties was substantial. However, when households could not be interviewed during the first visit, they had to be revisited at a later date, thus increasing the distance between the target households later in the survey time frame and reducing the number that could be completed in a day. A second notable – and unexpected – trend was an increase in the number of interviews completed per day with greater average distances from headquarters (Figure 3). ⁸ This may be because several of our survey teams remained in the field overnight when they were conducting interviews far from headquarters, increasing the number of hours they devoted to interviewing per day.

Figure 2. Number of interviews completed by survey day.

Figure 3. Number of surveys completed by distance from headquarters.

Finally, we estimate the influence of these combined factors on efficiency with an ordinary least squares regression and find the following:

1

where ‘Number’ refers to the number of questionnaires completed in a day; ‘day’ and ‘day²’ refer to the survey day and day squared, respectively; ‘ratio_ns’ is the ratio of interviews completed from the new sample relative to the total number completed that day; ‘ratio_ts’ is the ratio of interviews completed from the tracked sample relative to the total number completed that day; and ‘distance’ is the average distance from headquarters for the interviews completed that day. These estimation results confirm the nonlinear effect of survey day and the negative impact of completing surveys from the tracked and new samples. Distance from headquarters is not a statistically significant determinant after controlling for these other factors. This summary regression model confirms that the addition of new settlements and the tracking of households did reduce interviewer efficiency and ultimately resulted in higher costs per completed interview than would have been the case if we had simply maintained a time-series cross section. On the other hand, the incremental cost of interviewing a tracked household was only 18% higher than the cost of interviewing a household that remained on its original lot.

6. Descriptive statistics

Descriptive statistics for the households from the original stratified random sample (not including household members and households that were tracked to new locations) are reported for each of the survey years in Table 4. Overall, there has been little change in the demographic characteristics of households with the exception of significant improvements in education and a reduction in household size. The average age of the household head did not change significantly, remaining approximately 49 years, whereas the average years of schooling went up from 2.5 to 2.89, reflecting turnover to newer generations of household heads who have more education. On the other hand, there have been notable changes in assets, income, and land use over the time period. Cattle ownership increased over 70% between the survey years from 72 to 125 head of cattle per household lot. According to these data, there were even greater increases in income and vehicle ownership, increasing 90 and 118%, respectively. Over the same time period forest cover diminished by 50% (Table 4).

Table 4. Descriptive statistics for random sample of household lots in the original survey region

		1996		2000		2005
Variable	Definition	Mean	SD	Mean	SD	Mean	SD
Household characteristics
Family	Number of household members living on the lot	8.42	6.02	7.40	5.76	5.76	3.68
Age household	Average age of the household heads, (years)	46.36	12.94	49.08	12.45	51.25	13.03
Edu household	Average education level of the household heads, (years)	2.50	2.47	2.49	1.62	2.89	2.12
Year acquire	Year the household head or family acquired the lot	1985	6.70	1986	7.63	1987	9.53
South	= 1 if the origin of the household head is one of the states in the south or southeast region of Brazil; 0 otherwise	0.82	0.38	0.85	0.36	0.79	0.41
Unions	Number of unions, farming associations or co-ops household members participate in	0.71	0.82	1.02	0.85	1.21	0.99
Inputs for agriculture
Chemical inputs	Total cost of fertilizer, herbicides, and pesticides used by the household within the survey year, 2000R$	140.82	192.11	575.78	1963.76	485.61	1992.98
Labor payments	Labor paid for by household, 2000R$	172.84	818.56	103.92	231.28	670.22	2015.19
Harvest diversification	Number of annual and perennial crops, nuts, and fruits harvested on the lot	5.52	3.00	6.01	3.60	8.24	5.55
Income and assets
Cattle	Number of cattle owned on the lot	71.44	83.73	96.64	94.59	125.04	127.15
Total income	Income from annuals, perennials, milk, off farm labor, honey and fish, beef and calves, 2000R$	7957.34	9418.57	15,689.00	16,394.30	17,560.50	24,182.20
Vehicle value	Value of all vehicles owned by household, 2000R$	5152.16	13,420.10	5298.91	8892.69	8548.37	15,889.30
Vehicles	Number of vehicles (including motorcycles, cars, trucks and tractors) owned by the household	0.55	0.92	0.86	1.04	1.31	1.17
Lot characteristics
Lot size	Lot size, in hectares as reported by household	70.64	46.12	63.49	34.97	73.76	66.39
Forest	Number of hectares in primary forest as reported by household	16.67	18.88	11.86	14.20	9.67	17.51
Non-forest	Number of hectares in agroforestry, pasture, and agriculture, as reported by household	53.97	37.76	51.62	31.17	63.67	57.66
Distance city center	Distance to the city center (of Ouro Preto do Oeste), kilometers	37.73	19.35	37.65	19,22	37.88	19.68
Soil	Dominant soil type on lot, characterized by initial ability to support agriculture (1 – good, 2 – moderate, 3 – restricted, 4 – unsuitable)	2.27	0.74	2.28	0.74	2.29	0.75
Average slope	Average slope gradient on the lot	5.48	3.36	5.47	3.37	5.52	3.34
n		171		170		173

One point of interest is that these changes in welfare and forest cover are not independent of survey sample. Table 5 presents descriptive statistics for three sub-samples of survey respondents in 2005: (i) the original household sample: households from the original stratified random sample, not including household members and households that were tracked to new locations; (ii) the expanded household sample: all households in (i) plus the new random sample added in 2005 from both the original and new settlements; and (iii) the expanded and tracked household sample: all households in (ii) plus household members and households that were tracked to new locations. A previous study (Sills et al. 2007) suggests that migrants to the new settlements are relatively young, more educated, and less wealthy (own fewer assets) compared with established residents. Similar differences between migrant and established residents are evident in our data and reported in Table 5. Households living on our original surveyed lots are older and less educated but own significantly more cattle, more vehicles, and have higher levels of income compared with our complete sample of households that includes the expanded and tracked samples.

Table 5. Descriptive statistics for different household samples collected in 2005

	Original sample		Expanded sample		Expanded and tracked
	Households in the original random sample N = 173		Households in the original random sample plus those added randomly from original and new settlements N = 312		Households in the expanded sample plus those household members and entire households that were tracked N = 372
Variable	Mean	SD	Mean	SD	Mean	SD
Household characteristics
Family	5.76	3.68	5.46	3.38	5.23	3.35
Age household	51.25	13.03	47.94	13.97	46.56	14.61
Education household	2.89	2.12	3.12	2.16	3.17	2.18
Year acquired	1987	9.53	1990	10.22	1991	10.34
South	0.79	0.41	0.75	0.44	0.74	0.44
Unions	1.21	0.99	1.23	1.03	1.14	1.01
Agriculture inputs and harvest
Chemical inputs	485.61	1992.98	365.13	1610.59	321.72	1485.37
Labor payments	670.22	2015.19	596.85	1643.32	525.51	1527.87
Harvest diversification	8.24	5.55	8.12	5.17	7.23	5.39
Income and assets
Cattle	125.04	127.15	104.32	109.92	93.17	107.28
Total income	17,560.50	24,182.20	15,057.10	19,856.40	14,463.00	22,545.60
Vehicle value	8548.37	15,889.30	7375.68	16,886.60	6682.92	15,702.30
Vehicles	1.31	1.17	1.18	1.17	1.11	1.15
Lot characteristics
Lot size	73.76	66.39	62.70	62.09	55.89	60.54
Forest	9.67	17.51	7.95	15.47	7.09	14.69
Non-forest	63.67	57.66	54.23	53.82	48.32	52.34
Distance city center	37.88	19.68	39.66	19.04	39.63	18.76
Soil	2.29	0.75	2.33	0.74	2.34	0.74
Average slope	5.52	3.34	5.51	3.31	5.47	3.25

Table 6 summarizes the average deforestation per year on sample lots in settlements that were established in different years. These data are also summarized in Figure 4. These statistics suggest that households deforest most rapidly (in terms of hectares per year) in the first 5 years of occupation of a new (forested) lot, averaging around 6 ha per year. The deforestation rate falls to approximately 2 ha per year after 20 years of occupation. Households require large expanses of deforested land in order to grow crops and raise cattle and therefore clear more forest per year when they initially occupy forested lots, regardless of whether those lots are relatively small (i.e., 25 ha) or large (i.e., 100 ha).

Table 6. Deforestation (non-forest) estimations for municipalities in Ouro Preto do Oeste, Rondônia

		1996			2000			2005
	Year ofsettlement	Deforestation peryear (ha)	Lot deforested peryear (%)	Obs.	Deforestation peryear (ha)	Lot deforested peryear (%)	Obs.	Deforestation peryear (ha)	Lot deforestedper year (%)	Obs.
Municipality
Mirante da Serra	1992	6.90	0.16	24	3.25	0.08	25	3.04	0.05	45
Nova União	1970	2.30	0.03	36	1.89	0.03	35	1.81	0.02	48
Ouro Preto do Oeste	1970	3.34	0.03	15	2.61	0.03	15	2.16	0.03	37
Teixeirópolis	1970	2.82	0.03	25	2.31	0.03	25	2.05	0.02	28
Urupá	1981	1.79	0.05	36	1.23	0.04	36	0.89	0.03	47
Vale do Paraíso	1970	2.52	0.03	35	2.32	0.03	35	2.11	0.02	46
Settlement
Palmares	1997							2.61	0.11	24
Martim Pescador	2002							6.08	0.24	13
Padre Ezekiel	2001							3.55	0.23	9

Figure 4. Deforestation levels per year for 1996, 2000 and 2005 (n = 639).

Using our survey data, we are able to investigate land use in finer categories than can be deciphered from Landsat imagery. Although the land cover classification cannot distinguish between pasture and crops, households were able to answer questions concerning the land use with a high degree of confidence. Figure 5 summarizes 2005 land use for the original household sample, the expanded household sample, and the expanded and tracked household sample. The three samples appear to have similar divisions in land use. Approximately 11% of the lot was in primary forest, 1% in agroforestry, 6% in annual and perennial crops, and 83% in pasture or degraded pasture. Moreover, Figure 6 provides similar information for the subgroups of tracked individuals and those who moved to new settlements. Households residing within the new settlements have higher levels of forest, higher annual and perennial crops (i.e., crop area), and lower levels of pasture, most likely reflecting the household life cycle noted by several researchers (Perz 2001; Perz and Walker 2002; Walker et al. 2002; VanWey et al. 2007; Browder et al. 2008). The household life cycle posited by these researchers reflects both demographic changes (e.g., changing dependency ratios) and development of lots (e.g., taking advantage of initial soil fertility to plant crops and later investing surplus in cattle and pasture).

Figure 5. Pie charts of household land use for various 2005 samples.

Figure 6. Pie charts of household land use for 2005 subsamples.

7. Assessing reliability and representativeness of the data

The reliability and representativeness of the survey data are evaluated through comparison to two other sources: (1) remote sensing data on land cover on the same lots and (2) census data from rural areas of same municipalities. For the first, we calculate reliability indices to test the convergent validity of direct reports of land use and satellite-derived estimates of land cover. For the second, we apply variance tests to determine whether our sample could represent a random draw from the population of rural households.

7.1. Cross-referencing survey data with GIS estimates

To estimate reliability indices that cross reference our survey and satellite-derived land cover/use data, we use data on mature forest and non-forest [combining pasture, agricultural crops, and agroforestry for the survey responses and second growth forest, pasture, green pasture, urban/soil, burn (pasture), and rock/savanna for the satellite-derived land cover] from the three survey rounds (1996, 2000, and 2005). As one might expect, the estimates from these two sources do not match exactly for any of the observations. There are several reasons why these independent sources of data could be inconsistent or contain errors. Households may have incentives to over-report forest because of laws requiring 50% of each lot to be preserved, or they may provide inaccurate responses simply because of lack of precise information and/or rounding the number to hectares. In contrast, land cover derived from TM is reported in square meters. Thus, for example, a household that reported 40 ha (400,000 m²) of deforestation might have 399,600 m² of cleared land according to the classified satellite image. On the other hand, the precision of land cover estimated from satellite images is limited by the platform's (TM) spatial resolution of 30 m. Thus, pasture, secondary forest, and primary forest fragments less than 900 m² are not differentiated and are represented by a single land use category. In addition to such misclassification errors, there could be coverage misalignments, differences in property maps and the area that households consider to be part of their lot.

To assess the accuracy of household-reported land use relative to remote sensing classifications, we calculate a reliability index (R) (Marquis, Marquis, and Polich 1986; Bound and Krueger 1991). For example, for non-forest, R is calculated from the households' survey reports of cleared land (S) and the non-forest determined through remote sensing (R), both of which are subject to error. Specifically, the survey reported value of non-forest for household i, S_i , is equal to the true value, T_i , plus an error, e_i . In the case of classical measurement error, it is assumed that e_i is the random response error, uncorrelated to the true and criterion values of the variables, and has an expected value of zero and variance of _:

2

The remote sensing-derived criterion values of deforestation for household i, R_i are assumed to equal the true value, T_i plus an error, v_i .

3

Again, these errors are expected to have a value of zero and variance of .

In this context, R is the ratio of the variance of the ‘true’ values of deforestation and the variance of the values reported in the survey. When no error exists, this value is equal to 1. Because the true values are unknown, it is assumed that (the variance of the true values) in order to estimate as the ratio of the covariance of the criterion (remote sensing) and survey data () and the variance of the values reported in the survey ().

4

Thus, represents the difference in variance between the two sources of data, ranging between 0 and 1. The smaller the difference in the two measurements, the larger is the value of .

In addition to assessing the reliability of our survey responses, the calculation of these reliability indices enabled us to evaluate the remote sensing time series. For example, it was brought to our attention that different procedures were used in the classification as the images were processed in different years, resulting in much lower reliability indices than reported below. For this reason, all images were reclassified under a common algorithm and derived with the same digital masks. More specifically, the remotely sensed land covers for the three survey years were generated using a decision tree classifier applied to standardized remotely sensed variables derived from Landsat 5 data for all of the years from 1983 to 2008 (Roberts et al. 2002). These images were first coregistered and georeferenced to a UTM-projected base map with a SAD69 datum. Next, data were intercalibrated using temporally invariant targets then processed using a spectral mixture modeling approach to generate sub-pixel abundance estimates of green vegetation, non-photosynthetic vegetation (litter, stems, branches), and soil and shade (Roberts, Smith, and Adams 1993). Fraction images were fed into a single decision tree designed to map eight land-cover classes, including pasture, second growth forest, and upland forest. The time series data were used to reduce disallowed transitions, such as pasture reverting to upland forest within a few years. This procedure was also used to replace cloud-contaminated pixels with a land-cover class if the cover type did not change in the years before and after the cloud. As a final step, several digital masks were applied to each scene including an edge mask and rock/savanna mask (Roberts et al. 2002). Although we only report on cover estimated for 1996, 2000, and 2005, the procedure utilized the entire time series (1983–2008) to improve the cover maps.

Results indicate considerable consistency between household responses and the GIS coverage created from remote sensing classifications for all years (Table 7), especially for the deforestation (non-forest) levels. R ranges from 83% in 1996 to 90% in 2005 for the total amount of deforestation on the lot. On the other hand, there is considerably less consistency in estimates of primary forest, with R values ranging from a low of 55% in 2005 to 67% in 2000.

Table 7. Reliability index calculating for GIS derived and survey responses for land use

	1996	2000	2005
Non-forest (Deforestation)	82.85	89.27	89.97
Forest	61.31	66.67	54.56
Note: 70% cutoff suggested to identify problem variables (Marquis et al. 1986; Bound and Krueger 1991).

7.2. Cross-referencing survey data with census data

In our second test of data quality, we compare several socioeconomic characteristics of households elicited in our survey to the average census values for the study region. In the interest of space, we limit the discussion here to the characteristics that best represent the changes noted in the survey years and those that we can match best across sources. The Brazilian population census takes place every 10 years, with an agricultural census occurring at 5- to 10-year intervals. The most recent population census took place in 2000 and the most recent agricultural censuses in 1996 and 2007.

Cattle herd per lot is estimated from the 1996 agricultural census and government-projected values for 2000 and 2005 using the total herd reported per municipality and the total count of rural lots in each municipality (Table 8). Household head's education level and average household size are derived from the 2000 population census data on rural tracts in each municipality. Because variances are not reported by the census, χ² and similar statistical tests are not possible. Instead, we test whether the averages reported by the census fit within the 95%, and the narrower and more precise 80% confidence interval (CI), of our survey data. ¹⁰

Table 8. Estimations of cattle per household lot from census data

	1996			2000			2005
	Number of lots	Cattle	Cattle per lot	Number of lots	Cattle	Cattle per lot	Number of lots	Cattle	Cattle per lot
Mirante da Serra	1140	35,590	31.22	1141	55,466	48.61	1376	103,407	75.15
Nova União	NA	NA	NA	1217	84,159	69.15	1212	127,248	104.99
Ouro Preto do Oeste	2986	31,6175	105.89	2087	259,615	124.40	2085	359,948	172.64
Teixeirópolis	NA	NA	NA	387	60,696	156.84	387	90,519	233.90
Urupá	1561	55,166	35.34	1782	79,722	44.74	2717	153,665	56.56
Vale do Paraíso	935	66,147	70.75	937	95,591	102.02	937	160,569	171.36
Region	6622	473,078	71.44	7551	635,249	84.13	8714	995,356	114.22
Source: IBGE (2007) IBGE – Pesquisa Pecuária Municipal, http://www.ibge.gov.br/home/estatistica/pesquisas/default.shtm, accessed August 2007.

Census-reported herd size per municipality is divided by the number of rural lots in each municipality in order to obtain the number of cattle owned per lot (Table 8). These values are compared by year (and sample type for 2005) to the household-reported values. For 1996 all census values for cattle per lot fit within the 95% CI, but not the 80% CI, in particular for Mirante da Serra (Table 9). The government projections for 2000 and 2005 are less congruent with our survey data, with half of the census-based estimates falling outside of the 95% CI in 2000, and with one falling outside the 95% CI in 2005 (Table 9). This may be because our expanded sample in 2005 includes settlements that did not exist in 1996, the base year for the projections – and thus those projections may not reflect the settlement of these new areas.

Table 9. Confidence tests for cattle estimations per lot

	Census estimate	Survey estimate	N	95% confidence interval		Within confidence	80% confidence interval		Within confidence
1996 n = 171
Ouro Preto do Oeste	105.89	98.95	84	80.69	117.21	√	87.09	110.81	√
Vale do Paraíso	70.75	85.30	20	58.37	112.23	√	68.21	102.39	√
Nova União	NA	NA	NA	NA	NA		NA	NA
Teixeirópolis	NA	NA	NA	NA	NA		NA	NA
Urupá	35.34	42.03	35	7.65	76.40	√	19.92	64.14	√
Mirante da Serra	31.22	22.75	32	12.13	33.37	√	15.93	29.57	X
2000 n = 170
Ouro Preto do Oeste	124.40	159.91	44	127.36	192.46	X	138.90	180.92	X
Vale do Paraíso	102.02	109.50	20	80.50	138.50	√	91.10	127.90	√
Nova União	69.15	119.75	28	83.43	156.07	X	96.50	143.00	X
Teixeirópolis	156.84	136.36	11	225.79	27.56	√	81.29	191.43	√
Urupá	44.74	38.49	35	27.56	49.42	√	31.46	45.51	√
Mirante da Serra	48.61	34.38	32	20.93	47.82	X	25.74	43.01	X
2005 – original sample (n = 169)
Ouro Preto do Oeste	172.64	182.86	43	132.92	232.80	√	150.64	215.08	√
Vale do Paraíso	171.36	173.50	20	128.59	218.41	√	145.01	201.99	√
Nova União	104.99	142.20	30	97.12	187.28	√	113.30	171.10	X
Teixeirópolis	233.90	184.17	12	68.92	299.41	√	112.78	255.55	√
Urupá	56.56	51.23	30	37.01	65.45	√	42.12	60.35	√
Mirante da Serra	75.15	53.38	34	43.74	63.02	X	42.62	64.15	X
2005 – expanded sample (n = 286)
Ouro Preto do Oeste	172.64	162.89	66	128.03	197.76	√	140.29	185.50	√
Vale do Paraíso	171.36	139.00	32	104.86	173.14	√	117.08	160.92	X
Nova União	104.99	103.51	65	78.05	128.96	√	87.01	120.01	√
Teixeirópolis	233.90	178.94	16	95.40	262.47	√	126.40	231.48	√
Urupá	56.56	48.48	52	39.06	57.90	√	42.39	54.57	X
Mirante da Serra	75.15	56.29	55	41.75	70.83	X	46.88	65.70	X
2005 expanded and tracked (n = 358)
Ouro Preto do Oeste	172.64	141.94	80	110.78	173.09	√	121.71	162.17	X
Vale do Paraíso	171.36	139.94	36	104.35	175.54	√	117.04	162.85	X
Nova União	104.99	95.90	79	79.20	124.93	√	87.22	116.91	√
Teixeirópolis	233.90	178.94	16	95.40	262.47	√	126.40	231.48	X
Urupá	56.56	39.08	75	31.04	47.12	X	33.86	44.30	X
Mirante da Serra	75.15	51.69	72	38.02	65.37	X	42.82	60.56	X

Finally, we compare census and survey values for education of household head and average household size using the 2000 survey round and population census. All census estimates of the education level of the male household head fall within the 95% CI, whereas all but one are within the 80% interval (Table 10). Meanwhile, half of the census estimates of the education level of the female household head fall within the 95% CI, and none are within the 80% interval (Table 10). The household size differs significantly between the census and our survey, with none of the census values for household size falling within the 95% CI (Table 10). Instead of reflecting any blatant error, these discrepancies result from different definitions of household size. Although the survey data capture all residents residing on the lot, the census data reflect only the members of the main household. Because there are often multiple (and often related) households residing on any single lot, the survey data values are consistently higher than the census comparisons.

Table 10. Confidence tests for household education and size estimations by municipality; 2000

	Census estimate	Survey estimate	N	95% confidence level		Within confidence	80% confidence level		Within confidence
Average years of education of male household heads
Ouro Preto do Oeste	3.08	2.57	37	3.12	2.02	√	2.92	2.22	X
Vale do Paraíso	2.59	2.11	18	2.91	1.31	√	2.62	1.61	√
Nova União	2.72	2.85	20	3.73	1.97	√	3.41	2.29	√
Teixeirópolis	2.79	2.00	8	3.34	0.66	√	2.80	1.20	√
Urupá	2.77	2.53	30	3.29	1.78	√	3.02	2.05	√
Mirante da Serra	2.83	3.08	25	3.91	2.25	√	3.61	2.55	√
Average years of education of female household heads
Ouro Preto do Oeste	2.28	2.50	38	3.25	1.75	√	2.98	2.02	X
Vale do Paraíso	2.23	2.11	18	2.96	1.26	√	2.65	1.57	X
Nova União	2.49	1.58	26	2.24	0.91	X	2.00	1.15	X
Teixeirópolis	1.79	2.64	11	3.69	1.58	√	3.28	1.99	X
Urupá	1.74	2.38	34	3.01	1.76	X	2.78	1.98	X
Mirante da Serra	1.89	3.10	30	3.83	2.37	X	3.57	2.63	X
Average number of household members
Ouro Preto do Oeste	3.96	7.81	43	9.95	5.68	X	9.19	6.44	X
Vale do Paraíso	4.1	7.65	20	9.70	5.60	X	8.95	6.35	X
Nova União	4.2	8.25	28	10.39	6.11	X	9.62	6.88	X
Teixeirópolis	4.04	7.73	11	11.35	4.11	X	9.96	5.50	X
Urupá	4.33	6.14	35	7.33	4.96	X	6.91	5.38	X
Mirante da Serra	4.14	7.44	32	10.02	4.85	X	9.10	5.78	X

8. Forest clearing estimations

In this section we estimate 2005 forest clearing or deforestation levels (also termed non-forest for consistency between the survey and remote sensing estimates). In the interest of space we rely on previous work to specify a dynamic model of forest cover (Mertens and Lambin 1997; Pichon 1997b; Walker 2004). Our explanatory variables include indicators of distance and market access that many have found to be important determinates of deforestation (Pfaff 1999; Barbier 2001), biophysical conditions of the lot (such as soil type and slope) that have been identified as important determinants of land use, and household characteristics that reflect stage in life cycle and labor availability as well as wealth and productive assets because households are both consumers and producers (Singh, Squire, and Strauss 1986). To construct these variables we utilize the remote sensing time series and survey panel data. We add these data to our regression analysis in a stepwise fashion, first controlling for forest cover in 1990 and other biophysical conditions of the lot, and then adding socioeconomic variables (lagged to reduce endogeneity concerns). In all the cases, we control for unobserved differences across municipalities (e.g., in governance) by including dummy variables for all except the central municipality of OPO. We estimate models for four different balanced 2-year panels, using the years 1996 and 2005 to take advantage of the full span of the survey data.

The first panel is titled the ‘original lot panel’, including only the original stratified random sample of lots (Model 1, Table 11). Next, the ‘expanded lot panel’ incorporates all lots surveyed in 2005, including lots selected through additional random sampling of the original and new settlements (Model 2, Table 11). For both the lot panels, we analyze only geospatial variables (not relying on the household survey) and do not include any of the household and household members that we tracked to new locations within or outside of the original survey region, although the dependent variable (land cleared since occupation or non-forest area) is the same for each of these estimations. This method replicates a removed time-series approach or a random draw of households that is chosen from a map while additional locations are added in a random way to increase the survey sampling frame. No contact is necessary with the household to obtain these data. The exclusion of survey variables is what makes the expanded lot panel possible, because the analysis requires lagged values, and we have remote sensing data, but not survey data, for all of these lots for both 1996 and 2005 because of sample attrition.

Table 11. Estimations of deforestation (non-forest)

	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6
Variable	Original lot panel (all lots interviewed in 1996 and 2005)	Expanded lot panel (lots interviewed in 2005 plus lots added to sample from original and new settlements)	Original household panel (all households interviewed in 1996 and 2005 on original lots)	Original household panel (all households interviewed in 1996 and 2005 on original lots) selection model	Expanded household panel (all households interviewed in 1996 and 2005 plus original lots and tracked)	Expanded household panel (all households interviewed in 1996 and 2005 plus original lots and tracked) selection model
Constant	59.00∗∗∗ (19.89)	73.12∗∗∗ (9.87)	1839.91∗ (1122.30)	2972.96 (2758.35)	1509.09∗ (811.62)	1566.50∗ (835.70)
Deforestation on lot 1990	0.66∗∗∗ (0.18)	0.54∗∗∗ (0.08)	0.41∗∗ (0.21)	0.75 (0.55)	0.48∗∗∗ (0.14)	0.52∗∗ (0.22)
Distance city center	0.22 (0.18)	0.14 (0.30)	0.14 (0.34)	0.80 (1.05)	0.22 (0.23)	0.19 (0.23)
Soil	0.85 (4.44)	− 1.16 (3.06)	3.28 (4.85)	− 6.74 (13.94)	3.02 (3.39)	2.88 (3.37)
Average slope	− 0.95 (0.90)	− 0.45 (0.63)	− 0.96 (1.40)	− 2.78 (3.36)	− 0.88 (0.96)	− 0.90 (0.94)
Mirante da Serra(municipality dummy)	− 15.48 (17.20)	− 18.12∗ (9.77)	− 22.49 (19.06)	− 14.47 (43.13)	− 16.67 (12.69)	− 12.95 (16.58)
Nova Uniao (municipality dummy)	16.86∗ (9.59)	− 6.61 (6.33)	23.73∗∗ (911.56)	14.97 (27.41)	21.37∗∗∗ (8.37)	22.01∗∗∗ (8.19)
Teixeiropolis (municipality dummy)	0.39 (11.63)	3.19 (9.15)	3.09 (15.38)	22.23 (38.47)	2.44 (12.71)	− 0.25 (13.37)
Urupa (municipality dummy)	− 32.49∗∗∗ (12.64)	− 37.82∗∗∗ (7.54)	− 37.13∗∗ (15.47)	− 64.29∗ (39.12)	− 28.23∗∗∗ (9.70)	− 26.13∗∗ (11.40)
Vale do Pariso (municipality dummy)	− 1.03 (10.36)	− 2.70 (7.50)	1.58 (11.05)	19.52 (29.72)	3.41 (8.52)	5.01 (10.57)
Age household			0.15 (0.31)	2.73 (2.12)	0.18 (0.21)	0.20 (0.32)
Edu household			0.61 (1.89)	3.75 (5.01)	0.83 (1.31)	0.58 (1.26)
South			19.68∗∗ (9.71)	10.32 (23.00)	11.20∗ (6.36)	11.93∗ (6.43)
Cattle (in 1996)			− 0.07 (0.05)	− 0.01 (0.14)	− 0.05 (0.03)	− 0.05 (0.04)
Vehicle value, thousands (in 1996)			1.29∗∗∗ (0.25)	1.51∗∗∗ (0.59)	1.28∗∗∗ (0.16)	1.33∗∗∗ (0.25)
Total income (in 1996)			0.01 (0.45)	0.11 (1.08)	0.00 (0.00)	− 0.07 (0.43)
Harvestdiversification(in 1996)			− 3.52∗∗∗ (1.26)	− 1.59 (3.30)	− 3.13∗∗∗ (0.94)	− 2.88∗∗∗ (1.15)
Year lot acquired			− 0.89 (0.57)	− 1.58 (1.42)	− 0.73∗ (0.41)	− 0.76∗ (0.43)
Lambda				110.13 (83.48)		7.98 (30.22)
R	0.40	0.38	0.58	0.60	0.64	0.64
R ^2	0.37	0.37	0.51	0.53	0.60	0.60
F	12	20.75	9.08	9.36	17.99	16.85
n	173	312	132	132	190	190
Note: ∗, ∗∗, and ∗∗∗indicate significance at the 10, 5, and 1% levels, respectively; SE in parenthesis.

Estimation results for these ‘lot panels’ reveal that the amount of deforestation on the lot in 1990 (our pre-survey benchmark) alone explains approximately 27% of the deforestation in 2005 (results not reported here). The addition of biophysical conditions of the lot, distance to the city center, and municipality dummy variables increases the explanatory power of the model by over 10 percentage points (Models 1 and 2, Table 11). In both lot panels, only a few of the municipality dummy variables are significant determinates of total land clearing in 2005, after controlling for 1990 deforestation levels.

The remaining models reported in Table 11 are based on two household panels: (i) the original household panel: all households interviewed in 1996 and 2005 on their original lots and (ii) the expanded household panel: all households interviewed in 1996 and 2005, including those residing on their original lots and those that were tracked to new locations. Households that moved onto lots since 1996 are therefore excluded from both these samples. In addition to geospatial variables, these models include household characteristics such as age, education, and origin of household heads. We also include lags of various indicators of wealth and income, again restricting us to households contacted in both survey years.

The estimation results based on household panels (Models 3–6) show that in addition to the initial conditions on the lot (soil, slope, and prior deforestation), household characteristics such as wealth, origin, and diversification are significant drivers of deforestation. The addition of these variables increases explanatory power relative to Model 2, despite the substantially smaller sample size. Households who are from the South and Southeast of Brazil (and therefore probably with higher levels of initial capital), have more assets (as reflected in value of vehicles), and pursue more specialized production processes tend to have significantly higher levels of deforestation. For example, according to Model 3, for every additional R$1000 of vehicle ownership in 1996, 1.29 additional hectares are deforested by 2005. Similarly, every additional crop type harvested on the lot results in 3.5 fewer hectares deforested.

Not only are wealth and initial holdings important determinants of future deforestation, the inclusion of these variables reduces the size of the impact attributed to the biophysical conditions of the lot. For example, according to Model 3, each additional hectare of deforestation in 1990 leads to about 0.4 additional hectares deforested in 2005. This value is equal to 0.5 when only the biophysical and other characteristics of the lot are included. ¹¹ In other words, when not accounting for household level information, the result is an overestimation of the impact of conditions of the lot.

It is also worthwhile comparing estimations with the original household panel against those with the expanded household panel. Tracking expanded our original household panel from 132 to 190, reducing the attrition rate from 25 to approximately 2%. ¹² Although there is a 5% increase in explanatory power when the model is re-estimated with the expanded panel (Models 3 vs. 5), most of the coefficients are of the same sign and similar size. The exception is that the year of acquisition becomes significant. The negative sign indicates that those lots that were acquired later have significantly less deforestation.

Finally, Models 4 and 6 are Heckman selection models run for the two household panels in order to examine possible attrition bias. ¹³ These results do not indicate any significant bias for the estimates of forest clearing, as the inverse Mills ratio (reported as lambda) is not significant in either of these estimations. We attribute this lack of estimated bias to the low attrition rate in our sample and the systematic stratification of the originally selected households.

9. Conclusion

This article addresses key methodological issues central to advancing land use science and understanding of deforestation processes. We have also described in detail the development of one of the few publicly available panel data sets suitable for modeling socioeconomic and biophysical determinants of LULC change in the tropics. We demonstrate the feasibility and value of panel tracking, as well as periodically expanding the panel, to LULC analyses. In our study site, a pre-survey registry of households residing on lots in our panel was critical to reducing attrition by efficiently integrating tracking of households into our field plan. By estimating models of LULC with different panels, we clearly demonstrate the value of panel survey data from interviews with households who are the agents of land use change. Specifically, we find that accumulation and variation in household wealth impacts forest clearing, even after controlling for initial biophysical conditions of the lot and municipality-fixed effects. It is unlikely that a similar deduction could be made solely based on geo-spatial, macro-level, or even cross-sectional household survey data. These conclusions naturally lead to the reciprocal question of how LULC impacts wealth. Our survey data are available and well-documented for researchers wishing to pursue such questions.

Another interesting finding is that the panel unit and data type influence our estimations of LULC change. We confirm that the average values and variation for most of our household characteristics and land use categories are quite different between those in our original and expanded samples, thus, concluding that it is important to recognize the limitations of the sampling frame. For example, the tracked household sample – including all families in the original household sample who could be located plus a sample of individuals who had left those families – represents the situation of the settlers who were in the region in 1996. This sample can be used to analyze welfare dynamics over time, but we cannot argue that it is representative of the rural population or land owners/managers in 2005. On the other hand, our original household sample likely represents those households that are comparatively successful given that they did not relocate to new lots because of failure to meet subsistence needs. Whereas our expanded household sample that includes the original household sample plus the addition of a random sampling of households with the survey region (and including new settlements) is likely the best representation of land owners/manager in 2005.

Finally, we estimate total land clearing (or deforestation) since lot occupation to evaluate the impacts of sampling decisions. We confirm that household survey data are important for modeling deforestation, adding both to the explanatory power of our model and to the number of identified significant drivers. We test for attrition bias in these models with a Heckman selection model for both our household panels and do not find evidence of such bias. In sum, these results suggest that there are significant benefits to collecting panel data for understanding the complex drivers of LULC change. We note a fundamental difference in our understanding of the determinants of deforestation that can be attributed to household survey data. We believe such study and the provision of complementary public data are necessary components to advancing the field of land use science.

Acknowledgments

This research was funded by the National Science Foundation, under grant SES-0452852. We thank our survey team: Stella Maris de Souza Freitas, Eliane S. Pedlowski, Ivone Holz Seidel, Taís Helena Akatsu, Luciana Bussolaro Baraba, and Tânia Rodrigues Luz for their tireless efforts to complete the household surveys in 2005, as well as the local residents of OPO for their participation, and the Associação de Produtores Alternativos for logistical support. We gratefully acknowledge Crisanto Lopes de Oliveira for all of his hard work on our survey registry. The groundwork that he provided for our survey teams by visiting and marking each of the lots in our sample, permitted our survey administration to be more efficient. We also thank Carlos José da Silva for serving as a driver and guide to our GIS team, and Niklas Hebron for assistance with data entry. Carlos' local knowledge was invaluable. Previous rounds of data collection were supported by the National Science Foundation, grant SES-0076549 in 2000, and the National Security Education Program, the Organization of American States, the Institute for the Study of World Politics, and the McClure Fund Foundation in 1996. Partial support was also provided by the Perdue School of Business, Salisbury University.

Notes

 1. The data used in the analysis can be found at the archive of social science data for research and instruction at the Inter-university Consortium for Political and Social Research of the University of Michigan. All location identifiers have been removed.

 2. This method can create problems if weights are linked to the original sample without comparisons to the current.

 3. Precision component required – D largest acceptable difference between the value estimated from the sample and the true population value K is the measure of confidence with which it can be stated that the result does lie within the range represented by ±D The higher the value of K the greater the degree of confidence K = 2 often choices (95% confidence or odds of 19 to 1) K = 1 (odds of 2:1) (Casley and Lury 1982)Suppose yield has variance of 0.5 and satisfied with sample estimate within 10% of the true population D = 0.1

 4. Colored plaques were placed at each household to assist with the identification of properties.

 5. Car rental and fuel expenses for four vehicles and enumerator pay for six individuals are divided equally for each day within the first 4 weeks of the survey period. Car rental and fuel expenses for one vehicle and pay for two enumerators are applicable to the remaining 2 weeks devoted to completing unfinished work and tracking individual to new locations.

 6. The original stratified random sampling was applied for these households: target properties were identified every ‘x’ number apart. If none of the adult household members were in residence at the time of the interview, the enumerators were instructed to go to the property next door to conduct the interview.

 7.  i.e., survey  +  registry expenses.

 8. Distance is calculated as the average daily distance for all surveys completed in a particular day.

 9.  Standard errors are in parenthesis; ∗, ∗∗, ∗∗∗ indicate significance at the 90, 95, and 100% levels, respectively.

10. Note that decreasing the desired confidence level (i.e., from 95 to 80%) will tighten the CI. This decrease in width increases the precision of the estimates around the mean. This is because the selection of a confidence level for an interval determines the probability that the CI will contain the true parameter value. These levels correspond to percentages of the area of the normal density curve. For example, a 95% CI covers 95% of the normal curve whereas an 80% interval covers only 80% of the normal curve a smaller area that more precisely represents the true value.

11. Model 2 is estimated with the household panel (used in Model 3 estimation) to draw this conclusion.

12. The number of observations increased from the original 171 to 190 as households split and individual members moved to new locations or subdivided lots amongst smaller family units.

13. The probability of remaining in the sample, or not attriting, is estimated with the same household and lot characteristics as in the estimation of forest clearing.