Methodology Directory UMM :Data Elmu:jurnal:E:Ecological Economics:Vol31.Issue1.Oct1999:

government policies Moran, 1993. With a suc- cessful stabilization program now beginning to bear economic fruit in Brazil, figures for 1995 and 1996 reveal an alarming increase in deforestation rates. 8 Indeed, currently 16 million people live and work in legal Amazonia, with over 1.4 million living in the city of Manaus Andersen et al., 1996. The process of deforestation has, in a sense, taken on a life of its own responding to local and national economic forces of population growth, landlessness, and the need for producer and household goods, food and building supplies. Despite recent national ‘environmentally friendly’ legislation, further land clearing and local road building occurs due to these endogenous eco- nomic pressures. The co-movement of deforesta- tion and national economic performance underlines this point: the process of environmen- tal degradation may no longer respond primarily to changes in national land policy although these are certainly important but rather is part of the complex interaction of the local and national economies. While it is beyond the scope of this paper to elucidate the complexities of this relationship, we do believe that a necessary ingredient of a dy- namic, macroeconomic model of the economics of deforestation is some aggregate parameter esti- mate of how crop land degrades over time. The methodology described below is presented as an alternative form to estimate this parameter.

3. Methodology

3 . 1 . Description of the data The data available for this study was derived largely from Brazilian National Agricultural Cen- sus. The original database included municipality- level figures on economic, demographic, ecological and agricultural variables collected for the years 1970, 1975, 1980, and 1985 for 316 municipalities in the Brazilian Legal Amazonia. The data were cleaned, standardized and merged with data from other sources 9 in a painstaking exercise undertaken by Dr. Eustaquio Reis of the Institute of Applied Economic Research IPEA in Rio de Janeiro, without whose work this paper could not have been written. 10 For each munici- pality in each time period the variables 11 that are used for this analysis include total crop land, total planted pasture land, total fallow land, total labor force, value of total crop output, density of roads, population density, percentage land of high qual- ity soil, relative price of land, and finally, the state to which the municipality belongs. The variables and their definitions are reproduced for conve- nience in Table 2. Figs. 1 – 9 show the aggregate evolution from 1970 to 1985 of the percentage of Table 2 Variable definitions Total crop land in municipality i in time t crop it pasture it Total planted pasture land in municipality i in time t fallow it Total fallow land in municipality i in time t labor it Total labor force in municipality i in time t Value of total crop output in municipality i in output it time t road it Density of roads in municipality i in time t den it Population density in municipality i in time t Percentage land of high quality soil in munici- soil it pality I relpr it Relative price of land in municipality i in time t state it The state to which municipality i belongs 9 Other variables in the original data set at IPEA covering satellite deforestation measures and migration come from the population census and IBGE as well as some other govern- ment agencies. For this paper the data that was used were from the Agricultural Census, however. 10 For an extensive discussion of the entire database see Andersen et al. 1996. 11 Some variables appear as logs of the described data in later parts of the paper and are so noted. 8 See Andersen et al. 1996 for 1995 figures. Statements on 1996 deforestation rates are based on newly released statistics from IBGE, Rio de Janeiro Brazil. Fig. 1. Brazilian legal Amazonia. land used out of total land area for crop land, planted pasture and fallow land, respectively, for each of the eight states represented in the analysis. 3 . 2 . O6er6iew of the proposed methodology The purpose of this paper is to estimate the extent to which the value of crop output declines with time after forest has been cut and the land planted in crops. As described above, data is available in each period on the area of land used for crops and other agricultural activities in each municipality. However, in order to estimate the rate of land degradation directly we would have to know exactly what land was used for each activity, not just the share of the municipality dedicated to each use. For example, if crop land increases from 20 to 25 between 1980 and 1985, we cannot know how much of the old crop land is still being used and how much is newly cleared or converted land. Thus it is necessary to estimate Fig. 2. Acre. Fig. 3. Amazonas. Fig. 5. Roraima. land that was in crops in time t − 1, land that was in pasture in time t − 1, or land that was fallow in time t − 1. Analogously both pasture and fallow land in time t must also come from these four sources, although the proportion from each source may vary. The land use model is thus: crop it = b 1 Dclear it + b 2 crop it − 1 + b 3 pasture it − 1 + b 4 fallow it − 1 1 pasture it = f 1 Dclear it + f 2 crop it − 1 + f 3 pasture it − 1 + f 4 fallow it − 1 2 fallow it = h 1 Dclear it + h 2 crop it − 1 + h 3 pasture it − 1 + h 4 fallow it − 1 3 where b j , f j , and h j are parameters that indicate, at time t, what proportion of crop land, pasture land, and fallow land, respectively, come from source j, where j is an index that maps to Dclear j = 1, crop j = 2, pasture j = 3 and fallow j = 4 in time t − 1. The closed nature of the model implies that b j + f j + h j = 1 for j = 1, 2, 3, 4. Eq. 1 holds in all time periods t so we can lag all variables by one period to obtain: crop it − 1 = b 1 Dclear it − 1 + b 2 crop it − 2 + b 3 pasture it − 2 + b 4 fallow i t − 2 4 the pattern of land use and land vintage i.e. to estimate the proportion of crop land that derives from newly cleared land, from old crop land, from pasture land or from fallow land in each municipality in order to compute the rate of productivity decline. This estimation is accomplished by first con- structing a land use transition model. 12 We distin- guish between natural land which is defined as planted forest, virgin forest and natural pasture, and cleared land which is comprised of crop land, planted pasture and fallow land. As this is a closed system the definitions of cleared land clear, and the change in cleared land 13 Dclear, can be constructed as: clear it = crop it + pasture it + fallow it Dclear it = clear it − clear it − 1 Crop land in municipality i at time t must come from four possible sources: newly cleared land i.e. land that was in a natural state in time t − 1, Fig. 4. Maranha˜o. Fig. 6. Mato Grosso. 12 This model was originally proposed by Clive Granger and explored in Andersen and Granger 1995 for the Amazon. 13 i.e. newly cleared land, converted from its natural state. Fig. 7. Amapa´. Fig. 9. Para´. If we then substitute Eq. 4 into Eq. 1 it is then possible to expand the crop land Eq. 1 14 to get: crop it = b 1 Dclear it + b 2 b 1 Dclear it − 1 + b 2 crop it − 2 + b 3 pasture it − 2 + b 4 fallow it − 2 + b 3 pasture t − 1 + b 4 fallow it − 1 5 We then collect terms and interpret the compo- nents of Eq. 5 as corresponding to different vintages of crop land in time t. For example, the first term b 1 Dclear it is defined as the area of crop land that comes from land that has been newly cleared during the current 5-year period We note that since our data periodicity is every 5-years time ‘t ’ actually corresponds in practice to a 5- year period. Thus time t − 1 would correspond to the preceding 5-year period, and so on. We call this NEWLAND. The second term, b 2 b 1 Dclear it − 1, corresponds to land that was newly cleared in the previous 5-year period time t − 1 and planted in crops at that time, and which has remained in crop land to the present. Thus this land has been in crops for at least 5-years and we denote it 5 YRLAND. The third term is b 2 2 crop it − 2 which is the current crop land area that was planted in crops in both the previous 5-year pe- riod and the 5-year period before that. This land has been cultivated for at least 10 years and we thus denote it 10 YRLAND. We define the remain- der of the components of the expanded Eq. 5 in a similar fashion, with the definitions summarized in Table 3. Thus the coefficient estimates from the crop land Eq. 1 in the land transition model defined above can be used to construct an estimate of the proportion of total crop land that comes from newly cleared land, NEWLAND, from land that was cleared in the previous 5-year period, 5 YRLAND, from land that was cleared two 5- year periods ago, 10 YRLAND, and from land that has had other uses in the previous time periods, CRPPAS, CRPFAL, PAS and FAL as described in Table 3. We refer to these categories as different 6intages of crop land. We then assume Cobb – Douglas agricultural production: CROPOUTPUT = ALABOR y 1 LAND j y j where A is a technological constant and the n j denote elasticities that give the percentage change in output that results from a unit percentage change in input j. Taking logs of both sides and defining lower case variables as the log of upper- case variables, we can then estimate these elastic- ities from the regression: output it = a + y 1 labor it + y 2 newland it + y 3 5yrland it + y 4 10yrland it + y 5 crppas it + y 6 crpfal it + y 7 pas it + y 8 fal it + o it 6 Fig. 8. Goia´s. 14 It would, of course, be possible to expand out the last two equations as well, which would enable us to estimate the dynamics of pasture productivity from data on cattle herd. However, this is not necessary for the task at hand. Table 3 Constructed land type ‘vintage’ definitions Variable name Term from Eq. Variable definition 5 NEWLAND b 1 Dclear it Land that has been newly cleared sometime during the current 5-year period and planted in crops b 2 b 1 Dclear it−1 Land that was newly cleared sometime during the previous 5-year period and been used 5 YRLAND as crop land continuously since 10 YRLAND b 2 2 crop it−2 Land that has been in crops for the past 10 years continuously. It is not known what the original form of this land was before that b 2 b 3 pasture it−2 CRPPAS Land that was used as crop land for the current and previous 5-year periods, but which had been used as pasture land before that b 2 b 4 fallow t−2 CRPFAL Land that was used as crop land for the current and previous 5-year periods, but which had been used as fallow land before that Land that was pasture during the previous 5-year period but has now been converted to b 3 pasture t−1 PAS crop land Land that was left fallow during the previous 5-year period but has now been converted b 4 fallow t−1 FAL to crop land The estimated coefficients from this regression, 6 1 to 6 7 , give us an estimate of the percentage change in output for a 1 change in the corresponding land area. We can easily calculate the percentage change represented by an increase of a given area unit in one land category and the consequent change in output, which will give us an idea of the productivity of that land. 3 . 3 . Estimation procedure for the land use transition model As described in the previous subsection, the land use transition model has the theoretical property that b j + f j + h j = 1. Also, it is clear that theoretically all the coefficients should lie between zero and one since no less than 0 and no more than 100 of the land of a given type can be converted to crop land. When the basic model is estimated by a heteroskedasticity consistent gener- alized least squares procedure 15 henceforth sim- ply GLS separately on each equation, the coefficients do indeed sum to one. Two problems emerge, however. First, Andersen and Granger 1995 show that the heterogeneity in the panel can lead to bias in the GLS estimates. Second, the summing up property is an artifact of the data, and in the presence of heterogeneity comes at the cost of coefficient estimates that occasionally fall outside the [0, 1] bound. We will discuss the het- erogeneity problem first and then how it relates to the second problem of negative coefficient values. Land transition patterns could be expected to vary from municipality to municipality depending on any number of characteristics such as soil quality, population density, land type savanna etc., land area, and distance to nearby markets. In addition, the coefficients might be expected to change through time and from state to state. We thus try to control for these factors in the regres- sions. In addition to the above characteristics, we also control for the average price of land. Land prices serve as an excellent proxy for many unob- served or imperfectly observed characteristics that impact the desirability of land, such as pres- ence of a rural infrastructure or economic proxim- ity to an urban center. In particular, Andersen et al. 1996 have found that land prices are very Table 4 Productivity estimates of the value of additional crop output from a one hectacre increase of each given land vintage a NEWLAND 5 YRLAND 10 YRLAND 993.42 919.76 Estimate 4955.26 41.18 103.06 St. Dev. 1832.9 a Sample mean, 1671.60; sample SD, 1569.15. 15 We use White’s Generalized Heteroskedasticity-Consis- tent Feasible Least Squares see White 1980. strongly correlated with subsidized credit and other fiscal incentives for the region. The data on land prices is in current prices so in order to avoid any biases from using a deflator, only relative prices in each year are used, so that each price represents that municipality’s share of total land prices for that year. It is clearly necessary to take all of this infor- mation into account if meaningful estimates of land degradation are to be calculated. In addition, Andersen and Granger 1995 show that estimates will be biased if the heterogeneity is ignored. However, if we allow the b’s to vary according to different characteristics of the municipality, this multiplies the coefficients to be estimated and increases the probability of coefficient estimates falling outside the theoretical [0, 1] bound. 16 This in turn may exacerbate a second problem, which is that the Cobb – Douglas production function that will be estimated must have only non-nega- tive inputs. However GLS on the land transition model does not put any constraints on the coeffi- cient estimates, leading in some cases to negative values. Thus, although GLS yields the ‘summing up’ property referred to earlier, this is an artifact of the data construction and is achieved at the cost of coefficient estimates that occasionally lie out- side of the [0, 1] interval. While in theory coeffi- cients shouldn’t fall outside this interval, there could be measurement problems and actual cir- cumstances not controlled for in our simple model that could lead to negative GLS estimates. For example, the original census land use data is compiled from people who own agricultural es- tablishments and report to the census taker how the land is used. In some cases the boundaries between the municipalities is not well defined, and agriculturists and census takers may attribute land to one municipality that actually lies in a neigh- boring municipality. This over-measuring of cleared land in one period may lead in subsequent periods to negative values for the change in cleared land for a particular municipality. This problem occurs most frequently where it would be expected: in the smaller municipalities. Negative values for the change in cleared land is a suffi- ciently common problem that simply eliminating all observations with this property can lead to severe sample selection bias in the estimates. We call this problem ‘shrinking’ and introduce a ‘shrinking’ dummy variable that interacts with the primary variables to allow the coefficients to vary between municipalities which display this prop- erty and those that do not. We thus allow both the land use coefficients and the elasticity esti- mates from the Cobb – Douglas production func- tion to vary according to whether or not a municipality exhibits ‘shrinking’ over time. In sum, the estimation process thus entails us- ing GLS to estimate the initial land use transition system. Not truncating the coefficient estimates so as to enforce the [0, 1] bound actually accentuates the final results presented below. Therefore, to save space and in the interests of keeping the estimation consistent with the theory, only those results using truncated coefficients are presented in Table 5. 17 After the land use transition coeffi- cients have been estimated, they are used to con- struct the different vintages of crop land as described in Table 3, and these are in turn used as inputs into a time- and municipality-varying Cobb – Douglas production function in which the elasticities are allowed to vary according to mu- nicipality specific characteristics. Thus we have: output it = a + y 1it labor it + y 2it newland it + y 3it 5yrland it + y 4it 10yrland it + y 5it crppas it + y 6it crpfal it + y 7it pas it + y 8it fal it + o it 7 16 Previous versions of this paper have used nonlinear least squares to impose the restriction of the 0,1 bound on the coefficients. There were several problems associated with this method, however. The models were very difficult to estimate and often did not converge, leading to artificially restricted sets of explanatory variables. The estimates were also very sensitive to the choice of explanatory variables and often produced estimated betas with extremely bimodal distributions. A num- ber of trials were run in which betas estimated with GLS gave similar results to the betas estimated with nonlinear least squares with similar variables. However, by allowing for greater heterogeneity, more complex GLS estimation yields coefficients which are distributed with a unimodal bell shape. 17 Primarily the untruncated productivity estimates differ in that the productivity of NEWLAND is considerable higher. Table 5 Land use regression results: dependent variable = crop land 1985 a,b T for H0: Parameter Variable Variable Parameter T for H0: estimate parameter = 0 parameter = 0 estimate − 1.589 CROPS8 DCLEAR 0.571888 − 0.072620 6.014 0.506 CROPL1 PASSAV 0.127855 0.021368 3.828 2.852 PASSOI 0.170512 − 0.029589 FALLOL1 − 2.020 0.015924 PASTUL1 0.739 PASDEN − 0.000083341 − 0.186 − 0.066628 CLEARSAV − 4.055 PASARE − 0.001302 − 0.769 0.095 PAS5 0.003818 0.009524 CLEARSOI 2.571 1.285 CLEARDEN PAS80 0.000869 − 0.011756 − 2.518 2.858 PAS85 0.010913 − 0.006182 CLEARARE − 1.335 0.015403 CLEAR80 1.375 PASPR 1.697775 0.642 − 0.033775 CLEAR85 − 2.608 PASSQ 17.702235 0.074 1.478 PASTURS2 7.112195 0.040851 CLEARPR 0.593 0.041 PASTURS3 CLEARSQ − 0.095735 15.281006 − 1.983 1.628 PASTURS5 0.072107 0.000803 DCLEARS2 0.094 0.084827 DCLEARS3 2.901 PASTURS6 − 0.022572 − 1.116 0.011887 DCLEARS5 0.664 PASTURS7 0.003730 0.317 0.127 PASTURS8 0.003950 − 0.011261 DCLEARS6 − 1.937 1.386 FALSAV DCLEARS7 − 0.018554 0.028833 − 0.795 3.962 FALSOI 0.047039 0.040522 DCLEARS8 0.978 0.143347 CROPSAV 1.285 FALDEN − 0.000540 − 1.242 − 0.197411 CROPSOI − 1.451 FALARE − 0.013747 − 2.738 − 0.512 FAL5 − 0.000796 − 0.010793 CROPDEN − 0.877 3.983 FAL80 CROPARE 0.028101 0.073664 1.907 − 3.583 FAL85 − 0.204976 − 0.059743 CROP5 − 4.081 − 0.101815 CROP80 − 1.325 FALPR 22.797138 3.621 − 0.051464 CROP85 − 0.738 FALSQ − 304.697179 − 0.894 − 1.906 FALLOWS2 − 24.030320 0.008875 CROPPR 0.085 0.932 FALLOWS3 − 0.017157 − 0.485 CROPSQ 233.998902 − 0.010 FALLOWS5 − 0.003254 − 0.016179 CROPS2 − 0.641 0.294712 CROPS3 2.252 FALLOWS6 0.096211 0.979 0.260649 CROPS5 2.164 FALLOWS7 − 0.031654 − 1.271 0.359 FALLOWS8 0.161712 − 0.093095 CROPS6 − 3.837 2.660 CROPS7 0.332501 a Regression N, 945. b R-square, 0.9497, Adj R-sq, 0.9462 Finally, the productivity estimates are derived us- ing the estimated elasticities from this regression. For example, we can calculate the additional out- put that would be obtained from increasing each land type input by one hectacre and compare these yields across the different vintages of land. The difference between the yields on newly cleared land and land that has been in crops for longer periods will give us an estimate of the rate at which agricultural productivity is changing over time. 4. Results 4 . 1 . Construction and estimation of the land use transition model As discussed earlier we wish to allow the coeffi- cients to vary from municipality to municipality and over time depending on various characteris- tics of the location and land. Thus each coefficient in the model is defined as varying by year, state, share of savanna land, Table 6 Cobb–Douglas regression results a,b Dependent variable = logreal crop output Parameter Variable T for H0: Variable Parameter T for H0: estimate estimate parameter = 0 parameter = 0 23.950 YR105 INTERCEPT − 0.305727 7.462146 − 4.139 − 0.025811 T85 − 0.331 YR106 − 0.276502 − 1.263 13.414 YR107 0.687467 − 0.062654 LLABOR − 0.958 − 0.018486 LABSH − 0.310 YR108 − 0.160449 − 2.383 0.021742 NEWLAND 0.925 CRPPAS2 − 0.197825 − 0.777 0.204 CRPPAS5 0.005104 − 0.066042 YR5LAND − 1.340 − 0.003567 YR5SH − 0.179 CRPPAS7 0.019352 0.445 4.894 CRPPAS8 0.323511 0.126619 Y10LAND 2.199 0.009743 YR10SH 0.125 CRPFAL2 0.124595 0.384 − 0.021759 CRPFAL − 0.478 CRPFAL3 0.179939 2.610 0.889 CRPFAL5 0.031160 0.305736 CRPFSH 4.636 0.186 CRPFAL6 PAS − 0.183971 0.007579 − 1.017 − 0.365 CRPFAL7 − 0.012335 0.022055 PASSH 0.408 0.012004 CRPPAS 0.285 CRPFAL8 0.082920 1.558 0.038276 CRPSH 1.268 PAS2 0.142451 0.686 0.501 PAS5 0.011520 − 0.033912 FAL − 0.725 FALSH 0.027226 1.549 PAS7 − 0.039818 − 0.912 − 0.383 PAS8 − 0.029738 0.042607 NEW2 0.726 0.010081 NEW3 0.268 FAL2 0.078883 0.687 0.045140 NEW5 1.418 FAL3 − 0.016650 − 0.529 1.467 FAL5 0.109607 0.050125 NEW6 1.673 − 0.053 FAL6 NEW7 0.488438 − 0.001562 2.491 2.286 FAL7 0.082432 − 0.021756 NEW8 − 0.875 − 0.016530 YR52 − 0.356 FAL8 − 0.090620 − 2.653 0.003639 YR53 0.116 NEWSOI 0.049929 1.249 1.267 YR5SOI 0.041157 − 0.028029 YR55 − 0.705 − 3.473 YR56 YR10SOI − 0.294741 0.049763 0.489 0.716 CRPPSOI 0.020344 − 0.212787 YR57 − 2.166 0.094037 YR58 2.902 CRPFSOI − 0.129868 − 1.595 YR102 − 0.428 − 0.060895 PASSOI − 0.031603 − 0.349 − 1.863 FALSOI 0.078401 1.564 − 0.133088 YR103 a Regression N, 629. b R-square, 0.8093, Adj R-sq, 0.7888. share of good soil, population density, area, whether or not the municipality crop land is ‘shrinking,’ and by relative price of land and the square of the relative price of land in that municipality. GLS regression results from the land use tran- sition model are presented in Table 6. Variable names are easily deciphered as the first part cor- responds to the type of land new, crop, fallow, or pasture and the second part to the interac- tion variable. Percentage of good soil is denoted by soi, percentage of savanna land is sa6, log of area of the municipality is are, population den- sity is den, relative prices are denoted by pr and the square of relative prices is sq. Each state is denoted by its corresponding number, and time shifts are clearly named by the year. The model explains about 94 of the variation in crop land although since there is no intercept in the model due care must be exercised in interpreting the R-squared statistics. These coefficient esti- mates are then used to calculate the land use transition b’s for each municipality in each sam- ple period. The estimated b’s are in turn themselves used to construct the land categories described in Section 3. If the technique is accurate, the sum of all seven land categories ought to be close to the actual area of crop land in 1985. For both truncated and untruncated b’s the percentage error between actual and predicted area of crop land is quite similar. The mean percentage error is − 0.308 for the truncated b ’s and − 0.2216 for the untruncated b’s, with respective standard deviations 1.954 and 1.717. These seemingly high and variable percentage errors are almost completely determined by a very few outliers most caused by large errors in the estima- tion of CRPPAS and CRPFAL, however, as the corresponding medians are 0.038 and 0.068, respec- tively. If the worst offenders of these outliers are deleted, the mean errors fall dramatically. This is clearly illustrated in the case of the untruncated b‘s, whose mean error, when the ten largest percentage error outliers are deleted, falls to − 0.040 with a standard deviation of 0.735. At any rate, although some of these large outliers may lead to large errors in the estimates of productivity of CRPPAS and CRPFAL, they have very little impact on the estimated productivity of NEWLAND, 5 YRLAND or 10 YRLAND. The fundamental results on pro- ductivity reported below are extremely robust to the inclusion or exclusion of any of these outliers. 4 . 2 . Estimation of the loss of crop producti6ity The estimated b’s are different for each munic- ipality and over time. These estimated coefficients are in turn used to estimate the land categories described in Section 3 for each municipality. Taking logs of both the land categories, labor and output a heteroskedasticity-consistent 18 GLS estimation of a Cobb – Douglas production function Eq. 7 produces estimates of the relative elasticities for each land category. As with the land transition model, it could be expected that productivity of various land types could vary across municipalities. For this reason the elasticity estimates are allowed to vary by soil type, state, and whether or not the municipality displays the measurement problem we call ‘shrinking’. In addition each coefficient is allowed to vary over time to capture any general shift between sample periods. The relative price variable was not included in this regression due to the possibility of endogeneity between the value of crop output and land prices. For example, it could be the case that higher levels of agricultural output led to higher land prices rather than the reverse. If this is the case then including land prices on the right-hand side of the regression could induce bias in the coefficient estimates. Table 6 presents the results of the Cobb – Douglas GLS regression of the log of crop output on the logs of the different land types, allowing the coefficients to vary according to the variables described above. Although the names of the variables seem a bit difficult, again in fact, they are simply a compound of the two variables compromising the interaction variable. The first part of the name corresponds to the land type, the second part to the interaction variable. Shrink is denoted by sh, percentage of good soil is soi and each state is represented by its corresponding number. The time shift dummy variable is represented by t 85 . The final elasticities are calculated for each municipality in each time period using the estimated coefficients. Each of these elasticities is multiplied by 100land j which gives the percentage change in output produced by an increase of 1 area units of land in any given land category j. This percentage is then converted into an actual monetary change in output for the 1 area unit increase in crop land. These productivities are then averaged over all municipalities for each land category. In order to avoid letting a few outliers impact the mean exces- sively, each observation that falls outside of 2 SDs of the mean is deleted before averaging, leaving a mean that is closer to the median. 19 The percentage change in output estimates are multiplied by the actual real output in each munic- ipality to give a monetary unit estimate of the increase in output for a 1 area unit increase in each type of land i.e. marginal output. The main results are presented in Table 4 which focuses on 19 Including the outliers essentially increases the productivity estimates of NEWLAND but does not impact much the esti- mates of 5 YRLAND or 10 YRLAND. 18 White’s heteroskedasticity consistent methodology. the calculated productivity estimates after delet- ing outliers as described above for NEWLAND, 5 YRLAND and 10 YRLAND, and provides the sample standard deviation as a point of compari- son. 20 In particular, we find that: “ an additional hectacre of newly cleared land planted in crops will increase agricultural out- put an average of almost 5000 R 1985 prices; “ increasing 5-year-old crop land area by one hectacre will increase output by an average of only 920 R; “ increasing 10-year-old crop land area by one hectacre will increase output by an average of about 1000 R; it is important to note, how- ever, that this figure is not statistically different from the estimate of 920 R from increasing 5-year-old crop land. The results imply a dramatic fall in productivity in the first 5-years after initial land clearing. The additional output from a 1 unit area increase in 5-year-old crop land is only 20 of the value of an additional unit of newly cleared land. Since the time span between periods is 5-years, these figures yield an annualized rate of just over 30 decline in agricultural productivity per year. This figure corresponds closely to the rates predicted by field research and to the general consensus of experts in the region of what the average rate of land degradation has been during the sample period. The productivity decline levels out, however, between 5- and 10-year-old crop land. One expla- nation for this is that the land that was cultivated first i.e. the oldest land was the best quality land settlers could find. In addition, the result could also be expected given the common land use patterns that prevailed at the time. Early colonists would clear land and practice low-capital shifting agriculture for a few years until the land was exhausted, then abandoning or selling the land to better-endowed second-wave settlers who used more intensive agricultural practices to keep yields at sustainable levels.

5. Conclusions