Accounting for the water purification service of tropical rainforests

We analyzed the effects of land use on water treatment cost and water treatment chemicals in the Malaysian state of Perak by using land-use variables from two data sources: forest inve[r]

Accounting for the water purification service of tropical rainforests

Jeffrey R Vincenta, Ismariah Ahmadb, Norliyana Adnanb,

Jie-Sheng Tan-Sooa, and Kyle Thomasa

Prof Jeffrey R Vincent

Nicholas School of the Environment

purification service provided by tropical forests to WTPs, and to illustrate how an estimate of this value can be used to account for this service in the system of national accounts (SNA) We crudely estimate that the value of this service is equivalent to about 7% of value added in the waterworks sector, and 5% of value added in the forestry and logging sector, in the Perak state product account

Acknowledgments

This study was funded by the Global Environment Facility through the United Nations

Development Programme (MAL/04/G31), with additional support provided by the Government

of Malaysia (through the Ministry of Natural Resources and Environment and the Forest

Research Institute Malaysia) and by the Center for International Forestry Research (CIFOR) The cooperation of numerous Malaysian government agencies is gratefully acknowledged: the

Forestry Department Peninsular Malaysia; the Departments of Agriculture, Drainage and

Irrigation, and Statistics; the National Water Services Commission; and the Perak Water Board

1 Introduction

Water purification might be the most frequently invoked example of an economically

valuable service provided by ecosystems to human society A 1998 commentary in Nature

reported enormous savings that New York City reaped by investing in watershed management in the Catskills in lieu of constructing a new water-filtration plant (Chichilnisky and Heal, 1998).1The 2010 TEEB report, hosted by the United Nations Environment Programme and supported by multiple European and Japanese government agencies, is just one of many publications that have highlighted this “celebrated” case (TEEB, 2010, p 20) The influential 2005 Millennium

Ecosystem Assessment included water purification on the short list of regulating services in its canonical diagram of the linkages between ecosystems and human well-being (2005; see Figure

A on p vi)

Well before these publications, a spate of econometric studies in the late 1980s found that water treatment costs in the United States were lower when the raw water processed by water treatment plants (WTPs) was less turbid, i.e., contained lower levels of suspended and dissolved solids (Forster, Bardos, and Southgate, 1987; Moore and McCarl, 1987; Holmes, 1988) These studies found that a 1% decrease in turbidity reduced treatment costs by 0.07–0.33% Similar findings were reported by subsequent U.S studies (e.g., Dearmont, McCarl, and Tolman, 1998) These findings, when combined with even older evidence that runoff from forests tends to be cleaner than runoff from other land uses (Dunne and Leopold, 1978, Ch 17; Hewlett, 1983, Ch

8, 10), imply that ecosystems, especially forests, can indeed be expected to provide an

economically valuable water purification service (Brauman et al., 2007) Conte et al (2011) provide a recent example of an integrated analysis that values this service by relating land use in

1 The facts of this case have been disputed See Sagoff (2002) for a rebuttal, and see Kenny (2006) for a rebuttal of Sagoff

the northwestern U.S to water quality and water treatment cost

A 2002 study of 27 U.S water suppliers by the Trust for Public Land reported the rule of thumb, “for every 10 percent increase in forest cover in the source area, treatment and chemical

costs decreased approximately 20 percent” (Ernst, 2004, p 7; italics in the original).2

A joint report by the World Bank and WWF (Dudley and Stolton, 2003) similarly argued that improved watershed management could enable fast-growing municipalities in developing countries to supply clean drinking water to their populations more cost-effectively, a point that has been reiterated by other studies (e.g., Postel and Thompson, 2005) A comprehensive literature review

on the effects of forest cover on drinking water quality concluded, however, that the “Effects of multiple land uses that overlap in time and space across large watersheds are difficult to predict with current knowledge” (Dissmeyer, 2000, p ix) Simple relationships between treatment costs and forest cover are therefore unlikely to hold in large watersheds with complex patterns of land use

Information on the economic value of this service is especially scant in tropical

developing countries There is little doubt that, as in temperate regions, runoff from tropical forests tends to be cleaner than runoff from other land uses (Bruijnzeel, 2004), but little

economic analysis has been done on the effect of tropical forests on water treatment costs Based

on limited observations from a single WTP in Thailand, Sthiannopkao et al (2007) report an increase in treatment cost when raw water is more turbid Abdul Rahim and Mohd Shahwahid (2011) report a similar relationship for six WTPs in the Malaysian state of Kelantan We are unaware of other direct evidence from tropical developing countries

In this paper, we present econometric evidence on the effect of forest cover on water

2 Ernst, Gullick, and Nixon (2004, p 4) stated that this rule could not be verified for forest cover greater than 60%,

as “Not enough data were obtained on suppliers that had more than 65 percent forest cover in their watersheds to draw conclusions.”

treatment cost for a sample of 33 WTPs in the Malaysian state of Perak during 1994–2005 Perak

is located in the northwestern part of Peninsular Malaysia, which is the most developed and most populous of Malaysia’s three major regions.3

We analyzed monthly data on treatment cost (specifically, operating cost) and monthly use of the three primary treatment chemicals used by WTPs in Perak: alum, which is a coagulant used to reduce turbidity; lime, which is added to adjust pH; and chlorine, which is a disinfectant We related these variables to GIS-based

information on land use in the WTPs’ catchments We had access to two sources of land-use data: National Forest Inventories (NFIs) conducted by the Forestry Department Peninsular

Malaysia, and national Land-Use Surveys (LUSs) conducted by the Department of Agriculture These sources enabled us to disaggregate forest cover into two types, virgin forests and logged forests, and nonforest land use into four types, urban, rubber, oil palm, and other We were thus able to investigate how water treatment cost was affected by not simply the amount of forest but also the type of forest and the nonforest land uses to which forests were converted

In addition to demonstrating how data from WTPs can be combined with land-use data to estimate the marginal value of water purification provided by tropical forests, we demonstrate how such an estimate can be used to incorporate this service into the system of national accounts (SNA) It has long been known that GDP implicitly incorporates the ggregate, current effects of environmental resources on national income (Mäler 1991) Making the value of these services explicit requires reallocating value added from the sector that uses the service to the one that supplies it (Vincent 1997), with adjustments made to both the use table and the supply table in the product account Because these adjustments constitute a reallocation of monetary flows that are already in the product account, they do not change aggregate value added (i.e., GDP)

To illustrate, suppose that the annual value of the water purification service supplied by

3 The other two are the states of Sabah and Sarawak on Borneo

forests to the water treatment sector is V Because the water treatment sector does not pay for this service, the use table in the SNA implicitly records V in the sector’s operating surplus, OS w The use table should instead record it as an intermediate input “purchased” from the forest sector

Using IC f-w to denote the cell in the intermediate consumption portion of the use table that

records purchases of inputs by the water treatment sector from the forest sector, two adjustments

are required: OS w should be replaced by OS w – V, and IC f-w should be replaced by IC f-w + V

By the same logic, operating surplus in the forest sector should be adjusted to OS f + V

Because operating surplus is a component of value added, the offsetting adjustments to operating surplus for the two sectors imply that aggregate value added does not change To balance the

accounts, the addition of V to IC f-w requires offsetting adjustments for the forest sector in the supply table: specifically, the purification service should be added as a secondary output to the

table (i.e., a row should be inserted), with V recorded in the cell for the forest sector Total output

of the forest sector, TO f , should then be increased by the same amount, thus becoming TO f + V

The accounts thus remain in balance

The next section begins by presenting the model we estimated and discussing associated statistical issues, and it then describes the definitions of the variables in the model and data sources The subsequent section presents our econometric results and illustrates how they could

be used to estimate the amount required to adjust the state product account for Perak (V in the

previous two paragraphs) The final section of the paper recaps our main findings

2 Materials and methods

2.1 Econometric approach

Our econometric analysis aimed at identifying, in a causal sense, the mean marginal

effect of forests on water treatment cost As mentioned in the introduction, in addition to

analyzing cost we also analyzed quantities of the three major chemicals used by WTPs in

Malaysia: alum, lime, and chlorine The latter analyses helped us identify parts of the water treatment process that are affected by the quality of raw water treated by WTPs, which in turn helped us explain the effects of forests on treatment costs

We estimated the following fixed-effects regression model separately4 for each dependent variable (cost, alum, lime, chlorine) (Wooldridge, 2002, Ch 10–11):

y it is treatment cost or chemical use for WTP i in time period t, which is a given month m of a

given year y L iy is a matrix of land-use variables, which varied by year but not months within a

year q it and r it are treated water volume and rainfall; they varied by both year and month Land

use and rainfall refer to the WTP’s catchment β , α, and γ are parameters to be estimated, and u it

is the error term We assumed that the variance of u it might vary across observations, and we used robust standard errors to correct for any such heteroskedasticity (Huber, 1967; White, 1980)

The panel structure of our data enabled us to include c i , θ y , and θ m as fixed effects that controlled, respectively, for time-invariant WTP characteristics (e.g., catchment area,

topography, soils, geology), WTP-invariant annual characteristics (e.g., water treatment and water management policies, labor market conditions, chemical prices, land-use survey methods), and WTP-invariant monthly characteristics (e.g., seasonality) Fixed effects were more

appropriate than random effects because the sample included nearly all WTPs in Perak, not a random selection of them (Wooldridge, 2002, p 250–252; Kennedy, 2008, p 291)

4 By the time of the World Congress, we expect to have estimated the four equations as a system, using seemingly unrelated regression

When treatment cost is the dependent variable, eq (1) can be viewed as a WTP’s cost function.5 The theory of using cost functions to value environmental inputs is well-understood (McConnell and Bockstael, 2005) Cost functions for firms that use unpriced environmental inputs include four types of variables: (i) the firm’s output level, (ii) prices paid by the firm for treatment chemicals and other market inputs, (iii) the quantity of capital and other fixed factors used by the firm, and (iv) the quantity of environmental inputs used by it (Vincent, 2011) In eq

(1), variable type (i) is represented by q, and variable type (iv) is assumed to be correlated with

the land-use variables in L Rainfall (r) is another environmental input that potentially affects

cost; we included it as a control (see section 2.2), but it is not our environmental input of interest

Variable type (iii) is represented by the WTP fixed effects, c i Finally, our contacts in the Perak Water Authority reported that input prices did not vary across WTPs, and so the year fixed

effects, θ y , can be viewed as representing variable type (ii), with the month dummies (θ y)

controlling for any seasonal variation in input prices that occurs in a typical year

An important conceptual point is that cost savings typically underestimate the benefits of environmental improvements to firms that use environmental inputs (Vincent, 2011) This is because the increased supply of environmental improvements induces firms to increase output, but output is held constant in a cost function Our estimates of cost savings by Perak WTPs that result from enhanced water purification services can thus be expected to underestimate the actual total value of those services to the WTPs

An individual variable L in L gave the percentage of the surface area of a catchment that

was in a particular land use in a particular year Because the percentages summed to 100% across all land uses in each year, one of the land-use variables had to be excluded during estimation

5 Similarly, when the dependent variable is the quantity of a treatment chemical, then eq (1) can be viewed as a conditional input demand function See Huang and Smith (1998) for theory on using input demand functions to value environmental inputs, which is not a path we follow here due to lack of information on input prices

The excluded variable became the reference land use against which the effects of the others were defined If the excluded variable was the aggregate percentage in nonforest uses, then the

parameter estimate ˆ on a particular forest variable in L indicated the effect of avoiding

converting that type of forest to an average nonforest use Conversely, if the excluded variable was instead the aggregate percentage in forest, then the parameter estimate on a particular

nonforest variable in L indicated the effect of converting an average forest to that type of

nonforest use

In either case, ˆ represents a relative measure of the mean marginal effect The mean

marginal effect of L on y is given by differentiating expected treatment cost from eq (1) with respect to L,

m y i m

y i

c r q y L

c r q y

,,,,E

This implies that ˆ can be interpreted as a semielasticity (Wooldridge, 2002, pp 15–18): if 1%

of the area of a catchment changes from the reference land use to the land use represented by L,

then expected cost or chemical use changes by 100ˆ% An elasticity can be obtained by multiplying ˆ by the mean of L: if L increases by 1%, then the expected cost or chemical use

changes by  ˆL %

A statistical analysis of the relationship between treatment cost and land use could have objectives other than identifying mean marginal effects, such as forecasting, investigating

hydrological processes, or developing a parsimonious model for explaining the variation in cost

We would have needed to use different modeling approaches if these were our objectives If the objective were to develop a forecasting model, then we would have needed to investigate several

additional issues, including the out-of-sample predictive accuracy of the model for forecasts of varying lengths and the possibility that future deforestation might result from land uses that were not well-represented in our dataset (e.g., new crops)

If the objective were to investigate hydrological processes, then we would have needed to rely less heavily on fixed effects as controls for potentially confounding factors While fixed effects serve this purpose well from a purely statistical standpoint, they provide no information

on the identity of those factors A model for investigating hydrological processes would include such factors directly as covariates, instead of using fixed effects to sweep away their effects

If the objective were to develop a parsimonious model for explaining variation, then we would have needed to use a goodness-of-fit statistic, such as the Akaike or Bayes-Schwarz

information criterion (AIC, BIC), to guide model specification We did not do this, for two

reasons First, the AIC and BIC should not be used when data have a clustered structure (Hilbe

2011, p 69), which was the case with our data, as we will explain in section 2.2 The second reason was more fundamental Given our objective of identifying the causal effect of land use on treatment cost, our overriding concern was to minimize omitted-variables bias: the risk that the land-use variables could be proxying for factors omitted from the model We therefore erred on the side of including a large number of controls, primarily through the three groups of fixed

effects but also through q and r This approach can cause overfitting in small samples, but the

number of degrees of freedom in our models was large both absolutely and relative to the

number of observations: the smallest number of degrees of freedom in any of our models was nearly 1,400 (more than 95% of the observations in the sample).6

6

One type of overfitting in panel models occurs when fixed effects for cross-sectional units are included at a level that is more finely disaggregated than necessary for identifying the effect of interest; it causes standard errors to be underestimated but does not affect the consistency of parameter estimates (Ritschl, 2009) In our models, this type of overfitting would occur if WTPs could be classified into homogeneous groups, in which case fixed effects should be

Inclusion of a large number of controls can also cause multicollinearity, but given our objective this was a lesser evil than omitted-variables bias Multicollinearity inflates standard errors, but it does not bias parameter estimates (Kennedy, 2002, p 193–194) Hence, it

contributes to a conservative estimation strategy, in that it reduces the risk of overestimating parameter significance

2.2 Definitions of variables and data sources

The regression sample spanned 1994–2005, with 1996, 1997, 1999, and 2000 omitted These gaps were determined by data availability Observations were also missing for some months of some years for some WTPs, mainly due to incomplete data on cost or chemicals

We obtained monthly data on cost, water volume, and chemical use (disaggregated by alum, lime, and chlorine) for 44–46 WTPs (the number increased over time) from spreadsheets provided by the Perak water authority Cost referred to total operating cost: the sum of wages and benefits, treatment chemicals, and power (mostly electricity, but fuel in some cases) It was expressed in the Malaysian currency, the ringgit; we used the Malaysian GDP deflator to convert

to 2005 price levels Water volume referred to production of treated water and was expressed in cubic meters Chemical use was expressed in kilograms The final sample included 33 WTPs (Fig 1), whose catchments ranged in area from 44 ha to 146,190 ha, with a mean of 16,889 ha (standard deviation = 28,773) The other 13 WTPs were excluded because their cost data were too incomplete or the locations of their water intakes could not be verified, which was necessary for determining their catchments The number of observations in the regression models ranged from 1,431 to 1,945

included for groups instead of individual WTPs The WTPs do not fall into obvious groups, however, as their catchments differ greatly in area, aspect and form (Fig 1), topography, and other fixed characteristics Moreover, this type of overfitting occurs only if the number of fixed effects is large relative to the number of observations, which was not the case in our models (33 WTP fixed effects vs more than 1,400 observations in all models)

We had two sources of land-use data, and we estimated complete sets of models for each one The Forestry Department Peninsular Malaysia has conducted National7 Forest Inventories (NFIs) approximately every ten years since 1971–72 It granted us restricted access to GIS layers from the 1992–93 and 2004–5 inventories This enabled us to measure the area of each

catchment in two types of forest—virgin (unlogged) forest, and logged forest—with a third, residual category of all types of nonforest uses The forests in the catchments are tropical

rainforests located mostly on hilly or mountainous terrain The coastal plains of the Peninsula were once covered by lowland dipterocarp rainforests, but most of this forest type had already been converted to nonforest land uses by the 1980s (Vincent and Hadi, 1993; Vincent and

Mohamed Ali, 2007)

We interpolated the areas of the two forest types for intervening years between the NFI dates, and we expressed them as percentages of a catchment’s area The detail on forest type enabled us to estimate a pair of models that revealed the marginal effect of avoided deforestation

on cost and chemicals in a progressively more detailed manner The first model included a single, aggregate forest variable It provided an estimate of the marginal effect averaged across the two forest types and all nonforest land uses to which forests could have been converted The second model included the two forest types as separate variables It provided estimates of

marginal effects differentiated by the type of forest where deforestation was avoided, but still averaged across all types of nonforest land uses to which they could have been converted

Comparing the results of these two models sheds light on the importance of accounting for logging status when valuing the water purification service provided by forests

In parallel, the Malaysian Department of Agriculture has conducted mid-decadal land-use surveys (LUSs) since 1966–67 (Wong, 1971) While the NFIs provide detail on areas covered by

7 Though named “national,” the inventories do not include information on forests in Sabah or Sarawak

forest but no detail on nonforest areas, the LUSs do the oppositive The department granted us restricted access to GIS layers from the 1984–85, 1997–98, and 2004–5 LUSs This enabled us to measure the area of each catchment in four nonforest land uses—oil palm, rubber, urban,8 and other—with a fifth, residual category of all types of forest (a mix of virgin and logged, which the LUSs do not distinguish) As with the NFIs, we interpolated the areas of the four nonforest land uses for intervening years, and we expressed them as percentages of a catchment’s area Given that we had data from three LUSs but only two NFIs, the interpolated variables from the LUSs are probably more highly correlated with actual land-use trends than the interpolated variables from the NFIs On the other hand, the latter provide more detail on the logging status of forests

Also as with the NFIs, the detail on nonforest land uses from the LUSs enabled us to estimate two models The first one corresponded to the first model in the NFI analysis: a model that included a single, aggregate forest variable The second model included the four nonforest land uses as separate variables It provided estimates of marginal effects—now for deforestation, not avoided deforestation—differentiated by the type of land use to which forests were

converted, but averaged across all types of forest These estimates shed light on the importance

of accounting for the type of use to which forests are converted when valuing losses in water purification services associated with deforestation

Table 1 provides summary information on the land-use variables in the regression

models.9 As can be seen, the WTPs’ catchments varied greatly in terms of both area and land use On average, forests accounted for a majority of the surface area of the catchments, but most

of the forests were logged, not virgin Nonforest uses increased over time, with the principal

trend being an expansion of oil palm area and a retraction of rubber area This is consistent with trends in other parts of the Peninsula (Vincent and Hadi, 1993; Vincent and Mohamed Ali, 2007)

The LUSs report a higher forest percentage than the NFIs, which is due to the former including two types of land use excluded from the latter.10 One type is scrubland, which is

heavily degraded land that has woody vegetation but not enough to be considered forest by the Peninsular Malaysia Forestry Department Scrubland is typically land that is recovering through natural succession from either shifting cultivation or partial land-clearing for abandoned

agricultural conversion projects The second type is swamps and marshland, which can in

principle include forests (e.g., mangroves and peat-swamp forests) but in the catchments for WTPs, which are located in interior regions, are instead abandoned tin-mining ponds Despite these discrepancies, the aggregate forest percentages from the NFI and the LUS in 2004–5, which is the only year in the sample with estimates from both sources, were highly correlated

so it follows that interpolation causes our estimates of the effects of land use on treatment cost and chemical use to be conservative (i.e., underestimated) The attenuation bias is likely greater

10 By the time of the World Congress, we expect to have reestimated the models using a definition of the forest variable from the LUSs that more closely matches the definition used in the NFIs

11

This correlation uses a single annual observation for each WTP, not 12 monthly ones

12 Similarly, given that the Forest Department Peninsular Malaysia and the Department of Agriculture conduct the NFIs and LUSs using the same procedures in all parts of the Peninsula, there is no apparent reason to expect

measurement error in the NFIs or LUSs to be nonrandom across catchments or land uses

in models that used the NFI data, which were interpolated from just two point estimates instead

of three as in the case of the LUS data

The annual frequency of the interpolated land-use variables differs from the monthly frequency of the cost and chemicals variables Conventional robust standard errors can severely underestimate true standard errors when a covariate varies at a lower frequency than the

dependent variable (Moulton, 1986) Unlike the underestimation of parameters, this is not a conservative bias, as it exaggerates the significance of the covariate’s effect on the dependent variable This problem can be addressed by clustering standard errors at the lower frequency, which in our case meant clustering them by year for each WTP Clustering is an asymptotic correction that requires a large number of clusters, with 40–50 clusters being the rule-of-thumb (Angrist and Pischke, 2009, Ch 8) The number of clusters was more than 100 in all of our models Clustering also corrected for serial correlation in the errors between the months within a given year (Zeger and Liang, 1986).13

Although fixed effects control for many factors that could confound the marginal effect

of forests on water quality, they obviously do not control for all of them In particular, they do not control for factors that vary over time in different ways across WTPs The most obvious such factor is treated water volume, which is expected to have a positive effect on both cost and chemical use Water volume in the sample ranged from 2,272 m3/mo to 9,419,862 m3/mo, with a mean of 622,378 m3/mo (standard deviation = 1,091,939 m3/mo) Fig 2 shows a scatterplot of cost against water volume for the observations in the sample As expected, the plot shows a positive relationship

A second factor is rainfall, which affects water quality, and thus cost and chemical use,

13 A possible problem with the error structure yet to be addressed is the correlation of errors between WTPs with nested catchments (e.g., a WTP upstream from another WTP on the same river) We will try to investigate this by the time of the World Congress

through its effects on soil erosion and runoff from different land uses For this reason, controlling for rainfall helps explain variation and thus makes our estimates of land-use effects more precise (smaller standard errors) Controlling for rainfall is also important because rainfall might affect deforestation, for example by impeding the burning of woody debris when forests are cleared Omitting rainfall from the regression models could thus lead to biased estimates of the effect of land use on cost and chemical use

To construct the rainfall variable, we first divided Perak into four zones: a western

coastal zone, a southeastern interior hilly zone, and two zones for the northern and southern portions of the Perak River basin, which accounts for most of the state’s area We downloaded monthly data for all available rainfall stations in these zones operated by the Malaysian water authority, the Drainage and Irrigation Department (2010), and we obtained data for several additional stations directly from the department We calculated the simple average of the

readings across the stations for a given month of a given year in a given zone, and we used these zonal rainfall values as the rainfall variables for the WTPs located in the zones Rainfall in the sample ranged from 9 mm/month to 570 mm/month, with a mean of 193 mm/month (standard deviation = 102 mm/month)

With cost, chemicals, water volume, and rainfall all being log-transformed in the models, the parameter estimates on water volume and rainfall are directly interpretable as elasticities

3 Results

3.1 Effects of land use on water treatment cost

Table 2 presents results from the fixed-effects models for water treatment cost In the models with land-use variables from the NFIs, the aggregate forest variable shows weak

evidence of a negative effect (P = 0.066) Disaggregating this variable into its two constituent forest types reveals a much more significant effect of virgin forest (P = 0.0467) than logged forest (P = 0.137) This suggests that, compared to the average nonforest land use, virgin forest

provides a water purification service that significantly reduces water treatment cost but logged forest does not The fact that the parameter estimate on logged forest (-0.0130) is not much smaller than the estimate on virgin forest (-0.0168) leaves open the possibility, however, that logged forest does in fact provide this service but measurement error, insufficient variation, or multicollinearity prevented us from estimating it very precisely The estimate on virgin forest indicates that converting 1% of a WTP’s catchment from virgin forest to the average nonforest use increased cost by 1.68% This is very similar to Ernst’s (2004) the rule-of-thumb, which implies a 1%:2% relationship Expressed as an elasticity, increasing virgin forest area by 1% reduced cost by 0.48%

In contrast, neither of the models with land-use variables from the LUSs shows evidence

of a significant effect of land use on cost A possible explanation for this difference compared to the models just discussed is the inclusion of scrubland and swamps in the forest variable from the LUSs This variable is thus a blend of forest and nonforest land uses This can be expected to increase the difficulty of detecting differences between its effect and the effects of either the aggregate nonforest variable that the first model excludes or the disaggregated nonforest

variables that the second model includes

As expected, costs are increasing in treated water volume in all four models, with a

positive and highly significant (P < 0.01) elasticity of 0.24–0.27 The elasticity being below

unity implies increasing returns to scale in water treatment (i.e., declining average cost), which is expected and has been reported by prior econometric studies on water treatment costs (e.g.,

Forster, Bardos, and Southgate, 1987; Moore and McCarl, 1987; Holmes, 1988; Dearmont, McCarl, and Tolman, 1998) None of the four models shows evidence of a significant effect of rainfall on cost At least four explanations are possible: rainfall had no effect; our zonal rainfall variables were not sufficiently precise to identify the effect; the most important effects were due

to differences in rainfall across years and between months within a year, and the fixed effects for years and months fully absorbed these effects; and rainfall’s effect occurred through an

interaction with land use, which our models excluded.14

3.2 Effects of land use on use of water treatment chemicals

Tables 3–5 show corresponding results for chemical use WTPs add alum to remove suspended and dissolved solids from raw water (i.e., to reduce turbidity) WTPs in Perak use a trimer current detector to measure turbidity The detector displays a higher positive charge when raw water is more turbid, and the WTPs accordingly adjust the amount of alum added.15 The pattern of significance for the NFI-based and LUS-based models is the opposite of that for cost, with the LUS-based models now being the ones that exhibit significant land-use effects Results for these models indicate that alum use was decreasing in forest area and increasing in urban use,

rubber, and oil palm, with significance levels of P < 0.01 for all four variables Among the

nonforest uses, the largest effect was for urban use, a semielasticity of 35.4% The

semielasticities for rubber and oil palm were similar to each other, which makes sense given that they are both tree crops grown in broadly similar locations (i.e., lowlands), but their effects were only about half as large as for urban use (semielasticities of 16.6% and 17.1%, respectively) Overall, these results imply that reduced use of alum is one source of the cost savings enjoyed by WTPs that process raw water from catchments with less forest conversion to rubber, oil palm,

14 We intend to investigate interaction effects by the time of the World Congress

15 Information on use of alum, lime, and chlorine by WTPs in Perak was obtained from the Perak Water Board (Lembaga Air Perak) on February 10, 2014