Journal of Health Economics 42 (2015) 151–164

Contents lists available at ScienceDirect

Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase

Health and agricultural productivity: Evidence from Zambia Günther Fink a,∗ , Felix Masiye b a b

Harvard School of Public Health, USA Department of Economics, University of Zambia, Zambia

a r t i c l e

i n f o

Article history: Received 19 August 2014 Received in revised form 12 April 2015 Accepted 21 April 2015 Available online 28 April 2015 JEL classification: I15 J24 J43

a b s t r a c t We evaluate the productivity effects of investment in preventive health technology through a randomized controlled trial in rural Zambia. In the experiment, access to subsidized bed nets was randomly assigned at the community level; 516 farmers were followed over a one-year farming period. We find large positive effects of preventative health investment on productivity: among farmers provided with access to free nets, harvest value increased by US$ 76, corresponding to about 14.7% of the average output value. While only limited information was collected on farming inputs, shifts in the extensive and the intensive margins of labor supply appear to be the most likely mechanism underlying the productivity improvements observed. © 2015 Elsevier B.V. All rights reserved.

Keywords: Investment Health Productivity Agriculture Malaria

1. Introduction Despite the rapid speed of urbanization over the past decades, rural small-scale farming remains the primary source of food and income for a majority of the population in developing countries (World Bank, 2007). In most settings, the degree of agricultural mechanization is limited, so that agricultural production remains primarily dependent on the availability and productivity of human labor. While labor is abundant in principle in most developing countries (Pitt and Rosenzweig, 1986), labor inputs can be compromised by episodes of ill health and can result in output losses if absent labor cannot be replaced immediately. In this paper we investigate the economic impact of short-term morbidity on agricultural output in the context of small-scale farming in Zambia. The study setting is representative of many rural areas in the developing world both in terms of the general lack of advanced farming technology and in terms of the dominant role of farming as source of nutrition and income. With farming land available free of charge in most communities, a large majority of the working-age population engages in agriculture, while formal

∗ Corresponding author at: Harvard School of Public Health, Boston, MA 02115, USA. Tel.: +1 6174327389. E-mail address: gfi[email protected] (G. Fink). http://dx.doi.org/10.1016/j.jhealeco.2015.04.004 0167-6296/© 2015 Elsevier B.V. All rights reserved.

sector jobs are generally scarce. Despite major government efforts to reduce the burden of the disease in recent years (NMCC, 2010; Zambia Ministry of Health, 2006), malaria continues to be the primary cause of short-term morbidity in the country, with children and adults experiencing up to five episodes of malaria per year (NMCC, 2010; WHO, 2009). Since the planting season tends to overlap with the malaria season, health related absences from field work are frequent, and are commonly cited by local farmers as primary cause of lost field work and income.1 To evaluate the degree to which health affects agricultural productivity, we conducted a cluster-randomized field experiment with 516 farmers in Katete District, Zambia, from December 2009 to August 2010. As part of the experiment, farmers were randomly selected for bed net programs, which allowed them to obtain long-lasting insecticide treated nets (LLITNs) through agricultural loan program schemes at differentially subsidized prices. The basic intuition underlying the experiment is relatively straightforward: as long as household labor and consumption decisions are nonseparable from household production decisions2 (Benjamin, 1992),

1 On average, farmers surveyed at baseline claimed that their harvest would increase by 30% if field work was not interrupted by episodes of ill health. 2 If consumption and production decisions were perfectly separable, family labor could be perfectly substituted for by hired labor.

152

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

decreased exposure to malaria should increase the time and energy farmers can spend on their fields, and thus also increase the final harvest amounts. In a first paper based on this experiment, we analyzed the impact of the additional LLITNs distributed on self-reported morbidity (Fink and Masiye, 2012). In this paper, we analyze the impact of the net programs on agricultural productivity, the main outcome variable of the trial. In the first part of our analysis, we analyze the impact of the interventions on net ownership and usage. Consistent with recent work by Tarozzi et al. (2014), we find a substantial fraction of farmers to be willing to purchase LLITNs at full or partially subsidized prices when financing options are provided. On average, farmers in the loan group acquired 0.9 nets, resulting in a 24% point increase in the average fraction of sleeping spaces covered at the household level. In the second part of the paper, we estimate the impact of the bed net programs on agricultural production. In order to facilitate a rapid distribution of bed nets, treatments were randomly assigned at the cluster level prior to the collection of baseline data in the experiment. The non-stratified cluster-level randomization resulted in a rather unbalanced sample, with treated farmers on average both larger and more productive than farmers in the control group. To address these imbalances, we focus on analyzing changes in production outcomes between the 2009 (preintervention) and the 2010 (post-intervention) farming seasons. The point estimates from our preferred specification suggest that the returns to bed nets in the study sample were large: on average, we find that access to free bed nets (three nets for a typical household) increased agricultural output by US$ 76, which corresponds to 14.7% of the average annual harvest value. To address omitted variable bias concerns, we include a large set of covariates in our empirical models, and run an extensive series of robustness and heterogeneity checks. Overall, treatment effects appear largest among more educated farmers as well as farms with more diversified portfolios, and larger for cotton (as the more labor intensive crop) than for maize. In the last part of our analysis, we explore potential mechanisms underlying the productivity impacts observed. Unfortunately only limited and self-reported data on malaria incidence (and no data on parasitemia or asymptomatic malaria) was collected as part of this project. However, the general patterns observed in the data suggest that the programs likely induced substantial reductions in the days of field work lost due to ill health. Given that full recovery from acute malaria is often slow, reduced exposure to malaria can increase the marginal product of labor (Nur, 1993), particularly in cases where malaria induces anemia (Ehrhardt et al., 2006). While there is theoretically also the possibility that the reduced exposure to ill health may have been associated with a reduction in direct medical expenditure, most malaria treatment in the area appears to be provided for free, so that no evidence of lower health expenditure was found. Even though this paper is to our knowledge the first one using experimental data to evaluate the productivity effects of malaria, several studies have analyzed agricultural output in the context of nutrition and other diseases. Following the initial work by Strauss (1986) as well as Pitt and Rosenzweig (1986), Behrman et al. (1997) document a rather robust association between nutritional improvements and production in agricultural settings. Loureiro (2009) and Ulimwengu (2009) find positive associations between health and productivity using stochastic frontier regression techniques. Audibert and Etard (2003) examine the effect of schistosomiasis among rice-growers, and find that exposure to schistosomiasis reduces production by 26%. Fox et al. (2004) analyze the productivity declines associated with HIV positivity, and find that HIV-positive workers earn on average 16–17% less over a two year period. Similarly, Baranov et al. (2012) show that maize

production increases by up to 31% with HIV treatment, and attribute this increase to increased overall labor supply and improved physical and mental health. Similar effects were, however, not found for iron supplementation and deworming among tea pluckers in Bangladesh (Gilgen et al., 2001). Most similar to the results presented in this paper are two cross-sectional studies using harvest data to compute the agricultural output effect of malaria: Girardin et al. (2004) analyze vegetable farming in Côte d’Ivoire, and find that farmers who reported being sick more often had 47% lower yields. Morel et al. (2008) use total farming output to quantify the agricultural loss generated by work days lost due to malaria in Vietnam, and find an average cost of US$ 11 per case of malaria, suggesting returns to malaria prevention similar to the ones identified in this paper. Conceptually, an overwhelming majority of this literature suggests strong links between health and agricultural production in low-income setting; this suggests that household production, labor and consumption decisions are generally not separable (Benjamin, 1992), a finding which is also supported by recent evidence from Zambia (Fink et al., 2014). While this study primarily focuses on household-level outcomes, the results presented here naturally also link to the broader literature on the relation between health and income. Most of the micro-level literature in this area has focused on the long term benefits of improved childhood health in terms of education and labor market outcomes (Bleakley, 2007; Bleakley and Lange, 2009; Clarke et al., 2008; Kremer and Miguel, 2004). This paper highlights a more immediate and direct effect of health on income similar to the results shown in Thomas et al. (2010) for iron supplementation; this effect will clearly not apply in all low resource settings, but may be of particular importance among rural and frequently impoverished populations. The rest of the paper is structured as follows: we provide a detailed description of the study site and local agriculture practices in Section 2. In Section 3, we present the study design and provide details on study implementation. In Section 4, we analyze the effects of the bed net programs on net ownership and net usage. In Section 5, we estimate the impact of the net programs on productivity. Section 6 shows some evidence on the mechanisms underlying the main productivity results. We conclude with a short summary and discussion in Section 7.

2. Study background Fig. 1 shows the geographic location of the study site within Zambia. Katete district is one of eight districts within Zambia’s Eastern Province. Eastern Province is one of the least developed regions of Zambia, with a majority of the population living below the onedollar-per-day poverty line, and an estimated under-5 mortality rate of 151 per 1000 live births (Macro International, 2007). Katete district is similar in its topography to the Western part of Malawi, which is located about 100 km east of the district. The current district population is estimated at 250,000, approximately half of which live in the urban centers of Sinda and Katete (Zambia Central Statistic Office, 2011a). Malaria is endemic in most parts of Zambia, and the primary cause of short term morbidity in the country (Zambia Ministry of Health, 2012). The regional climate displays pronounced seasonal fluctuations, with virtually no rainfall from May to November, followed by a period of major rainfall from December to April. The strong seasonal patterns are directly reflected in the seasonal fluctuations of malaria. Malaria in the area is considered endemic and seasonal, with a majority of the transmission occurring between December and May, when continued rainfalls support the breeding of the Anopheles mosquito larvae. According to the latest round of the Malaria Indicator Survey, Eastern region is among the areas

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

153

Fig. 1. Zambia (white) and Katete District (shaded red). (For interpretation of the references to color in this text, the reader is referred to the web version of the article.)

with the highest parasite prevalence in the country, with parasites detected in 22 percent of children under the age of five in early 2010 (NMCC, 2011). Fig. 2 shows the clinical burden of Plasmodium falciparum malaria in terms of the number of clinical cases per year

and per 5 km2 as computed by the Malaria Atlas Project in 2007 (Hay and Snow, 2006). Katete district ranks among the most highly exposed areas of the country, with an estimated annual burden of approximately 500 cases per 5 km2 . Since 2006, Zambia has made major efforts to reduce the burden of malaria (Ashraf et al., 2010; Zambia Ministry of Health, 2006). As part of the internationally supported Rollback Malaria Initiative, four principal strategies have been employed by the country through the National Malaria Control Centre: indoor-residual spraying (for densely populated and primarily urban areas), mass distribution of long-lasting insecticide treated nets (LLITNs), intermittent preventive treatment of malaria in pregnancy (IPTp) and case management through diagnostics and artemisinin-based combination therapies (NMCC, 2007, 2009, 2010; Zambia Ministry of Health, 2006). Between 2006 and 2008, 96,000 LLITNs were distributed in Katete district (NMCC, 2011). At the time of the 2008 Malaria Indicator Survey, the average number of LLITNs in Eastern Province was 0.96 nets (NMCC, 2009), a level very similar to the one observed in the study area at the beginning of the study in December 2009. The rural part of Katete targeted in this study is sparsely populated, with clusters of family-run farms grouped into small villages. With an average size of approximately 4 ha (10 acres), the typical farm is small, and most planting and harvesting done without machinery. Farm land is generally owned by communities, who allocate the land to families via local headmen and chiefs. The amount of farm land families can get access to is – at least theoretically – not limited; any individual can claim additional land from the chief as long as they can show they have the manpower and skills to use the land (Nolte, 2012). 3. Study design, randomization and descriptive statistics 3.1. Intervention background

Fig. 2. Regional distribution of malaria incidence. Source: Malaria Atlas Project.

As stated above, small-scale farming constitutes the primary source of nutrition and income in the region. Cotton is the main cash crops in the area, and sold to multinational cotton buyers (“ginners”), who process the cotton and sell it on international markets.

154

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Fig. 3. Cluster-level group assignment.

To enhance productivity and strengthen commercial links with farmers, cotton ginners have set up a variety of agricultural loan schemes, which allow farmers to receive cotton seeds and “chemicals” (fertilizer and pesticides) as well as agricultural machinery throughout the planting and growing season on a loan basis. Upon receipt of the cotton, ginners deduct the outstanding loan amount from the final sales price, and pay out the remaining balance in cash. Agricultural loans are generally provided free of interest, and are given under the assumption that farmers will sell their harvest to the cotton company offering them the agricultural loan. At the time of the study, our partner organization (Dunavant Cotton) was providing loans to approximately 80,000 cotton farmers across Zambia. According to the latest national estimates, 67% of the Zambia labor force is employed in agriculture (Zambia Central Statistic Office, 2011b), which corresponds to approximately 1.3 million farming households. The share of farmers working with Dunavant is relatively small (approximately 6% of all farming households), both because Dunavant is only one of about 10 cotton ginners in the country and because only about 25% of small-scale farmers in the country grow cotton (Goeb, 2011). 3.2. The intervention In order to assess farm’s willingness to pay for nets within existing agricultural loan schemes, two different net programs were implemented as part of the experiment: a free net program and a bed net loan program. Under the free net program, selected farmers were allowed to obtain one free bed net for each uncovered sleeping space in the household. Clusters in the loan arm were assigned to one of two loan types: a “full price loan”, which required farmers to repay the full price3 of the net (Zambian Kwacha (ZK) 25,000, US$ 5) at the end of the harvesting season,

3 “full price” we charged reflects the current wholesale price, which is about 30% below regional retail prices of about ZK 35,000 (US$ 7).

and a “subsidized price loan”, which required farmers to repay ZK 12,500 (US$ 2.5) at the end of the harvesting season. Both prices are substantially higher than what households reported to have paid for bed nets in the past; 90% of bed nets found in households at baseline were received through free governmental or NGO programs (most likely received as part of the large national programs); 7% had paid ZK 3000 (US$ 0.6) – the price public health facilities charged prior to the national mass distribution, and the remaining 3% reported to have paid a price between ZK 5000 and ZK 10,000 acquiring nets through door-to-door sales. All nets were distributed at the beginning of the weeding season (late December) in order to provide protection for farmers during the peak malaria season (December–May). 3.3. Sampling frame, enrollment and program assignment The sampling frame for the study was provided by Dunavant Zambia. Dunavant has nine regional offices, which distribute farming inputs and acquire cotton through local sheds and distributors. Each distributor handles between 10 and 50 farmers in his community. At the time the study was launched, Dunavant was working with 96 distributors in the study area. In order to reduce the risk of local spillovers, we restricted the sample to distributors operating in spatially separated areas, with a minimum distance of 3 km between any two locations. This left us with a final sample of 49 distributors and their respective villages. On average, each of the 49 distributors was working with about 20 farmers, a listing of which was provided to the study by Dunavant. We randomly selected 11 farmers from each distributor, and visited them for a baseline interview in December 2009. Only the 11 farmers selected for the study were allowed to receive the free or subsidized nets; with an average village size of about 50 households, this means that the program covered about 20% of the average village population. Out of 539 farmers invited to participate in the study, 516 farmers (95.7%) were enrolled in the study, and completed the baseline interview in December 2009.

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

155

Table 1 Descriptive statistics.

Farmer age Farmer is married Farmer years of education Members under age 5 Members age 5–14 Members age 15–59 Members age 60+ Chicken, geese and ducks Goats, pigs and sheep Cows Bicycles Mobiles TVs Mosquito nets Cars, tractors and trucks Maize planting area (ha) Cotton planting area (ha) Other crops (ha) Cotton harvest 2009 (bales) Maize harvest 2009 (bags) Total harvest value 2009 (US$)

Control (N = 153)

Loans (N = 185)

Free nets (N = 155)

Equal means test (p-value)a

Mean

St. dev.

Mean

St. dev.

Mean

St. dev.

Control vs. Loans

Control vs. Free nets

Loan vs. Free nets

All Means equal

39.24 0.83 4.34 0.87 1.40 2.44 0.18 4.58 2.57 1.47 0.77 0.23 0.07 1.09 0.00 1.78 1.36 0.58 9.46 20.13 453.2

12.59 0.38 3.41 0.88 1.36 1.22 0.49 6.20 3.40 3.51 0.56 0.45 0.25 1.33 0.00 1.17 1.09 0.77 8.59 15.62 304.0

41.32 0.83 4.14 0.95 1.88 2.76 0.22 6.89 3.41 2.34 0.93 0.36 0.14 1.53 0.06 2.16 1.38 0.73 12.16 29.83 649.0

13.24 0.37 3.66 1.00 1.50 1.53 0.51 8.45 4.18 4.09 0.81 0.80 0.37 1.25 0.43 3.16 1.25 0.84 11.12 35.96 589.1

39.70 0.88 4.21 0.86 1.92 2.84 0.17 6.26 3.00 1.92 0.87 0.35 0.06 0.48 0.03 1.86 1.17 0.83 11.60 28.16 613.8

12.90 0.32 3.88 0.86 1.55 1.58 0.44 7.96 3.68 3.24 0.60 0.75 0.27 0.69 0.20 1.13 0.73 0.93 11.33 25.98 476.0

0.27 0.96 0.63 0.55 0.00 0.09 0.54 0.01 0.12 0.07 0.02 0.14 0.04 0.08 0.05 0.19 0.94 0.32 0.09 0.01 0.00

0.80 0.21 0.81 0.93 0.00 0.03 0.97 0.07 0.35 0.32 0.22 0.17 0.98 0.01 0.05 0.71 0.21 0.15 0.19 0.00 0.00

0.31 0.15 0.90 0.39 0.82 0.71 0.51 0.49 0.41 0.37 0.44 0.95 0.05 0.00 0.33 0.27 0.32 0.55 0.75 0.68 0.63

0.50 0.30 0.89 0.66 0.00 0.07 0.78 0.01 0.29 0.19 0.08 0.23 0.08 0.00 0.03 0.38 0.33 0.39 0.19 0.00 0.00

Notes: Based on 493 observations with complete information. All variables reflect baseline conditions as collected in December 2009. a p-Values based on cluster-bootstrapped standard errors. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor.

In order to accelerate the distribution of bed nets,4 bed net loan programs were randomized prior to the collection of baseline data at the distributor level. Randomization was done using a simple random number draw generated by Stata. Out of the 49 eligible distributors, 15 were assigned to the control group (30%), and 15 distributors (30%) were selected for the free net program. Since we were particularly interested in the loan group and wanted to assess differences in uptake with and without subsidy, a slightly larger number of distributors (20% of distributors in each loan program) were randomized into the net loan programs. The spatial distribution of treatment assignment is illustrated in Fig. 3. All farms in the net program arms were informed about the programs at the end of the baseline interview, and given 48 h to decide on the number of nets they wanted to receive. Ordered nets were delivered within 10 days of the baseline survey, between December 20 and December 31, 2009. A first follow-up or midline survey was conducted in April 2010, during which information on recent illness episodes was collected. The endline survey was conducted in July and August 2010 with a primary focus on harvest outcomes and harvest sales. Out of the 516 farmers initially enrolled in the study, 510 (98.8%) were followed up successfully throughout the subsequent farming and harvesting season. One farmer passed away, three farmers moved, and two farmers refused to participate in the follow-up surveys. An additional two surveys were excluded from analysis due to missing planting and harvesting information. Sixteen further surveys have missing values on at least one of the extended list of covariates used in some of the specifications, resulting in a final analytical sample of 493 farmers. 3.4. Descriptive statistics Table 1 shows descriptive statistics for the households enrolled and followed up in the study by study arm. Average plot size was

4 Due to delays in funding and IRB approval, baseline surveys got pushed back to December, so that we decided to do the distribution of nets immediately after baseline to make sure farmers would benefit from them during the peak rainy season.

4.15 ha (median 3.1) in 2009, and average harvest value in 2009 was US$ 577 (median US$ 463). With an average household size of close to six members, this implies average per-capita resources of approximately US$ 0.26 per day, placing the majority of these households well below the international US$ 1.25 dollar per day poverty threshold, even when input-related expenses (such as cotton loans) are not accounted for. Cotton farmers are on average slightly larger than other farms;5 at the national level, the average plot size among small- and medium scale (non-commercial) farmers is 3.1 ha (Jayne et al., 2008). On average, farms owned one bed net at baseline; with a mean of three sleeping spaces per household, this implies that two thirds of household members were not covered by nets at the beginning of the study. While the randomized assignment of net programs across clusters generated a fairly balanced sample with respect to household head characteristics, the same was unfortunately not true for farm size, with farms in the free net and net loan arms on average larger and more productive than farms in the control group. As Table 1 shows, the largest and most productive farms were found in the net loan group, followed by the free net group. Detailed data collected on 2009 harvest outcomes suggests that the average value of farm production (sum of all crops harvested multiplied with the median sales prices of the respective crops in the area in 2009 – see Table 1 for further details) was US$ 453 in the control group, US$ 614 in the free net group, and US$ 649 in the net loan group. Similar differences were found for household size: on average, households in the loan and free net groups had 0.8 and 0.9 more household members in the 5–59 age range than households in the control group. These differences between farmers in the control and farmers in the two intervention groups are large and statistically significant, and complicate statistical inference, since endline differences will be at least be partially attributable to differences observed at baseline. The imbalance does not appear to be driven by differences in the spatial distribution or by spatial clustering of farms, but rather

5 Cotton ginners recommend to use at least one hectare of land for growing cotton, which means that farms growing cotton rarely use less than two hectares of land.

156

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Table 2 Program effect on ownership and coverage panel a: ownership and usage of nets. Panel A: ownership and usage of nets

Loan program Free nets Control group average Controls included Observations R-squared

Number of nets received through program

Number of nets owned at endline

Number of nets used at endlinea

Number of nets received through programb

Number of nets owned at endlineb

Number of nets used at endlineb

0.811*** (0.147) 2.413*** (0.121) 0.00

0.992*** (0.199) 1.626*** (0.190) 1.05

0.823*** (0.189) 1.517*** (0.185) 0.96

0.676*** (0.150) 2.052*** (0.113) 0.00

0.710*** (0.168) 1.571*** (0.137) 1.05

0.555*** (0.157) 1.461*** (0.141) 0.96

No

No

No

Yes

Yes

Yes

493 0.46

493 0.23

493 0.20

493 0.64

493 0.45

493 0.46

Fraction of nets used at endlinec

Fraction of sleeping spaces covered

Probability of not having any space covered

Fraction of nets usedb

Fraction of sleeping spaces coveredb

Probability of not having any space coveredb

−0.0479 (0.0318) −0.0176 (0.0310) 0.93

0.243*** (0.0687) 0.438*** (0.0625) 0.41

−0.279*** (0.0700) −0.366*** (0.0654) 0.39

−0.0459 (0.0364) −0.00924 (0.0359) 0.93

0.248*** (0.0667) 0.472*** (0.0587) 0.41

−0.258*** (0.0688) −0.383*** (0.0674) 0.39

No

No

No

Yes

Yes

Yes

419 0.01

492 0.20

493 0.15

419 0.07

492 0.25

493 0.21

Panel B: bed net coverage

Loan program Free nets Control group average Controls included Observations R-squared a

Usage is defined as the number of nets observed hanging by interviewers as part of the endline survey in July 2010. Specifications control for plot sizes used for cotton, maize and other crops in 2009 as well as in 2010, family members under 5, family members 5–14, family members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status, total maize harvest 2009, total cotton harvest 2009 and total economic value of production 2009. c Fraction of nets used not defined for households without a bed net. Cluster-bootstrapped standard errors in parentheses. p-Values based on cluster-bootstrapped standard errors. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor. *** p < 0.01, ** p < 0.05, * p < 0.1. b

seems to reflect presumably random variations in productivity level across farmers.6 While a large number of meta-reviews in the medical literature suggest that results from imbalanced trials controlling for baseline characteristics do on average not yield different results from fully balanced trials (Berger, 2010; Knottnerus and Tugwell, 2012; Riley et al., 2013), there is clearly a concern that observable differences may be correlated with other unobservable characteristics such as malaria knowledge or farming skills. To deal with these concerns, we follow the approach proposed by Glennerster and Takavarasha (2013) as well as by Bennett et al. (2014) and estimate both models where we control for lagged dependent variables and models where we use differences in outcome measures between the 2009 (pre-intervention) and 2010 (intervention) seasons as the dependent variable. These models exclusively identify changes in the outcome measures over time, and thus directly eliminate confounding or omitted variable bias concerns due to time-invariant farm-specific differences prior to the intervention. The resulting estimates will yield unbiased program impact assessments as long as the random treatment assignment is not correlated with changes in unobservable characteristics between baseline and endline conditional on initial values, which seems reasonable given the relatively short time period analyzed.

6

The baseline table looks virtually the same when specific areas (Eastern, Western or central parts) or villages are excluded.

4. Program impact on bed net ownership, usage and coverage Table 2 shows the main results for bed net ownership and usage. As described above, endline surveys were conducted in July and August 2010. During both visits, net status was verified by interviewers. We consider a bed net “used” if the net was observed hanging by interviewers during the visit.7 In Panel A of Table 2, we show the impact of the net programs on the number of nets received (columns 1 and 4), the number of nets owned at endline (columns 2 and 5) and the number of nets in use at endline (columns 3 and 6). In Panel B of Table 2 we show the impact of the two treatments on bed net coverage in the household, i.e. the number of nets hanging relative to the number of sleeping spaces used by the household. In columns 1–3 of both panels we show unadjusted differences between the three groups; in columns 4–6, we show estimates with a full set of baseline covariates to control for pre-treatment differences in household size and bed net ownership. As documented in previous studies (Cohen and Dupas, 2010; Dupas, 2014; Tarozzi et al., 2014), the demand for bed nets is highly price elastic; on average, households in the loan group obtained 0.8 nets, compared to 2.4 nets in the free net group. It is worth

7 Nets are generally tied to a knot during the day to keep them clean, which means that observing a net as hanging does not necessarily mean it was used the previous night.

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

157

highlighting that despite the high price elasticity, demand is strictly positive in this setting with credit financing even when the full price of the net is charged. This is rather different from the zero demand found for nets with prices over US$ 1 in settings where net acquisition requires an upfront cash payment (Cohen and Dupas, 2010; Dupas, 2014), but similar in magnitude to recent work by Tarozzi et al. (2014) who find that 52% of household purchase bed nets when financing options are provided. Given that households in the control group had more nets at baseline, the differences in ownership at endline are smaller than the differences in the number of nets received; on average, households in the loan group owned 0.9 more nets than households in the control group, while households in the free net group owned 1.6 nets more than households in the control group. Consistent with Cohen and Dupas (2010) as well as Tarozzi et al. (2014), we found no effect of net pricing on usage. On average, 90% of nets owned were actively used by the household during our second follow-up, with no differences in utilization rates between intervention and control groups. The remaining nets were generally found stored for future use in the households, which is consistent with the generally high levels of appreciation of nets in this population. In terms of sleeping space coverage, both treatments had a sizeable impact. As shown in Panel B of Table 2, 41% of sleeping spaces were covered on average at endline in the control group; this fraction increased to 65% in the loan group, and to 88% in the free net group. Along the same lines, the likelihood of no sleeping space being covered by a bed net decreased from 39% in the control group to 13% in the loan group and to less than 1% in the free net group.

where qi is the quantity of crop i and P50,i is the median price for one unit of the respective crop.9 In the study sample, eight crops were grown: cotton, maize, ground nuts, sweet potato, sunflower, soy beans, tomato and cassava. In general, price variations across farmers were low: for the two major crops (cotton and maize) prices are negotiated and established at the national level by the government. For some of the minor crops like sunflowers and soy beans, prices are established at local markets. For all crops, individual prices reported rarely deviated by more than 10% from the most commonly reported market prices. For our analysis, we focus on cotton and maize as the two most commonly grown crops as well as the aggregate TEVP variable. Fig. 4 shows kernel density estimates of all three variables in 2010 on an absolute (top panel) and logarithmic (bottom panel) scale. The average cotton harvest in 2010 was 6 bales, which corresponds to about 480 kg of cotton harvested with an average plot of 1.3 ha. Average maize harvest was 28 bags, which implies an average yield of about 1400 kg for an area of 2 ha. Both yields are small when compared to industrial farmers, who frequently achieve yields of over 10 tons of maize per hectare (13 times the sample average) and over 2 tons of cotton per hectare (about 9 times the sample average).10 The average total economic value of all crops harvested in 2010 was 2.6 million Kwacha (US$ 517).

5. Impact on agricultural production

5.1. Empirical strategy

The main hypothesis investigated in this experiment is whether short-term fluctuations in labor supply generated by ill health lead to lower agricultural output. To measure production, we focus on three different measures: maize harvest, cotton harvest, and total production value. Maize is by far the most common crop in the country, and grown on approximately 50% of all plots in the sample. Maize is traded in standard bags of 50 kg, which makes measuring total production of maize relatively easy. The second most common crop in our sample is cotton. Cotton is used as a cash crop, and, as described above, sold to cotton ginners at the end of the harvesting season. While ginners pay by kilogram (2009 prices were US$ 0.30 per kg), farmers generally delivery cotton in large bags, which are referred to as “bales”, and generally contain about 80 kg each. Even though cotton and maize account for the large majority of farming land and production in this sample, most farmers use small plots to grow a variety of other plants such as sunflowers, beans, groundnuts, and sweet potatoes. The diversity in crop portfolios means that crop-specific quantities cannot easily be compared across farmers or treatment groups. On the other hand, not accounting for the resources generated by these crops would clearly mean that the effects of additional labor inputs may not be fully captured. To measure total farm production, detailed harvest information was obtained from all crops, and then converted into monetary values using the median 2009 market prices reported among farmers who sold the respective crops. In theory, one may wish to use farmspecific crop prices to account for differences in market access or production quality; in practice, this is unfortunately not feasible since a large number of farms do not sell specific crops at all, but rather use it for their own consumption.8

The basic empirical (intention-to-treat) model we estimate to identify the treatment effects of interest is given by

8 The only cash crop in the sample is cotton; all other crops are mostly used for own consumption, with some occasional sales to cover additional cash needs.

In order to have a comprehensive measure of farm productivity, we defined the total economic value of production (TEVP) as TEVP =

8 

qi P50,i

(1)

i=1

yij = ˛ + Tj ˇ + Xij  + εij ,

(2)

where yij is the harvest outcome of interest observed for farm i in cluster j in 2010, Tj is a vector of indicator variables capturing the distributor-level treatment assignment, and Xij is a vector of baseline covariates. Given the limited use of fertilizer in the region, yields tend to display mean-reverting patterns over time, with good yield years depleting the soil and being followed by less productive years. Fig. 5 illustrates these patterns, showing the year-over-year change in output as a function of total output in the 2009 (preintervention) period. To control for baseline differences as well as the observed meanreverting patterns, we first estimate models where we including the lagged (2009) value of the outcome variable of interest. With lagged outcome variables, the main estimated equation becomes yijt = ˛ + yijt−1 + Tj ˇ + εij ,

(3)

where yi,t−1 is the lagged value of the dependent variable, i.e. the harvest outcome in the 2009 season. Following the approach taken in Bennett et al. (2014), we also test an alternative model where we take changes in the outcome variables as dependent variable, and control for an extensive set of baseline covariates, which is given by yij = ˛ + Tj ˇ + Xij  + εij .

(4)

9 We assume that the quality of the produced crops are comparable across farmers; this assumption is empirically always true for cotton and maize (were prices are fixed), but may not necessarily reflect prices for local small-quantity trades. 10 See http://www.indexmundi.com/agriculture/ for a country ranking for crop productivity.

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Maize (bags)

0

20

40

60

0

kernel = epanechnikov, bandwidth = 0.9652

100

200

300

0

kernel = epanechnikov, bandwidth = 4.4401

Ln(bales of cotton)

5000 10000 15000 20000

kernel = epanechnikov, bandwidth = 355.2049

Ln(Bags of maize)

Ln(Total Value (ZKR))

Density .4 .2 0

0

0

.2

Density .4

Density .2 .4

.6

.6

.6

.8

Total value (ZKR)

0

0

0

Density .02 .01

Density .02 .04 .06 .08

.1

.03

Cotton (bales)

Density .0001 .0002 .0003 .0004

158

0

1

2

3

4

kernel = epanechnikov, bandwidth = 0.1381

0

2

4

6

kernel = epanechnikov, bandwidth = 0.1921

4

6

8

10

kernel = epanechnikov, bandwidth = 0.1668

Fig. 4. Agricultural outcomes: Kernel density estimates. Notes: total values are in rebased Zambian Kwachas (ZKR); one rebased Kwacha corresponds to 1000 “old” Zambian Kwachas.

200 100 0 -100

Year-over-year change in output (%)

300

In this specification, yij is the change (difference) in the outcome between the 2009 (pre-intervention) and the 2010 harvests, and Xij is a vector of baseline covariates. Table 3 shows the main treatment effect results. In Panel A of Table 3 we show unconditional differences across the three groups. The observed differences

0

500

1000

1500

Total harvest value 2009 in USD Fig. 5. Agricultural outcomes: year-over-year mean reversion. Notes: the percentage (year-over-year) changes in output value between the 2009 and 2010 harvest seasons as a function of the total 2009 harvest value in US$.

in total harvest value across the three groups are large, with farms in the loan and free net groups showing additional yields of US$ 180 and US$ 156, respectively (Panel A, column 3). Given the large differences in total yields at baseline, a substantial fraction of this differential is clearly attributable differences in baseline covariates. In Panel B of Table 3, we directly control for these baseline differences in productivity by including lagged dependent variables in our model as outlined in Eq. (2). As expected, the lagged dependent variables are highly significant in all models, and explain a substantial fraction of the unadjusted differences observed in Panel A. Consistent with the convergence patterns seen in Fig. 5, all coefficients on lagged variables are strictly positive and smaller than one. The adjusted model suggests that the loan program increased total harvest value by an average of US$ 69, while the free net program increased yields by about US$ 65 – only the latter is marginally significant (Panel B, column 3). Once we estimate Eq. (4) and control for baseline covariates in Panel C of Table 3, we get smaller point estimates for maize, and larger point estimates for cotton. Average cotton yields declined substantially from 2009 to 2010 (from 11.1 to 6.3 bales), which appears to be mostly attributable to less favorable rain patterns. Declines were substantially smaller in the intervention groups. In relative terms, this implies that the net programs increased cotton yields by about 25% (1.3 additional bales relative to the control group average of 5.2), while maize yields increased by 6% (loans) and 12% (free nets) compared to the control group average of 22 bags. The relatively larger impact on cotton is consistent with the idea of cotton being the more labor-intensive crop; it could

159

0

0

Production Value 2010 (US$) 500 1000

Production Value 2010 (US$) 500 1000 1500

1500

2000

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

0

500

1000 1500 Production value 2009 (US$) Control 95% CI

2000

0

500

Loans

1000 1500 Production value 2009 (US$) Control 95% CI

CONTROL VS. LOAN

2000

Free nets

CONTROL VS. FREE NETS

Fig. 6. Fractional polynomial predictions: farm yields 2010 as function of 2009 farm yields.

(2) 1.917* (1.101) 1.179 (0.751) 5.2 0.014 No

(3) 179.9*** (54.71) 155.9*** (44.38) 202.0 0.032 No

Panel B: Controlling for lagged dependent variables Dependent

Bags of maize

Bales of cotton

Total harvest value

Loan program

4.107 (2.839) 3.521 (2.394) 0.633*** (0.0916) 22.1 0.369 No

1.197 (0.799) 0.610 (0.693) 0.266*** (0.0899) 5.2 0.188 No

69.25 (44.34) 65.08* (36.83) 0.565*** (0.0731) 202.0 0.409 No

Free nets Lagged dependent Control group average R-squared Baseline covariates

Panel C: Differences in outcomes with full set of covariates Dependent

Bags of maize

Bales of cotton

Total harvest value

Loan program

1.439 (2.972) 2.747 (2.701) 1.99 0.331 Yes

1.371* (0.737) 1.366* (0.733) −4.23 0.688 Yes

45.82 (42.83) 75.60** (35.85) −51.19 0.438 Yes

Free nets Control group average R-squared Baseline covariates

Notes: Specifications in Panel A do not include covariates. Specifications in Panel B control for lagged dependent variables only. Specifications in Panel C control for lagged dependent variables as well as for plot sizes used for cotton, maize and other crops in 2009, family members under 5, family members 5–14, family members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status. Cluster-bootstrapped standard errors in parentheses. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor. *** p < 0.01, ** p < 0.05, * p < 0.1

300

(1) 10.25*** (3.826) 8.611*** (3.057) 22.1 0.022 No

200

Control group average R-squared Baseline covariates

Total harvest value

100

Free nets

Bales of cotton

0

Loan program

Bags of maize

-100

Dependent

YoY Change in Yields (Cluster Average)

Panel A: Unadjusted

also be interpreted as evidence of households prioritizing maize as the primary food crop, and therefore reducing (increasing) labor inputs on cotton rather than maize in cases where labor constraints become (less) binding. Overall, the fully adjusted models shown in Panel C suggest that the net loan program increased total harvest value by (a not statistically significant) US$ 45, while the free net program increased harvest value by US$ 76 (Panel C, column 3). To provide a more direct sense of how overall production was affected by the program, we plot 2010 yields as fractional polynomial functions of 2009 yields by treatment arm in Fig. 6. As the figure illustrates, farmers in the loan and free net groups did consistently better in 2010 than farmers in the control group. These differences appear to be larger for very small farmers (less than US$ 300 production value) and appear to be particularly large for farms at the upper end of the yield distribution with more than $ 1000 production value in 2009. Given that the treatment assignment was made at the distributor level, one may find it more intuitive to analyze year-over-year changes at the distributor (cluster) level. In Fig. 7, we show the cluster-level distribution of year-over-year changes. Each observation in the plot corresponds to the average absolute change in production value in the cluster. The boxes at the center of the figure show the median, 25th and 75th percentile for villages in each arm;

-200

Table 3 Productivity impact.

Control

Loans

Free nets

Fig. 7. Changes in farm yields at the cluster level.

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

the outside whiskers show the lower and upper adjacent values11 (Frigge et al., 1989) While the large majority of clusters in the control arm fared worse in 2010 than in 2009 (negative year-over-year change), year-over-year changes were positive for a majority of villages in the loan group, and were positive for close to 75% of villages in the free net group. In order to make sure these results are not driven by individual farmers or clusters, we run a series of robustness checks in Table 4, based on the fully adjusted specifications shown in Panel C of Table 3. In column 1 of Table 4, we exclude households with heads who did not receive any schooling. In column 2, we restrict our analysis to farms with a medium degree of crop diversification (3–5 crops) to ensure that the year-over-year changes are not affected by differential exposure to price and crop risk. In order to deal more directly with pre-existing differences in productivity, we exclude the 20% least productive farmers in the baseline season from the regressions in column 3; in column (4) we exclude the most productive farmers in 2009, and last, in column 5, we exclude both the least and the most productive farmers from the analysis. Overall, the results appear rather robust, with average program estimates between US$ 44 and US$ 79 for net programs, and estimates between US$ 54 and US 110 for free net programs, both within the confidence intervals of our main impact estimate of US$ 46 and US$ 76 (Table 3, Panel C). In terms of program impact, largest effects are observed for more educated and diversified households; in terms of baseline yields, effects seem to be larger at the very bottom and the very top of the distribution, a pattern which is consistent with the non-parametric estimates shown in Fig. 6. 6. Mechanisms and discussion The primary mechanisms through which the programs were intended to affect productivity was household health. In order to monitor the health of households, two follow-up visits were conducted, one in April 2010, and one in July/August 2010. During both follow-up visits, interviewers went over the household roster collected during the baseline assessment, and asked respondents to indicate for each household member whether they had been sick since the last interview. In a previous paper (Fink and Masiye, 2012), we used data on self-reported morbidity outcomes from both follow-ups to identify the health impact of the additional nets. Given the concerns with longer-term recall data highlighted in the recent literature (Arnold et al., 2013; Das et al., 2012), we restricted our analysis to morbidity episodes in the two weeks preceding the interview. Consistent with the estimates reported in the bed net literature (Lengeler, 2004), we found large health benefits attributable to the additional nets, with treated households reporting reductions in the incidence of health problem of any kind of about 40% and in the incidence of confirmed malaria of at least 50%.12 In order to directly assess the impact of ill health on farm labor inputs, farmers were asked to report the number of days of field work list for each illness episode. Specifically, farmers were asked to report how many days of field work were lost because (i) the sick person was not able to work and (ii) because somebody else in the household had to take care of the sick person. Fig. 8 shows the average number of days of field work lost per malaria episode13

11 Adjacent values are defined as the most extreme values within 1.5 interquartile ranges of the lower and upper quartile, respectively. See Frigge et al. (1989) for further details. 12 “Confirmed malaria” was defined as a fever episode where respondents indicate that they were tested for malaria and that the test was positive. 13 Malaria episodes include all recent illnesses attributed to malaria by respondents (based on the answer provided to the question “What health problem did the person have”).

Average days of field work lost per bout of malaria

160

8 7 Days of caretaker field work lost 6

Days of patient field work lost

5 4 3 2 1 0 Under 5

Age 5-14

Age 15-59

Age 60+

Fig. 8. Work days lost by malaria episode.

by age group.14 The average number of days of field work lost by malaria episode is 4.7; the burden is substantially lower for children (slightly less than 4 days taken off for taking care of children on average) than for adults, where the average number of days lost per episode is about 7, with a majority of days lost directly attributable to the patient’s absence from the field. Given that the second follow-up was done after the harvest with most farmers not working, we further restrict our analysis to illness episodes reported in the first follow-up. In total, only 85 households reported illness episodes in the 2-week period preceding the interview in April, which means that we have only very limited power to detect effects, despite the substantial differences in the likelihood of reporting an illness episodes: in the control group, 32 out of 153 (21%) of farms reported an illness episode in the twoweek period prior to the first follow-up. In the loan group, 33 out of 185 (18%) reported a health problem, and last, in the free net group, only 20 out of 155 (13%) of households reported an illness episode; in relative terms, this means that the likelihood of experiencing an acute morbidity episode was reduced by about 40% in the free net group. In absolute terms, these reductions are of course relatively small. Over a six-month period, the observed differences would correspond to approximately one more illness episode per household. In Table 5, we analyze the program impact on field work labor, as well as total health expenditure. In the first four columns, we show unadjusted treatment effects. In columns 5–8, we show the same estimates when a full set of household covariates is included; different from the productivity impact estimates, the inclusion of covariates does not make much of a difference in these specifications. On average, we find that net programs reduce the number of days of field work lost by 0.3 days over a two-week period, which corresponds to about 3.6 days over the agricultural season. Given the small number of illness episodes the confidence intervals around this estimate are wide however, and the estimated coefficients not statistically significant from zero. In terms of the estimated effects, about 60% of the field work losses appear to result from the person sick not being able to complete field work, and 40% appear to result from field workers having to take care of other household members. It is worth pointing out here that these measures only capture complete absence from field work; no data was collected on worker strength or ability to work on the fields. In

14 Since no field work was reported during the second follow-up (scheduled in the post-harvest season in July), all numbers reflect the morbidity reports provided during the first follow-up round in April.

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

161

Table 4 Robustness checks and heterogeneous treatment effects. Change in production value 2010–2009 (US$) Net loans

64.60 (47.27) 110.4** (45.21) Yes

43.83 (48.78) 100.8** (44.85) Yes

61.11 (53.41) 67.75 (44.82) Yes

41.46 (48.49) 60.23** (30.19) Yes

77.51 (59.06) 53.70 (39.66) Yes

Sample

Education of head > 0

Observations R-squared

341 0.467

Medium diversification: 3–5 crops 395 0.456

Excluding bottom 2009 productivity quintile 386 0.470

Excluding top 2009 productivity quintile 400 0.392

Excluding clusters in top or bottom quintile in 2009 293 0.414

Free nets Full set of covariates

Notes: All specifications control for lagged dependent variables as well as for plot sizes used for cotton, maize and other crops in 2009, family members under 5, family members 5–14, family members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status. Cluster-bootstrapped standard errors in parentheses. Each cluster corresponds to one distributor and 11 randomly selected farmers working with the distributor. *** p < 0.01, ** p < 0.05, * p < 0.1

Table 5 Labor supply and health expenditure effects. Days of field work lost due to own sickness in last 2 weeks

Days of field work lost due to other sickness in last 2 weeks

Total days of field work lost in last 2 weeks

Total health expenditure last two weeks

Days of field work lost due to own sickness in last 2 weeks

Days of field work lost due to other sickness in last 2 weeks

Total days of field work lost in last 2 weeks

Total health expenditure last two weeks

Any programa)

−0.174 (0.149)

−0.129 (0.129)

−0.303 (0.258)

0.0783 (0.0604)

−0.170 (0.154)

−0.0818 (0.117)

−0.252 (0.234)

0.0740 (0.0573)

Loan program

−0.206 (0.157)

−0.0764 (0.136)

−0.283 (0.257)

0.0842 (0.0676)

−0.182 (0.145)

0.00153 (0.125)

−0.180 (0.239)

0.0972 (0.0668)

Free net program

−0.134

−0.192

−0.326

0.0713

−0.154

−0.196

−0.350

0.0423

(0.170)

(0.138)

(0.273)

(0.0850)

(0.192)

(0.152)

(0.292)

(0.0905)

Control included

No

No

No

No

Yes

Yes

Yes

Yes

Control group mean

0.391

0.360

0.752

0.072

0.391

0.360

0.752

0.072

Observations

493

493

493

493

493

493

493

493

Notes: All specifications control for plot sizes used for cotton, maize and other crops in 2009 as well as in 2010, family members under 5, family members 5–14, family members 60 and older, household wealth, mosquito nets at baseline, ownership of chickens, goats and sheep, farmer age, education and marital status, total maize harvest 2009, total cotton harvest 2009 and total economic value of production 2009. Cluster-bootstrapped standard errors in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1.

highly endemic areas like the one studied, asymptomatic malaria infections are common, and likely to substantially reduce farmers’ ability to complete field work tasks even when they do not suffer from acute infections (Nur, 1993), As part of the follow-up interviews, we also collected information on health expenditure. Zambia’s health sector – in particular in rural areas – is dominated by public health facilities which generally provide basic health services (including malaria testing and drugs) for free. In our sample, 90% of respondents indicated zero out-of-pocket expenditure for household members getting treated when sick. Given this, the very small and not significant estimated program impact on total health expenditure found in Table 5 is not surprising.

7. Discussion The estimates presented in this paper suggest a rather large positive impact of malaria prevention on agricultural productivity. While the data collected as part of the project does unfortunately not allow us to precisely identify the mechanism underlying this impact, increased labor inputs appear to be the most plausible causal pathway. Given that labor is in principal abundant in Zambia, one might argue that short-term labor supply shocks should not matter at all for farm production; in settings where labor can be freely hired, morbidity should not affect farming output, since

any lacking labor could be hired in local markets. During the first follow-up rounds, we directly questioned farmers about labor substitution. In total, 167 instances were reported where the head of household was not able to work on the field because he or she was sick.15 Out of these 167 episodes, substitute labor was hired only in 10 cases (6%). In 7 out of these 10 substitution cases, farmers found somebody to work for free; only three farmers reported to pay for labor, with wages ranging between 0.5 and 11 dollars per day. Anecdotally, local piece work (“ganyu”) labor is widely available (Fink et al., 2014); in practice, most small-scale farmers do however not appear to have the resources to hire such labor. Even if one is willing to accept that hiring short-term labor may be hard for farmers, the estimated impact numbers appear large. As shown in Goldberg (2014), agricultural wages for day labor are frequently less than US$ 1 in rural Malawi. In the Zambia settings, wages appear slightly higher, with median daily wages varying between KR 12.5 and KR 25 (US$ 2.5–5) reported for the peak labor season in focus groups. Even at US$ 5, direct labor costs are unlikely to account for the full differences in output observed.

15 Note that the total number of episodes here related to the four months period between the baseline and the midline survey, and thus is different from the 85 illness episodes reported for the two-week period preceding the interview analyzed in the previous section of the paper.

162

G. Fink, F. Masiye / Journal of Health Economics 42 (2015) 151–164

Our estimates suggest that having all sleeping spaces covered with bed nets will save the average farm approximately 3–5 working days, which translates to about US$ 15–25, and therefore to 30% of the estimated effects at most. Two factors may at least partially explain the remaining gap: first, recovery from malaria is a slow process, frequently taking more than two weeks, during which farm workers are likely less productive (Nur, 1993) even if they report back to work on the field and thus would not be counted as “able to work” in our analysis. A second possibility is that farmers in the study may have only reported major health events, so that the reported numbers do not fully capture the true program impact. In the 2007 Zambia Demographic and Health Survey (Macro International, 2007), 20% of children under-5 were reported to have suffered from fever or diarrhea over the 2-week period preceding the survey. Even if adults are substantially less prone to be sick, the reported morbidity prevalence seems very low: 32 episodes across 153 households in the control group implies about one episode for every 25 individuals or a fever prevalence of about 4%, which is about one fifth of the under-5 prevalence in the DHS. One alternative explanation for the relatively large effects observed are local spillover effects, which could potentially also undermine the internal validity of the study: as demonstrated by Apouey and Picone (2014), social interactions in the realm of malaria and malaria prevention are likely, and may lead to stronger associations between health behaviors and health outcomes at the village or regional level. Even though it is quite likely that farmers interact with other farmers outside the 3 km radius chosen for the randomization, large spillover effects do not seem very likely in our setting: first, by working only with small-scale farmers having an outgrower contract with Dunavant, we covered only a small fraction (