3 Yield Response of Dryland Cereals to Fertilizer on Smallholder Farms in Mali

Abstract

In Mali, over 60% of the population lives in rural areas, and about half of them live under the poverty line. Since farming is the primary livelihood of those in rural areas, raising productivity is crucial for decreasing poverty. This chapter explores the effectiveness of nitrogen fertilizer in raising yields of dryland cereals on smallholder farms in Mali, using regional and national datasets. Simple econometric analysis suggests there is minimal effect of nitrogen fertilizer use on crop yields. However, when accounting for the endogeneity of fertilizer use, we find yield response rates within the range reported in the literature. As expected, sorghum has lower response rate to fertilizer than maize. Yield responses are stronger in the Sudan Savanna region than nationwide, highlighting the importance of agroecological factors and the farming system. Soil texture and practices (anti-erosion structures) affect both yields and estimated marginal effects of fertilizer. We also find phosphorus to be a binding constraint in increasing agricultural productivity. While most emphasis in the literature is placed on understanding nitrogen fertilizer use, as recommended by agronomists, it is crucial to promote balanced use of fertilizers so that other complementary nutrients are available in the soil.

Keywords: yield response, cereals, Mali, smallholder farmers

Introduction

In Mali, over 60% of the population is rural, and half that segment lives below the national poverty line (World Bank, 2017). Most rural people depend on agriculture as their main source of livelihood. Dryland cereal crops, such as maize, millet, and sorghum, account for two-thirds and three-quarters of all cultivated land, depending on the years. Over recent decades, the production of dryland cereals has grown primarily through the expansion of cultivated areas rather than intensification, which is unsustainable. Despite the continued release of improved varieties, millet and sorghum yields have stagnated, with national averages hovering below 1 t/ha. Meanwhile, national average maize yields have risen from 1.4 t/ha in 2001/05 to 2.6 t/ha in 2016/18 (http://www.fao.org/faostat/en/#data/QC, November 2, 2020). Mali is West Africa’s third-largest producer of maize even though it stands fifth in the area harvested, with the highest average yields among all 15 maize-producing countries in the region (Abate et al., 2015). Nonetheless, as is common throughout Sub-Saharan Africa, maximum yields for improved maize varieties in farmers’ fields remain substantially below yield potential based on experimental conditions (4-6 t/ha according to Coulibaly 2008; Macauley & Ramadjita, 2015) given the challenges of growing conditions and incomplete input markets.

Inadequate use of mineral fertilizer has often been pinpointed as a cause of stagnating productivity in dryland cereals in Sub-Saharan Africa (NEPAD, 2003). In 2008, the Malian government decided to reinstate an input subsidy program to boost cereal productivity through improved access to fertilizer while contributing to food and nutrition security via higher income and lower consumer prices (Kone et al., 2019; Smale et al., 2012). Fertilizer subsidies now constitute the largest expense item, accounting for about 25% of government spending on rural development (Theriault et al., 2018). Given that extremely small amounts of fertilizer are currently used on either sorghum or millet, we do not expect the fertilizer subsidy program to have a generalized effect on production decisions nationwide. However, Theriault et al. (2018) found a significant impact on fertilizer use and yields for sorghum and maize in the Sudan Savanna, a region suitable for agriculture. A more resounding critique of policy regarding mineral fertilizer is that the soils of Mali are generally deficient in specific nutrients and that soil organic matter is necessary for effective integration of nutrients (Dicko et al., 2016; Mason et al.,2014). In addition, long-term losses due to soil erosion have been documented (Bishop and Allen 1989). However, these have been offset in some areas by successful resource management programs (Tappan & McGahuey, 2007).

Despite the policy emphasis on mineral fertilizer, few studies have systematically examined the cereal yield response to fertilizer in Mali—which is fundamental for evaluating program impacts. An important exception is the analysis of farm experimental data by Dicko et al. (2016), who estimated response functions for nitrogen (N), phosphorus (P), and potassium (K) on maize, rice, and millet across the four bioclimate-based agro-ecological zones of Mali. In general, the authors found that economic optima of nutrient application rates diverged from recommended application rates, which remain uniform throughout the country and vary only by crop.

We know of only two other studies that address the productivity of dryland cereals in Mali, both of which use data collected in farm household surveys rather than under experimental conditions. Using a 12-year farm household panel dataset (1994-2006) from Mali’s Sikasso region, Foltz et al. (2015) found a strongly significant response of maize yields to fertilizer, concluding that increasing fertilizer use has driven most of the maize productivity growth. Sikasso is a region of high productivity potential for maize, where it is grown in rotation with cotton. Applying a stochastic production frontier to nationally representative data from the Living Standards Measurement Survey-Integrated Survey of Agriculture (LSMS-ISA), Ahmed et al. (2017) found no significant response of yields to fertilizer across crops and regions. We know of no published analyses of yield response to fertilizer in sorghum or millet that employ farm household data.

We thus contribute to sparse literature by estimating dryland cereals response functions (maize, millet, and sorghum) using two farm household survey datasets. The first dataset, the Living Standards Measurement Survey-Integrated Survey of Agriculture (LSMS-ISA), is nationally representative and includes information on all three crops (maize, millet, and sorghum). The second dataset, collected by a research team from the Institut d’Economie Rurale, was collected only in the Sudan Savanna and focused on sorghum and maize. Both datasets were collected during the 2014/2015 growing season. In estimating our yield response functions, we employ a combination of econometric approaches to compare and check the robustness of our findings.

Methods

1. Data Sources

We utilize data from the Living Standards Measurement Survey-Integrated Survey of Agriculture (LSMS-ISA), conducted in Mali in two visits during the 2014-15 growing season. Summary information about the survey is provided in a document compiled by the Planification and Statistics Unit (2016). With the probability of selection proportional to population size as of the 2009 Census, the statistical sample is nationally representative of rural and urban areas, excluding the region of Kidal. The total sample size was limited by the inability to collect data in some regions because of political insecurity, with the largest sample losses in the Mopti, Tombouctou, and Gao regions. The final sample includes about 3,804 households compared to the planned sample of 4,218. The number of standard enumeration areas (SEs, or grappes) was 1070, with 80% in rural areas, including 2-3 households per grappe. Compared to LSMS surveys which focused on household consumption, expenditures and income, the LSMS-ISA survey also contains plot-level data on input use and crop production. One-third of all plots inventoried by households in each SE were randomly sampled after grouping them by crop and crop association.¹ This procedure was necessary given that in Mali, large numbers of plots may be simultaneously cultivated by extended family farms, augmenting respondent burden and survey costs. We conduct our analysis only on data from the main rural agricultural regions of Mali, excluding Tombouctou, Gao, and some observations around Bamako. Therefore, our analytical sample covers the regions of Kayes, Koulikoro, Sikasso, Segou, and Mopti.

The second data source is a case study undertaken in the sorghum belt of Mali, which we use as a comparison since it is more focused on a specific farming system. Survey details are provided in Smale et al. (2015). The sample was drawn from a baseline census of all sorghum-growing households in 58 villages in the Cercles of Kati, Dioila (both Koulikoro Region), and Koutiala (Sikasso Region) of Sudan Savanna, within the 800 mm isohyet. Villages surveyed included fewer than 1000 persons listed as sites where the national research program and farmer associations had implemented activities since 2009. The multi-visit survey was conducted in four rounds from August 2014 through June 2015 by a team of experienced enumerators employed by the Institut d’Economie Rurale. The sample is representative of areas in the Sudan Savanna with some exposure to agricultural research outreach activities. For cereals, many sorghum growers also grow maize and millet in this region. Millet is also grown, but due to budget constraints, detailed plot information was collected only for sorghum and maize.

The sample of households was drawn with simple random sampling and augmented by five percent to account for possible non-responses, leading to a total of 623 households and an overall sampling fraction of 25%. Enumerators listed all plots operated by each sampled household. One plot was randomly sampled per crop and management type per household. After removing yield and fertilizer use outliers, the total analytical sample employed here is 1,086 plots, including 421 sorghum plots and 665 maize plots.

In addition to the household survey data, this dataset includes soil nutrient indicators measured in laboratory tests conducted on soil samples by the Institut d’Economie Rurale, Sotuba, Mali. Due to budget constraints, soil samples could not be collected from all sorghum and maize plots. Plots were subsampled randomly within crop (maize and sorghum) and plot management (collective, individual) groups. After the harvest, soil samples were obtained following a standard protocol with 8 sub-samples per plot collected in a zig-zag pattern to assure overall plot representation. Laboratory analysis followed Sparks et al. (1996). The analytical sample for soil nutrients is 643 plots.

The two household datasets are not directly comparable with respect to sampling methodology or statistical representation of the farming population. We exploit both datasets because each provides information on nitrogen use and response on cereal crops grown in farmers’ fields in Mali. Information of this type is scarce. The Sudan Savanna dataset is structured on a village list reflecting organizational structure and agroecology. The LSMS-ISA dataset is structured according to administrative units with census-based, standard enumeration areas. No georeferenced identifiers enable us to link one dataset to the other and directly compare results. Finally, variables are not identical between the two datasets.

Rainfall data were downloaded and compiled from the Climatology Resource for Agroclimatology site of the National Air and Space Administration.²

2. Econometric Approach

Our objective is to quantify the effect of fertilizer use on the yields of dryland cereal crops using household survey data. First, we estimate the yield response model:

Y_i = β₁F_i + β₂I_i + γX_i + ε_i .                            (1)

The dependent variable Y_i denotes the crop yield (kg/ha) on plot i. The key explanatory variable is F_i, the quantity of fertilizer applied on plot I, with coefficient β₁. Fertilizer quantity is measured in nitrogen kg/ha to standardize different fertilizer types. Quantities are summed across the nitrogen content of the urea (46%), NPK (14%), DAP (18%), and other fertilizer (16.5%) applied on plot i.

Other than fertilizer, agricultural inputs I_i are applied to plots, and these are typically included in yield response functions estimated with either experimental or survey data. Accounting for other inputs is important because fertilizer is often used in conjunction with other inputs. For example, plots that apply more fertilizer may have more labor allocated. If we do not control for labor quantity, the coefficient on fertilizer will include the effect of labor and will overestimate the effect of fertilizer on crop yield. X_i is a vector of factors other than inputs that affect crop yield, such as plot characteristics.

The estimation strategies for equation (1) are tailored to the two data sets. Despite controlling for plot characteristics, X_i, the estimate of β₁ may be inconsistent if other omitted variables explain crop yield and correlate with the fertilizer quantity applied. For example, wealthier households may have higher crop yields because they can acquire more efficient agricultural practices from radio, television, or agricultural extension agents. They may also apply more fertilizer because they have money or credit available to purchase fertilizer. Farmers may also apply fertilizer on plots with soils that they observe to be more responsive.

To eliminate household level confounding factors, in the LSMS model, we use household fixed effects to estimate model (1). In addition, since the survey is cross-sectional and contains data from a single year, we compare yields on plots within the same household with varying fertilizer levels applied. This allows us to control for any unobserved household-level factors that explain crop yields and correlate with the fertilizer quantity applied.

While this estimation strategy can provide reliable estimates for β₁, we may still be concerned about omitted variables bias. Unobserved plot-level characteristics are likely to affect crop yields. We can measure such variables, as in the case of soil samples collected in the Sudan Savanna, and include them as explanatory variables. For example, the soil’s organic matter partly determines the responsiveness of nitrogen fertilizer (Marenya & Barrett, 2009).

With the LSMS data, we use instrumental variables estimation, combined with household fixed effects, to provide more robust estimates of the effect of fertilizer on crop yields. The instrument captures the general diffusion of fertilizer in the region where the household is located. We use the average fertilizer rate used across all other plots of households growing that crop in the grappe (1 to 30 plots, with an average of 3.5 plots).

Since the identification strategy combines household fixed effects with instrumental variables estimation, the instrument must vary within households. The diffusion rate of fertilizer differs across crops, and hence the instrument varies within households when we consider more than one crop. If we estimate a yield-response function for each crop separately, the instrument will not vary within households, and we are unable to use household fixed effects combined with instrumental variables estimation. Thus, we estimate separate regressions for millet, sorghum, and maize with only household fixed effects.

We expect the instrument to be correlated with the potentially endogenous variable Fi, because it captures the general availability of fertilizer in the locality. We do not expect the fertilizer allocation of households to directly influence the crop yields of other households – except through the increased use of fertilizer. Under these assumptions, we obtain consistent estimates of β₁. The instrument will not be valid if it affects yields through mechanisms other than more fertilizer being applied on that plot. For example, it would be problematic if the instrument was correlated with more intensive use of other inputs. Since we include the quantity of other inputs allocated to the plot as explanatory variables, we overcome this problem by controlling for the instrument’s effect through other inputs.

In the regressions estimated with nutrients measured in laboratory tests on soil samples, we combine both dryland cereals in one regression, given the even smaller number of observations. Binary and interaction variables are included to control crop effects on grain yield and yield response to nitrogen nutrients per ha. We also test a quadratic term for nitrogen nutrients per ha, which expresses whether a turning point in yield response to nitrogen is observable in the data. In the set of regressions estimated with farmer-perceived soils classes, along with the combined regression, we have also estimated separate regressions for maize and sorghum because sample sizes are larger.

As a robustness check, we also test the final combined model with farmer-perceived soil classes (the largest sample) with the Control Function Approach. A Control Function Approach is applied instead of instrumental variable methods because of the potential endogeneity of multiple variables (nitrogen fertilizer applied, interaction, and squared terms). As instruments, we utilize whether the plot manager has benefited from the fertilizer subsidy and the village share of plot managers who are registered cooperatives. Theriault et al. (2018) found that Malian farmers who are members of cooperatives have better access to fertilizer than non-members. Participation in the subsidy program and cooperative is likely to affect fertilizer use but unlikely correlated with unobserved variables. All Sudan Savannah regressions are estimated with robust standard errors, clustered by household.

3. Variables

In the two analyses, explanatory variables differ somewhat to reflect underlying differences in the data. Variables used in the LSMS analysis are listed in Table 1. The plot area, measured in hectares, allows us to examine whether productivity differs across plot sizes. Vector I includes manure, compost, other organic fertilizer (i.e., crop residues), pesticides, herbicides, fungicides, other protecting liquids, improved and local seed. The distance of the plot from the homestead is related to the time taken to reach the field for crop management. It is also, potentially, a measure of fertility since those located farther away may have been more recently cleared and brought into cultivation. Also, households may choose to invest more in nearby plots since they are more secure (Gebremedhin & Swinton 2003) and easier to reach. A dummy variable is included to control for whether there is an anti-erosion structure on the plot, such as stone contour bunds or dikes—which have been promoted in certain regions of the country to offset the heavy loss of soil nutrients during the rainy season and enable farmers to retain moisture (Tappan & McGahuey, 2007).

Location in the toposequence (lowland, plain, slope, plateau) and soil texture are important indicators of soil quality in this region and are highly correlated with the crops grown (Guirkinger et al., 2015; Udry, 1996). Bazile et al. (2008) explain that farmers define soil type according to the position of the field in the toposequence. Farmers distinguish the shallow soils of the plateaus or higher areas from medium-depth soils and alluvial, low-lying soils (‘bas-fonds’). Soil differentiation observed within and among farms explains growing multiple varieties and crops per farm and across a landscape. To capture the location of the plot in the toposequence, we include binary indicators of location in the plain, lowland, on a slope, or plateau. Dummy variables for soil type are also included, representing farmer-perceived soil classes (sandy, silty, clayey).

Table 1
Descriptive Statistics, LSMS-ISA Data

Variable	Mean	S.D.	Min	Max
Millet Yield (kg/ha)	695	629	0.660	3759
Sorghum Yield (kg/ha)	735	686	0.530	3930
Maize Yield (kg/ha)	1492	1221	1.26	6000
Material Inputs
Nitrogen Fertilizer (N nutrient kg/ha)	6.74	23.6	0.000	288
Manure (kg/ha)	1594	3800	0.000	29850
Compost (kg/ha)	25.8	247	0.000	4889
Other Organic Fertilizer (kg/ha)	5.64	60.6	0.000	1708
Pesticides (liter/ha)	0.053	0.51	0.000	10.5
Fungicide (liter/ha)	0.033	0.42	0.000	19.7
Herbicide (liter/ha)	0.270	1.17	0.000	19.1
Other Protecting Liquids (liter/ha)	0.007	0.11	0.000	3.14
Local Seed (kg/ha)	10.3	15.4	0.000	236
Improved Seed (kg/ha)	1.00	4.45	0.000	50.7
Labor
Total Labor (no. of days/ha)	45.1	84.8	0.000	1031
Plot Characteristics
Plot Area (ha)	3.07	6.12	0.020	52.7
Distance to plot from house (km)	2.81	4	0.000	60.0
Plain (0/1)	0.7	0.464	0.000	1.00
Plateau (0/1)	0.147	0.355	0.000	1.00
Lowlands (0/1)	0.034	0.18	0.000	1.00
Sloped (0/1)	0.13	0.34	0.000	1.00
Soil Sandy (0/1)	0.536	0.498	0.000	1.00
Soil Clay (0/1)	0.357	0.479	0.000	1.00
Soil Lateritic (0/1)	0.11	0.31	0.000	1.00
Anti-Erosion Structure (0/1)	0.043	0.2	0.000	1.00

Source: Authors, based on LSMS-ISA, Mali. Number of plot observations=3733

Explanatory variables used in the analysis of the Sudan Savanna data are listed in Table 2. We estimate maize-sorghum yield (grain harvested per ha) response to nitrogen but do not include millet plots, for which we do not have production data. Fertilizer application is computed in the same way as in the LSMS analysis. All regressions include conventional inputs, a common set of plot characteristics, and a rainfall indicator at the village level. Manure application is measured as a binary variable because of difficulties in measuring quantities reliably. Labor days, liters of herbicides, and hours of equipment use are computed per ha. Common plot characteristics are the distance (in minutes walking) from the homestead to the plot, presence of a soil erosion structure on the plot, association of the primary crop with a leguminous (groundnut, cowpea) intercrop. Average rainfall during the period of fertilization in the survey year is recorded at the geographical scale of the village.

We test two sets of soils characteristics. In the first, soils characteristics are by soil nutrient content as tested in the laboratory: percentage soil organic matter (C), sand, silt, and clay, the percentage total nitrogen (N), assimilable phosphorus (P), exchangeable potassium (K), and soil pH (KCI). Since total carbon content (C) changes over centuries and active carbon changes over 3-5 years (Weil et al., 2016), these are not affected by recent fertilizer applications. Similarly, recent fertilizer use does not affect total nitrogen content (N), which includes the nitrogen in the soil organic matter. Nor can farmers deduce the specific nutrient content of their soils (P, K). In the second, we use farmer-perceived soils characteristics. Binary variables are entered for the location of the plot in the toposequence and farmer-perceived soil classes.

Table 2
Descriptive Statistics, Sudan Savanna Data

Variable	Mean	S.D.	Min	Max
Maize yield (kg/ha)	1497	945	12.5	4730
Sorghum yield (kg/ha)	642	681	0.000	4286
Nitrogen fertilizer (nutrient kg/ha)	19.5	25.7	0.000	100
Sorghum plot (0/1)	0.609	0.488	0.000	1.00
Manure (0/1)	0.641	0.479	0.000	1.00
Labor (days/ha)	68.0	66.6	0.000	800
Herbicide (liters/ha)	1.68	2.24	0.000	25.0
Equipment (hours/ha)	475	474	0.000	5294
Distance from House (minutes)	17.4	17.5	1.00	160
Soil Erosion Structure	0.188	0.391	0.000	1.00
Legume Intercrop 0/1)	0.112	0.316	0.000	1.00
N (% total nitrogen)	0.028	0.023	0.010	0.200
C (% organic matter)	0.522	0.334	0.020	2.63
P (assimilable phosphorus)	1.29	1.31	0.210	15.9
K (exchangeable K)	0.246	0.210	0.020	1.87
Ph (KCI)	5.34	0.400	3.15	7.25
0.05)"}”>Sand (% > 0.05)	59.6	12.8	7.00	90.0
Silt (%0.05-0.002 mm)	36.2	12.3	8.00	90.0
Clay (% < 0.002 mm)	4.26	2.88	0.000	23.0
Rainfall (mm, period of fertilization)	220	31.0	164	299
Plain	0.865	0.341	0.000	1.00
Lowlands	0.015	0.122	0.000	1.00
Slope	0.119	0.324	0.000	1.00
Sandy (0/1)	0.381	0.486	0.000	1.00
Silty (0/1)	0.203	0.403	0.000	1.00
Clayey (0/1)	0.269	0.444	0.000	1.00
Gravelly (0/1)	0.147	0.354	0.000	1.00

Source: Authors, based on Sudan Savanna case study data. n=1222 for all except manure (1096).

Results

1. Descriptives

Millet, sorghum, and maize are grown on 39, 32, and 29 percent of the plots in the LSMS sample, respectively. Fertilizer application rates differ significantly across crops. Table 3 shows average use rates calculated from the LSMS and Sudan Savanna datasets by cereal crop and region, compared with recommended rates and economically optimal rates estimated with response functions based on experimental data (Dicko et al., 2016). They also differ in major ways by region.

Agronomic recommended rates of N per ha are 32 for sorghum and millet, and 84 for maize, throughout the country (Dicko et al., 2016). These correspond to 100 kg/ha on all three crops for cereal complex, 100 kg/ha of NPK (16-16-16) for millet and sorghum, 250 kg/ha of NPK (23-10-5) for maize, 100-150 of DAP on cereals, and 50-400 kg/ha for other crops and fertilizers (see also Thériault et al., 2016). The overall mean for N kg per ha on millet, sorghum, and maize in the LSMS data is 6.7, but use rates on maize are considerably higher except for the region of Kayes. For all three cereals and all five regions except millet in Sikasso, estimated mean rates of use are but a fraction of economically optimal rates. In Koulikoro, Segou, and Mopti, average N use rates on sorghum and millet are in the single digits or lower, while economic optima range from 21 to 26. Again, the mean rate of N applied per ha to maize in Sikasso (36.2) is closest but is still far from the economic optimum (56-65). Mean phosphorus application rates for the entire LSMS sample are only 1 kg/ha and 1.9 kg/ha for millet and sorghum, respectively, but 6.7 kg/ha for maize—which is close to recommended levels. Recommended use rates for P are 10 for sorghum and millet, and economically optimal rates are estimated to be higher (Dicko et al., 2016). Overall, applying their fertilizer optimization tool, Dicko et al. (2016) found that the economically optimal rates of N were well below recommendations for maize, sorghum, and millet, varying by bioclimate.

Mean applications rates of N per hectare on maize and sorghum are also shown in Table 3 for the Sudan Savanna. In this relatively high potential area, mean rates of use on sorghum (6.41) are closer to rates in Sikasso in the LSMS (9.38) than in Koulikoro at (<1), and even roughly the same on maize (39.8) as in Sikasso (36.2). Again, both are but a fraction of the economically optimal rate estimated by Dicko et al. (2016), which is, in turn, but a fraction of the nationally recommended rate.

2. Yield Response Functions

Tables 5-8 present yield response models estimated with the LSMS data. The yield and input variables, including the quantity of fertilizer, are included in logarithms to smoothen their distributions, which are concentrated in lower values and skewed in shape. β₁ and β₂ are interpreted in terms of percentage changes of yield. The coefficients have been converted into marginal products by computing the marginal change in yield (in kg/ha) from a one percent increase in the quantity of fertilizer (N kg/ha) at the mean and are indicated in the bottom rows of the tables.

In preliminary regressions, the coefficients on nitrogen fertilizer across the household fixed effects models that do not use instrumental variables suggest that a one percent increase in the quantity of nitrogen fertilizer applied to the plot results in only a 0.04-0.07 percent increase in yields of dryland cereals. These elasticities correspond to marginal physical products of only 5.2 to 8.8, on average, for dryland cereals (maize, millet, and sorghum).

FE-IV estimates for more robust inference of the effect of fertilizer on yields are shown in Table 4. In all specifications, the first-stage F-statistic is much greater than 10, which is often used as a rule of thumb for the inclusion restriction. The first-stage F-statistic is also greater than the Stock-Yogo 10 percent maximal IV size critical value of 16.38, suggesting that the inclusion restriction is satisfied. The FE-IV estimates for fertilizer are several times as large as those reported above, suggesting that not controlling for endogeneity may diminish estimates of the yield response to fertilizer. In all specifications, the effect of nitrogen fertilizer applied is statistically significant. A one percent increase in the quantity of fertilizer applied to the plot results in a 0.1-0.2 percent increase in yields of dryland cereal crops. This translates to a 17-27 kg/ha increase in dryland cereal crop yield for an additional nitrogen kg/ha fertilizer.³

Other control variables described above are included sequentially to see whether the coefficient on fertilizer is sensitive to these. We find that yields are decreasing in plot size – consistent with the inverse productivity relationship that has also been observed in this region (Guirkinger et al., 2015; Kazianga & Wahhaj, 2013; Udry 1996). Plots where manure has been applied, and those with anti-erosion structures, achieve greater yields. Yields are increasing in the quantity of local seed and increasing even more in the quantity of improved seeds. More labor allocated to a plot also raises yields. This effect is strong throughout and with a relatively high elasticity—suggesting that labor constrains productivity. We find no effects of pesticides, fungicides, or other protecting liquids, which are used in very limited amounts. The presence of anti-erosion structures has a meaningful effect on yields. Controlling for toposequence and soil type reduces the marginal product attributable to fertilizer. Models were also estimated with specifications that contained squared terms for fertilizer and interaction terms between fertilizer and crop dummy variables. The point estimates of these square and interaction terms were close to zero and not statistically significant, so we retained more parsimonious specifications.

Table 5
Millet Yield Response to N Fertilizer Applied, Household Fixed Effects Model

Variables	(1)	(2)	(3)	(4)
Nitrogen Fertilizer	0.0392	-0.00945	0.00449	-0.00372
	(0.0724)	(0.0685)	(0.0642)	(0.0714)
Manure		0.0271**	0.0222*	0.0156
		(0.0122)	(0.0122)	(0.0137)
Compost		-0.0774	-0.0518	-0.0751
		(0.0754)	(0.0744)	(0.0862)
Other Organic Fertilizer		0.0832	0.0806	0.0803
		(0.0720)	(0.0676)	(0.0657)
Fungicide		-0.668	-0.630	-1.261
		(1.004)	(0.938)	(0.923)
Herbicide		2.565	0.960	2.138
		(10.59)	(9.893)	(9.408)
Total Labor		0.470***	0.314***	0.156***
		(0.0481)	(0.0529)	(0.0602)
Local Seed			0.409***	0.360***
			(0.0652)	(0.0698)
Improved Seed			0.550***	0.350
			(0.208)	(0.259)
Plot Area				-0.0310***
				(0.00828)
Distance to plot				0.0334**
				(0.0150)
Plaine				0.0543
				(0.186)
Plateau				0.0202
				(0.314)
Lowland				-0.220
				(0.383)
Sandy				-0.00658
				(0.411)
Clay				0.411
				(0.448)
Anti-Erosion Structure				1.789***
				(0.369)
Constant	5.996***	4.633***	4.267***	4.635***
	(0.0282)	(0.139)	(0.150)	(0.466)
Observations	1,376	1,182	1,018	813
Number of households	771	688	585	476
Marginal Effect of N	10.37	-2.34	0.99	-0.81
Nutrient Applied

Overall, analysis of the LSMS data demonstrates that nitrogen fertilizer has positive and statistically significant effects on yields of dryland cereals. When we account for the endogeneity of fertilizer use, predicted magnitudes raise the range of elasticities to 17-23 for the three crops combined. For maize, in particular, these fall within the expected range. Ahmed et al. (2017) found no statistically significant effects of mineral fertilizer on crop yield. Foltz et al. (2012) estimated significant maize yield elasticities of 0.2-0.3 for fertilizer—higher than we found here. However, they used total fertilizer kg and utilized data only from the highly productive region of Sikasso.

The Sudan Savanna has the greatest agricultural potential in Mali to produce both sorghum and maize (Dicko et al., 2016). Despite this, anecdotally, farmers surveyed reported that sorghum yields were lower than expected due to declining soil fertility, moisture problems, and pest damage—encouraging a switch from sorghum to maize. The data indicate a very modest yield response rate to fertilizer for either crop. One reason why, in sorghum, could be the extremely low rate of application—application rates per ha on maize plots in our survey averaged 158 kg total of fertilizer, or 39.8 N nutrients/ha, compared with only 27 total kg of fertilizer on sorghum (6.4 N nutrients/ha). There are many zeros in our sample for sorghum (66%), compared with only 14% on maize plots.

Three response function specifications are shown in Table 8, each including soils characteristics measured in laboratory tests on samples. Model 8(1)⁴ is a simple linear regression, with sorghum plot entered only as a binary variable affecting overall yields. The effect on yield is strong and negative, reducing average grain yields by about 600 kg/ha relative to maize, controlling for other factors. Model 8(2) includes an interaction effect between N nutrient kg/ha and sorghum plot. The effect is negative but not statistically significant. In Model 8(3), the squared term is added for N applied and is negative in sign but not statistically significant. The interaction effect becomes significant, indicating that growing sorghum reduced yield response by 9 N nutrient kg/ha relative to maize. On average, Model 8(3) suggests that an additional kg of N nutrients per ha contributes 10.4 kg of maize grain per ha. This suggests a response rate of only about 1.3 for sorghum combined with the interaction effect.

Consistent across the three specifications, other coefficients of interest include a positive and significant effect of labor and equipment use and a negative and significant effect of distance to the plot and legume intercrop on yields. The magnitudes and significance of the input effects suggest that these may constrain productivity. This is supported by the negative effect on time walking to the plot from the homestead. The inverse relationship between yield and the legume intercrop is explained by the fact that we were unable to control for the area planted to primary and secondary crops—leading to a downward bias in the yield of the primary crop. On the other hand, any long-term, positive effects of intercropping would be difficult to discern in a single year’s survey data of this type. Similarly, erosion structures were often constructed earlier and may not be repaired. Most are stone contour lines to control erosion on slopes, but most of the plots in the sample are on the plain. The insignificance of the manure variable may reflect the fact that while most farmers apply manure (64%), there is considerable variation in the quantity and quality applied.

Among measured soil nutrients, the effect of P is strongly significant (at 1%), suggesting that it poses a constraint to productivity. In many of Mali’s sorghum-growing areas, we believe that phosphorus is a more binding constraint than N (Dicko et al., 2016; Kihara et al., 2016).

Table 9 shows models for maize and sorghum in combined and separate regressions. The combined results in Model 9(3) are similar to those shown in Table 4 (4), both in terms of response magnitudes and significance and in terms of other main inputs (labor, equipment) influencing productivity. However, they cannot be directly compared because of differences in sampling and covariates. The yield response rate for maize taken separately is 14.4 and significant but is insignificant at 3.6 for sorghum. Some differences appear in key factors across the regressions reported in Table 9. Statistically, the maize regression is statistically weaker, with far fewer observations than the sorghum regression.

Table 10 shows the coefficients from the second stage, yield response function estimated with the Control Function Approach. The yield regression is based on the combined maize-sorghum model with farmer-perceived soils classes to benefit from as many observations as possible. In the first-stage regression, which tests and controls for potential endogeneity of fertilizer use in yield response, both the coefficient on the binary variable indicating that the plot manager benefited from the fertilizer subsidy and the coefficient on the village proportion of plot managers who belong to a registered cooperative, are statistically significant at the 1% level. So too is the residual entered in the yield regression, failing to support exogeneity of fertilizer use in the yield response function. As was true in the LSMS estimates, the marginal product of nitrogen fertilizer rises meaningfully when we control for endogeneity.

The estimated response rates for maize reported for the Sudan Savanna fall within the range of other estimates for the same crops based on data collected from farmers’ fields in Sub-Saharan Africa. A review conducted by Yanggen et al. (1998) shows that maize’s response rates to nitrogen are generally lower in West Africa than in East and Southern Africa, with most in the 10-15 range. Based on nationally representative cross-sectional and panel datasets, Koussoubé and Nauges (2017) and Theriault et al. (2017) estimated a yield response rate of about 19 kg/ha to nitrogen on maize in Burkina Faso, respectively. By contrast, Marenya and Barrett (2009) estimated a marginal product of 40-44 kg/ha in Western Kenya, while Sheahan et al. (2013) reported marginal products ranging from 14 to 25 kg/ha across the agro-ecologies of Kenya. Xu et al. (2009) reported response rates for Zambia varied from under 10 to 30 kg/ha, with a median of 16.

Estimates of sorghum yield response to fertilizer are statistically insignificant. Analysis of trial data by Institut de Recherches Agronomiques Tropicales (IRAT) from 1978-82 in Burkina Faso showed experimental responses of 10.3 kg grain of sorghum per N nutrient kg, with much lower figures measured in farmers’ fields (Matlon, 1983). In an early review of literature on this topic, Yanggen et al. (1998) found that the marginal physical product of nitrogen nutrients in sorghum production was similar in Sub-Saharan Africa to other regions of the sorghum-producing world such as India, but were lower in West Africa, were most reported rates were in the 4-5 range. In a recent analysis conducted in Nigeria, Omonona et al. (2016) found response rates of only around 1 kg of sorghum in cereal-root crop and agro-pastoralist farming systems.

Table 10
Maize-sorghum Yield Response to Nitrogen Applied, Control Function Approach

	(1)
Nitrogen fertilizer	23.98***
	(5.035)
Residual, stage 1	-16.33***
	(3.916)
Sorghum plot	622.1***
	(236.6)
Sorghum plot x N nutrients/ha	-5.565*
	(3.110)
(N nutrients/ha)2	0.00795
	(0.0439)
Manure	180.7***
	(60.61)
Labor	3.365***
	(0.777)
Herbicides	-3.601
	(15.01)
Equipment	0.351***
	(0.101)
Distance to plot	-3.118*
	(1.816)
Soil erosion structure	115.5
	(70.55)
Legume intercrop	-524.4***
	(65.38)
Plain	-100.5
	(119.9)
Slope	-120.4
	(144.2)
Sandy	-10.72
	(76.59)
Silty	35.67
	(80.92)
Clayey	173.7**
	(77.86)
Rainfall	-0.54
	(0.979)
Constant	-18.04
	(262.7)
Observations	1,086
R-squared	0.424

Source: Authors, based on Sudan Savanna data. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Conclusion

In Mali, raising the production of dryland cereals (maize, millet, and sorghum) to improve food security must be achieved through higher yields rather than a further extension of cultivated area. Although maize yields on farms have increased in recent years, they are far below their potential. Generally, yields of millet and sorghum on farms have remained low. Inadequate use of inorganic fertilizer has been pinpointed as a cause of low agricultural productivity in these crops. To encourage fertilizer use and spur productivity, the Malian government has implemented a fertilizer subsidy program since the global food price crisis in 2008. Beginning in rice, the program now also targets dryland cereals. Yet, little is documented about the responsiveness of those crops to fertilizer under farmer’s conditions. This study aims to fill this gap by examining dryland cereal yield responses to fertilizer using two farm household datasets. The first, the LSMS-ISA, is nationally representative. The second, collected in the Sudan Savanna region, is representative of a relatively high-potential zone for sorghum and maize production.

We applied a combination of econometric techniques to control for potential endogeneity and check the robustness of the results. Four key findings emerge. First, it is important to control for endogeneity to avoid underestimating the effect of fertilizer use on yields. Second, soil texture and practices (anti-erosion structures) affect both yields and estimated effects of fertilizer. Third, sorghum yields have a lower response to fertilizer than maize yields. Fourth, dryland cereal yield responses to fertilizer are stronger in the Sudan Savanna region than nationally representative data, highlighting the importance of agroecological factors and farming systems. Together, these findings suggest that the use of mineral fertilizer can boost productivity, especially for maize, but in complementarity with other practices that reduce soil erosion and improve soil quality.

One key aspect that we have not addressed is the profitability of fertilizer use. Related work by Dicko et al. (2016), which supports the need for varied recommendations, also suggests that economic optima are generally lower than agronomic optima recommended by national programs. Given that maize yields do respond to fertilizer use, can inadequate application rates be explained by low economic incentives? How does the subsidy program affect economic incentives, and at what social cost? When response rates are so low, does it make sense to include millet and sorghum in the subsidy program? Further research is needed to tackle those important questions and make sound policy recommendations on the subsidy program and other mechanisms to promote agricultural intensification in Mali.

Acknowledgements

The authors thank anonymous reviewers for their comments and suggestions and gratefully acknowledge the financial support of the Bill & Melinda Gates Foundation under the project titled Guiding Investments in Sustainable Agriculture in Africa (GISAIA), and USAID/Mali under the project titled “Projet de recherche sur les politiques de sécurité alimentaire au Mali (PRePoSAM) awarded under the Food Security Innovation Lab’s Cooperative Agreement Number AID-688-A-16-00001.

References

Abate, T., Coulibaly, N., Menkire, A., & Wawa, B. (2015). Maize in Mali: Successes and opportunities. Drought Tolerant Maize in Africa Quarterly Bulletin, 4(1), 1-2.

Ahmed, M., Gaskell, J., & Gautam, M. (2017). Is there potential for dryland agriculture?  Some evidence from Mali. The World Bank.

Bazile, D., Dembélé, S., Soumaré, M., & Dembélé, D. (2008). Utilisation de la diversité varietale du sorgho pour valoriser la diversité des sols du Mali. Cahiers Agricultures, 17(2), 86-94.

Bishop, J., & Allen, J. (1989). The on-site costs of soil erosion in Mali. (Environment Department Working Paper, No. ENV 21). The World Bank.

Coulibaly, N. (2008). Fiche technique sur les variétés de maïs au Mali. Institut d’Economie Rurale.

Dicko, M., Koné, M., Traoré, L., Diakité, C.H., Kamissoko, N., Sidibé, B., Kouyaté, Z., Sogodogo, D., Dioni, L., Konaré, H., & Gakou, A. (2016). Optimizing fertilizer use within the context of integrated soil fertility management in Mali. In S. Wortmann & K. Sones (Eds)., Fertilizer Use Optimization in Sub-Saharan Africa. (pp 100-112). CCAB International.

Foltz, J., Aldana, U., & Laris, P. (2012). The Sahel’s silent maize evolution: Analyzing maize productivity in Mali at the farm-level. (Working Paper 17801). National Bureau of Economic Research. http://www.nber.org/papers/w17801.

Gebremedhin, B., & Swinton, S. M. (2003). Investment in soil conservation in northern Ethiopia: the role of land tenure security and public programs. Agricultural Economics, 29(1), 69-84.

Guirkinger, C., Platteau, J.-P., & Goetghebuer, T. (2015). Productive inefficiency in extended agricultural households: Evidence from Mali. Journal of Development Studies, 116(2015), 17-27.

Kazianga, H., & Wahhaj, Z. (2013). Gender, social norms, and household production in Burkina Faso. Economic Development and Cultural Change, 61(3), 539-576.

Kihara, J., Nziguheba, G., Zingore, S., Coulibaly, A., Esilaba, A., Kabambe, V., Njoroge, S., Palm, C., & Huising, J. (2016). Understanding variability in crop response to fertilizer and amendments in sub-Saharan Africa. Agriculture, Ecosystems, and Environment, 229, 1-12.

Kone, Y., Kergna, A., Traore, A., Keita, N., Theriault, V., Smale, M., & Haggblade, S. (2019). Regard sur les origines de la subvention des engrais au Mali. FSP Bulletin Politique 95. Michigan State University.

Koussoube, E., & Nauges, C. (2017). Returns to fertilizer use: Does it pay enough? Some new evidence from Sub-Saharan Africa. European Review of Agricultural Economics, 44(2), 183-210.

Macauley, H., & Ramadjita, T. (2015). Cereal crops: rice, maize, millet, sorghum and wheat (Feeding Africa Background Paper).

Marenya, P., & Barrett, C. (2009). State-conditional fertilizer yield response on western Kenya Farms. American Journal of Agricultural Economics, 91(4), 991-1006.

Mason, S. C., Ouattara, K., Taonda, S. J. B., Pale, S. B., Sohoro, A., & Kabore, D. (2014).

Soil and cropping system research in semi-arid West Africa as related to the potential for conservation agriculture. International Journal of Agricultural Sustainability. doi: 10.1080/14735903.2014.945319.

Matlon, P. (1983). The technical potential for increased food production in the West African semi-arid tropics. Paper presented at the Conference on Accelerating Agricultural Growth in Sub-Saharan Africa.

New Partnership for Africa’s Development (NEPAD). (2003). Comprehensive Africa Agriculture Development Programme. African Union and NEPAD.

Omonona, B. T., Liverpool-Tasie, L. S. O., Sanou, A., and Ogunleye, W. (2016). The profitability of inorganic fertilizer use in sorghum production: Evidence from Nigeria. Guiding Investments in Sustainable Agricultural Intensification in Africa (GISAIA).

Planning and Statistics Unit, Ministry of Rural Development, Mali. (2016). Basic Information Document. Enquête Agricole de Conjoncture Intégrée aux Conditions de Vie des Ménages.

Rattunde, H. F. W., Weltzien, E., Diallo, B., Diallo, A. G., Sidibe, M., Touré, A. O., Rathore, A., Das, R. R., Leiser, W. L., & Touré, A. (2013). Yield of photoperiod-sensitive sorghum hybrids based on Guinea-race germplasm under farmers’ field conditions in Mali. Crop Science, 53(November-December), 1-8.

Sheahan, M., Black, R., & Jayne, T. S. (2013). Are Kenyan farmer under-utilizing fertilizer? Implication for input intensification strategies and research. Food Policy, 41, 39-52.

Smale, M., Assima, A., Kergna, A., Traoré, A., & Keita, N. (2015). Survey research Report: diagnostic survey of sorghum production in the Sudanian Savanna (FSP Innovation Lab Working Paper No. Mali-2015-1). Michigan State University.

Smale, M., Diakité, L., & Keita, N. (2016). Location, vocation, and price shocks: Cotton, rice and sorghum-millet farmers in Mali. In M.J. Cohen & M. Smale (Eds.), Global food-price shocks and poor People (pp. 36-149). Development in Practice Books.

Sparks, D. L., Page, A. L., Helmke, P. A., Loeppert, R. H., Soltanour, P. N., Tabatabai, M. A., Johnston, C.T., & Sumner, M.E. (1996). Methods of soil analysis. Part 3 – Chemical methods. SSSA book series 5.

Tappan, G., & McGahuey, M. (2007). Tracking environmental dynamics and agricultural intensification in southern Mali. Agricultural Systems 94, 38-51.

Thériault, V., Smale, M., & Assima, A. (2018). The Malian fertilizer value chain post-subsidy: An analysis of its structure and performance. Development in Practice, 28(2), 242-256.

Theriault, V., Smale, M., & Haider, H. (2017). Maize yield response to fertilizer under differing agro–ecological conditions in Burkina Faso (International Development Working Paper 155). Michigan State University.

Traore, K., Sidibe, D. K., Coulibaly, H., & Bayala, J. (2017). Optimizing yield of improved varieties of millet and sorghum under highly variable rainfall conditions using contour ridges in Cinzana, Mali. Agriculture and Food Security 6, 11. doi: 10.1186/s4066-016-0086-0

Udry, C. (1996). Gender, agricultural production, and the theory of the household. Journal of Political Economy, 104(5), 1010-1046.

Weil, R. R., & Brady, N. C. (2016). The Nature and Properties of Soils (15^th edition). Pearson Education Limited.

World Bank. (2017). Database. http://databank.worldbank.org/data/home.aspx

Xu, Z., Guan, Z., Jayne, T. S., & Black, R. (2009) Factors influencing the profitability of ffertilizer use on maize in Zambia. Agricultural Economics, 40, 437-446.

Yanggen, D., Kelly, V., Reardon, T., & Naseem, A. (1998). Incentives for fertilizer use in sub-Saharan Africa: A review of empirical evidence on fertilizer response and profitability (MSU International Development Working Papers, No. 70). Department of Agricultural, Food, and Resource Economics, Michigan State University.

ENDNOTES

[1] Pers. Comm. Assitan Traoré, Cellule de Planification Cellule de Planification et de Statistiques du Secteur #sdfootnote1ancDéveloppement Rural (CPS/SDR), pers. Comm, June 15, 2017 and November 15, 2017.

[2] Accessed from http://power.larc.nasa.gov/cgi-bin/cgiwrap/solar/agro.cgi?email=agroclim@larc.nasa.gov

[3] These figures are calculated by multiplying the estimated coefficients by the average yield and fertilizer quantities of the regression sample.

[4] The first number refers to table and second number refers to model.