Data compilation

We compiled a large data set on maize grain yield (in t ha⁻¹) with N, P and K fertilizer and the corresponding yield from the control (without any external nutrient inputs) on 542 sites across 23 countries (Table 1) accounting for over 90% of the maize area harvested in SSA. The sites covered all agroecological zones (Figure 1a) and most of the soil types on which maize is grown (Table 1; Figure 1b). We retrieved the data from trials conducted under 23 projects supported by the Alliance for a Green Revolution in Africa, eight legacy data sets, and 147 peer-reviewed studies (Supplementary Table S1). The legacy data sets included trials implemented by the Food and Agriculture Organization (FAO), the World Agroforestry Centre (ICRAF) and the International Center for Tropical Agriculture (CIAT) across SSA. When extracting data, we only included studies that used improved cultivars of maize grown with fertilizer and the same cultivar grown without fertilizer. We also included data only when yield data came from N application rates of at least 48 kg N ha⁻¹ and 10 kg P ha⁻¹ on elemental basis. These minimum N and P rates were required because levels below those were deemed to be inadequate to satisfy the requirement of a maize cultivar with yields exceeding 3 t ha⁻¹. Improved cultivars of maize on average use 24 kg N ha⁻¹ and 12 kg P ha⁻¹ to produce 1 t ha⁻¹ of maize grain (Zingore et al., Reference Zingore, Njoroge, Chikowo, Kihara, Nziguheba and Nyamangara2014). Therefore, N and P rates below these will not have enabled us to detect the yield response of improved maize cultivars. The 3 t ha⁻¹ yield target corresponds to the minimum required to kick start the African green revolution (Sánchez, Reference Sánchez2015). In all trials, all agronomic practices, including weeding, planting date, density, were applied optimally following the recommended practices for the site.

Table 1.

Soil type, the number of countries and sites covered in this analysis (details in supplementary material tables S1), total number of observations (N), resilience to erosion, clay activity, P fixation capacity, clay content, sand content, cation exchange capacity (CEC in cmol_c kg^-1), base saturation percentage (BSP), pH and soil organic carbon (SOC) content

NA = not available.

* Site represents geographic areas at district or comparable level, where several farms were covered.

† Clay activity classes are as defined in Jones et al. (Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013).

‡ P fixation categories are as defined in Batjes (Reference Batjes2011).

^§ Resilience to degradation following Stocking (Reference Stocking2003) and IUSS (2014).

^‖ Parent material is defined in Supplementary Table S2.

Figure 1.

(a) Map of agroecological zones and trial sites across Africa and (b) the percentage area of continental Africa covered by each of the soil types included in this study (Data from Jones et al., Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013).

For each trial site, we extracted grain yield, the amount of N, P and K fertilizer applied (in kg ha⁻¹), the geographic coordinates, agroecological zones (AEZ), elevation and information on soil type where available. Across the sites, the median N rate, P rate and K rate were 80, 20 and 25 kg ha⁻¹, respectively. The K fertilizer was applied only on 45.9% of the sites, and therefore our primary focus is on N and P response. On some of the sites where K fertilizer was applied, the rates were not mentioned. Therefore, we created an additional category for K application coded as ‘applied’ or ‘omitted’. Soil pH, SOM, sand and clay content of the till layer (0–20 cm depth) of the trial sites were also extracted from the studies as these are considered to be fundamental indicators of the agricultural potential of a soil (Gray et al., Reference Gray, Humphreys and Deckers2009). Since some studies had reported soil organic carbon (SOC), to facilitate analysis we converted all SOM to SOC by dividing SOM by 1.72. We have presented the SOC, pH, clay and sand content of the different soil type in the form of box and whisker plots in Supplementary Figures S1 and S2. In addition, we extracted the base saturation percentage (BSP), cation exchange capacity (CEC), pH and SOC content from ISSS (1998). We also extracted information about clay activity of the different soil types from Gray et al. (Reference Gray, Humphreys and Deckers2011) and ISSS (1998).

The soil type strictly refers to the WRB Reference Soil Groups (RSG) and the Soil Atlas of Africa (IUSS, 2014; Dewitte et al., Reference Dewitte, Jones, Spaargaren, Breuning-Madsen, Brossard, Dampha, Deckers, Gallali, Hallett, Jones, Kilasara, Le Roux, Michéli, Montanarella, Thiombiano, van Ranst, Yemefack and Zougmore2013; Jones et al., Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013). The third edition of the WRB published in 2014 (IUSS, 2014) provides the most up-to-date classification of soils of the world. This edition has replaced Albeluvisols of earlier edition with Retisols and Lithosols with Leptosols. It has also broadened the definition of Gleysols and Arenosols and narrowed the definition of Acrisols, Alisols, Luvisols and Lixisols (IUSS, 2014). Therefore, we assigned older soil names and those given in other systems to the updated WRB RSG through pro parte matching. When soil type was not given for a site, we extracted the information from the harmonised Soil Atlas of Africa (Jones et al., Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013) using the locational information. Accordingly, we identified a total of 14 RSGs from the trial sites (Table 1; Supplementary Table S1). These are briefly described below following the mineralogy of their parent materials (Supplementary Table S2) and soil-forming processes as described in Gray et al. (Reference Gray, Humphreys and Deckers2011) and (IUSS, 2014). Based on their parent materials, soils are often classified as calcareous, silicic or mafic (Gray et al., Reference Gray, Humphreys and Deckers2011; Jones et al., Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013). Calcareous parent materials include limestone, dolomite, calcareous shale and sands with >50% CaCO₃ or MgCO₃ (Gray et al., Reference Gray, Humphreys and Deckers2011). Silicic (also called acidic) refers to rocks that contain significant amounts of silica (>68% Si), while intermediate parent materials contain 52–68% Si. The term mafic (basic) is applied to rocks with relatively low amounts of silica (45–52% Si) (Gray et al., Reference Gray, Humphreys and Deckers2011). The 14 RSGs of the trial sites in alphabetic order are Acrisols, Alisols, Andosols, Arenosols, Cambisols, Ferralsols, Fluvisols, Leptosols, Lixisols, Luvisols, Nitisols, Phaeozems, Plinthosols and Vertisols together covering 85% of the land area of continental Africa (Figure 1b).

Acrisols are weathering products of a wide variety of intermediate parent material under humid climates (Gray et al., Reference Gray, Humphreys and Deckers2011). They typically occur on old land surfaces with hilly or undulating topography (Gray et al., Reference Gray, Humphreys and Deckers2011; IUSS, 2014). They are acidic, strongly leached and have low base status (Table 1). Acrisols also have low resilience to erosion, they become degraded very quickly (Stocking, Reference Stocking2003) and lose their productivity rapidly when cultivated without adequate soil cover (Sileshi et al., Reference Sileshi, Akinnifesi, Debusho, Beedy, Ajayi and Mong’omba2010). They require substantial applications of fertilizer to produce satisfactory crop yields.

Alisols evolve from intermediate parent material under humid climates in hilly or undulating topography (Gray et al., Reference Gray, Humphreys and Deckers2011; IUSS, 2014). They are characterised by very low pH and SOC content (Table 1; Supplementary Figure S1), high-activity clay rich in exchangeable Al (IUSS, 2014) and Al toxicity. They are also prone to drought and have an unstable surface, which makes them susceptible to erosion (IUSS, 2014).

Andosols are relatively young soils formed from glass-rich volcanic ejecta and pyroclastic materials (IUSS, 2014). Andosols may also develop from silicate-rich materials under acid weathering in humid climates. They occur in undulating to mountainous regions. They have high SOC content, high pH, high clay activity (Table 1; Supplementary Figure S1) and are generally fertile. But they have the highest P fixation capacity, which is caused by active Al and Fe and their amorphous clay (allophane) mineralogy (Batjes, Reference Batjes2011; IUSS, 2014).

Arenosols are deep sandy soils that develop on extremely silicic (>85% Si) parent materials. They can also evolve from weathering of quartz-rich sediments or from Ferralsols after kaolinite weathering (IUSS, 2014). They develop under a wide range of annual rainfall, varying from very dry (<500 mm) to extremely wet (>2000 mm) climates (Gray et al., Reference Gray, Humphreys and Deckers2011). They represent the largest soil type covering 21.5% of the land area of Africa (Figure 1b). Arenosols are characterised by high permeability, low water and nutrient storage capacity (IUSS, 2014) and low SOC content (Supplementary Figure S1). They are also easily erodible.

Cambisols are derived from a wide range of intermediate parent material of medium to fine texture under very dry to wet climates (Gray et al., Reference Gray, Humphreys and Deckers2011). They represent the third largest soil type covering 10.8% of the land area of Africa (Figure 1b). They are relatively young soils with slight to moderate weathering (IUSS, 2014). Cambisols are characterised by high resilience, high SOC content, high clay activity and low P fixation (Table 1; Supplementary Figure S1). They also do not contain appreciable amounts of Al and Fe compounds (IUSS, 2014).

Ferralsols represent deeply weathered red or yellow soils of the humid tropics. They develop on intermediate silicious parent material under very wet climates on level to undulating land (Gray et al., Reference Gray, Humphreys and Deckers2011). Ferralsols represent the fourth largest soil type covering 10.3% of the land area of Africa (Figure 1b). They are dominated by low-activity clay minerals, mainly kaolinite (IUSS, 2014). They have good physical properties including great soil depth, permeability and stable microstructure making them less susceptible to erosion (IUSS, 2014). However, they are inherently acidic (Table 1; Supplementary Figure S1), nutrient-poor and suffer from high P fixation caused by Fe and Al oxides (IUSS, 2014).

Fluvisols are young soils in fluvial, lacustrine or marine deposits. They have good natural fertility because they occur in periodically flooded areas receiving fresh sediment (IUSS, 2014). The have high SOC content, high pH, high clay activity and base saturation (Table 1; Supplementary Figure S1).

Leptosols are very thin soils that are extremely rich in coarse fragments over continuous rock (IUSS, 2014). The have low SOC content, high pH and high base saturation (Table 1; Supplementary Figure S1). They occur in all climate zones, in particular in strongly eroding mountainous regions. They are the second largest soil type in SSA, and they cover 17.5% of the land area of continental Africa (Figure 1b). Erosion is the greatest threat to Leptosols in cultivated areas.

Lixisols are weathering products of highly silicic parent material (65–85% Si) in seasonally dry tropical regions on flat land to sloping land (Gray et al., Reference Gray, Humphreys and Deckers2011). They form in a variety of strongly weathered and fine textured materials (IUSS, 2014). Lixisols have low resilience to erosion. When degraded, they are prone to slaking and erosion. Tillage and use of heavy machinery compact the soil causing structural deterioration (IUSS, 2014).

Luvisols evolved from a wide variety of intermediate silicious parent material in regions with distinct dry and wet seasons (Gray et al., Reference Gray, Humphreys and Deckers2011; IUSS, 2014). They are characterised by high-activity clays and a high base saturation (Table 1). Most Luvisols are fertile and suitable for a wide range of agricultural uses. However, Luvisols with a high silt content are susceptible to structural degradation when tilled with heavy machinery, while Luvisols on steep slopes are prone to erosion (IUSS, 2014).

Nitisols are deep, well-drained and clay-rich soils formed of mafic parent material under wet to very wet and warm conditions on level to hilly land (Gray et al., Reference Gray, Humphreys and Deckers2011; IUSS, 2014). Their clay is dominated by kaolinite and/or (meta) halloysite. Nitisols have good internal drainage and fair water holding properties. Nitisols also have clay content >30% and relatively high SOC content in surface soils (Supplementary Figure S1). Their deep and porous soil and their stable structure permit deep rooting and make them resistant to erosion (IUSS, 2014). However, they suffer from high P fixation (Batjes, Reference Batjes2011) and acidification (Agegnehu et al., Reference Agegnehu, Amede, Erkossa, Yirga, Henry, Tyler, Nosworthy, Beyene and Sileshi2021).

Phaeozems are weathering products of mafic parent material in warm to cool climates on flat to undulating land (Gray et al., Reference Gray, Humphreys and Deckers2011). They have dark, humus-rich surface horizons, and porous and fertile soils. But wind and water erosions are serious hazards (IUSS, 2014).

Plinthosols are soils with plinthite, petroplinthite or pisoliths. The later develop from plinthite by hardening. Plinthite is a Fe-rich and/or Mn-rich, humus-poor mixture of kaolinitic clay and other products of strong weathering such as gibbsite and quartz (IUSS, 2014). Plinthosols occur on level to gently sloping areas with seasonally fluctuating groundwater or stagnating surface water. They have poor natural soil fertility caused by strong weathering, and they are prone to concretions or formation of hardpan (IUSS, 2014). They also have very low SOC content (Supplementary Figure S1).

Vertisols are heavy clay soils with a high proportion of swelling clays. Smectite is the major clay mineral, while kaolinite is of secondary importance. Vertisols are weathering products of mafic parent material (Gray et al., Reference Gray, Humphreys and Deckers2011) occurring in depressions and level plains in climates varying from semiarid to humid with an alternation of distinct wet and dry seasons (IUSS, 2014). Although they have good chemical fertility, their use is limited by their susceptibility to waterlogging when wet and moisture stress when dry.

Criteria and metrics

In the past, non-responsiveness was evaluated using different metrics (e.g., yield gain vs response ratios). Kihara et al. (Reference Kihara, Nziguheba, Zingore, Coulibaly, Esilaba, Kabambe, Njoroge, Palm and Huising2016) separated non-responsive soils into two clusters based on the yield in the fertilized plots: non-responsive poor (<3 t ha⁻¹) and non-responsive fertile soils (>3 t ha⁻¹). Similarly, Njoroge et al. (Reference Njoroge, Otinga, Okalebo, Pepela and Merckx2018) defined soils as responsive when yield increases by 2 t ha⁻¹ above the control. On the other hand, Roobroeck et al. (Reference Roobroeck, Palm, Nziguheba, Weil and Vanlauwe2021) used yield gain of <0.5 t ha⁻¹ to categorise farms as non-responsive to NPK fertilizer in maize and <0.15 t ha⁻¹ in soybean. Ichami et al. (Reference Ichami, Shepherd, Sila, Stoorvogel and Hoffland2019) used the response ratio (RR) to categorise soils as responsive and non-responsive to fertilizer. To achieve a more rigorous classification, we applied two criteria, namely, mean effects and probabilities of no response. We also distinguished between metrics of (1) overall non-responsiveness and (2) non-responsiveness to specific nutrients. We used the RR as defined by Ichami et al. (Reference Ichami, Shepherd, Sila, Stoorvogel and Hoffland2019) and yield gain of <0.5 t ha⁻¹ as defined by Roobroeck et al. (Reference Roobroeck, Palm, Nziguheba, Weil and Vanlauwe2021) to judge the overall non-responsiveness of soils to NPK fertilizer. The RR is the ratio of some measured quantity in experimental and control groups, which quantifies the proportionate change that results from an experimental manipulation (Hedges et al., Reference Hedges, Gurevitch and Curtis1999). We calculated RR by dividing the treatment yield by the control yield and. We transformed the RR values to their natural logarithm, i.e., ln(RR), to satisfy the assumptions of normality and homogeneity of error variance (Hedges et al., Reference Hedges, Gurevitch and Curtis1999).

To quantify the lack of response to one nutrient, we used the agronomic use efficiency of N (AEN), P (AEP) and K (AEK) as the effect size metrics. Negative values of AEN and AEP are common in empirical studies (Jama et al., Reference Jama, Kimani, Kiwia, Rebbie and Sileshi2017; Kihara et al., Reference Kihara, Nziguheba, Zingore, Coulibaly, Esilaba, Kabambe, Njoroge, Palm and Huising2016; Vanlauwe et al., Reference Vanlauwe, Kihara, Chivenge, Pypers, Coe and Six2011), but these are often thought to be inexplicable (Kihara and Njoroge, Reference Kihara and Njoroge2013; Vanlauwe et al., Reference Vanlauwe, Kihara, Chivenge, Pypers, Coe and Six2011). We calculated AEN, AEP and AEK following Ladha et al. (Reference Ladha, Pathak, Krupnik, Six and van Kessel2005) as

(1) $${\rm{AEN}} = \left( {{{\rm{Y}}_{\rm{t}}} - {{\rm{Y}}_{\rm{c}}}} \right)/{{\rm{N}}_{\rm{i}}}$$

(2) $${\rm{AEP}} = \left( {{{\rm{Y}}_{\rm{t}}} - {{\rm{Y}}_{\rm{c}}}} \right)/{{\rm{P}}_{\rm{i}}}$$

(3) $${\rm{AEK}} = \left( {{{\rm{Y}}_{\rm{t}}} - {{\rm{Y}}_{\rm{c}}}} \right)/{{\rm{K}}_{\rm{i}}}$$

where Y_t and Y_c are the grain yields (in kg ha⁻¹) in the treatment and the control, whereas N_i, P_i and K_i represent N, P and K inputs (in kg ha⁻¹), respectively. We defined ‘no response’ as zero agronomic response, i.e., a case where the yield differences between fertilized and unfertilized treatments is 0 or lower than 0 (Nziguheba et al., Reference Nziguheba, van Heerwaarden and Vanlauwe2021) resulting in RR ≤ 1 (Ichami et al., Reference Ichami, Shepherd, Sila, Stoorvogel and Hoffland2019) or AEN ≤ 0, AEP ≤ 0 or AEK ≤ 0. We also took AEN, AEP and AEK below their first quartiles as indicative of “low response” to N, P and K. In the data analysed here, the first quartiles were 10, 30 and 20 kg kg^-1 for AEN, AEP and AEK, respectively. AEN < 10 kg kg^-1 is also considered too low (Vanlauwe et al., Reference Vanlauwe, Kihara, Chivenge, Pypers, Coe and Six2011). Accordingly, we took AEN ≤ 10, AEP ≤ 30 and AEK ≤ 20 as additional criteria to represent “low response”.

Statistical analysis

We applied exploratory analysis using general linear models to evaluate the relative contribution of agroecological zones, soil type, trial location (on research station vs farmers’ fields) and K application (applied vs omitted) to the variations in RR, AEN, AEP and AEK. Then, we calculated the percentage of variation explained from the type 3 sum of squares (Supplementary Table S3). Soil type contributed the largest proportion of the explained variations in RR, AEN, AEP and AEK, whereas the contributions of agroecological zone and trial location were very small (Supplementary Table S3). In addition, we implemented multiple regression analysis to evaluate the contribution of SOC, pH, clay content, N rate, P rate and K rate to the variations in RR, AEN, AEP and K. These variables together explained only ∼7.8, 2.8, 15.2 and 20.3% of the variations in RR, AEN, AEP and AEK, respectively (Supplementary Table S4). SOC did not significantly contribute to the variation in RR and AEP and AEK (Supplementary Table S4). Although soil pH and clay content had statistically significant effects, their partial R ² values were <0.05 (Supplementary Table S4). Given the relatively small contributions of soil pH and SOC to the various metrics, we focused our attention on soil type. We undertook two sets of analyses to get a deeper understanding of the variations in non-responsiveness.

Our first set of analyses involved linear mixed effects models (LMM) implemented via restricted maximum likelihood (REML) estimation in the SAS system. To accommodate imbalances in sample sizes among predictors, we used the Kenward–Roger method for approximating the denominator degrees of freedom (Spilke et al., Reference Spilke, Piepho and Hu2005). Initially, we tested various models by adding or removing variables until an optimal model was achieved. At each step, we examined improvements in model fit using the Akaike Information Criterion (AIC). We used agroecological zone (A), soil type (S), trial location (L), N fertilizer rate (N), P rate (P), K application (K), soil clay content, pH and SOC content as the fixed effects. We entered country as the random effect because data from one country are likely to be more correlated than data from another country. We calculated the intraclass correlation coefficient (ICC) as the ratio of the covariance estimate of the random effect (i.e., country) to the covariance estimate of the residual + random effect. The ICC estimates how much of the total variation in the outcome is accounted for by the random effect. Our initial analyses revealed that entering soil pH, SOC and clay content in the same model with soil type was problematic for two reasons: (1) pH, clay and SOC measurements were available for only 29–30% of the sites and (2) most soil type have characteristic soil pH, SOC and clay contents that vary in a narrow range (Table 1; Supplementary Figure S1). Taking these facts into consideration, we included a reasonable set of variables (Supplementary Table S5) in our mixed models for the control yield (CY), treatment yield (TY), yield gain (YG), RR, AEN, AEP and AEK specified as follows:

(4) $${\rm{C}}{{\rm{Y}}_{ijk}} = {A_i} + {S_j} + {L_k} + {\epsilon _{ijk}}$$

(5) $${\rm{T}}{{\rm{Y}}_{ijklmn}} = {A_i} + {S_j} + {L_k} + {N_l} + {P_m} + {K_n} + {\epsilon _{ijklmn}}$$

(6) $${\rm{Y}}{{\rm{G}}_{ijklmn}} = {A_i} + {S_j} + {L_k} + {N_l} + {P_m} + {K_n} + {\epsilon _{ijklmn}}$$

(7) $${\rm{ln}}{({\rm{RR}})_{ijklmn}} = {A_i} + {S_j} + {L_k} + {N_l} + {P_m} + {K_n} + {\epsilon _{ijklmn}}$$

(8) $$AE{N_{ijklm}} = {A_i} + {S_j} + {L_k} + {P_l} + {K_m} + {\epsilon _{ijklm}}$$

(9) $$AE{P_{ijklm}} = {A_i} + {S_j} + {L_k} + {N_l} + {K_m} + {\epsilon _{ijklm}}$$

(10) $$AE{K_{ijklm}} = {A_i} + {S_j} + {L_k} + {N_l} + {P_m} + {\epsilon _{ijklm}}$$

To avoid confounding of AEN, AEP and AEK with N rate, P rate and K application, we omitted the respective nutrient rates from equations 8, 9 and 10.

We implemented a second set of analyses with the premise that information on the probability (ϕ) of falling below a critical level or exceeding a desired target yield is more useful than mean values for decision making. Here, we implemented stepwise logistic regression of the forms given in equations 6–10 using the same variables used in the LMM (Supplementary Table S6) to predict the probabilities of no response in terms of YG, RR, AEN, AEP and AEK. We analysed binary responses after dummy-coding the response variables. For example, we coded YG as ‘0’ when YG values are <0.5 t ha⁻¹ and ‘1’ when they exceed 0.5 t ha⁻¹. We coded RR as ‘0’ when its values are ≤1 and ‘1’ when they exceed 1. Similarly, we coded AEN, AEP and AEK as ‘0’ when their values are ≤ 0 and ‘1’ when they exceed 0. We implemented stepwise logistic regression using the Logistic procedure of the SAS system. Using the best model thus selected, we generated the predicted probabilities (ϕ) for each observation, and then ϕ for each soil controlling for the other independent variables retained in the final model. Our analytical procedure uses the average marginal effects (AME) of both categorical and continuous predictors to estimate individual, cross-validated, predicted probabilities. The AME was preferred over other marginal effects as it represents the effect for a case picked at random from the sample (Breen et al., (Reference Breen, Karlson and Holm2018) and as such being the most representative. For the continuous predictors (e.g., N rate, P rate), the AME is the partial derivative of the event probability (i.e., no response) with respect to the predictor of interest. For the categorical predictors (e.g., agroecology, soil type, location, K application), the AME is the change in event probability when each predictor is changed between its levels. Accordingly, we estimated the probability of ‘no response’ as well as ‘low response’ as defined above.

In the same manner, we also estimated ϕ of the various metrics for the different classes of parent materials, resilience to erosion, clay activity and P fixation capacity of soils. Here, we replaced soil type with each of these variables in equations 6–10. Following Gray et al. (Reference Gray, Humphreys and Deckers2011), we grouped parent materials of the soils on the trial sites into “silicic,” “intermediate” and “mafic” before analysis (Table 1). None of the study sites were on calcareous parent material. We also grouped the soil types based on their clay activity as “high” and “low” activity soils (Table 1). High-activity clay minerals are less weathered, have high effective surface area and high cation exchange capacity (CEC) ≥ 24 cmol/kg clay with 2:1 clay lattice (2 silica sheets per 1 alumina sheet, e.g., montmorillonite, smectite, vermiculite, illite and chlorite) or amorphous (non-crystalline) materials like allophane, inmogolite and halloysite. Low-activity clay minerals are highly weathered and have low CEC (≤ 24 cmol/kg clay) and 1:1 lattice (one silica sheet per alumina sheet, e.g., kaolinite or halloysite) (Jones et al., Reference Jones, Breuning-Madsen, Brossard, Dampha, Deckers, Dewitte, Hallett, Jones, Kilasara, Le Roux, Micheli, Montanarella, Spaargaren, Tahar, Thiombiano, Van Ranst, Yemefack and Zougmore2013). In addition, we grouped the soil types into classes of resilience to erosion as “low” or “high” following Stocking (Reference Stocking2003) and P fixation capacity as “low,” “moderate” or “high” following Batjes (Reference Batjes2011) before analyses.