Average Oga treatment effects in the on-farm trials
In all 3 years and across all on-farm trials conducted, the treatment (Oga + OM in 2014 and 2015, Oga solely in 2016) resulted in a positive and significant yield effect (2014: +254 kg ha−1, 2015: +193 kg ha−1, 2016: + 288 kg ha−1, Fig. 3).
The control yield differed not only due to inter-annual climate differences (Fig. 2) but also due to the fact that different sites and villages contributed to the testing in the different years (Table 1). The low coefficient of variation in 2015 in contrast to 2014 is probably forced by the fact that only sites in Maradi region were sampled. Since the average treatment effect is positive and of similar magnitude in all 3 years, we can assume that the major effect is triggered by the Oga application that delivers, in contrast to OM, readily available, i.e., water-soluble nutrients. This conclusion is supported by a multivariate analysis, where the treatment type appeared as insignificant variable (F value = 1.870, Prob. = 0.173).
In tendency, the relative yield effect of the Oga (+OM) application decreased with increasing control yield (Fig. 4, Table 2 sites Serkin Haussa and Say). Since a 3-year data were used here that represent the climatic variability, the control yield is interpreted as a major function of the chemical soil fertility. However, factors such as planting time and previously grown crops (legumes vs cereals) may affect the productivity levels in general.
Only in the relatively high yielding Dan Tsunsu village in 2014 (i.e., the two data points each with an arrow), the average Oga (+OM) treatment effect was insignificant. In all other village-year combinations, the Oga (+OM) technology was, on average, efficient and did neither present a risk for the yield nor for the monetary investment. It appears that up to a level of about 1000 kg ha−1 panicle yield, the Oga technology can be recommended without hesitation. Above this threshold, the yield effect is facultative and other fertilizer strategies might be more promising.
This view is deepened by Fig. 5, presenting the single observation points. While in 2014 still a relative high number of trials ended up with a failure (24 %), this changed with increasing experience of the animators and participating women farmers in on-farm testing to a negligible amount (≤ 2% in 2015–2016).
The major problem in 2014 was that not all women farmers strictly followed the experimental protocol and applied to the control other additional management measures than to the treatment plot. This has to be taken into account when interpreting the 2014 data. Looking at the 2015 and 2016 data (Fig. 5), the rare cases where the Oga (+OM) treatment effect was significantly negative appear to be in the higher yielding environments (> 1000 kg ha−1 panicle yield).
The 3-year aggregated means show for Serkin Haussa (driest site, wide-spread sandy soils) the lowest average panicle yield (Table 2). Control yield and absolute yield effect of the OGA (+OM) treatment decrease consistently from 2014 to 2016. However, the relative treatment effect increases in the same course from +26% to +38%.
The Safo site farther south receives more rainfall and has, due to geological reasons, more fertile soils. In contrast to Serkin Haussa, all yield averages are higher, and absolute treatment effects increased from 2015 to 2016 with higher yield differences between villages in 2016. The relative yield increment is stable with about 40% difference in both years (Table 2).
The site Say lies about 500 km away. In contrast to the Maradi region (major ethnic group Haussa), settlers represent here a mixture of Djerma and Peulh people. The latter have easier access to animal manure due to their origin as herders. Soil conditions appear more variable. Climate is more comparable to Safo. Remarkable is the high yield for the Tchela village in 2014. Otherwise, the yield range is comparable to the Maradi region.
With respect to data variability, the sites distinctively differ. This is presented in Fig. 6 using the 2016 data. Bokki represents an example with a consistent low control yield on all fields, a distinct treatment effect, and a complete separation of control and treatment yields with small data spread within the treatments (Table 2). Arraouraye data, in contrast, show a distinctly smaller treatment effect but a higher spread within the treatments. Garin Mai Gari finally exhibits a high control yield, a distinct treatment effect, and a large absolute as well as relative data spread within the treatments.
Neither the number of observation points (Table 2) nor the soil diversity can directly explain these differences. Although the principally more fertile soil type Damba (in general higher Corg, Nt, and CEC) occurs in Garin Mai Gari, it did not produce higher yields than the other soil types in that village. However, a study by Geiger (2017) showed that pearl millet control plot yields in the Maradi region were mainly and significantly explained by village (Spearman r = 0.77) and local soil type (Spearman r = 0.41).
Soil as a factor
Looking at the global annual averages, the treatments lead for most soil types to an absolute yield increase in the order of 200–300 kg ha−1 (Table 3).
Judging 2014 as an average year with respect to rainfall, the control yields confirm the farmer judgment on productivity of the soils for pearl millet, i.e., Jigawa > Gueza > Jampali. In that year, the treatment effect was highest on the soil with the lowest fertility (Jampali, Table 3). On the more loamy Gueza, soil treatment effects were positive but the lowest observed. In the more difficult year 2016, the treatment effects ranged between 244 and 323 kg ha−1. In this case, the lowest absolute yield increase was recorded for the local soil type Jampali. Stating this, we must take into account that the total number of observation points as well as their relative shares changed between the years. This fact leads to limited interpretability and is a typical shortcoming of trials in diverse and uncontrollable environments.
With respect to soil effects in 2016 (Fig. 5), GuezamiGuezami and Jambali have an absolute yield limit at about 1000 kg ha−1. Damba yields start where the former ones cease. The most frequent soils—Jigawa and Gueza—show an extreme widespread.
Another observation is that the treatment effect converges to about 30% panicle yield increase towards a higher number of observations (Fig. 7). This convergence effect is also visible for single soil types where a 3-year data are available. The highest relative effect was achieved with the presumably least fertile soil (i.e., Jampali). Gueza, as preferred sorghum cropping site, shows lower relative yield increase than the other local soil types. The observations achieved (N) somehow reflect the surface share of soils present in the intervention area, i.e., Jigawa > Gueza > GuezamiGuezami > Jampali > Damba. These reasons in combination explain the convergence effect. A shortcoming was that not all farmers or animators were able to deal with the soil names. To that for a share of the data, yield information was collected but no soil names stated. Looking at the specific soil-related data, we conclude that, per soil type, a minimum of 50 observations should be collected. This is hardly achievable in most R4D environments.
Anyhow, this average of 30% yield increase matches with data from other large-N trials (Nwankwo et al.—submitted) on low-level nutrient supplementation via seedballs to pearl millet in Niger.
The effect of other contextual variables
Age group, too, has an influence on the yield level (Fig. 8). Elderly women farmers without assistance harvest the lowest average yield and never attain the highest yield levels. This can be easily explained by the lower resource endowment, the lower available work force, and lower educational level. In contrast, young and elderly women farmers with assistance show a wide spread of yields, with the younger ones usually staying ahead.
Additional significant yield effects could be determined for the type of weed management (Fig. 8). Partial weeding, i.e., weeding only around the planting pockets during the first weeding phase (about 2 weeks after sowing), showed lower control yields in 2014 but an over-compensation under the Oga + OM treatment, leading to higher yields than the complete weeding variant in both observation years. It appears that without nutrient amendment, the remaining weeds in the partial weeding plots exhibit a competition effect in years with lower rainfall.
Once nutrients are locally applied to the planting pocket to which the weeds have limited or no access, there can be a positive weed effect in the form of erosion control and seedling protection against sand storms, reducing the damaging effect of saltating sand grains on the young and tender seedlings (Michels et al. 1995) so that the latter can make full use of the fertilizer.
Agronomic lessons learned from the field trials
Reported yield levels are within the expected order of magnitude for on-farm trials (e.g., Bationo et al. 1992). After a first year (2014) of on-farm trials with highly variable Oga (+OM) treatment effects of ca 100 kg (Fig. 5), continued farmer training on how to correctly apply the trial protocol, treatment effects were throughout positive in the second and third year, i.e., in 2015 and 2016. Although also in the latter years’ treatment effects showed certain variability, the global panicle yield increase ranged from +170 to +313 kg ha−1 for the final 2016 season. Although no other study on human urine application to pearl millet in Niger is scientifically documented, sheep urine application to pearl millet in Niger (Powell et al. 1998) and human urine application to sorghum in Ghana (Germer et al. 2011) in researcher trials showed even higher yield effects. Comparing the yield effects of sole Oga (applied in 2016) with combined application of OM and Oga (applied in 2014 and 2015) reveals that Oga is responsible for the main yield effect. This does not mean that solid organic manure application should be neglected, since SOM values are in general low (< 1 wt%) in the surveyed soils (Geiger 2017) and plant available phosphorus is absolutely deficient (≤ 4 mg kg−1 Bray1 P). The advantage of Oga application is, however, that the nutrients applied are directly available—in contrast to the ones in OM—and that the application of the fluid might help the plant during dry spells, as they often occur in the early season (Sivakumar 1988). In short, Oga application does not only supply nitrogen but also water and thus provides two important factors for crop survival under drought conditions. An additional plus can be the provision of potassium (K) with Oga, since this element is important for cell turgor stabilization and in consequence water use efficiency.
There seems to be a geographical trend with respect to the yield level. The farther south the intervention zone the higher the yield level (Serkin Haussa ≤ Safo, Say). This trend can be mainly explained by climatic effects. The average annual rainfall and vegetation period length increase southward; this in turn increases the potential for biomass production. At least in the Maradi region, farmers in the southern intervention zone are perceived as the wealthier ones (source: farmer organization’s own database on social stratification), having more financial resources for agricultural inputs like fertilizer and having at the same time easier access to these by the close-by frontier to Nigeria and the markets situated there. In addition, in the south, more “Dallol soils” are present stemming from ancient and recent river valleys, leading, i.a., to groundwater access in lower depth.
In tendency, the relative treatment effect depends on the control yield level. The lower the control yield level the higher the relative treatment effect. This can be mainly explained by the multifold nutrient limitations detected in Sahelian soils (Voortman 2010). In contrast to mineral fertilizers, organic fertilizers are characterized by a multi-nutrient content. However, even when applying those, biomass production is finally limited by water deficiency (Hofer et al. 2017).
Pearl millet yields showed a clear seasonal effect. This is to be expected in the Sahel that is generally characterized with highly variable rainfall and frequent intra-seasonal droughts. Taking the climate station at ICRISAT Sahelian Center (Sadoré) as indicative for the intervention site—Say in SW—Niger, 2015 was a normal year with evenly distributed rainfall; 2014 had below average rainfall and extreme unimodality with peak rainfall in August 2016 was a drought year with a serious pre-season rainfall in April. For the interpretation of the general yield level, it is not sufficient only to consider the total annual rainfall but also its distribution defining the vegetation period and intra-seasonal droughts. Rainfall onset times and peak rainfall can also lead to pest calamities (synchronization of pest and crop cycle), pollen outwash from the open-pollinating pearl millet crop, or massive fungal attack. For the presented trials, rainfall amount and distribution seem to be the major argument behind seasonal yield variability. However, this was not the case for Serkin Haussa in 2014. There, the reasons for a relative high yield are not clear. One could be the June sowing in 2014, while sowing was done later in July in the years 2015 and 2016. Another reason could be the outbreak of shoot flies (Atherigona soccata) in 2015 and 2016, which caused significant damages to the pearl millet crop particularly at the Serkin Haussa site and in some trials at Safo.
Averaging the control yields over the 3-year period reflects the information gathered from group interviews in the village of Warzou (Serkin Haussa). Farmers there indicated Jigawa being the preferred pearl millet soil. Gueza, having a slightly more loamy texture, showed relative higher surface shares cropped with sorghum. GuezamiGuezami represents a transition form between the two. Jampali is an extremely sandy soil with the lowest organic matter and chemical fertility status. So, the averaged control yield follows farmers’ soil assessment: Jampali (500 kg ha−1) < GuezamiGuezami (612 kg ha−1) < Gueza (647 kg ha−1) < Jigawa (715 kg ha−1). The absolute treatment effect shows—with the exception for Gueza—the inverse order (Jampali 346 kg ha−1 > GuezamiGuezami 253 kg ha−1 > Jigawa 244 kg ha−1> Gueza 169 kg ha−1). Therefore, the general trend depicted in Fig. 7 is supported, taking the soil type as variable. Here, it is interesting to note that on the poorest soil, the treatment effect was highest in the normal rainfall year 2015, while for the other soils, it was always in the drought year 2016. Overall, there rests an unexplained treatment × rainfall interaction effect.
A village effect is also observed. However, neither soil type counting nor rainfall effects could consistently explain the diverse Oga (+OM) treatment responses. So, there must be other processes and variables involved that have not been collected as data. What has been observed is the influence of age group. Concerning the age group, elderly people without helpers never reach the highest yield spectrum. In contrast, younger people reach higher average and the maximum yields (Fig. 8). Also, an interaction between crop management measures can occur. In our case, partial weeding (in contrast to traditional complete weeding) did reduce the control yield, but together with OM + Oga produced the highest yields. Partial weeding (i.e., either in stripes or only around the sowing pocket) is a measure to protect young seedlings from wind erosion damage and to reduce workload in a time where available work force is limited. The effect is thought to be season-depending (number and timing of wind erosion events). However, this was not investigated in detail in this study.
From these observations, we can formulate an optimum data set to be collected in order to ease interpretation of heterogeneous on-farm trial results. It would include data on the following: soil (local soil type, texture, Corg, Nt, Bray1 P + K; pre-seasonal sampling), climate data (daily rainfall, intra-seasonal droughts), crop (sowing date, early crop vigor, biomass components at harvest, previously grown crop), pest and fungal attack (locusts, head miners, stem borers, shoot flies, birds at harvest, downy mildews), pollen outwash (coincidence of flowering and high intensity rainfall), farmer (wealth status, age, gender), and field (history, management measures).
Lessons for R4D research
This research profited from a long-term collaboration of stakeholders in agronomic research, i.e., international and national research as well as two farmer federations. The long-term collaboration (> 10 years) provided a sound and reliable framework and trust for conducting the relatively large number of on-farm trials on an agronomic innovation. Farmers must have the confidence that the innovation is not setting them at risk. Research stakeholders need trust that the data are collected in the right manner. So, we have a case of mutual dependency that needs to be dealt with in an open manner. Mutual respect and continuous learning on both sites are a pre-requisite for success. Therefore, it should be allowed to change the trial protocol between seasons if recent results or pertinent arguments propose it.
Large-N trials are usually characterized by a non-equilibrated data set and in consequence a lower statistic force. In particular if—on the run—new villages or farmers are introduced in the trial design—point data loss is programmed and needs to be dealt with.
In long-term trials—as the present case—it makes sense to give more freedom to farmers with respect to management of the trial sites. While it is convenient for the researcher to work with ceteris paribus conditions at the beginning in order to reduce heterogeneity in the trials, more variance in later stages improves the potential for site adapted recommendations. E.g., for farmers who practice urea fertilization at booting stage anyway, it does not make sense to apply Oga for a second time. Sometimes farmers make their own observations and then change the “protocol”. E.g., it was observed that Oga application reduces certain pest attacks (according to farmers presumably due to the odor). In consequence, timing of the application might be changed, or the application is done for other reasons than nutrient application. Researchers should be aware of this possibility and should not try to block these opportunities, because otherwise they obtain falsely reported data that can hardly be interpreted. Also, these farmer observations can result in other local innovations that benefit smallholder farmers.
Gender aspects deserve a crucial consideration in research for development. They impact on the livelihood of the community through access to inputs, financial resources, and services. More women live in poverty than men and they do take more risks than men. A low-cost innovation is crucial for these women and more likely to be successful. Therefore, in this framework, Oga was introduced just for women. The number of women (about 160) who applied the innovation in the very first beginning was encouraging. This number increased to > 230 in the subsequent trial years. At first, men were more skeptical. They plainly judged Oga as what it is: a low-level innovation adapted particularly to women. However, having seen the results, until present, many men have opted for its application in their farmlands. A difference with respect to inputs, financial resources, and services can exist within the same gender, too. Therefore, the target group should be well defined. In our example, the poorest were targeted, and for these people, innovations should generally exert a low risk, be simple, and effective in order to increase the chance of auto-spreading and adoption.
Large-N trials will always suffer from the above-mentioned shortcomings like non-equilibrated data sets, missing values, and sometimes missing data from whole villages. On the other hand, their advantage is that they deliver data on the potential of an innovation in the real world that is the multitude of factors influencing and decisions taken by the farmer. The valuable outcome is that a more complete picture arises of how and when an innovation can be reasonably recommended. Furthermore, scaling is embedded in the process. Looking at the reality in the field, the researcher-loved ceteris paribus conditions are only another form of bias. And this latter form of bias (with normally much too high projected yield effects) leads to recommendations that do neither contain risk-related nor contextualized how-and-when information. Dealing with subsistence farming conditions, we should use on-station trials only as a first step to understand mechanisms, but large-N farmer-led on-farm trials in order to evaluate the potential of an innovation in the “real world” and subsequently develop site- and (socio-economic) environment-adapted recommendations.