Experimental design
The trials were conducted in five regions (Boro, Ugunja, Ukwala, Wagai and Yala) of Siaya County in western Kenya. Siaya County is located at 00°08.468′ N, 34°25.378′ E, and at an altitude of 1,336 m above sea level. The experimental sites were in lower midland 1(LM1) and lower midlands 2 (LM2) agro-ecological zones, which experience bimodal rainfall with long rains (LR) starting in March to July and short Rains (SR) starting in Late August to December41 and receive average annual rainfall of 1,500 mm42. The soils are mainly Ferralsols and Acrisols in the higher areas and Vertisols in the low areas.
Trials were conducted in 48 randomly selected villages, using stratification at the sub-county level. Half of them (randomly selected) participated in the trials in the long and short rain of 2014 and the long rain of 2015. The other half started in the short rain of 2014 and continued throughout the long and short rain of 2015. In each village 10 farmers participated in the trial. Half of them were specifically selected for participation in a community meeting. In those meetings, the researchers explained the objectives of the trials and asked the community members to nominate 5 farmers (as well as 5 potential substitutes), including two women, thought to be good farmers and interested in participating in the trials. Such non-random selection of farmers is common practice in research trials (Supplementary Table S1 online). The other half was selected randomly from the list of all the farmers in the village. All selected farmers were visited to obtain consent for the trials and identify the potential trial parcel (chosen by the farmer, conditional on fitting with some criteria for suitability to the research trials). A small number of replacements was done (but always keeping 5 selected by the community and 5 random). In each village 4 random farmers (2 random and 2 selected) were assigned to participate in the maize trial, while 3 random farmers (at least 1 random and 1 selected) were assigned to participate in the soybean trial. The other 3 farmers participated in a maize-soybean intercrop trial. During implementation, the assignment of inputs for the intercrop trial was, however, contaminated. As a result none of the plots in the intercrop trial received a best-bet input package, making the agronomic findings from that trial hard to interpret, and therefore not necessarily of interest for the decomposition proposed in this paper. Nevertheless, for completeness, results for these intercrop trials are shown in Supplementary Table S14 online.
Researcher-designed and farmer-managed trials
The trials would qualify as researcher-designed and farmer-managed (under the supervision of the researchers). The research team had full control over the design of the trials, from the choice of inputs to spacing and other management practices. All inputs were provided by the research team, with the exception of the local maize seed tested in two out of the six plots. A researcher (local expert agronomist) was present and led planting, gapping and thinning, all fertilizer applications, and harvesting. In these activities, labor was typically provided by the farmer. Planting dates were mostly decided by the researchers to best target the onset of rains, also responding to the farmers’ feedback on beginning of rains and availability to schedule the visit for planting. The farmer was in charge of land preparation, weeding and other management, with the researchers providing guidelines on those practices. In each village, a contact person (typically one of the ten farmers) visited the trials weekly to verify that the farmers fulfilled their responsibilities. Farmers were also asked to inform the contact person in case of any pest or disease, in which case the researcher provided the required pesticide or fungicide.
Treatment structure and application
Supplementary Table S2 online presents the full factorial designs of the multi-locational trials for maize and soybean, including details on crop varieties and quantity of inputs. The plot sizes were 4.5 × 5 m and the treatments were completely randomized between the six plots on each parcel. Plot sizes are of a similar order of magnitude as those found in other recently published work. A 1 m inter-plot spacing was planted with sweet potatoes to act as a buffer between plots to prevent inter-plot contamination. The sweet potatoes were planted at 50 cm from each plot, and border rows of the maize and soya plots were excluded for yield estimations to limit any edge effect. Hence the area harvested was 12.9 m2 for maize and 13.5 m2 for soybean. The experiments were repeated for three seasons, and plot layout and treatments were maintained for three seasons.
For the soybean trials, a soybean rhizobia inoculant was tested alone, with Minjingu hyper phosphate (0-30-0 + 38CaO) or Sympal (0:23:15 + 10CaO + 4S + 1MgO + 0.1Zn) in a full factorial design. Phosphorus rate of 30 kg P ha−1 was used to determine the quantity of Sympal and Minjingu hyper phosphate to be applied. On each farm only one replicate was used; hence, 6 plots were installed on each farm. Inoculation was done at planting as a seed coating using the directions for use in the respective product labels. Each plot had 6 soybean lines of 5 m in length each spaced at 5 cm from plant to plant within row and 50 cm from row to row. Inoculation was done on all the rows. Soybean variety TGx1740-2F with medium maturity (95–100 days)43 was used as the test crop. The spatial variability of the soybean response is studied in44. The soybean trials demonstrated that the combination of rhizobia inoculant and P-source led to important yield gains44.
The choice of inputs resulted from prior research conducted as part of the Compro project. Soybean was chosen as test crop mainly because in the prior phase of the project it had shown good response to rhizobia inoculation45 and was agro-ecologically suitable to the region. Kenya is an importer of soybean and multiple efforts are geared towards raising local production. In Compro I, the two rhizobia inoculants were tested and shown to be effective in increasing nodulation, nitrogen fixation and yield when inoculated on the tested soybean variety. Minjingu and Sympal were chosen based on their formulation with respect to the chemical characteristics of the soils in the test sites and results of earlier research46. The soils generally lack phosphorus and are acidic. A mapping study47 specifically identified Western Kenya as a potential K deficient area, and soil acidity has long been identified as a constraining factor in Western Kenya48 hence the importance of CaO. Results from soil sampling of the trial plots confirmed that more than 56.87% of soils were acidic (pH < 5.5), and more than 20.46% have low or very low K (pH < 0.160). Minjingu provides P but also reduces soil acidity because of the liming effect of the CaO, whereas Sympal also addresses potassium as well as other micronutrient deficiencies. A combination of inoculant and fertilizer was expected to perform better than the individual components as the soils are lacking both P and N, hence the combination was considered a best bet package to improve yield performance. Two different soybean inoculants were tested. Legumefix soybean from Legume technology (UK) containing Bradyrhizobium japonicum strain 532c45 and Biofix soybean from MEA Ltd (Kenya) containing Bradyrhizobium diazoeficiens strain USDA11049. Randomly selected villages were assigned to test one of the two inoculants.
For the maize trials, a package of Mavuno (10N:26P:10K: 5S: 14CaO + micronutrients) was tested at the rate of 45 kg P ha−1. This rate of P gave a rate of 39.6, 39.6 and 55 kg ha−1 of N, K and CaO) respectively. As for soybean, the use of fertilizer with K and CaO was motivated by reports of deficiency of K and soil acidity. The gains from adding K to NP for maize plots in Western Kenya have been demonstrated50. Other work more generally indicates that addressing limitations in secondary and micronutrients, and increasing soil carbon can improve response to fertilizer51. Topdressing was done with Mavuno topdressing fertilizer at the rate of 35 kg N ha−1. Phymix was applied at the rate of 250 kg ha−1 as recommended on product label. Phymix is a vermicompost with total N (0.88%), organic C (7.31%), available P (0.39%), Ca (0.29%), Mg (0.1%), K (0.22%), and a pH that is approximately neutral (6.7%). Three maize varieties were tested: local seeds, imidazolinone-resistant (IR) maize seed or a hybrid seed, randomly allocated to the 6 plots in a full factorial design. See Supplementary Table S2 online. Two different hybrid seeds were tested. KSTP94 and DH04. Randomly selected villages were assigned to test one of the two hybrid seeds.
The maize trials aimed to test some of the most viable solutions that an integrated soil fertility management combination of inputs can bring to the conditions faced by farmers in the region. A pilot phase with 60 trials in the Wagai region of Siaya county showed relatively low yields for maize because of a widespread Striga infestation problems, hence the decision to include seeds that were expected to perform better under such conditions. IR maize is marketed as a viable option to farmers with high Striga infestation problems; KSTP is a Striga tolerant variety that has high yield while DH04 is another high-yielding hybrid variety commonly grown in the area. Mavuno planting is a balanced planting fertilizer blend allowing for a split application of N as top dressing to increase N use efficiency and reduce losses at the early crop growth. The Mavuno top dressing was used for top dressing in accordance with the recommendation of split N application, as the rate used at planting cannot meet the N needs of the crop. Phymix was used as the soils are very low in organic matter and it was not possible to obtain the organic matter from sources such as compost in the quantity and consistent quality required for the trials. Being a commercial product, it ensured availability and similar quality across trials.
Data
Soil samples were taken once before any treatments were applied and sent for analysis. Available P was determined using the Mehlich 3 method (Mehlich, 1984) while pH (H2O) was done as described in52 and exchangeable K, Ca, Mg, Total N organic carbon were determined as in53,54.
Detailed observations were also collected regarding crop management (quality of land preparation, absence of weed, and absence of Striga) on trial and neighbouring non-trial parcels of the participating farmers. The agronomist and contact person were asked to evaluate quality of land preparation, Striga prevalence and weed prevalence on a scale of 1 to 3. The information was filled both for the research trial area and for the area directly bordering the trial plots and cultivated by the same farmer (referred to here as the neighbouring non-trial parcel). The agronomist completed the information during 4 visits per season (planting, top dressing, biomass assessment, harvest). The local contact person visited the trials weekly and completed the form 4 times (before planting, between planting and top dressing, between one and two months, and after two months but before harvest). Agronomists and contact persons were trained for standardized application of the forms. Observations of the agronomists and contact person were averaged across all visits and all seasons, separately for trial and non-trial parcels, to create indices capturing weed absence, Striga absence, and quality of land preparation. After averaging the grades, they were rescaled so that 1 corresponds to the highest possible score and 0 to the lowest possible score (hence Striga and weed prevalence were inversed). While the amount of weed and Striga could also result from inherent plot conditions, most of those are factored out in the comparison of the trial parcel with the neighbouring parcel. We use these data sources to measure observed management.
A farmer survey was conducted prior to the first season, and again after each season. At baseline, detailed information regarding the farmer’s skills were collected, which are aggregated in separate standardized indices for cognitive, non-cognitive and technical skills, following methods explained in detail in55. Data on asset ownership were also collected and aggregated in a standardized wealth index using principal component analysis. The units of the skill and wealth indices correspond to standard deviation changes. We also collected data on whether the main farmer in the household (who was to be responsible for the trial) was a woman and was the household head. Data were also collected regarding plot characteristics, allowing to characterize both the parcels selected for the trials, and all other parcels cultivated with annual crops, with questions capturing subjective soil quality (on a scale from 1 to 5), inclination (scale from 1–3), area (in acres), distance to homestead (in minutes), past crops and past practices used (fertilizer, manure, erosion control)—see Supplementary Table S3 online for the specific questions. Values for non-trial parcels are averaged across parcels. Note that while self-reports on past practices suggest more past fertilizer on the trial plots of community selected farmers, the soil sampling suggested no observed differences in soil properties at the start of the trials. Possibly fertilizer was only applied in small quantities as it is commonly reported in the smallholder farming systems56,57.
Agronomists also collected information about pre-trial presence of Striga (this is captured by the agronomists observation of Striga infestation on the parcel, and questions to the farmer on the prevalence of Striga in the main season of 2013, and on the number of years with Striga on the parcel), which is also used to characterize the trial parcel at baseline. This information, like the soil sampling, is only available for the trial parcels.
After each trial season, parcel-level and farmer-level data were collected to characterize input use and practices on trial and non-trial parcels. Data were collected for all cereal and legume parcels cultivated by the farmer. We use this data to obtain alternative proxies of management, by constructing variables capturing whether the farmer harrowed the parcel at least twice, weeded the parcel at least twice, and did gapping or thinning to adjust for uneven germination after planting. As management practices can differ by plot, and to reflect the best management practices farmers themselves chose for their own plots, management practices for the non-trial parcels are coded based on the best practice observed among all cereal and legume parcels for each farmer. If anything, the positive differences in Supplementary Table S3 online hence should underestimate differences between trial and non-trial parcels. By using the values of the best management plot (rather than the average for mono-cropped maize and/or (soy)bean plots—which are lower), we implicitly assume that farmers would adjust practices upwards to the best they currently do on any plot. This may still, however, underestimate the behavioural response if, when adopting the best-bet package, they would be able to overcome internal or external constraints and further improve practices. In that case, the magnitude of the behavioural adjustment would be smaller, but the direction would stay the same (as long as there is no complete adjustment).
Harvest quantities, collected in each season through crop cuts, are used to calculate agricultural production on each of the 6 trial plots for each farmer. Harvesting for both soybean and maize was done at physiological maturity. Plant populations, effective area, total fresh weight before taking sub-samples were recorded. Maize grain production was determined by weighing the fresh weight of all the cobs in the inner four rows of the plot (effective area). Sample fresh and dry weights of cobs were taken and then shelling was done. This was then used to calculate grain yield per effective area and then extrapolated to hectare basis. Soybean grain production was determined by threshing grain from the effective area and determining the weight. A sample was taken, and its weight determined before drying. Thereafter these dry weights were used to calculate yield per plot and extrapolated to hectare basis. For both maize and soybean, plant density adjustments were made by multiplying the yield by the expected plant population (based on spacing at planting) with the effective plant population (number of stands in the plot before harvesting). Plots without production are omitted for the agronomic calculations.
Yield increments for maize are calculated using three pairs of treatment–control plots: T2–T1, i.e. subtracting yields in the plot with local seeds and no fertilizer (T1) from yields in the plot with local seeds and the full fertilizer package (Mavuno, planting and top dressing, and Phymix). Similarly, calculations are done for hybrid seed plots with and without fertilizer with (T4–T3), and for IR seeds plots with and without fertilizer (T6–T5). See Supplementary Table S2 and S15 online.
Yield increments for soybean are calculated using two pairs of treatment–control plots: (a) by subtracting yields in the control plot from yields in plots containing a soybean inoculant and Sympal (T6); and (b) by subtracting yields in the control plot from yields in plots containing a soybean inoculant and Minjingu (T5).
We do not use yield of the other soybean plots in the main analysis, as they contain only one of the fertilizer types and hence don’t reflect a best-bet input package. These subplots were included in the experiment as part of the factorial design of the agronomical trial, analysed in44. For completeness the decomposition results for the paired comparison of each of these subplots with the control plot is included in Supplementary Table S14 online, showing that the different adjustments broadly lead to qualitatively similar changes for different subplot pair comparisons.
Related, to enhance comparability, we focus the analysis on yield increments obtained from the fertilizer packages for both maize and soybean. Although the maize trials also allow comparing yield estimates of different crop varieties, we do not analyze these separately, because the differences between plots with different seeds are relatively small, which limits the statistical power of any comparison between treatment and control plots. This is consistent with earlier findings showing that IR varieties are not necessarily doing better than local varieties in absence of Striga and (ii) that the chemical that is coating the maize seed is often lost quite rapidly, thus not allowing the expression of the IR properties on maize yield58,59.
Supplementary Table S14 online also reports results of the intercrop trials, which were meant to follow the same factorial design and input mix for intercropped soybeans as the soybean trials. However, input application was contaminated, making the agronomic results hard to interpret and therefore of less interest for the decomposition (as illustrated by low agronomical yield increments). Decomposition results are nevertheless presented here for completeness. They show a broadly similar pattern as for the soybean trials.
Statistical analysis
Trials were repeated for 3 consecutive seasons, to gain power and avoid results being excessively subject to weather variations in a single season. The yield increments are obtained by taking the difference between the yields in the paired treated (with the input package) and control plots. Supplementary Table S15 online displays the yield by subplot averaged over the 3 seasons, for all farmers together and separately for community selected and randomly selected farmers. When data are missing for a given season because the trial was not planted (typically because of refusal from the farmer to continue the experimentation in his/her field), we predict the value of the missing season before averaging the three seasons. The prediction uses the average value of the yield increments during the seasons for which observations for that plot are available and adjusts it for the seasonal variation. Hence to calculate the yield of a plot planted in the first and second seasons but not in the third season, we first calculate the difference in yield between the third season and the average of the first two seasons for the subsample of farmers who planted the trial in the three seasons. Then, we add this difference to the average yield from the first and second season of this plot to impute yield for that plot in the third season. This way of imputing missing yields accounts for how well the farmer’s trial is doing compared to other trials cultivated in the same seasons, and accounts for how much higher or lower yields were in the missing season(s) compared to the season(s) available.
Calculation of adjustments in yield increments
For the yield increment predictions underlying Fig. 1, we regressed the yields increments based on all the paired best-bet treatment–control plot pairs (i.e. T6–T5; T4–T3 and T2–T1 for maize; and T6–T1 and T4–T1 for soybean) on all possible covariates listed in Supplementary Table S3 online the variable indicating whether farmer was selected by the community, and binary variables to control for constant differences between pairs. We then restricted the predictor set by conducting a stepwise selection of variables with backward elimination and using the adjusted R2 as information criteria. Supplementary Table S11 online shows the regressions keeping only the variables that survived the stepwise selection. We use the point estimates of that regression to obtain predictions of yield increments for different values of the regressors. As there are multiple plot-pairs per household in each regression, standard errors are clustered at the household level in the main specification. Column 3 and 4 show alternative estimates, with standard errors clustered at the village level. Comparing the results with those of column 1 and 2 shows that results are robust to this alternative specification.
The first two predictions of Fig. 1 use the values of management, soil and farmers’ characteristics of the community selected farmers as observed in the trials. The calculation adjustment uses the same value of regressors, but changes the outcome variable of the prediction by taking away the plant density adjustment and re-inserting observations with failed crops (with zero yield rather than missing). Supplementary Table S12 online shows adjustment for plant density and crop failures separately, and shows that the plant density adjustment accounts for most of the changes, as discussed above. For crop failure, farmers could not only indicate the specific reason in about half of the cases, with animal damage (32%), lack of germination (10%) and weather (9%) being the most common reasons (in LR 15). For the other half of cases, farmers either explicitly indicated they didn’t know the reason for crop failure, or indicated symptoms (e.g. crop being yellow) rather than underlying reasons. As the data hence does not include comprehensive information on the causes of crop failures all are adjusted for equally. However, given the nature of the events causing crop failure, the frequency of these events over 3 years and across 48 randomly-selected villages should give a good estimate of the expected frequency in the region of study, and hence are taken to be reflective of probabilities of such events occurring in real life conditions.
All the other adjustments illustrated in Fig. 1 reflect the changes in yield increment predictions that result from changing the values of the covariates and recalculating the yield predictions with these different values of covariates using the point estimates from Supplementary Table S11 online. The differences in covariates underlying these estimations can be found in Supplementary Table S3 online.
Specifically, the management adjustment changes the value of all management variables from the measure of management of community-selected farmers in trials to the values of management of representative farmers in their own parcels. The average value for representative farmers is obtained by using the weighted average of randomly selected farmers and community-selected farmers. For a village of size n, community-selected farmers are assigned a weight of 5/n, and farmers randomly drawn from the rest of the village are assigned a weight of (n-5)/n (i.e. the share of randomly selected farmers over the total). As average village size is 85 (ranging from 36 to 160), the value of the representative farmer is close to that of the randomly selected farmers. This adjustment can be interpreted as the effect of changing the level of management from the one in trial to the average level of management in real-life conditions.
Similarly, the soil adjustment replaces plot characteristics (according to survey answers) by the average ones of representative farmers in their own parcels and replaces the values of the soil property variables for community selected farmers by the average ones for representative farmers. But as information from soil sampling and information obtained during identification of the trial parcels is not available for non-trial parcels, we are not able to account for differences between trial and non-trial parcels of representative farmers that are not captured by the observed plot characteristics. As expected, however, several of the observed plot characteristics are strongly correlated with soil properties. For example, quality of the plot according to the farmer’s assessment, an indicator variable for prior use of manure in the parcel and an indicator variable for prior use of chemical fertilizer are significantly correlated with respectively 7, 7 and 12 out of the 14 variables that measure soil properties. This implies that differences in observed plot characteristics between trial and non-trial parcels partly capture differences in soil properties and hence that the predictions also only partially correct for the effect due to parcel selection.
Finally, the adjustment for skills and other socio-economic characteristics is done by replacing the averages of these variables for the community-selected farmers by the ones of the representative farmers. In all cases, Fig. 1 incudes the 90% confidence interval of each prediction, and Supplementary Table S13 online provides the value of the changes in prediction due to each adjustment and the p-value of its significance (in a two-sided test).
Calculation of the returns of the input packages
The value-cost ratio presented in Table 1 is obtained by dividing the value of the additional yield by the cost of the input package used in the treatment plot but not in the control plot. Profits are calculated by subtracting the cost of the input package from the value of the additional yield. To obtain the cost of each input, we multiply the quantity of the input used in the trial plot by the price per kg at which the input was purchased. We then sum this across all inputs to obtain the cost of the input package. To obtain the value of the additional yield, we multiply the difference in production between the treated parcel and the control one by the price per kg of the crop, using the average price at which the crop was sold, obtained from the farmer survey data.
Labour costs were not included in the calculation of additional costs because labour effort is notably difficult to quantify and its valuation can vary between households. For the calculation before adjustments, labour costs will only affect the value-cost ratio if it differs between control and treatment plots. Practices (such as weeding, harrowing and gapping) were harmonized between plots so that labour costs for those tasks don’t differ. However, the application of the inputs and the time to harvest additional production would generate additional labour costs in treatment plots. Hence the value-cost provided can be interpreted as an upper bound. For the calculation after adjustments, further assumptions would be needed in order to put a value on the differences in labour costs between plots. As such assumptions could be hard to defend, we instead report costs and profits without labour costs, while being explicit about their exclusion.
Discussion of the choice of specification
In the main specification used for Fig. 1, all variables can directly contribute to the predictions of yield increments. This satisfies the purpose of predicting variations in yield while being agnostic about how these changes are triggered. A possible alternative approach is to estimate the physical-biological yield function based on agronomic variables and then predict variations in agronomic variables as a function of management and socio-economic variables. Such a model would assume that human capital variables only affect yields through changes in agronomic variables. If this specification is correct then one should find that, when predicting yield increments with agronomic variables, the addition of management and socio-economic variables do not increase the predictive power. Supplementary Table S16 online sheds light on this question by presenting the adjusted R2 in the predictions of yields and yield increments with different sets of variables. Compared with a model that only includes the agronomic variables, the results show that the inclusion of management variables adds between 4 and 9 percentage points to the explanatory power of the regression, which is a substantial improvement. The inclusion of socio-economic characteristics adds 4–5 percentage points to the explanatory power. Hence socio-economic characteristics and/or the quality of management adds additional elements of information not captured by agronomic characteristics. We can therefore reject a specification that assumes that management and socio-economic characteristics would only affect yield increments through agronomic characteristics.
Review of recently published papers
To provide evidence that the sources of discrepancies highlighted in this paper are common in the on-farm trial literature, and to complement insights from the references cited in the introduction, we reviewed articles published in 2018 in two top field journals that frequently publish results from on-farm trials: Field Crops Research and Agriculture, Ecosystems and Environment. We also reviewed articles published in 2018 PNAS, Nature and Nature Plants but did not identify any articles reporting yield results from on-farm trials Among the 88 papers that report on agronomic trials, we kept those that satisfy the following criteria: (1) being in a low or middle income country, (2) reporting yield results for annual crops, (3) being implemented on farmers’ fields with at least some farmer involvement, (4) not having a complex design with multiple treatments and replications on same plot (which are more akin to on-station trials); and (5) not being a paper of which the results from trials were described in more detail in previous papers. The majority of papers that were not further considered are on-station trials, which by definition are researcher managed and not of interest to the arguments in this paper.
In 2018 there are 16 articles that fit these criteria, including44, which reports on the agronomic findings of the soybean trials used for the current paper. Supplementary Table S1 online summarizes information for those 16 articles44,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74 relating to the 5 possible sources of discrepancies, and documents that these are indeed common practices in recently published research. Column 3 shows that random selection of farmers is the exception, and also shows that many papers don’t specify selection criteria (discrepancy 4). The later is even more the case for the selection of plots (discrepancy 3). When papers describe the non-random selection criteria, it typically involves selection by extension agents, or it is based on visibility and farmers interests in trial participation. Such selection criteria are likely to generate a sample of farmers that are not representative of other farmers in the area.
For all agronomic trials, researchers provided (as expected) the inputs and technical guidelines regarding the specific innovation that was tested. However, column 5 and 6 document that, in the vast majority of cases, researchers also provide complementary inputs and technical assistance on practices that were not being tested. This confirms that management and effort are likely to differ substantially between trial plots and the typical plot of a representative farmer (discrepancy 2). Less than half of the reviewed papers account for these additional inputs and effort through (partial) profit or cost–benefit calculations.
Column 7 documents that it is common to omit certain plots from the analysis, with the indicated reasons varying between crop failures, abnormal harvest values and management mistakes (discrepancy 1). Certain plots were excluded in almost half of the cases, and it is further unclear for some of the others whether all plots were taken into account. None of the papers specifies whether any adjustment for plant density was done, so that it is difficult to evaluate the frequency of this practice.
In the case of multi-season trials, it is often the case that not all farmers continue after the first season (column 9), but none of the reviewed papers accounts for this selection.
Finally, column 10 indicates the recommendations formulated at the end of the papers often are targeted to extension services or farmers themselves, confirming that trials with different levels of researcher involvement are being used to make recommendations for farmer practices without further testing in farmer-managed trials.
Ethics review
The research obtained IRB approval both from the IRB at JPAL-Europe at PSE and from the Maseno University Ethics Review Committee in Kenya. All research was performed in accordance with relevant guidelines and regulations, and informed consent was obtained from all participants.
Source: Ecology - nature.com
