Embedding stakeholders’ priorities into the low-emission development of the East African dairy sector

A growing body of evidence shows that more intensive dairy systems can be good for both nature and people. Little research considers whether such systems correspond with local priorities and preferences. Using a mixed methods approach, this study examined the effects of three intensification scenarios on milk yield and emission intensities in Kenya and Tanzania. Scenarios included (a) an incremental change to feed management; (b) adaptive change by replacing poor quality grass with nutrient-rich fodder crops; and (c) multiple change involving concurrent improvements to breeds, feeds and concentrate supplementation. These scenarios were co-constructed with diverse stakeholder groups to ensure these resonate with local preferences and priorities. Modelling these scenarios showed that milk yield could increase by 2%–15% with incremental changes to over 200% with multiple changes. Greenhouse gas emission intensities are lowest under the multiple change scenario, reducing by an estimated 44%. While raising yields, incremental change conversely raises emission intensities by 9%. Our results suggest that while future interventions that account for local priorities and preferences can enhance productivity and increase the uptake of practices, far-reaching shifts in practices are needed to reduce the climatic footprint of the dairy sector. Since top-down interventions does not align with local priorities and preferences in many situations, future low-emission development initiatives should place more emphasis on geographic and stakeholder heterogeneity when designing targeting and implementation strategies. This suggests that in low-income countries, bottom-up approaches may be more likely to improve dairy productivity and align with mitigation targets than one-size-fits-all approaches.


Introduction
Dairy production is a source of smallholder revenue, nutrition, and can function as a safety net particularly for women . Some studies suggest that increasing dairy productivity can positively influence smallholder dairy farm (SHDF) income and help overcome market access constraints (Westermann et al 2018). In many developing countries, smallholders fail to realize yield potentials, and are often confronted by adoption barriers arising from lack of capacity, resources, and incentives (Orr et al 2018). In particular, the agricultural sector in Sub-Saharan Africa (SSA) continues to be disproportionally affected by climate change, which discourages investment in better dairy practices and technologies (Vermeulen et al 2012, Taylor et al 2017. Compared to more industrialized regions, the dairy sector in SSA accounts for a particularly high proportion of total greenhouse gas (GHG) production (Valin et al 2013, FAO andGDP 2018). Recognizing this, many donors and development agencies in SSA now actively promote low emission development (LED), also because of the dairy sector's significance to food and nutritional security and rural incomes.
It is widely claimed that delivering on dairy sector mitigation targets demands a reduction of GHG emissions intensities. This is generally achievable through milk yield-oriented interventions (Gerber et al 2011, Forabosco et al 2017. However, this often necessitates significant capital and labour investments into more sustainable management practices (Odhong' et al 2019). With enteric methane accounting for approximately 60% of sectoral emissions in East Africa (Mottet et al 2017), much mitigation can be achieved through improvements in feed quality (Caro et al 2016). Doing so, could also increase milk yield (by approximately 51%-96%) and smallholder dairy income (Ortiz-Gonzalo et al 2017, Brandt et al 2020. To achieve this, a large-scale food system transition is needed (Rufino et al 2013), but is hindered by smallholder capacity, resource constraints, and pervasive feed availability issues. Interventions more responsive to what are often highly variegated, addressing feed adoption barriers are therefore sorely needed.
The success of LED interventions in dairy is contingent on local buy-in and responsiveness. This demands more participatory and bottom-up strategies that depart from technologist one-sizefits-all approaches (Schoneveld et al 2019). Multistakeholder planning is increasingly regarded as an essential first step to designing locally appropriate solutions sensitive to context and group-specific adoption challenges (Dunnett et al 2018). Smallholders prioritizing milk outputs in land scarce areas, for example, may prefer adopting nutrient-rich feedstuffs (Cameron et al 2018) or introducing higher yielding grasses (e.g. Pennisetum purpureum) (Maleko et al 2018), while more subsistence-oriented farmers are likely to be more receptive to practices and technologies that help manage seasonal water and feed shortages (Campbell et al 2014, Gebremeskel-Haile et al 2019. Complex land management decisions typically underpin smallholder adoption dilemmas, which may range from preserving native grasslands under increasing cattle densities to food-feed crop production trade-offs within small production units (Herrero et al 2014). In some cases, adopting improved feeding practices may be achievable through incremental changes that seek to minimize disruptions on dairy farms (Garnett et al 2013). Other dairy farms may require highly integrated and more radical approaches involving multiple concurrent changes to several aspects of farm management (e.g. introducing exotic cattle breeds alongside nutrientrich diets and improved health care) (Notenbaert et al 2018).
Given the complexity of designing LED interventions that account for complex SHDF decisionmaking patterns, this study explored how future LED strategies can be better grounded in stakeholder priorities whilst not losing sight of mitigation targets. We did this by positioning LED within diverse decision-making spaces. Specifically, through multistakeholder workshops held in Kenya and Tanzania, we co-constructed LED scenarios that resonate with diverse stakeholder priorities and preferences. Three scenarios were constructed, which included an incremental change scenario (ICS), an adaptive change scenario (ACS) and a multiple change scenario (MCS). By means of a mixed-methods analytical strategy that draws on large-scale survey data, we subsequently analysed how these three scenarios were likely to impact milk output, land use and GHG emissions. In doing so, we showed that stakeholders' priorities varied across the study area and demonstrated how more locally appropriate and evidencebased LED interventions could be designed.

Study area
The study covered important dairy production areas in the highlands of Kenya and Tanzania, covering approximately 70 000 km 2 (dark grey, figure 1). Dairy households were surveyed in three dairy counties in Kenya (Bomet, Nandi and Murang'a) and four dairy districts in Tanzania (Njombe, Rungwe, Mvomero and Mufindi). Drawing on survey results, stakeholders' workshops were held in Bomet, Nandi, Njombe and Rungwe. These areas were selected because of the significance of dairy in local livelihoods and economies. Scenario models were then implemented at the administrative unit and the regional level.

The baseline and household surveys
A baseline scenario was constructed using data collected through a household survey conducted under the IFAD-funded greening livestock project in 2018-2019. A total of 2250 SHDFs were randomly sampled across the study areas in Kenya and Tanzania (details included in appendix 1). The survey instrument captured data on, amongst other things, on-farm feed production, feed purchases and seasonal feed shortages (details in table S1). Feed shortage periods (i.e. seasonality) was used to model feed available from native grasslands in the baseline and three scenarios (figure S1, tables S2 and S3 (available online at stacks.iop.org/ERL/16/064032/mmedia))

Mapping stakeholders' priorities
To understand the adoption potential for improved management practices, we organized and facilitated four multi-stakeholder workshops in 2019. These workshops involved a total of 55 participants in Tanzania and 38 in Kenya. Participants were selected based on their direct engagement in the dairy sector (i.e. dairy farmers, farmer organisations, developmental organisations, milk buyers and processors)  or their regulatory functions (see table 1 for an overview). Prior to the workshops, a literature review of global and SSA case studies was conducted to identify best dairy practices and their impact on farm output (figure 2, appendix 2). This review guided the framing of questions on management practices and promising LED strategies. Stakeholders' priorities and preferences were identified through a questionnaire on various feed and dairy management practices. Participants were specifically asked about the importance of selected practices to SHDF milk output on a fivepoint Likert scale (i.e. priorities), as well as preferred intervention strategies (i.e. preferences) on a sevenpoint Likert scale (see appendix 3 for details).
Analysis of the farm management practices revealed three priority areas: adoption of improved cows, better diets, and improved health (figure 2). Three priority areas related to feed management were also identified: zero-grazing (cattle confined all the time), grazing on native grasslands and supplementation with concentrates (figure 2). To analyse the data on stakeholders' preferences for future interventions, we employed the 'top-two box' and 'Z-score to percentile rank' approaches (Nielsen and Levy 1994). The analysis revealed a distinct preference for interventions that increase feeds purchased, the cultivation of nutrient-rich pasture on-farm, the allocation of land for grazing and the replacement of commercial feeds with on-farm feeds, and also involve market-oriented actions that raise milk prices (figure 2). Because this analysis revealed no significant differences between the different stakeholder groups, we then pooled these results on priorities and preferences at counties and districts level.

Modelling future dairy production under different scenarios
The results from the workshops were used to design three scenarios (table 2) that align with stakeholder priorities and preferences. These scenarios are summarized:

Incremental change scenario
Aligns with priorities on minimizing SHDF investment burden. It therefore quantifies the impact of implementing change to improve feeds only on milk yield and emission intensities (table 2). Under this scenario, there is no adoption of improved cows because of the high cost and lack of access to artificial insemination services (Murage and Ilatsia 2011). ICS therefore involves a retention of crossbred cows with low-genetic merit and changes to SHDF feeding strategies through the introduction of Napier grass (tables S2 and S3) and supplementation with 1 kg of concentrate per cow per day during early lactation.

Adaptive change scenario
Quantifies the impact of transitioning to crossbred cows with higher genetic merit (i.e. improved cows) and better diets. It also considers the replacement of low-quality native grass with a diverse mix of nutrient-rich pasture and the introduction of modest amounts of concentrate (1.5 kg per cow per day) (tables 2, S2 and S3). ACS promotes a reliance on locally cultivated and more nutritious feed resources to insulate farmers from market fluctuations. Furthermore, there is the assumption that intercropping with herbaceous legumes is more likely to be associated with improved feed management practices which may also improve the quality and yield of commonly used feeds under ACS (e.g., Napier grass) (Orodho 2006, Lukuyu et al 2009.

Multiple change scenario
Quantifies the effects of simultaneously improving cows, diets and concentrate use, as per best practices. Under MCS, SHDF are more exposed to input and output price fluctuations due to high amounts of concentrate fed (3 kg per cow per day).

Scenario modelling
The effect of each scenario on milk productivity was quantified using the dynamic livestock simulation (LIVSIM) model (Rufino et al 2009). LIVSIM quantifies the impact of farm resource allocation on lifetime productivity by simulating milk yield and faecal and urine excretion. It also quantifies energy and nutrient requirements for the maintenance of dairy cow as a function of mature body weights (appendix 5). For each scenario, 1000 simulations were run, with the mean of relevant parameters (e.g. milk yield) used to calculate GHG emissions.
GHG emission associated with milk production for a dairy cow were quantified using the IPCC Tier 2 approach (IPCC 2006). We used the Python programming language as a shell to incorporate LIVSIM outputs with several functions written for the calculation of emission factors from the different sources (appendix 6). These emission sources are methane (CH 4 ) due to enteric fermentation and manure management and direct and indirect nitrous oxide (N 2 O) from manure management, synthetic fertiliser  applications, manure on grasslands and crop residues arising from conversions of grasslands for feed cultivation (i.e. land use change (LUC)). Carbon dioxide (CO 2 ) emissions from feed cultivation were also quantified (table S6). In some scenarios, land demand for Napier grass, Stylosanthes spp and Trifolium repens (white clover) exceeded the available grassland area. As a result, CO 2 emissions associated with the conversion of grasslands and forests were quantified (appendices 5 and 6). Crop residues (i.e. maize stovers) were from croplands. Cropland and grassland availability were determined using data from the European Space Agency Climate Change Initiative and FAOs' Global Land Cover-SHARE (FAO 2013, ESA 2016. In modelling this, we made two assumption: (a) that all grassland areas were available for conversion and (b) current croplands were unavailable for additional feed production, this was to avoid compromising the food security in smallholder systems. To quantify emissions from concentrate use, we multiplied concentrate intake by a factor of 1.36 kg CO 2eq DM −1 (Weiler et al 2014). To determine the emission intensities for the dairy production regions and selected administrative locations, we upscaled the calculated emissions using cattle population data (figure 2, appendix 7, tables S7 and S8). We adopted a bottom-up spatial mapping approach to upscale emissions from livestock production systems to the dairy production regions, following Brandt et al (2018b). For the administrative unit analysis, we delimit livestock gridded data with polygon features in selected counties and districts (i.e. stakeholders' workshop locations), following spatial aggregation techniques described in Lloyd et al (2017). We compared productivity outcomes under the different scenarios against the baseline using two indicators of climate change mitigation, namely: emission intensities per kilogram of fat protein corrected milk (FPCM) and percentage change in the different emission sources relative to the baseline. Furthermore, we used Wilcoxon-Mann-Whitney test to determine whether there was significant (i.e. p < 0.05) increase or decrease in milk yield under the different scenarios compared to the baseline.

Stakeholders' priorities and preferences
Stakeholders' priorities for the different management practices were not always unanimous across locations (table 3). Zero-grazing and use of concentrates were considered the most important practices for raising smallholder milk yield, mentioned by 99% and 96% of respondents, respectively. Improved health was also prioritized by most (89%), as was better diets (85%), especially in Nandi and Njombe. Adoption of improved cows received least interest, with only 68% prioritizing genetic improvement. Geographical variation were observed with improved cows, for example, improved cows were considered unimportant in Rungwe and particularly important in Nandi. This certainly points to the importance of geographically-adapted LED strategies. The analysis of future interventions showed strong preference (>50%) for the cultivation of nutrient-rich feeds (including cultivated fodder and pastures) that could reduce the reliance on feed markets (table 4). Stakeholders had weak preference (<50%) for allocating more land to pasture and replacing on-farm feeds with commercial feeds. However, with the exception of Nandi, stakeholders did prefer to see SHDF use more commercial feeds, albeit not at the expense of on-farm feed production. This can be attributed to perceptions of on-farm feed availability constraints. Weak preferences towards reallocation of arable land to pasture furthermore highlights widespread food security concerns.

Milk production under scenarios
Improvements to management practices had a positive impact on milk production across all scenarios. Incremental change to feed practices increased milk yield per cow by 2%-15% compared to the baseline (table 5). Under the ACS, milk yield increased by 60%-130% to 6.2-10.1 kg cow −1 d −1 . Similarly, milk yield increased three-fold under the MCS compared to the baseline. This suggested that widespread adoption of the proposed practices could lead to increased nutrient supply and marketable surplus. In Table 4. Stakeholders' preferences for the implementation of intervention strategies on SHDFs. Preferences were determined using top-two box analysis and represents the proportion of stakeholders with strongly agree and agree responses. Bold numbers indicate strong preference (>50%) for any given strategy.  both countries, there was an increase in milk yield per cow under change scenarios compared to the baseline (p-value < 0.001).

Dairy carbon footprints at sectoral level
There were no significant differences between emission intensities at both regional and administrative level scales (tables S9 and S10). Nevertheless, the adoption of the proposed management practices by SHDFs would have a varied effect on total GHG emissions in Kenya. Specifically, emission intensities numerically increased from 3.43 kg CO 2eq kg FPCM −1 in the baseline to 3.75 kg CO 2eq kg FPCM −1 (9% increase) under ICS ( figure 3(A)). However, GHG emission intensities decreased by 28% and 44% in Kenya under ACS and MCS, respectively. Under the baseline and ICS low-genetic merit crossbreed cows continued to be kept, while under ACS and MCS genetic improvements were introduced. LUC was the largest source of emissions, increasing by at least 500% in all change scenarios (figure 3(C)). Emissions associated with fertiliser application on croplands decreased by more than 40% in all scenarios. Similarly, emission intensities due to enteric fermentation, methane from manure management and manure on grassland decreased in all scenarios. For the dairy production region of Tanzania, GHG emission intensities numerically increased from 3.22 kg CO 2eq kg FPCM −1 in the baseline scenario to 3.51 kg CO 2eq kg FPCM −1 (9%) and numerically decreased to 2.08 kg CO 2eq kg FPCM −1 (35%) under ICS and ACS, respectively ( figure 3(B)). The total emission intensity decreased by 30% under MCS. In all scenarios, LUC was the largest source for increased emission intensities ( figure 3(D)). Emission intensity decreased by more than 70% due to fertiliser application on croplands. Emissions from enteric fermentation and methane from manure management decreased for all scenarios ( figure 3(D)). The emissions from forest losses were negligible, accounting for less than 0.01% of total emissions in all locations.

County/district-level carbon footprint
Kenya's county level analysis of GHG emission intensities revealed differing patterns across change scenarios compared with the baseline (figure 4). For instance, the total emission intensity numerically decreased from 3.2 kg CO 2eq kg FPCM −1 in the baseline scenario to 2.73 and 2.16 kg CO 2eq kg FPCM −1 under ICS and ACS, respectively, in Bomet ( figure 4(A)). In Nandi, the GHG emission intensity decreased for all scenarios ( figure 4(B)). The largest source of increase in emissions was from LUC (>500%) in all scenarios (figures 4(C) and (D)). Emissions intensities from enteric fermentation, manure management and manure on grassland reduced under all scenarios (figures 4(C) and (D)). In both counties, emissions intensities from methane from manure management decreased by a minimum of 90% across scenarios.
In Njombe, the total emission intensity numerically decreased from 3.22 kg CO 2eq kg FPCM −1 under the baseline to 3.15 kg CO 2eq kg FPCM −1 under ICS ( figure 5(A)), representing a 2% decrease. However, the total emission intensity decreased by 16% and 28% under ACS and MCS, respectively ( figure 5(A)). In Rungwe, emission intensities also decreased for all scenarios compared with the baseline ( figure 5(B)), ranging from 12% under ICS to 40% under MSC. In Njombe and Rungwe, emissions intensities from enteric fermentation, fertiliser application and manure management decreased by at least 90% under all scenarios (figures 5(C) and (D)). Emissions from LUC increased by at least 500% for each scenario. Across the study area, interventions in feed management would reduce emission intensities from manure on grasslands (figures 5(C) and (D)).

Emissions from the sourcing of concentrates
Concentrate use increased GHG emission intensities across scenarios. GHG emission intensity increased by 21% under MCS in the dairy production region of Kenya (table S11). In Nandi and Bomet, concentrate use increased the GHG emission intensity between 3% and 30% (table S12). Concentrates increased emission intensities by approximate 3% under ACS due to the comparatively small amounts that were used under this scenario. Similarly, concentrate use led to the largest increase on emission intensities under MCS for the dairy production region of Tanzania (table S13). In Rungwe and Njombe, concentrate use increased the total emission intensity by 23% and 18% under MCS, while the least increase in total emission intensity due to concentrate use was under the ACS (table S14).

Discussion
Our study shows that stakeholders' priorities and preferences are more heterogeneous across space than across social groups. This points to a miss-match between a hegemonic vision of LED and many subnational development interests. However, what is preferred by stakeholders is not necessarily in line with LED objectives. For instance, we show that incremental changes to feed management, which are more relevant to some areas, increases emission intensities. Adaptive changes to native grasslands and cow genetics in contrast reduced total emission intensities by 28%, while multiple concurrent changes reduce these by an estimated 44%. However, notable intra-regional variations were observed. For instance, emission intensities did reduce under ICS in several administrative areas, suggesting that incremental change was not always incompatible with emission reduction goals. Therefore, in certain areas (in this case Rungwe), resource-constrained interventions may still be impactful. In other areas, LED interventions were only likely to raise yields and lower emission intensities simultaneously with a more encompassing approach that targets multiple practices concurrently. However, in some areas (notably Bomet and Rungwe in this case), some stakeholders were likely to resist such interventions, which could affect efficacy as a result of uptake problems and/or reduced local buy-in and legitimacy. This highlights that LED strategies should be tailored to local interests and practice preferences, or, alternatively, invest more in sensitization. While calling into question one-size-fits-all LED, top-down approaches may still have a place alongside more bottom-up approaches, albeit if the geographic focus of such interventions is restricted to amenable geographies. This, however, is unlikely to be truly transformational without more deliberate large-scale efforts to institutionalize more progressive and reflexive alternatives.
Accounting for stakeholders' priorities and preferences at scale in uncertain contexts clearly requires national-level LED design strategies that are responsive to changing and variegated geographic conditions. This can be achieved by embedding multi-stakeholder processes within all phase of the project lifecycle; from design to implementation and back to re-design. Institutionalizing such processes requires sectoral governance structures that facilitate inter-organizational coordination and helps overcome the fragmented policy space and development programming pervasive in many developing country contexts.
While results had important implications, they also offered some empirical evidence that could help delineate the technical direction of national LED strategies. This study demonstrated that replacing low-quality grasslands and reducing dependence on commercial feeds (e.g., under ACS), were likely to satisfy the interests of the many stakeholders that oppose the allocation of cropland to fodder production. Doing so increased milk yield, reduced emission intensities and aligns with 'food-first' priorities of most districts/counties. However, emissions from LUC increased significantly-as it does under the other scenarios. This study found that feed cultivation raised LUC emissions in particular (in line with Herrero et al 2014, Brandt et al 2018b). Upscaled results suggest that LUC account for between 25% and 42% of absolute emissions, peaking at 63% in sampled counties/districts. This suggested that LED viability assessments should place more emphasis on LUC-related emissions to increase adoption and effectiveness (e.g. FAO and NZAGGRC 2017, Michael et al 2018. While LUC emissions could be largely offset by the inclusion of other practices in the scenarios, the results of this study do give reason to caution against land-centric and -intensive intervention strategies, especially in areas experiencing land constraints or comprising ecologically-and socially-significant ecosystems. This is similar to common practices adopted by dairy farmers in high-income countries to reduce GHG emission intensities (van Meijl et al 2006). Greater use of concentrates (e.g. under MCS) could also increase milk yield, while reducing the risk of forest disturbance (Brandt et al 2018a). However, concentrate use could be associated with increase offfarm and overseas emissions (Styles et al 2018), as well as, increasing pressure on existing land capacity (Brandt et al 2020). This study also showed that concentrate use under MCS could increase total emission intensity by 17%-29%. While on-farm fodder production is generally more desirable from an emissionreduction perspective (Dawson et al 2014), this is considered socially detrimental and therefore, politically contentious by most sampled stakeholders. This further highlights that an acute understanding of contextual factors that condition socio-environmental trade-offs is needed in LED design.
Overall, the emission intensities calculated here were consistent with recent studies on East African smallholder production systems of East Africa (e.g. Mottet et al 2017, Wilkes et al 2020. However, total emission intensities were not substantially different between scenarios, which could be attributed to lack of drastic changes in proposed feeding interventions (e.g. supplying only Napier grass under ICS). Nevertheless, by calibrating feeding strategies using stakeholders' priorities and preferences, this study added an important social dimension. However, there is a risk that the scenarios co-developed here are biased in favour of the priorities of better represented groups such as dairy farmers and government officials. However, since there was no statistically significant difference in priorities and preferences between groups, therefore, the risk was considered low. In future studies, particularly in areas where priorities and preferences diverge, attention to balanced representation is needed; possibly through larger-scale surveying.

Conclusion
This article demonstrates the importance of accounting for spatially differentiated stakeholder priorities and preferences in future LED programming and policy-making. It further shows that there is need for caution when implementing mainstream top-down LED particularly when dealing with smallholder dairy farmers who exhibited different preferences for management strategies in this study. Rather dairy policies should include more bottom-up approaches that can help integrate local priorities and preferences that condition uptake and, by extension, intervention efficacy into national LED strategies. This can be achieved by improving multi-stakeholder participation structures and nesting different scales and aligning different thematic domains of governance. This is also more aligned with ongoing processes of devolution. While this article does show that socioenvironmental trade-offs are likely inherent to any feed-oriented LED strategy, these can be managed and optimized through participatory processes. In this regard, we show that land use change dynamics deserved particular emphasis and both social and environmental safeguards should be expressly considered.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors. Brandt P, Yesuf G, Herold M and Rufino M C