论文信息

A. I. Y. Lankapura / Sanjaya Fernando / K. P. P. Kopiyawattage 等 5 人

EarthArXiv (California Digital Library)

论文解读

AI 辅助整理
方法
对440户稻农问卷调查,基于跨理论模型与COM-B框架,用有序混合模型分析。
发现
约束因实践和阶段而异,集体与使用经验约束对意向者和采用者更严重。
机会点
可探索将阶段化行为模型嵌入推广服务设计,并检验集体灌溉制度对个体地块实践的跨边界影响。

论文摘要

气候智能型农业实践有助于小农户适应气候变化,但其采用率仍然较低。关于采用制约因素的研究很少比较不同实践或不同采用阶段。本研究考察了采用五项水稻气候智能型农业实践的制约因素是否因实践和采用阶段而异。调查覆盖了斯里兰卡阿努拉德普勒区高度气候脆弱地点社区管理季节性蓄水池灌溉下的440名稻农。对于每项实践,依据跨理论模型将农民分类为非意向者、意向者或使用者。农民对每项实践的十项制约因素进行评分,并对阻碍其最大的三项因素进行排序。制约因素采用能力、机会、动机和行为模型进行分组,并扩展了集体约束和使用经验约束两组,采用有序混合模型进行分析。对于认证气候韧性品种、综合杂草管理以及托盘育苗与抛秧移栽,物理机会制约因素排名第一。对于综合植物养分管理,知识缺乏排名第一;对于气候信息服务,对预报准确性的怀疑排名第一。对于四项实践,实践特定制约因素被评为比所有实践共有的制约因素更为严重,而只有成本是大多数实践共有的因素。每项实践的制约因素相对重要性在不同采用阶段之间存在差异。与非意向者相比,意向者和使用者将能力和动机制约因素评为较轻,但将集体约束和使用经验约束评为较重。在调整意识和农场家庭特征后,这一模式仍然存在。与预期相反,对于综合杂草管理——一项在个体田块内实施的实践——源于蓄水池管理的集体约束与受集体商定种植日历定时的实践一样严重。研究结果表明,基于阶段和行为模型的采用理论是互补的。因此,气候智能型农业的推广支持应与实践和农民的采用阶段相匹配。

全文

EarthArXiv (California Digital Library)

1
1Practice- and stage-specific constraints to adopting
2climate-smart agriculture practices in smallholder rice
3farming under community-managed seasonal tank
4irrigation: Evidence from Sri Lanka
5A.I.Y. Lankapura 1, Sanjaya Fernando1, K.P.P. Kopiyawattage2, A.M.K.R. Bandara1, N.M.
6Hung 3,4*
7 1 Department of Agricultural Systems, Faculty of Agriculture, Rajarata University of Sri
8Lanka, Anuradhapura, Sri Lanka
9 2 Department of Agricultural Extension, Faculty of Agriculture, University of Peradeniya,
10Peradeniya, Sri Lanka
11 3 Department of Environmental Economics, Institute for Environmental Studies, Vrije
12Universiteit Amsterdam, Amsterdam, the Netherlands
13 4 Faculty of Economics and Rural Development, Thai Nguyen University of Agriculture and
14Forestry, Thai Nguyen, Vietnam
15* Corresponding author
16E-mail: n.manhhung@vu.nl (NMH)
17Abstract
18Climate-smart agriculture practices can help smallholder farmers adapt to climate change, yet
19their adoption remains low. Studies of adoption constraints rarely compare practices or stages
20of adoption. This study examined whether the constraints to adopting five climate-smart
21agriculture practices for rice differ by practice and by adoption stage. A survey covered 440
22rice farmers under community-managed seasonal tank irrigation in highly climate-vulnerable
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

2
23locations of Anuradhapura District, Sri Lanka. For each practice, farmers were classified as
24non-intenders, intenders or users following the Transtheoretical Model. They rated ten
25constraints per practice and ranked the three that held them back most. Constraints were
26grouped using the capability, opportunity, motivation and behaviour model, extended with
27collective and use-experience constraint groups, and analysed with ordinal mixed models.
28Physical opportunity constraints ranked first for certified climate-resilient varieties, integrated
29weed management, and tray nursery and parachute transplanting. Lack of knowledge ranked
30first for integrated plant nutrient management, and doubt about forecast accuracy ranked first
31for climate information services. Practice-specific constraints were rated more severe than
32constraints common to all practices for four practices, and only cost was common to most
33practices. The relative importance of constraints differed across adoption stages for every
34practice. Compared with non-intenders, intenders and users rated capability and motivation
35constraints as less severe, but collective and use-experience constraints as more severe. This
36pattern persisted after adjustment for awareness and farm-household characteristics. Contrary
37to expectation, collective constraints arising from tank management were as severe for
38integrated weed management, a practice carried out within individual plots, as for practices
39timed by the collectively agreed cropping calendar. The findings indicate that stage-based
40and behavioural models of adoption are complementary. Extension support for climate-smart
41agriculture should therefore be matched to both the practice and the farmer’s adoption stage.
42Introduction
43Climate change has already slowed the growth of agricultural productivity, particularly in
44warmer, low-latitude regions [1,2]. Small-scale food producers in developing economies are
45among those most vulnerable to these impacts, as they depend on climate-sensitive farming
46and have limited means to adapt [1,3,4]. In South Asia, adaptation options for smallholder
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

3
47production systems are available, but the institutional arrangements needed to implement and
48disseminate them remain weak [5]. Identifying what prevents smallholders from adopting
49available adaptation practices is therefore central to effective adaptation.
50Climate-smart agriculture (CSA) provides one widely promoted set of such practices. It is an
51approach for transforming and reorienting agricultural systems to support food security under
52climate change [6,7]. Yet the adoption of CSA practices by smallholders remains limited [8]
53and uneven [9]. Constraint studies in Ghana, India and China [10,11,12] have identified
54recurring constraints to CSA adoption. These include cost, credit, access to inputs and land,
55and limited knowledge and skills [10,11]. A systematic review has also shown that CSA
56adoption depends on enabling institutions and policies, not on technologies alone [13].
57Moreover, the effects of most adoption determinants vary in sign and size across technologies
58and contexts [8,14]. Constraints identified for one practice or setting may therefore not apply
59to another. Extension programmes designed on such general findings risk targeting
60constraints that do not actually limit the adoption of a given practice.
61Research on constraints to CSA adoption has followed four main approaches. Constraint
62surveys ask farmers to rate predetermined constraints and then rank them by severity,
63commonly with a Problem Confrontation Index [10] or Garrett ranking [15]. These rankings
64show which constraints farmers perceive as most serious. However, they seldom examine
65how constraints relate to adoption outcomes [11] or whether they differ between practices or
66groups of farmers. Determinant models examine adoption outcomes directly, yet most
67measure adoption as a binary variable [8] and therefore compare only adopters with non-
68adopters [16]. Process perspectives, in contrast, view adoption as a learning process [17] that
69unfolds through successive stages [18]. Stage theories imply that people at different stages
70face different constraints [19]. The fourth approach uses behaviour-change frameworks to
71classify the content of constraints. The most widely used of these is the capability,
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

4
72opportunity, motivation and behaviour (COM-B) model, which attributes behaviour to a
73person’s capability, opportunity and motivation [20]. However, COM-B does not indicate
74which of these components should be addressed first [21]. It also does not treat institutional
75influences as a distinct component, but includes them only in general terms within
76opportunity [22]. These four approaches have developed largely in isolation from one
77another. As a result, it remains unclear how the content of constraints changes as farmers
78progress towards adoption.
79This separation has left three gaps in the evidence on constraints to CSA adoption. First, most
80constraint studies either pool several practices under the general label of CSA [10] or
81examine a single technology or region [12,15]. They rarely compare constraints
82systematically across practices [11]. Without such comparisons, constraints common to all
83CSA practices cannot be distinguished from those specific to a particular practice. Second,
84stage-specific constraints have rarely been tested quantitatively. Ha et al. [16] applied the
85Transtheoretical Model (TTM) to farmers, but they modelled the factors that determine a
86farmer’s adoption stage rather than the constraints that operate at each stage. The most
87comparable stage-based study of CSA constraints is qualitative and covers a single region
88and group of practices [23]. Without quantitative evidence of this kind, extension support
89cannot be targeted to farmers’ adoption stage. Evidence from health behaviour shows why
90such targeting matters. In an experiment on testing homes for radon, a naturally occurring
91radioactive gas, risk information was relatively more effective among undecided people,
92whereas a low-effort intervention was more effective among those who had already decided
93to act [24]. Third, the institutional context of CSA adoption, including the interactions among
94the actors involved, has received comparatively little attention [13]. This gap is particularly
95relevant to farmer-managed irrigation systems. These systems are common-pool resources
96governed through collective-choice arrangements [25], and collective decisions on the
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

5
97cropping calendar and water releases can directly constrain the practices of individual farmers
98[26]. Extension support aimed at individual farmers may therefore overlook constraints that
99arise from collective decisions, unless these are distinguished from the constraints that
100farmers face individually.
101This study addressed these gaps among smallholder rice farmers under community-managed
102seasonal tank irrigation in the dry zone of Sri Lanka. It examined whether the constraints to
103adopting CSA practices are common across practices or specific to each practice, and
104whether their relative importance differs across adoption stages. The study had four
105objectives. The first was to identify and prioritise the constraints to adopting each of five
106CSA practices. The second was to test whether constraints common to all practices weigh
107equally across practices and how they compare with constraints specific to each practice. This
108comparison included whether collective constraints weigh more on practices tied to the
109collectively agreed cropping calendar than on a practice carried out within individual plots.
110The third was to examine whether the relative importance of the same constraints differs
111across adoption stages within each practice. The fourth was to assess whether these stage
112differences persist after accounting for farmers’ awareness of the practices and their farm-
113household characteristics.
114The dry zone of Sri Lanka offers a suitable setting for addressing these gaps. Climate change
115has already lowered rice yields in Sri Lanka [27]. Much of the paddy land in the dry zone is
116irrigated from community-managed seasonal tanks, also known as village or minor tanks,
117which are linked in tank cascade systems. About 23% of Sri Lanka’s paddy land lies within
118these cascade systems [28]. Climate modelling projects that, by 2050, up to 96% of the paddy
119land in this zone that is currently highly suitable for rice may be at risk of losing this
120suitability. The reason is that temperature and rainfall conditions in the zone are projected to
121become less favourable for paddy [28]. For each tank, farmers and government officials
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

6
122collectively decide the cropping calendar before the season, and water releases are then
123managed collectively during the season [26]. Earlier studies of CSA adoption in the dry zone
124used a determinant model [29] or a participatory case study [30]. Neither compared
125constraints across practices or adoption stages. This study focused on the five practices that
126received the highest combined priority scores as CSA practices for rice cultivation in this
127setting, in an assessment of extension officers’ and farmers’ priorities in Anuradhapura
128District [31]. These are certified climate-resilient rice varieties, integrated plant nutrient
129management (IPNM), integrated weed management (IWM), tray nursery and parachute
130transplanting, and climate information services (CIS).
131The study makes three contributions to research on CSA adoption. First, it compares common
132and practice-specific constraints across five CSA practices within the same sample of
133farmers. Second, it provides a formal test of stage-specific constraints across these practices.
134This test combines severity ratings with farmers’ own priority rankings. It also compares
135constraint groups within the same farmers, which guards against differences in how farmers
136generally use rating scales. Third, it extends the COM-B model with constraints arising from
137use experience and from collective decisions. The collective group is relevant wherever farm
138decisions depend on collective irrigation governance. The empirical findings apply to
139climate-vulnerable seasonal tank systems in the dry zone, but the approach can be applied in
140other farmer-managed irrigation systems.
141Conceptual framework and hypotheses
142This study draws on three complementary perspectives. The TTM describes the stages
143through which people move when they adopt a new behaviour. The COM-B model classifies
144the content of the constraints that may hold them back. An institutional perspective, based on
145the governance of common-pool resources, explains how collective decisions in seasonal tank
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

7
146systems can constrain individual farmers. Combining these perspectives allows the study to
147ask which constraints matter, for which practice and at which stage of adoption.
148Adoption stages and the Transtheoretical Model
149The TTM explains intentional behaviour change as progress through a sequence of stages
150[32]. In its widely used form, it distinguishes five stages, namely precontemplation,
151contemplation, preparation, action and maintenance [33,34]. People in precontemplation do
152not intend to change in the foreseeable future. People in contemplation intend to change, and
153those in preparation plan to act soon. Action refers to a recent change in behaviour, and
154maintenance to a change that has been sustained. Movement from one stage to the next
155depends on the balance between the perceived advantages and disadvantages of changing
156[33].
157Applications of the TTM to farmers often use fewer stages. Otter and Deutsch [35]
158distinguished four stages in a study of German farmers and did not separate maintenance
159from action. Ha et al. [16] merged adjacent stages in a study of intercropping adoption.
160Following these studies, this study uses three stages. Stage 1 comprises non-intenders and
161corresponds to precontemplation. Stage 2 comprises intenders and combines contemplation
162and preparation. Stage 3 comprises users and combines action and maintenance. Farmers are
163assigned to a stage separately for each practice, as a farmer may use one practice while not
164considering another.
165Content of constraints and the extended COM-B model
166The COM-B model forms the core of the Behaviour Change Wheel. It holds that a behaviour
167occurs only when a person has the capability, the opportunity and the motivation to perform it
168[20,36,37]. Capability refers to the knowledge and skills needed for the behaviour.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

8
169Opportunity refers to factors outside the person that make the behaviour possible, and it may
170be physical or social. Motivation refers to the beliefs, intentions and habits that energise and
171direct behaviour [20,36]. The model has been used to analyse the factors that drive or
172constrain farmers’ adoption of agricultural practices [38,39].
173Following COM-B, this study classifies constraints into four groups. Capability constraints
174concern gaps in knowledge, skill and understanding. Motivation constraints include doubts
175about benefits, perceived risk or distrust, competing priorities, habits and beliefs. Physical
176opportunity constraints include cost, input availability, labour, time and field conditions.
177Advisory support constraints are a form of social opportunity and include a lack of training,
178demonstration and advice.
179Two further groups extend COM-B to fit the setting of this study. The first is a collective
180group, which is also a form of social opportunity. It covers constraints that arise from
181decisions made collectively for the tank command area, such as the dates of water releases.
182This group is needed because the opportunity component of COM-B does not fully capture
183behaviour shaped by several autonomous actors with divergent interests [22]. The second is a
184use-experience group. It covers performance problems that farmers learn of through their
185own use of a practice or by observing others, such as persistent weed pressure or inaccurate
186forecasts. This group draws on the implementation and confirmation stages of the innovation-
187decision process, in which adopters evaluate an innovation through use [18]. For CIS, the
188common constraint on risk or distrust took the form of doubt about the accuracy of forecasts
189[40]. This constraint was assigned to the use-experience group, as such doubt usually arises
190from experience of forecasts that proved wrong.
191Practices differ in attributes such as complexity and trialability [18] and in the behavioural
192factors that drive their adoption [41]. Some constraints may therefore apply to all practices,
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

9
193whereas others may apply only to particular practices. The study distinguishes these common
194constraints from practice-specific ones.
195Institutional context of community-managed seasonal tanks
196The institutional perspective addresses the collective setting in which farmers decide on these
197practices. Community-managed seasonal tanks are common-pool resources governed through
198collective-choice arrangements [25,26]. Before each cultivation season, the farmer
199organisation and government officials hold a pre-season cultivation meeting. At this meeting,
200they set the cropping calendar, the water-release dates and the allocation of water, and these
201decisions require official approval [26,42]. A decision taken at this meeting can therefore
202constrain an individual farmer, whatever that farmer’s own capability or motivation.
203Hypotheses
204The extended COM-B groups and the distinction between common and practice-specific
205constraints guide the identification and comparison of constraints in this study. Stage theories
206further hold that people at different stages face different constraints [19,24,43]. The same
207constraints should therefore differ in relative importance between adoption stages within each
208practice. Three hypotheses specify the expected pattern. Each compares Stage 2 and Stage 3
209separately with Stage 1.
210The first hypothesis (H1) states that capability constraints are less severe in later adoption
211stages. Adoption is a dynamic learning process [17], and knowledge has been associated with
212later adoption stages [16]. The second hypothesis (H2) states that motivation constraints are
213less severe in later stages. As people move towards action, the perceived disadvantages of
214change decline relative to the perceived advantages [33]. The third hypothesis (H3) states that
215physical opportunity, collective and use-experience constraints are not less severe in later
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

10
216stages. Practical matters of convenience and opportunity can strongly influence whether a
217decision is carried out, although they are often ignored when the decision is first made [43].
218Farmers with similar intentions can also face different physical opportunity constraints [39].
219Collective decisions do not ease as an individual farmer progresses towards adoption.
220Performance problems also become more visible once farmers try a practice or observe others
221using it [18]. No direction is hypothesised for advisory support constraints. The need for
222advice may decline as farmers learn, but it may also increase when they first implement a
223practice [18].
224The fourth hypothesis (H4) compares practices rather than stages. The pre-season cultivation
225meeting fixes the cropping calendar and the water-release dates [26]. Certified varieties, tray
226nursery and parachute transplanting, and CIS depend on this calendar, so they are referred to
227here as calendar-bound practices. IWM, in contrast, is largely carried out within individual
228plots. H4 therefore states that collective constraints are more severe for the calendar-bound
229practices than for IWM. For CIS, inflexible farm operations are a known constraint to the use
230of climate information [40].
231Differences between stages could also arise because farmers in later stages differ in other
232characteristics. Farmers who are aware of a practice, better educated or in more frequent
233contact with extension services may progress further and also report different constraints. In
234an earlier application of the TTM to farmers, education, knowledge and, to some extent,
235income differed between stages [16]. Awareness also deserves separate attention. The
236Precaution Adoption Process Model distinguishes people who are unaware of an issue from
237those who are aware of it but not yet engaged [43]. This study therefore treats awareness and
238five farm-household characteristics as potential confounders and tests whether stage
239differences persist after adjusting for them. Comparisons of people in different stages at one
240point in time cannot show that people move through the stages in sequence [19]. The
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

11
241framework therefore specifies associations between stages and constraints, rather than
242transitions between stages.
243Materials and methods
244Study area
245The study was conducted in Anuradhapura District, North Central Province, Sri Lanka. The
246district falls within the Low Country Dry Zone agro-ecological region DL1, with mean
247annual temperatures of 27 to 28 °C and annual rainfall of 1,200 to 1,400 mm [44]. Rainfall
248follows a bimodal pattern, with the Maha or major season from October to March and the
249Yala or minor season from April to September [44,45]. Rainfall varies greatly between years,
250and recurrent droughts, erratic monsoon onset and occasional floods regularly disrupt
251agricultural production [46,47].
252Anuradhapura District combines high vulnerability to climate change with a strong
253dependence on agriculture [47,48]. Rice is the main crop, and paddy yields in the dry zone
254are sensitive to changes in temperature and rainfall [45,46]. Smallholder farmers in such
255settings have limited capacity to adapt, as their access to capital, technology and institutional
256support is restricted [3,4]. Much of the district’s paddy land is irrigated from community-
257managed seasonal tanks. These are small reservoirs, with command areas generally below 80
258ha, that capture and store seasonal rainfall for irrigation [42]. Their limited storage,
259dependence on seasonal rainfall and susceptibility to drought expose tank-irrigated rice to
260climate risk [28,49]. Each tank is managed by a farmer organisation, which decides the
261cropping calendar and the schedule of water releases together with government officials at
262the pre-season cultivation meeting. The farmer organisation then manages water releases
263during the season [26,42].
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

12
264The study focused on the most climate-vulnerable locations in the district. A climate
265vulnerability and risk assessment of North Central Province identified five high-risk
266Divisional Secretariat (DS) divisions in Anuradhapura District and, within them, the most
267vulnerable Grama Niladhari (GN) divisions [50]. GN divisions are the smallest administrative
268units in Sri Lanka. The study covered all 22 high-risk GN divisions in these five DS
269divisions, each of which cultivates paddy under seasonal tanks. Their names and locations are
270given in Table S1 in S1 Appendix.
271Research design
272The study used a cross-sectional household survey design and is reported according to the
273Strengthening the Reporting of Observational Studies in Epidemiology guidelines for cross-
274sectional studies (S1 Checklist). Each farmer was assigned an adoption stage for every
275practice and rated the same set of constraints for each practice. Practices could therefore be
276compared within the same farmers, whereas adoption stages were compared between farmers.
277Constraints were measured in two complementary ways, as severity ratings and as farmers’
278own priority rankings.
279Study population and sampling
280The target population comprised paddy farmers cultivating under community-managed
281seasonal tanks in the high-risk GN divisions of Anuradhapura District. A multistage sampling
282procedure was used. In the first stage, the five high-risk DS divisions were selected
283purposively on the basis of the provincial climate vulnerability and risk assessment [50]. In
284the second stage, all 22 high-risk GN divisions within these DS divisions were included, so
285no sampling took place at this level. In the third stage, 20 paddy-farming households were
286selected from each GN division by simple random sampling from the membership registers
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

13
287of the farmer organisations. Households were eligible if they had cultivated paddy under a
288seasonal tank in at least one of the two cultivation seasons before or during the survey. In
289each household, the principal decision-maker for paddy farming was interviewed. This
290procedure yielded a sample of 440 farmers. With this sample size, a proportion of 50% can be
291estimated within ±4.7 percentage points at the 95% confidence level, assuming simple
292random sampling.
293Data collection
294The study used both primary and secondary data. Secondary data came from the provincial
295climate vulnerability and risk assessment and the membership registers of the farmer
296organisations, and the published literature informed the development of the constraint items.
297Primary data were collected mainly through a household questionnaire survey, supported by
298key informant interviews. The key informants were extension officers, rice subject matter
299officers, farmer organisation leaders and researchers, and their input helped to refine the
300constraint items described below. A structured, pretested questionnaire was administered by
301the authors in face-to-face interviews from 5 June 2025 to 28 November 2025. This period
302spanned the 2025 Yala season and the start of the 2025/26 Maha season. All interviews were
303conducted in Sinhala. At the beginning of each interview, the interviewing author read the
304consent statement to the respondent and recorded the respondent’s verbal consent on the
305survey form in the respondent’s presence. All respondents were adults. The questionnaire
306also recorded the farm-household characteristics defined in Table S4 in S1 Appendix. Gender
307was recorded as reported by the respondent, as woman or man.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

14
308Ethics statement
309Informed verbal consent was obtained in Sinhala from every participant before data
310collection. A standard statement explained the purpose of the study, the voluntary nature of
311participation, the anonymity of responses and the intention to publish the findings. Only
312consent status was recorded on the survey forms, with no personal identifiers, and the privacy
313rights of participants have been observed. This study was conducted in Sri Lanka in
314accordance with applicable national laws, institutional research governance requirements and
315the principles of the Declaration of Helsinki. The study was an anonymous, voluntary,
316minimal-risk survey of adult participants on non-sensitive agricultural topics, in which no
317personally identifying information was collected.
318Climate-smart agriculture practices and adoption stages
319The five CSA practices were those prioritised most highly among ten expert-validated CSA
320practices for smallholder rice farming in the community-managed seasonal tank irrigation
321systems of Anuradhapura District [31]. Current use was defined separately for each practice.
322For example, a farmer was considered a current user of certified varieties if they had sown
323certified seed of a climate-resilient variety on at least part of their paddy land. Table S2 in S1
324Appendix gives the standard description of each practice and the criterion that defined its
325current use.
326Awareness was recorded first, by asking whether the respondent had heard of the practice
327before the interview. The interviewer then read the standard description of the practice.
328Respondents next reported whether they had used the practice in the last two cultivation
329seasons. Those who had not used it stated whether they intended to use it within the next two
330seasons. Following the TTM stages described in the conceptual framework, these answers
331placed each farmer in one of three stages for each practice. Stage 1 comprised farmers who
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

15
332neither used the practice nor intended to use it. Stage 2 comprised farmers who did not use
333the practice but intended to do so, and Stage 3 comprised current users. Health applications of
334the TTM usually define intention over the next six months [33]. A horizon of two cultivation
335seasons was used instead, as decisions on rice practices are made season by season.
336Constraint items, ratings and rankings
337Constraint items were developed from a structured review of the related literature,
338supplemented by the key informant interviews. This process produced 74 candidate
339constraints. Ten constraints were then selected for each practice against six criteria specified
340in advance. These were the strength of the evidence, relevance to Sri Lanka and to seasonal
341tank systems, coverage of the COM-B components, non-redundancy, farmer-reportability and
342the potential to discriminate between adoption stages. This procedure follows recommended
343practice in item development, which combines a literature review with key informant input
344and judges candidate items against explicit criteria [51]. COM-B-based surveys of farmers
345have similarly derived their items from the literature on the practice concerned [38]. The full
346selection procedure is described in S2 Appendix.
347Five constraints were common to all practices and were worded to suit each practice. Lack of
348knowledge refers to insufficient understanding of how to apply the practice [10,11]. Doubtful
349benefit refers to uncertainty about whether the practice improves yield or income [17]. Risk
350or distrust refers to fear of loss from the practice or a lack of trust in it [17,40]. Cost refers to
351the expense of the inputs, services or equipment that the practice requires [10,11]. Lack of
352training or advice refers to insufficient training, demonstration or advisory support [12,13].
353The other five constraints were specific to each practice. They included constraints arising
354from collective decisions on tank management, such as water-release dates fixed for the
355whole command area. Before the survey, each of the 50 constraints was assigned to one of
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

16
356the six constraint groups defined in the conceptual framework. Table S3 in S1 Appendix lists
357the constraints and their groups. The common constraints were asked first, in the same order
358for every practice.
359Every respondent rated all ten constraints for each practice. For each constraint, respondents
360indicated how much it held them back from using, or continuing to use, the practice. The four
361response categories were not at all, a little, moderately and very much, scored from 0 to 3. A
362‘cannot say’ option was also available. This scale follows the four-point scale of the Problem
363Confrontation Index (PCI) [10]. Respondents then ranked the three constraints that held them
364back most, choosing only from the constraints they had rated 2 or 3. Respondents who had
365rated fewer than three constraints in these categories ranked only those constraints, and those
366who had rated none did not rank any.
367Rankings were collected in addition to ratings for two reasons. Severity ratings can be
368distorted by individual response styles, such as a general tendency to agree with statements
369[52]. Rankings, in contrast, show each farmer’s own order of priority. Ranking was limited to
370three positions, as respondents tend to rank the first few items carefully and later ones
371carelessly [53].
372Data analysis
373The analyses followed the four objectives of the study. Severity ratings are ordinal, so
374inference was based on cumulative link models with a logit link, also known as proportional-
375odds models [54,55]. These models assume that each observed rating reflects a continuous
376latent severity, which is divided into the four rating categories by three thresholds:
377 𝑦∗𝑖𝑘 = 𝜂𝑖𝑘 + 𝜀𝑖𝑘,  𝑌𝑖𝑘 = 𝑐  if  𝜃𝑐―1 < 𝑦∗𝑖𝑘 ≤ 𝜃𝑐 (1)
378where y *ik is the latent severity of constraint k for farmer i, Yik is the observed rating (c = 0, 1,
3792, 3), η ik is a linear predictor and εik follows a standard logistic distribution. The thresholds
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

17
380satisfy θ −1 = −∞ < θ0 < θ1 < θ2 < θ3 = ∞. It follows that logit Pr(Yik ≤ c) = θc − ηik, so positive
381values of η ik indicate more severe ratings. Ratings were clustered within farmers, and farmers
382within GN divisions. Cumulative link mixed models (CLMMs) therefore included random
383intercepts for farmers and for GN divisions [56]. The farmer random intercept absorbs each
384farmer’s general level of reported severity. All high-risk GN divisions were included rather
385than sampled, but a random intercept was a parsimonious way to absorb differences between
386them. Table S5 in S1 Appendix summarises all the analyses described below.
387‘Cannot say’ responses, which made up 2.3% to 13.3% of the ratings for a practice, were
388treated as missing (Table S6 in S1 Appendix). The likelihood-based models assume that these
389responses were missing at random. Eight priority rankings (0.4%) included a constraint that
390the farmer had rated 0 or 1. They were retained as recorded, and no observations were
391excluded.
392Nominal-effect tests rejected the proportional-odds assumption for several terms. The main
393interaction tests were therefore repeated with two models that do not require this assumption.
394The first was a binary mixed model of ratings of 2 or 3 against ratings of 0 or 1. The second
395was a linear model of farmer-centred ratings, that is, each rating minus the farmer’s mean
396rating, tested with farmer-cluster robust Wald tests [57]. Centring removes each farmer’s
397general rating level and thus reduces the influence of acquiescent response style [52]. The
398linear model treats ratings as interval data, however, so it was used only as a robustness check
399[55]. Odds ratios (ORs) from the proportional-odds models are interpreted as average shifts
400across the three thresholds.
401Multiple testing was controlled within each family of related tests, as specified below for
402each objective. Holm’s procedure controlled the familywise error rate [58], and the
403Benjamini-Hochberg procedure controlled the false discovery rate (FDR) [59]. Each pre-
404specified hypothesis addressed a separate question and was tested without adjustment. All
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

18
405tests were two-sided at α = 0.05. Analyses used R version 4.4.2 [60] with the ordinal [61],
406lme4 [62], emmeans [63], survival [64] and multcompView [65] packages. S1 Appendix
407gives implementation details, software versions and random seeds, and S1 Code contains the
408R code that reproduces the analyses from S1 Dataset.
409Identification and prioritisation of constraints within each practice
410The first objective was to identify and prioritise the constraints to each practice. Two
411complementary measures were used, one based on severity ratings and the other on priority
412rankings. The severity of constraint k of practice j was summarised as the percentage of
413farmers who rated it 2 or 3. It was also summarised with the PCI [10]. The number of valid
414ratings n jk differed between constraints, so the index was expressed as a percentage of its
415maximum:
416 PCI%𝑗𝑘 = 100 ×
∑𝑖 𝑦𝑖𝑗𝑘
3𝑛𝑗𝑘 (2)
417where y ijk is farmer i’s rating from 0 to 3 and the sum runs over valid ratings. Priority was
418measured with a Weighted Priority Index (WPI), which weights first, second and third ranks
419by 3, 2 and 1:
420 WPI𝑗𝑘 = 100 ×
3𝑛1𝑗𝑘+2𝑛2𝑗𝑘+𝑛3𝑗𝑘
3𝑁 (3)
421where n rjk is the number of farmers who placed constraint k at rank r and N is the number of
422farmers analysed. The WPI ranges from 0, when no farmer ranks the constraint, to 100, when
423every farmer ranks it first. Constraints were ordered by WPI. Confidence intervals (CIs) for
424the WPI, PCI% and ranks were obtained from 2,000 bootstrap samples. Each sample
425resampled farmers within GN divisions, in line with the sampling design. Constraints with
426overlapping rank intervals were treated as tied.
427Severity differences among the ten constraints of each practice were tested with a likelihood-
428ratio test of the constraint effect in a CLMM. These differences were also described with
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

19
429Friedman tests [66], Kendall’s coefficient of concordance W [67] and Wilcoxon signed-rank
430tests [68] with letter displays [69]. The Wilcoxon tests were Holm-adjusted across the 45
431pairs of constraints within each practice. Agreement between the WPI and PCI% orders was
432measured with Spearman’s rank correlation ρ. Sensitivity checks restricted the WPI to
433farmers aware of the practice and resampled whole GN divisions [70].
434Comparison of common and practice-specific constraints
435The second objective was to test whether the common constraints weigh equally across
436practices and how they compare with practice-specific constraints. The five common
437constraints were analysed jointly with the following CLMM:
438 logitPr 𝑌𝑖𝑗𝑘 ≤ 𝑐 = 𝜃𝑐 ― 𝜋𝑗 + 𝜅𝑘 + (𝜋𝜅)𝑗𝑘 + 𝑢𝑖 + 𝑣𝑔(𝑖) (4)
439where Y ijk is farmer i’s rating of common constraint k for practice j, πj and κk are practice and
440constraint effects, (πκ) jk is their interaction, and ui and vg(i) are the farmer and GN random
441intercepts. A likelihood-ratio test of the interaction, with 16 degrees of freedom, assessed
442whether the severity profile of the common constraints differed between practices. Latent-
443scale intraclass correlation coefficients (ICCs), based on the logistic residual variance π 2/3
444[71], measured how consistently the same farmers reported a constraint across practices.
445They were computed from Eq. (4) and, for farmers, from a separate model for each common
446constraint. Predicted probabilities of a rating of 2 or 3 were obtained for a typical farmer,
447with random effects set to zero, together with delta-method 95% CIs. Within each common
448constraint, practices were compared with Holm-adjusted ORs and with Friedman and
449Wilcoxon tests. A common constraint was classified as broadly common if at least 50% of
450farmers rated it 2 or 3 for four or five practices. It was classified as practice-contingent if this
451applied to two or three practices, and as limited otherwise.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

20
452Within each practice, each farmer’s mean ratings of the practice-specific and common
453constraints were compared with Wilcoxon signed-rank tests, with the effect size r = |Z|/√n.
454The share of first-ranked constraints that were practice-specific was tested against 0.5 with an
455exact binomial test. These within-practice comparisons were not adjusted for multiplicity. Six
456recurring practice-specific themes, such as local input availability and peak-season labour
457scarcity, were compared across practices, with Holm adjustment across themes.
458H4 was tested on the four practices that included a collective constraint, as IPNM had none.
459For each farmer, the collective constraint ratings for the three calendar-bound practices were
460averaged, after first averaging the two CIS constraints. This mean was compared with the
461IWM collective constraint, insufficient early-season tank water, using a two-sided Wilcoxon
462signed-rank test. Each calendar-bound constraint was also compared with the IWM
463constraint, with Holm adjustment across the four comparisons.
464Differences in constraints across adoption stages
465The third objective was to test whether the relative importance of the same constraints differs
466across adoption stages within each practice. Following the criterion that constraints differ
467between stages [19], the ten constraints of each practice were modelled as follows:
468 logitPr(𝑌𝑖𝑘 ≤ 𝑐) = 𝜃𝑐 ― 𝛼𝑘 + 𝛿𝑠(𝑖) + (𝛼𝛿)𝑘,𝑠(𝑖) + 𝑢𝑖 + 𝑣𝑔(𝑖) (5)
469where s(i) and g(i) are the stage and GN division of farmer i. The terms α k, δs and (αδ)k,s are
470the constraint, stage and interaction effects, with the first constraint and Stage 1 as references.
471The farmer and GN random intercepts are u i ∼ N(0, σ2u) and vg ∼ N(0, σ2v). The primary test
472was the likelihood-ratio test of the constraint-by-stage interaction, with 18 degrees of
473freedom, Holm-adjusted across the five practices. A significant interaction indicates that the
474relative importance of the constraints differs between stages. The test was repeated with the
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

21
475binary and farmer-centred models and among farmers aware of the practice, without further
476adjustment.
477For each constraint, the OR for Stage 2 or Stage 3 against Stage 1 was exp[δ s + (αδ)k,s]. The
478stage effect of each constraint was tested with a Wald test with two degrees of freedom,
479Holm-adjusted within practice. The WPI was also computed within each stage using Eq. (3),
480with N as the number of farmers in the stage, and the priority orders of the stages were
481compared with Spearman’s ρ.
482Priority rankings were analysed with a rank-ordered logit model [53,72]. The model assumes
483that the importance farmer i attaches to constraint k has a systematic part η ik and a random
484part with an extreme-value distribution. Each ranking step is then a choice of the most
485important constraint among those not yet ranked:
486 Pr(ranking𝑖) = ∏𝑅𝑖𝑟=1 exp(𝜂𝑖,𝑘𝑟)∑𝑚∈𝐶𝑖𝑟 exp (𝜂𝑖𝑚),  𝜂𝑖𝑘 = 𝛼𝑘 + (𝛼𝛿)𝑘,𝑠(𝑖) (6)
487where R i ≤ 3 is the number of constraints that farmer i ranked, kr is the constraint ranked in
488position r and C ir is the set of constraints rated 2 or 3 that remained unranked at that step.
489Terms that are constant within a farmer, such as the stage main effect, cancel out. The model
490was estimated as a conditional logit. The constraint-by-stage terms were tested with a farmer-
491cluster robust Wald test with 18 degrees of freedom [57], Holm-adjusted across the five
492practices.
493H1 to H3 concern constraint groups rather than single constraints. Within practice j, the log
494OR of a group for a given stage contrast, ψ j, was the mean of the log ORs of its constraints.
495The same farmers rated all practices, so the group estimates of the P practices were pooled by
496generalised least squares:
497 𝜓 = 𝜄T𝑆―1𝜓𝜄T𝑆―1𝜄 ,  Var 𝜓 = (𝜄T𝑆―1𝜄)―1,  𝑄 = 𝜓 ― 𝜄𝜓 T𝑆―1 𝜓 ― 𝜄𝜓 (7)
498where ψ = (ψ 1, …, ψP)T, ι is a vector of ones and S = DRD. Here D is the diagonal matrix of
499the standard errors of ψ j from each CLMM, and R is the between-practice correlation matrix.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

22
500This matrix was estimated from a joint linear model of all 50 constraints with farmer-cluster
501robust covariance, so the pooled standard errors are approximate. The statistic Q generalises
502Cochran’s heterogeneity statistic [73] and was compared with a χ 2 distribution with P − 1
503degrees of freedom. H1 and H2 were supported when the pooled OR was significantly below
5041. As pre-specified, H3 was supported unless the pooled OR was significantly below 1. This
505criterion cannot demonstrate equivalence, so a post hoc non-inferiority margin of 0.80 was
506also applied to the lower limit of the 90% CI. Relative contrasts subtracted the mean log OR
507of all ten constraints of the practice from each group estimate. They therefore compare
508constraint groups within the same farmers and remove any stage difference in the general
509level of ratings.
510Robustness to awareness and farm-household characteristics
511The fourth objective was to assess whether stage differences persist after accounting for
512farmers’ awareness and farm-household characteristics. Five characteristics were selected
513before the analysis as potential confounders, as each has been associated with both adoption
514and the constraints that farmers face [8,14,16]. They were years of formal education, the
515number of contacts with extension or Agrarian Services officers in the previous 12 months,
516cultivated paddy extent, the number of household members regularly working in paddy and
517average monthly household income. These variables, which were recorded in the
518questionnaire, are defined in Table S4 in S1 Appendix. Awareness of each practice was also
519included. Awareness and extension contact may partly result from adoption, so adjusting for
520them may remove part of a genuine stage difference [74]. The adjusted test is therefore
521conservative. Age, gender, paddy-farming experience and farmer organisation role were used
522only to describe the sample.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

23
523Differences in these variables across stages were tested with Kruskal-Wallis tests [75] for the
524five characteristics and with Fisher’s exact test for awareness, with Holm adjustment across
525the 30 tests. The characteristics were standardised before modelling, and income was log-
526transformed. Collinearity was checked with Spearman correlations and variance inflation
527factors.
528Stage differences after adjustment were tested in two ways. First, the farmer-centred model
529was extended with constraint-specific effects of awareness and the five characteristics, and
530the constraint-by-stage terms were tested with a farmer-cluster robust Wald test. Second, Eq.
531(5) was refitted with constraint-specific awareness effects and tested with a likelihood-ratio
532test. Each of the 50 constraints was then modelled separately, with and without adjustment
533for all six covariates:
534 logitPr(𝑌𝑖𝑘 ≤ 𝑐) = 𝜃𝑘𝑐 ― 𝛿𝑘,𝑠(𝑖) + 𝑥T𝑖 𝛾𝑘 + 𝑣𝑔(𝑖) (8)
535where x i is the vector of covariates of farmer i and γk contains their coefficients for constraint
536k. A farmer random effect cannot be identified when each farmer contributes only one rating,
537so these models included only the GN random intercept. Stage was tested with a Wald test
538with two degrees of freedom, Holm-adjusted within practice. ORs are non-collapsible [76],
539so adjusted and unadjusted results were compared on direction and significance rather than
540magnitude.
541Associations between the covariates and the constraints came from the adjusted models in Eq.
542(8). Thirty-nine tests corresponded to six expectations specified before the analysis, for
543example that higher income reduces cost constraints and that larger paddy extent increases
544input and labour constraints. These tests formed a confirmatory family with FDR control, and
545the remaining 261 tests formed an exploratory family. For each expectation, the constraint-
546level log ORs were also pooled by generalised least squares, as in Eq. (7), with Holm
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

24
547adjustment across the six expectations. These cross-sectional associations are not interpreted
548causally.
549Results
550Sample characteristics, awareness and adoption stages
551All 440 respondents were the principal decision-makers for paddy farming in their
552households (Table 1). Most were men. On average, they had 22.6 years of paddy-farming
553experience and cultivated 2.35 acres (0.95 ha) of paddy.
554Table 1. Characteristics of the sampled farmers and their awareness and adoption stage
555for the five climate-smart agriculture practices (N = 440).
Variable
Mean (SD) or
% Range
Panel A. Farmer characteristics
Men, % 88.0
Age (years) 49.1 (7.3) 33 to 65
Education (years of schooling) 9.4 (2.5) 3 to 13
Experience in paddy farming (years) 22.6 (9.3) 4 to 36
Cultivated paddy extent (ha) 0.95 (0.42) 0.20 to 2.02
Household members working in paddy
(persons)
2.1 (1.0) 1 to 5
Monthly household income (LKR) 76,643 (17,870) 37,000 to 140,000
Extension contacts in the previous 12
months (count)
3.4 (1.9) 0 to 10
Farmer organisation office-bearer, % 15.0
Panel B. Awareness and adoption stage by practice
Practice Aware (%) Stage 1, n (%) Stage 2, n
(%)
Stage 3, n (%)
Certified climate-resilient varieties 98.4 36 (8.2) 68 (15.5) 336 (76.4)
Integrated plant nutrient management
(IPNM)
65.2 258 (58.6) 150 (34.1) 32 (7.3)
Integrated weed management (IWM) 85.2 152 (34.5) 105 (23.9) 183 (41.6)
Tray nursery and parachute transplanting 77.3 280 (63.6) 130 (29.5) 30 (6.8)
Climate information services (CIS) 94.3 84 (19.1) 97 (22.0) 259 (58.9)
556Stage 1 refers to farmers who neither used the practice nor intended to use it within the next
557two cultivation seasons. Stage 2 refers to those who did not use it but intended to, and Stage 3
558to those who used it in the last two seasons. Aware refers to having heard of the practice
559before the interview. Extension contacts are contacts with extension or Agrarian Services
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

25
560officers. All respondents were registered members of farmer organisations. LKR = Sri
561Lankan rupees, SD = standard deviation.
562Adoption stages differed markedly between practices. The share of users was 76.4%
563(336/440) for certified varieties, 7.3% (32/440) for IPNM, 41.6% (183/440) for IWM, 6.8%
564(30/440) for tray nursery and parachute transplanting, and 58.9% (259/440) for CIS. Most
565farmers were non-intenders for IPNM (258/440, 58.6%) and for tray nursery and parachute
566transplanting (280/440, 63.6%). Awareness of the practice ranged from 65.2% for IPNM to
56798.4% for certified varieties.
568Severity and priority of constraints within practices
569The leading constraint differed between practices (Fig 1). For certified varieties, IWM, and
570tray nursery and parachute transplanting, it was a physical opportunity constraint, that is, a
571shortage of inputs, labour, time or money, or unsuitable field conditions. For IPNM, a
572capability constraint ranked first, namely a lack of knowledge. For CIS, a use-experience
573constraint ranked first, namely doubt about the accuracy of forecasts. Within every practice,
574the ten constraints differed significantly in severity (likelihood-ratio test, χ²(9) = 457.9 to
575995.6, p < 0.001, Table S7 in S1 Appendix). Agreement among farmers on the order of
576severity was weak, however, with Kendall’s W of 0.11 to 0.26. The priority order based on
577the WPI closely matched the severity order based on the PCI% (Spearman’s ρ = 0.95 to 0.98,
578Table S7 in S1 Appendix).
579Fig 1. Priority of constraints to adopting the five climate-smart agriculture practices,
580measured by the Weighted Priority Index (N = 440). Panels show (A) certified climate-
581resilient varieties, (B) integrated plant nutrient management (IPNM), (C) integrated weed
582management (IWM), (D) tray nursery and parachute transplanting, and (E) climate
583information services (CIS). Horizontal lines are bootstrap 95% confidence intervals. Colours
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

26
584show the constraint group, filled markers show common constraints and open markers show
585practice-specific constraints. Cap = capability, Mot = motivation, OppP = physical
586opportunity, Adv = advisory support, Coll = collective, Use = use experience.
587For certified varieties, the unavailability of certified seed near sowing time had the highest
588priority (WPI = 45.8, 95% CI 41.8 to 49.7). This constraint was rated 2 or 3 by 80.0% of
589farmers. The unsuitability of the variety for the field or the water-issue period (WPI = 32.2)
590and the cost of seed (WPI = 31.7) followed as a tied pair. Lack of knowledge was among the
591lowest priorities, with a rank interval of 8 to 10.
592For IWM, a lack of time to weed at the critical crop stage ranked first (WPI = 35.0). Three
593constraints followed with equal priority (WPI = 23.9). These were the unavailability of
594weeding labour, the cost of weed control and the difficulty of controlling weeds or weedy rice
595in broadcast fields. Their rank intervals overlapped with those of two further constraints,
596namely fear of yield loss when less herbicide is used and insufficient early-season tank water.
597For tray nursery and parachute transplanting, three constraints had overlapping rank intervals
598and thus similar priority. These were the cost of seedling trays (WPI = 36.8), the scarcity of
599labour in the peak season (WPI = 33.1) and a lack of skill in raising tray seedlings (WPI =
60031.2).
601For IPNM, not knowing how to combine organic and chemical fertilisers ranked first (WPI =
60244.1), and 84.2% of farmers rated it 2 or 3. The belief that more chemical fertiliser gives a
603higher yield ranked second (WPI = 30.4). The cost of buying or transporting compost and the
604lack of training ranked lowest, with rank intervals of 9 to 10. For CIS, doubt about forecast
605accuracy ranked first (WPI = 39.2), and 79.4% of farmers rated it 2 or 3. Three constraints
606followed with similar priority (WPI = 27.0 to 27.2, rank intervals 2 to 4). These were the lack
607of location-specific forecasts, forecasts that arrive after decisions are made at the pre-season
608cultivation meeting, and collective season plans that prevent farmers from acting on forecasts.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

27
609These priority orders were robust to two sensitivity checks. They agreed closely with the
610priority orders among farmers who were aware of the practice (Spearman’s ρ = 0.98 to 0.99,
611Table S11 in S1 Appendix). The leading constraint was also stable when whole GN divisions
612were resampled. Its rank interval remained 1 to 1 for four practices and widened only
613slightly, to 1 to 3, for tray nursery and parachute transplanting.
614Common and practice-specific constraints
615The five constraints common to all practices showed different severity profiles across
616practices (Table 2). The interaction between practice and constraint was significant
617(likelihood-ratio test, χ²(16) = 1517.7, p < 0.001) and remained so in the binary and farmer-
618centred models (Table S12 in S1 Appendix). For a typical farmer, the predicted probability of
619rating the lack of knowledge 2 or 3 ranged from 0.20 for certified varieties to 0.90 for IPNM.
620Cost was the only constraint rated 2 or 3 by at least half of the farmers for four practices, and
621it was therefore classified as broadly common (Tables S13 to S15 in S1 Appendix). This
622classification was marginal, as only 54.0% of farmers rated cost 2 or 3 for the fourth practice,
623CIS. The advisory support constraint, lack of training or advice, reached this level only for
624tray nursery and parachute transplanting.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

28
625Table 2. Predicted probability that a typical farmer rated each common constraint as
626moderately or very much constraining, by practice, and the commonality of each
627constraint across practices.
Common constraint
Certified
varieties IPNM IWM
Tray nursery
and parachute
transplanting CIS
Practices
≥ 50%
Farmer
ICC Classification
Lack of knowledge
(Cap)
0.20
(0.14 to
0.26)
0.90
(0.87 to
0.94)
0.39
(0.30 to
0.48)
0.82
(0.77 to 0.88)
0.24
(0.17 to
0.30)
2 0.41 Practice-
contingent
Doubtful benefit
(Mot)
0.29
(0.22 to
0.37)
0.71
(0.63 to
0.78)
0.34
(0.25 to
0.42)
0.52
(0.43 to 0.62)
0.55
(0.46 to
0.64)
3 0.58 Practice-
contingent
Risk or distrust (Mot) 0.31
(0.23 to
0.39)
0.76
(0.69 to
0.83)
0.67
(0.59 to
0.75)
0.41
(0.32 to 0.50)
0.86
(0.81 to
0.90)
3 0.36 Practice-
contingent
Cost (OppP) 0.75
(0.68 to
0.82)
0.37
(0.28 to
0.45)
0.73
(0.66 to
0.80)
0.90
(0.87 to 0.93)
0.57
(0.48 to
0.66)
4 0.49 Broadly
common
Lack of training or
advice (Adv)
0.40
(0.32 to
0.49)
0.42
(0.33 to
0.51)
0.36
(0.28 to
0.45)
0.63
(0.54 to 0.72)
0.33
(0.25 to
0.41)
1 0.60 Limited
628Probabilities and 95% confidence intervals (CI) are from the cumulative link mixed model in
629Eq. (4), with random effects set to zero. Practices ≥ 50% is the number of practices for which
630at least 50% of farmers rated the constraint 2 or 3 (Table S13 in S1 Appendix). Constraints
631were classified as broadly common for four or five such practices, practice-contingent for two
632or three and limited otherwise. Farmer ICC is the latent-scale intraclass correlation
633coefficient, and higher values indicate that the same farmers reported the constraint across
634practices. For CIS, the risk constraint, doubt about forecast accuracy, belongs to the use-
635experience group. Cap = capability, Mot = motivation, OppP = physical opportunity, Adv =
636advisory support, CIS = climate information services, IPNM = integrated plant nutrient
637management, IWM = integrated weed management.
638Farmers who reported advisory support or benefit constraints for one practice tended to report
639them for the other practices as well. The farmer ICCs of these two constraints were 0.60 and
6400.58, respectively. The risk or distrust constraint depended more on the practice, with a
641farmer ICC of 0.36. Overall, 48.6% of the unexplained variation in the ratings of the common
642constraints lay between farmers, and only 5.4% lay between GN divisions.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

29
643Practice-specific constraints received higher severity ratings than common constraints for
644four practices (Wilcoxon signed-rank tests, p < 0.001, Table S16 in S1 Appendix), with effect
645sizes of r = 0.29 to 0.53. The exception was tray nursery and parachute transplanting, for
646which the mean ratings of practice-specific and common constraints were similar (1.84 and
6471.81, p = 0.149). In the priority rankings, practice-specific constraints were ranked first more
648often than common constraints only for certified varieties and IWM, at 67.9% and 63.0% of
649first ranks, respectively (binomial p < 0.001).
650Constraints of the same type also differed in severity between practices. Six such recurring
651themes appeared in two or more practices, and for all six, ratings differed significantly
652between the practices concerned (Holm-adjusted p ≤ 0.006, Table S17 in S1 Appendix). For
653example, local unavailability of the key input was rated 2 or 3 by 80.0% of farmers for
654certified seed, 62.2% for organic material and 72.9% for seedling trays.
655The fourth hypothesis (H4) predicted that collective constraints would be more severe for the
656calendar-bound practices than for IWM. This hypothesis was not supported. The mean
657collective constraint rating did not differ significantly between the calendar-bound practices
658and IWM (1.71 and 1.73, Wilcoxon signed-rank test, p = 0.477, r = 0.03, Table S18 in S1
659Appendix). Collective constraints were rated 2 or 3 by 52.4% to 69.2% of farmers for the
660calendar-bound practices. The IWM collective constraint, insufficient early-season tank
661water, was rated 2 or 3 by 61.0%. Comparisons of individual collective constraints showed
662mixed directions. The constraint that the variety is decided at the pre-season cultivation
663meeting was rated similarly to the IWM constraint (Holm-adjusted p = 0.885). The constraint
664on seedling age relative to water-issue dates was rated lower (p < 0.001), whereas both CIS
665collective constraints were rated higher (p = 0.010 and p = 0.012).
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

30
666Differences in constraints across adoption stages
667The relative importance of constraints differed significantly across adoption stages for all five
668practices (likelihood-ratio tests, χ²(18) = 82.7 to 208.7, Holm-adjusted p < 0.001, Table S19
669in S1 Appendix). This result held in the binary and farmer-centred models and among
670farmers who were aware of the practice.
671At the level of individual constraints, 19 of the 50 constraints differed significantly across
672stages after Holm adjustment within practice (Table S20 in S1 Appendix). All 19 were
673capability or motivation constraints, and all received lower severity ratings in later stages.
674The largest difference concerned the lack of skill in raising tray seedlings. This constraint
675was rated 2 or 3 by 85.8% of Stage 1 farmers but by only 33.3% of Stage 3 farmers (OR =
6760.06, 95% CI 0.02 to 0.21). No physical opportunity, collective or use-experience constraint
677differed significantly across stages.
678Combining the estimates across the five practices revealed a clear pattern by constraint group
679(Fig 2A, Table S21 in S1 Appendix). Compared with Stage 1, capability constraints were
680rated lower in Stage 2 (OR = 0.39, 95% CI 0.29 to 0.52) and in Stage 3 (OR = 0.17, 95% CI
6810.12 to 0.24). Motivation constraints showed the same pattern, with ORs of 0.49 in Stage 2
682and 0.21 in Stage 3. These results support H1 and H2 for both later stages.
683Fig 2. Pooled stage effects by constraint group across the five practices. Panel A shows
684odds ratios of a more severe rating in Stage 2 (intenders) and Stage 3 (users) relative to Stage
6851 (non-intenders). Panel B shows the same contrasts relative to the average stage effect of all
686ten constraints of the practice. Values below 1 indicate lower severity in the later stage.
687Horizontal lines are 95% confidence intervals, and the number of practices pooled is shown
688in brackets. H1 to H3 are the hypotheses tested by each group.
689Physical opportunity constraints were rated higher in Stage 2 (OR = 1.78, 95% CI 1.34 to
6902.36) but did not differ significantly between Stage 3 and Stage 1 (OR = 1.25, 95% CI 0.90 to
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

31
6911.73). Collective constraints were rated higher in Stage 2 (OR = 2.21, 95% CI 1.51 to 3.24)
692and in Stage 3 (OR = 1.59, 95% CI 1.07 to 2.37). Use-experience constraints were also rated
693higher in Stage 2 (OR = 1.70, 95% CI 1.01 to 2.87) and in Stage 3 (OR = 1.89, 95% CI 1.19
694to 3.00). The Stage 2 estimate for use experience was marginal and was based on only two
695practices. None of these three groups was rated significantly lower in either later stage, so H3
696was supported under the pre-specified criterion. A post hoc non-inferiority check gave the
697same result, as the lowest lower limit of the 90% CI, 0.95 for physical opportunity in Stage 3,
698exceeded the margin of 0.80. Advisory support constraints did not differ significantly in
699Stage 2 (OR = 1.03) but were rated lower in Stage 3 (OR = 0.55, 95% CI 0.38 to 0.80). Stage
700effects did not vary significantly between practices (heterogeneity p ≥ 0.129).
701Within-farmer comparisons, which measure each constraint group against the average stage
702effect of the practice, showed the same pattern (Fig 2B). Compared with Stage 1 farmers,
703Stage 3 farmers rated capability and motivation constraints lower than constraints in general,
704with relative ORs of 0.29 and 0.43. They rated physical opportunity, collective and use-
705experience constraints higher than constraints in general, with relative ORs of 2.28, 2.89 and
7063.28 (all p < 0.001). Advisory support constraints changed in line with the practice average
707(relative OR = 1.06, p = 0.628). Estimates for each practice and model diagnostics, including
708the proportional-odds tests, are given in Tables S22 and S23 in S1 Appendix.
709The leading constraints in farmers’ priority rankings also changed with stage (Table 3).
710Capability and motivation constraints occupied 11 of the 15 leading positions in Stage 1,
711compared with 2 of 15 in Stage 2 and 1 of 16 in Stage 3, where one position was tied. These
712counts are descriptive, as some adjacent positions differed by less than one index point. In
713Stage 1, the leading constraints included a preference for own seed for certified varieties, fear
714of yield loss with less herbicide for IWM and lack of skill for tray nursery and parachute
715transplanting. In Stages 2 and 3, seed availability, time to weed, tray cost and forecast
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

32
716accuracy ranked first for the corresponding practices. IPNM was a partial exception, as the
717knowledge constraint ranked first in Stages 1 and 2 and second in Stage 3. Priority orders
718were similar between Stages 2 and 3 (Spearman’s ρ = 0.77 to 0.95) but not between Stages 1
719and 3 (ρ = −0.20 to 0.19). The rank-ordered logit model compared priority orders only among
720the constraints that each farmer was eligible to rank. In this model, the ordering differed
721significantly across stages only for certified varieties (robust Wald test, χ²(18) = 40.9, Holm-
722adjusted p = 0.008). For the other practices, unadjusted p-values ranged from 0.077 to 0.269.
723Table 3. Three highest-priority constraints at each adoption stage, by practice.
Practice Stage 1 (non-intenders) Stage 2 (intenders) Stage 3 (users)
Certified
varieties

  1. Preference for own or neighbours’
    seed (Mot, 30.6)
  2. Certified seed not available at
    sowing (OppP, 25.0)
  3. Doubtful yield advantage (Mot,
    23.1)
  4. Certified seed not available at
    sowing (OppP, 47.5)
  5. Certified seed too costly (OppP,
    35.3)
  6. Variety unsuited to field or water-
    issue period (OppP, 28.9)
  7. Certified seed not available at
    sowing (OppP, 47.6)
  8. Variety unsuited to field or water-
    issue period (OppP, 35.4)
  9. Certified seed too costly (OppP,
    32.7)
    IPNM 1. Do not know how to combine
    fertilisers (Cap, 46.8)
  10. More chemical fertiliser gives
    more yield (Mot, 34.6)
  11. Fear of yield drop (Mot, 24.7)
  12. Do not know how to combine
    fertilisers (Cap, 42.0)
  13. More chemical fertiliser gives
    more yield (Mot, 24.7)
  14. Organic material not available
    (OppP, 24.0)
  15. Bulky organics hard to handle
    (OppP, 34.4)
  16. Do not know how to combine
    fertilisers (Cap, 32.3)
  17. Organic material not available
    (OppP, 28.1)
    IWM 1. Fear of yield loss with less
    herbicide (Mot, 37.1)
  18. Unsure of herbicide choice, dose
    and timing (Cap, 26.3)
  19. No time to weed at critical stage
    (OppP, 23.5)
  20. No time to weed at critical stage
    (OppP, 39.0)
  21. Weed control too costly (OppP,
    31.1)
  22. Weeding labour unavailable
    (OppP, 29.2)
  23. No time to weed at critical stage
    (OppP, 42.3)
  24. Weeds or weedy rice hard to
    control (Use, 34.6)
  25. Weeding labour unavailable
    (OppP, 29.5)
    Tray nursery and
    parachute
    transplanting
  26. Lack skill to raise tray seedlings
    (Cap, 40.8)
  27. Seedling trays too costly (OppP,
    33.2)
  28. Labour scarce at peak (OppP,
    31.8)
  29. Seedling trays too costly (OppP,
    41.0)
  30. Labour scarce at peak (OppP,
    34.9)
  31. Trays not available locally (OppP,
    25.4)
  32. Seedling trays too costly (OppP,
    52.2)
  33. Labour scarce at peak (OppP,
    37.8)
  34. Trays not available locally (OppP,
    24.4)
    CIS 1. Forecasts not actionable (Mot,
    30.2)
  35. Messages hard to understand
    (Cap, 25.0)
  36. Trust traditional signs (Mot, 19.8)
  37. Doubt forecast accuracy (Use,
    39.5)
  38. Collective season plan prevents
    action (Coll, 32.6)
  39. Forecast comes after pre-season
    cultivation meeting (Coll, 28.9)
  40. Doubt forecast accuracy (Use,
    46.1)
  41. Forecasts not location-specific
    (Use, 32.2)
    3= Forecast comes after pre-season
    cultivation meeting (Coll, 29.5)
    3= Collective season plan prevents
    action (Coll, 29.5)
    724Entries give the rank, the constraint, its group and the Weighted Priority Index within the
    725stage, and ‘=’ marks tied ranks. Stage sizes are given in Table 1, and the ranks of all
    726constraints in each stage are given in Table S20 in S1 Appendix. Cap = capability, Mot =
    727motivation, OppP = physical opportunity, Coll = collective, Use = use experience, CIS =
    728climate information services, IPNM = integrated plant nutrient management, IWM =
    729integrated weed management.
    This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
    (CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
    license
    EarthArXiv

33
730Robustness to awareness and farm-household characteristics
731Farmers in later stages differed from non-intenders in several characteristics (Table S24 in S1
732Appendix). Education, extension contact and awareness differed across stages for every
733practice, and income for four practices, with higher values in later stages (Holm-adjusted p ≤
7340.026). Paddy extent and household labour did not differ significantly across stages (Holm-
735adjusted p ≥ 0.550). Correlations among the five characteristics were weak to moderate
736(Spearman’s ρ ≤ 0.44), and the maximum variance inflation factor was 1.55 (Table S25 in S1
737Appendix).
738These differences did not explain the stage pattern in constraints. The constraint-by-stage
739interaction remained significant for every practice after adjustment for awareness and the five
740characteristics (robust Wald tests, χ²(18) = 80.6 to 134.8, p < 0.001, Table S19 in S1
741Appendix). It also remained significant when awareness effects specific to each constraint
742were added to the CLMM (likelihood-ratio tests, χ²(18) = 74.1 to 195.8, p < 0.001). The
743separate models for individual constraints, which included only the GN random intercept,
744gave a similar result. Without adjustment, 18 constraints differed significantly across stages,
745and 15 did so after adjustment. The direction and significance of the stage effect were
746therefore unchanged for 47 of the 50 constraints (Table S26 in S1 Appendix). Three
747constraints lost significance after adjustment. These were the belief that more chemical
748fertiliser gives a higher yield, fear of yield loss with less herbicide and lack of skill in raising
749tray seedlings. Their estimated stage effects changed only slightly, but they no longer reached
750significance after Holm adjustment.
751None of the 39 pre-specified associations between farm-household characteristics and
752individual constraints was significant after FDR control (q ≥ 0.156, Tables S27 to S29 in S1
753Appendix). None of the 261 exploratory associations was significant either. When the
754constraint-level estimates were combined for each pre-specified expectation, two associations
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

34
755were significant. Higher household income was associated with less severe cost constraints
756(OR per standard deviation of log income = 0.77, 95% CI 0.65 to 0.90, Holm-adjusted p =
7570.007). Larger paddy extent was associated with more severe input and labour constraints
758(OR = 1.20, 95% CI 1.06 to 1.36, Holm-adjusted p = 0.016). The pooled associations for
759education, extension contact, household labour and awareness were not significant (Holm-
760adjusted p ≥ 0.304, Table S29 in S1 Appendix).
761Discussion
762Practice-specific constraint profiles
763The leading constraint differed between the five practices. It was a physical opportunity
764constraint for certified varieties, IWM, and tray nursery and parachute transplanting. For
765IPNM, it was a capability constraint, and for CIS, a use-experience constraint. The five
766common constraints also showed sharply different severity profiles across practices. Cost was
767the only constraint that met the criterion for being broadly common, and it did so only
768marginally. Practice-specific constraints were rated as more severe than common constraints
769for four practices. Constraints of the same type, such as local unavailability of the key input,
770also carried different weight in different practices. These findings show that a single pooled
771list of constraints to CSA adoption would have concealed the constraints that matter most for
772each practice. They are consistent with expert ratings showing that behavioural drivers vary
773between agricultural practices [41]. They also agree with meta-analytic evidence that
774adoption determinants vary widely by technology and context [14].
775The leading constraints are consistent with practice-level evidence from comparable settings.
776For certified varieties, the unavailability of certified seed at sowing time was the highest
777priority. Lack of awareness and poor availability of seed are the main reasons why stress-
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

35
778tolerant rice varieties remain under-adopted [77]. In a study of rice farmers in Anuradhapura
779District, Ilangathilaka et al. [78] found that 43% used informal seed because formal seed was
780unavailable. For IWM, lack of time to weed at the critical crop stage ranked first, followed by
781the unavailability of weeding labour. This is expected where labour scarcity at critical crop
782stages delays or limits manual weed control [79]. For IPNM, the leading constraints were
783lack of knowledge and the belief that more chemical fertiliser gives a higher yield. The same
784belief has been identified as a cause of fertiliser over-application in Anuradhapura District
785[80]. The yield losses that followed Sri Lanka’s ban on chemical fertilisers in 2021 [81] may
786have reinforced this belief. For CIS, the leading constraints were doubts about the accuracy
787and local specificity of forecasts. These are among the constraints most often reported in
788studies from developed countries [40]. They also match the limited local resolution of climate
789services reported in Nepal [82].
790Advisory support constraints behaved differently from the other constraint groups. Farmers
791who reported a lack of advice for one practice tended to report it for the others, as was also
792the case for doubts about benefit. Across adoption stages, advisory support constraints also
793changed in line with the average for the practice. They therefore appear to reflect a general
794characteristic of the farmer, such as confidence in trying new practices, rather than gaps
795specific to a practice.
796Constraints along the adoption pathway
797For every practice, the relative importance of the same constraints differed between adoption
798stages. This agrees with a defining property of stage theories, namely that people in different
799stages face different constraints [19]. Earlier stage-based studies of constraints in agriculture
800were qualitative [23] or covered a single practice [16]. The present study adds a formal test
801across several practices. The stage differences were also systematic and supported H1 to H3.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

36
802Capability and motivation constraints dominated among non-intenders, whereas physical
803opportunity, collective and use-experience constraints dominated among intenders and users.
804The same pattern appeared when constraint groups were compared within the same farmers.
805It is therefore unlikely to reflect a general tendency of farmers in later stages to rate all
806constraints differently. The pattern also did not vary significantly between practices.
807Lower capability and motivation constraints in later adoption stages are consistent with the
808view of adoption as a dynamic learning process [17]. They also agree with the TTM
809expectation that perceived disadvantages weaken as people move towards action [33]. On-
810farm trials of climate-resilient rice in Bangladesh illustrate this mechanism, as the judgements
811of host farmers matched the measured performance of the varieties [83]. However, learning
812through own use cannot explain why these constraints were already lower among intenders.
813Farmers with more knowledge or higher expected benefits may have been more likely to
814form intentions, and some may have learned from neighbours. The cross-sectional data
815cannot separate such selection from learning. The stage pattern also refines qualitative
816evidence on conservation agriculture from Northeast China [23]. In that study, limited
817knowledge characterised the awareness stage, whereas economic constraints and a lack of
818skills in operating machinery limited implementation. In the present study, implementation
819requirements appeared mainly as physical opportunity constraints. The knowledge constraint
820remained prominent among users only for IPNM, which requires procedural knowledge of
821how to combine fertilisers.
822The persistence of physical opportunity constraints is consistent with the gap between
823intention and behaviour. People are less likely to act on their intentions when they lack the
824opportunities, resources or control needed to do so [84]. Similarly, physical opportunity
825constraints distinguished European farmer types with similar intentions to adopt CSA [39].
826Use-experience constraints, such as difficulty in controlling weeds or weedy rice under IWM
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

37
827and doubts about forecast accuracy under CIS, are likely to become evident through use. This
828corresponds to the implementation and confirmation stages of the innovation-decision
829process [18]. Higher ratings in later stages may also partly reflect greater recognition of
830constraints. In Odisha, India, participation in a CSA programme raised farmers’ awareness
831and, where implementation was inadequate, also increased the constraints they reported [11].
832Both explanations imply that the nature of the constraints changes as farmers progress
833towards adoption. Distinguishing between them requires longitudinal data.
834Farmers’ priority rankings indicate where the main stage contrast lies. Stage differences in
835priority orders were significant only for certified varieties, whereas severity ratings differed
836by stage for all practices. This difference arises partly from the design of the ranking model.
837The model compares only the constraints that a farmer had already rated 2 or 3, so it does not
838capture stage differences in which constraints entered that set. Stage was thus associated
839mainly with which constraints farmers rated as severe, rather than with how they ordered
840them. Priority orders were similar for intenders and users but differed between non-intenders
841and users. The main contrast therefore lies between non-intenders and the two later stages.
842Collective constraints on within-plot and calendar-bound
843practices
844Contrary to H4, the mean rating of collective constraints was similar for IWM and for the
845calendar-bound practices, whose timing depends on the collectively agreed cropping
846calendar. One likely reason lies in the IWM collective constraint, which concerned
847insufficient early-season tank water. Standing water at crop establishment suppresses weed
848growth in direct-seeded rice [85]. In seasonal tank systems, early-season water depends on
849the water-release dates agreed at the pre-season cultivation meeting [26]. Thus, even a
850within-plot practice such as IWM depends on a collective decision. However, early-season
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

38
851water also depends on rainfall and tank storage, so this constraint may overstate the role of
852collective decisions. The overall result also combines comparisons in opposite directions. The
853two CIS collective constraints were rated higher than the IWM constraint, whereas the
854constraint on seedling age relative to water-issue dates was rated lower. The hypothesised
855difference may therefore hold for some calendar-bound decisions but not for others. In
856seasonal tank systems, the more relevant distinction may be whether a practice depends on
857collectively timed water, rather than whether it is carried out within individual plots.
858Collective constraints have received little attention in research on CSA adoption, although
859they can directly limit individual practices. In a systematic review of climate information use,
860institutional support was the least represented of three thematic areas [86]. In a study of
861adaptation intentions in the dry zone of Sri Lanka, village identification was related only to
862Bethma, a highly communal practice of temporarily sharing land within the command area
863[87]. That finding suggests that collective influences matter mainly for communal practices.
864The present results show that collective constraints can weigh just as heavily on a practice
865carried out within individual plots. Treating collective decisions as a distinct form of social
866opportunity, separate from advisory support, made this constraint visible. It also showed that
867collective constraints were rated more severe in later stages. This supports the argument that
868the opportunity component of COM-B needs to be extended when behaviour depends on
869decisions shared among several actors [22]. Tank governance in Sri Lanka has been described
870as consistent with Ostrom’s design principles for collective choice [26]. The present results
871suggest that even collective arrangements of this kind can restrict the timing of individual
872practices.
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

39
873Role of awareness and farm-household characteristics
874Farmers in later stages had more education, more extension contact and greater awareness
875and, for most practices, a higher income. Yet adjusting for these differences left the stage
876pattern in constraints largely unchanged. None of the associations between these
877characteristics and individual constraints remained significant after FDR control. This agrees
878with evidence from the dry zone of Sri Lanka that socio-psychological factors predicted
879farmers’ adaptation intentions better than demographic characteristics [87]. Two pooled
880associations were consistent with prior expectations. Higher income was associated with less
881severe cost constraints, and larger paddy extent with more severe input and labour
882constraints. Paddy extent, however, did not differ between stages. Farmers with larger farms
883may therefore face more severe input and labour constraints without being at a different
884adoption stage. Such scale effects may help to explain why the effect of farm size on
885adoption varies in sign across studies. This effect was negative among rice farmers in Nepal
886[88] but positive on average across developing countries [14].
887Theoretical and practical implications
888The findings have three theoretical implications. First, they indicate that the TTM and COM-
889B are complementary. The TTM describes the stage structure of adoption, whereas COM-B
890describes the content of the constraints that operate at each stage. Second, the practice-
891specific constraint profiles support treating the attributes of a practice as a determinant of
892which constraints limit its adoption [18,41]. Third, the stage pattern by constraint group was
893similar across practices. This suggests that the shift from capability and motivation
894constraints towards opportunity-related constraints may be a general feature of adoption.
895The findings also have practical implications. They support interventions that are matched to
896farmers’ adoption stage [19,24] and to the COM-B component that constrains their behaviour
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

40
897[20]. For non-intenders, education, training and field demonstrations can address the
898dominant capability and motivation constraints. This applies particularly to IPNM and to tray
899nursery and parachute transplanting, for which most farmers were non-intenders. In coastal
900Bangladesh, on-farm head-to-head adaptive trials increased the adoption of climate-resilient
901rice varieties, although limited seed distribution restricted wider uptake [89]. For IPNM, side-
902by-side demonstration plots could show that partly replacing chemical fertiliser with organic
903sources maintains yields. For intenders and users, enabling conditions matter more than
904persuasion. A reliable supply of certified seed [77], timed to the sowing dates agreed at the
905pre-season cultivation meeting, would address the leading constraint for certified varieties.
906Labour-saving services that do not require farmers to own machinery [90] could ease the
907labour and time constraints of IWM and of tray nursery and parachute transplanting. For CIS,
908location-specific forecasts could be issued before the pre-season cultivation meeting, at which
909seasonal decisions are made [26]. Reporting the past accuracy of these forecasts could also
910address the leading constraint, as farmers’ trust in forecasts depends on a record of accuracy
911[91].
912Limitations and future research
913Some limitations should be considered when interpreting these findings. First, the cross-
914sectional design compares farmers in different stages at one point in time, rather than
915following the same farmers as they move between stages. The stage differences therefore
916describe associations rather than transitions. Second, the measurement of stages was
917necessarily simplified. Stage 1 combined farmers who had not yet considered a practice,
918those who were undecided and those who had decided against it. The Precaution Adoption
919Process Model treats these as distinct groups [43]. Current use also did not capture partial use
920or later discontinuation. Third, three stage groups had fewer than 40 farmers. The absence of
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

41
921significant stage differences in individual opportunity constraints should therefore not be read
922as evidence that these constraints are equal across stages. Fourth, the use-experience group
923rested on three constraints in two practices. Finally, the study covered registered farmers in
924climate-vulnerable seasonal tank systems of one district, so the findings should be transferred
925to other systems with care.
926Future research could address these limitations in three ways. Panel studies that record stage
927transitions and discontinuation could show whether constraints change because farmers learn
928through use, or because farmers who already face fewer constraints are the ones who
929progress. Trials of stage-matched extension would test the practical value of the shift in
930constraints across stages. Multi-item measures of collective constraints, applied in other tank
931cascade systems and in major irrigation schemes, would show whether collective constraints
932are equally important in other irrigation settings.
933Conclusions and recommendations
934This study identified and prioritised the constraints to adopting five CSA practices for rice
935among 440 smallholder farmers under community-managed seasonal tank irrigation in
936Anuradhapura District, in the dry zone of Sri Lanka. It also examined whether these
937constraints were common across practices or specific to each practice, and whether their
938relative importance differed across adoption stages.
939The leading constraints differed between practices. Physical opportunity constraints ranked
940highest for certified varieties, IWM, and tray nursery and parachute transplanting. For tray
941nursery and parachute transplanting, however, tray cost had a similar priority to labour
942scarcity and lack of skill. Lack of knowledge was the leading constraint for IPNM. For CIS,
943the leading constraint was doubt about forecast accuracy, a use-experience constraint.
944Practice-specific constraints were rated as more severe than common constraints for four
This manuscript is a preprint and has not been peer reviewed. The copyright holder has made the manuscript available under a Creative Commons Attribution 4.0 International
(CC BY) license and consented to have it forwarded to EarthArXiv for public posting.
license
EarthArXiv

42
945practices. Cost was the only constraint common to most practices, and only marginally so.
946Contrary to expectation, collective constraints arising from tank management decisions were,
947on average, as severe for IWM, a within-plot practice, as for the calendar-bound practices.
948Comparisons of individual collective constraints, however, differed in direction. As
949hypothesised, the type of constraint that mattered most changed across adoption stages.
950Capability and motivation constraints dominated among non-intenders, whereas physical
951opportunity, collective and use-experience constraints dominated among intenders and users.
952This stage pattern persisted after adjustment for awareness and five farm-household
953characteristics. Higher income was associated with less severe cost constraints, and larger
954paddy extent with more severe input and labour constraints.
955The study makes three contributions. First, it compares common and practice-specific
956constraints across five CSA practices within one sample of farmers. This comparison shows
957that a single pooled list of constraints would conceal the specific constraints that limit the
958adoption of each practice. Second, it provides a formal test of stage-specific constraints
959across practices, combining severity ratings, priority rankings and within-farmer
960comparisons. The results indicate that the TTM and COM-B are complementary, as the TTM
961describes the stages of adoption and COM-B the content of the constraints at each stage.
962Third, it extends COM-B with collective and use-experience constraint groups, both of which
963were rated more severe by intenders and users than by non-intenders.
964These findings have several implications for policy and extension practice. Extension support
965for CSA should be matched to both the practice and the farmer’s adoption stage, rather than
966delivered as a uniform package. For non-intenders, particularly for IPNM and for tray nursery
967and parachute transplanting, demonstrations and practical training could address the
968dominant capability and motivation constraints. For intenders and users, existing ext

论文原文与出版信息

EarthArXiv (California Digital Library)
查看论文原文

继续探索