Development and validation of a multidimensional behavioural change index and its application to farmer producer organisations in India: implications for sustainable smallholder food systems

1 Introduction

Agriculture remains one of the most fundamental human enterprises, underpinning food security, rural employment and the broader economy of most developing nations (FAO, 2023). In India, agriculture occupies a position of singular importance, serving not only as the principal source of food and raw materials but also as a vital means of livelihood for a large proportion of the rural population (OECD, 2023). Agriculture in India is deeply interwoven with social structure, culture and the everyday survival of millions of households, making the sector’s health inseparable from the well-being of the country’s people.

Yet the very population that depends on agriculture faces persistent and structural challenges (Bisht et al., 2020). The agriculture sector in India is highly fragmented, with over 86% of operational holdings less than 2 ha in size and 89.40% of agricultural households having an average per capita holding of only 1.08 ha (MoSPI, 2021). This structural reality places small and marginal farmers at a steady disadvantage in the market, through high input prices, low bargaining power, exploitation by intermediaries and post-harvest distress sales, which eventually reduces their income and jeopardises the sustainability of farming as a livelihood (Patil et al., 2025).

Recognising these barriers, successive policy efforts have sought to organise smallholders into collective institutions so that, through aggregation and joint action, they can overcome the disadvantages of acting in isolation. Cooperatives, Self-Help Groups (SHGs) and, more recently, Farmer Producer Organisations (FPOs) represent successive expressions of this collective-action logic, each intended to strengthen smallholders’ market access, input procurement, credit linkage and overall livelihood security (Markelova et al., 2009; Barrett, 2008). International experience reinforces this rationale, with a substantial body of evidence indicating that producer organisations can enhance smallholders’ market participation, commercialisation and household welfare through collective mechanisms (Bernard and Spielman, 2009; Fischer and Qaim, 2012; Verhofstadt and Maertens, 2015; Bachke, 2019).

However, the mere creation of such interventions is not, in itself, a measure of success. Establishing a cooperative, an SHG or an FPO, enrolling members and providing institutional support are necessary first steps, but they do not guarantee that the intended transformation actually takes place. The real and lasting success of any such intervention lies not in its formation but in the change it brings about in the people it is meant to serve, specifically, in whether participation genuinely reshapes the way farmers think, decide and act in their agricultural enterprise.

Moreover, even where these interventions deliver measurable economic gains, higher incomes, lower input costs or improved access to credit, such gains do not by themselves secure a durable improvement in livelihood. Economic outcomes are often transient and may erode once external support is withdrawn, unless they are accompanied by a deeper change in how farmers operate. Sustainable livelihood improvement therefore depends less on the economic outputs recorded at a single point in time than on whether the intervention reshapes farmer behaviour in a lasting way.

Behavioural difference is, in this sense, the true marker of an intervention’s impact, because durable improvements in livelihood flow from durable changes in farmers’ attitudes and conduct rather than from institutional presence or short-term economic gain alone. It is the shift toward greater market orientation, more active communication and information-seeking, calculated risk-taking, responsiveness to feedback and confidence in enterprise and credit decisions that ultimately sustains the benefits of any such intervention (Rogers, 2003; Dessart et al., 2019). Indeed, farmers’ intention to join and actively participate in producer organisations has itself been shown to depend on such behavioural and attitudinal factors (Möllers et al., 2018; Gagana and Velmurugan, 2023; Pabba et al., 2026).

Recent international scholarship reinforces the centrality of behavioural and psychological factors to farmer decision-making. Working across Cameroon and Kenya, Tabe-Ojong et al. (2024) showed that farmers’ aspirations and aspiration gaps, rather than resource constraints alone, help explain the adoption of climate-smart practices such as crop rotation and organic soil amendment, indicating that behavioural antecedents operate alongside, rather than merely following from, structural conditions. Using an extended Theory of Planned Behaviour framework, Erekalo et al. (2025) further found that the behavioural drivers of climate-smart agriculture adoption differ systematically between non-adopters, practice-adopters and technology-adopters, with social norms motivating initial adoption while perceived compatibility and ease-of-use sustain continued use, evidence that behavioural change unfolds unevenly across adoption stages rather than as a single uniform process.

In a closely related application to producer organisations, Mina et al. (2026) developed a multidimensional, behaviour-based Commitment Index for rice cooperative members in the Philippines, explicitly noting that prior studies had relied on attitudinal self-reports rather than a validated multidimensional behavioural construct, a gap that closely parallels the one motivating the present study. At the systems level, a recent review of socially inclusive approaches to sustainable agriculture found that participatory, empowerment-based and knowledge-co-production strategies are most consistently associated with durable adoption of sustainable practices, in contrast to purely informational or resource-based interventions (Mamabolo et al., 2026). This growing international literature situates the present study’s behavioural focus within an active research agenda, while underscoring that behavioural constructs, and the instruments used to measure them, should be multidimensional and context-appropriate rather than borrowed wholesale from single-outcome adoption models.

Despite this, the prevailing evidence on such interventions has concentrated almost exclusively on economic outcomes, such as income, input costs, savings and credit access, while the behavioural transformation that underlies these gains has received comparatively little systematic attention. The limited studies that have examined the behavioural dimension have generally relied on Entrepreneurial Behaviour Indices (EBI) that were originally developed for individual farmers and that capture only a narrow subset of traits, such as risk-bearing, innovativeness and economic motivation (Singh et al., 2022; Jose et al., 2023; Gn et al., 2023; Arvind et al., 2025; Anand and Ghosh, 2023).

As a consequence, best of our knowledge there are few validated, multidimensional instrument suitable for assessing the behavioural status that follows a developmental scheme or policy intervention. The limited presence of a fit-for-purpose tool, one that measures the full range of behaviours relevant to a farmer rather than a fragment of entrepreneurial disposition, represents a clear and pressing measurement gap and leaves a central question about the value of such interventions only partially answered.

The present study addresses this gap by developing a Behavioural Change Index (BCI) designed to assess the Behavioural status of farmers following a developmental scheme or policy intervention. The need for such an index is practical as much as scholarly: a standardised, validated instrument renders otherwise intangible behavioural difference measurable, and thereby allows implementing and promoting agencies, as well as the participating institutions themselves, to monitor behavioural outcomes alongside the economic indicators already in use, to identify which behavioural dimensions remain weak, and to target interventions, such as capacity-building, mentorship and exposure visits, where they are most needed.

The BCI developed here contributes an multidimensional framework, integrating twelve (12) theoretically grounded behavioural dimensions, several of which (marketing, communication, feedback, business skills and credit-seeking behaviour) are absent from existing single-domain indices; it is designed specifically to capture intervention-induced behavioural difference rather than borrowed from generic individual-farmer instruments; it is expert-weighted rather than assuming equal importance across dimensions; and it is subjected to a complete contemporary psychometric validation pipeline (DeVellis, 2017; Boateng et al., 2018). Together these features establish it as a reliable, content-valid and structurally sound tool of direct practical value to the farming community.

To validate the BCI and demonstrate its application, it was tested on members of Farmer Producer Organisations (FPOs). FPOs were chosen for this purpose because they constitute a developmental intervention of national importance: under the Central Sector Scheme for the Formation and Promotion of 10,000 FPOs, launched by the Government of India and implemented through agencies such as NABARD and SFAC, FPOs have become the leading contemporary instrument for organising smallholders in the country (Gulati and Juneja, 2022). Constituted as a hybrid model that combines the participatory principles of cooperatives with the governance discipline of corporate enterprises, FPOs bring scattered smallholders together, secure economies of scale and connect them to formal input and output markets (Trebbin and Hassler, 2012; Trebbin, 2014). Equally important for the present purpose, FPOs maintain formal membership records, which makes it possible to compare members with otherwise similar non-member farmers from the same villages, providing the comparison required to examine differences in Behavioural status between FPO members and non-members. FPO members thus serve as the test case for validating and applying the index in this study.

Accordingly, the present study pursues two sequential objectives. The first is to develop, standardise and validate the BCI as a multidimensional, expert-weighted instrument for assessing farmer behavioural status across relevant behavioural dimensions, validated here in the context of FPO membership. The second objective is to apply this validated index to examine the association between FPO membership and farmer behavioural status, comparing members with non-members using propensity-score matching, doubly robust estimation and E-value sensitivity analysis to account for observable differences and assess sensitivity to unmeasured confounding. Together these objectives provide both a rigorously validated measurement tool and the selection-adjusted evidence on the association between FPO membership and farmer behavioural status, while delivering a practical instrument that agencies can use to monitor behavioural outcomes and strengthen the support extended to smallholder farmers.

2 Materials and methods

2.1 Overview and two-objective research design

The present investigation was conceived as a single, integrated study pursued through two sequential and mutually reinforcing objectives. The first objective was to develop, standardise and validate a multidimensional Behavioural Change Index (BCI) for assessing farmer behavioural status across multiple behavioural dimensions, tested in this study among farmer producer organization (FPO) members. The second objective was to examine the association between FPO membership and the behavioural statuscaptured by this index. The two objectives are methodologically distinct but logically dependent: the first delivers a rigorously validated measurement instrument, while the second applies that instrument to quantify adjusted differences in behavioural status between FPO members and non-members using a non-randomised observational design. For this reason, both objectives are presented in sequence within this single Materials and Methods chapter, with the construction of the instrument (Objective 1) preceding its application (Objective 2).

A cross-sectional, ex post facto research design was adopted, as the behavioural attributes of interest had already developed through respondents prior and ongoing FPO experience and could not be experimentally manipulated. Accordingly, the study assesses behavioural status and between-group differences at a single point in time rather than longitudinal within-farmer behavioural change. The study was conducted in the state of Odisha, India, during the agricultural year 2025–26. Data were collected using structured, bilingual interview schedules administered to four distinct respondent groups, expert judges, non-sampled FPO members, sampled FPO members, and non-member farmers, each engaged according to the specific methodological requirement of the study phase concerned.

2.2 Study area

The districts of Mayurbhanj and Nayagarh were selected to represent contrasting levels of FPO density in Odisha. District-wise FPO enrolment data obtained from the Odisha Horticulture Department dashboard, which reported a state total of 948 enrolled FPOs, were used as the basis for district categorisation. Based on the number of enrolled FPOs, districts were classified into density categories, with Mayurbhanj falling under the highest-density category (>40 FPOs) and Nayagarh under the lowest-density category (<20 FPOs). From these two extreme density categories, one district was selected randomly from each category, resulting in the selection of Mayurbhanj and Nayagarh. Thus, the selection procedure ensured that the study deliberately captured contrasting FPO-density contexts while incorporating random selection within the respective categories. The contrasting institutional settings of the two districts provided an appropriate context for examining the behavioural implications of farmer participation in producer organisations. This contrast is analytically valuable because producer organisations function as hybrid institutions that combine cooperative principles with corporate governance and are intended to connect smallholders to formal input and output markets (Trebbin and Hassler, 2012; Trebbin, 2014).

2.3 Rationale and novelty of the behavioural change index

Research on producer organisations and similar collective interventions has largely worked through an economic lens, consistently reporting gains in income, input costs, market access, credit and welfare (Bernard and Spielman, 2009; Fischer and Qaim, 2012; Verhofstadt and Maertens, 2015; Bikkina et al., 2018; Bachke, 2019; Pallavi et al., 2024), yet it documents these outcomes while leaving the behavioural mechanism that produces them unmeasured; the few studies that do address behaviour rely on Entrepreneurial Behaviour Indices developed for individual farmers (Singh et al., 2022; Jose et al., 2023; Gn et al., 2023; Arvind et al., 2025; Anand and Ghosh, 2023) that capture only a narrow band of dispositional traits and omit the collective, market-facing behaviours most relevant once a farmer joins a scheme, leaving a clear gap that, to the best of our knowledge, limited availability of psychometrically validated multidimensional tool specifically designed to measure intervention-related farmer behavioural outcomes currently bridges. To substantiate this measurement gap, representative behavioural assessment approaches were compared in terms of their primary purpose, behavioural domains, validation approach and methodological scope (Table 1). Existing entrepreneurial behaviour instruments primarily assess dispositional and motivational traits, while technology adoption measures focus on adoption outcomes and empowerment measures emphasize agency, participation and decision-making. These approaches provide valuable measures of specific behavioural constructs, but they do not combine the broader entrepreneurial, market-oriented, communication, managerial and institutional behavioural domains within a single intervention-focused framework. Developing such an index is therefore both a scholarly and a practical necessity, since a standardised instrument renders otherwise difficult to measure behavioural attributes measurable and enables implementing and promoting agencies such as NABARD and SFAC, as well as the participating institutions themselves, to monitor behavioural outcomes alongside the economic indicators already in use and to target support, such as capacity-building, mentorship or exposure visits, where it is most needed. The Behavioural Change Index (BCI) developed here addresses this need and is novel in being genuinely multidimensional, integrating twelve theoretically grounded dimensions, several of which (marketing, communication, feedback, business skills and credit-seeking behaviour) are absent from existing single-domain indices; in being designed to capture multidimensional farmer behavioural status relevant to participation in developmental interventions rather than borrowed from individual-farmer instruments; in being expert-weighted rather than assuming equal dimension importance; and in being validated through a complete contemporary psychometric pipeline of content validation, reliability testing, exploratory factor analysis, discriminant validity and common-method-bias assessment (DeVellis, 2017; Boateng et al., 2018), together establishing it as a reliable, content-relevance and multidimensional structure instrument with clear advantages over existing tools.

Table 1. Comparison of representative farmer behavioural assessment approaches with the proposed behavioural change index (BCI).

2.4 Sampling procedure and respondent groups

Because the two objectives drew on different respondents, the sampling design is described separately for the instrument-development phase and for the impact-assessment phase.

2.4.1 Respondents for index development and validation (Objective 1)

For the development of the BCI, an expert panel comprising 80 responding judges, drawn from a larger pool of 180 invited experts, was constituted for relevancy analysis, content validation and dimension weighting. The panel included agricultural scientists, extension specialists, FPO programme managers, social scientists and experienced FPO practitioners, and its composition was verified to ensure representation across disciplines and across the geographic regions of India. Internal-consistency reliability of the preliminary scale was then assessed on a separate sample of 40 non-sampled FPO members drawn from FPOs lying outside the main study area, so that reliability testing did not contaminate the impact sample.

2.4.2 Respondents for the impact assessment (Objective 2)

A multistage sampling procedure was adopted for the main survey. In the first stage, the districts of Nayagarh and Mayurbhanj were selected purposively on the basis of FPO density (Section 2.2). In the second stage, operational FPOs were identified from records maintained by Krishi Vigyan Kendras (KVKs), SFAC and NABARD, and a stratified random sample of FPOs was drawn in proportion to district-wise FPO density. The required sample size was determined using Yamane’s (1967) formula, and in the final stage active member farmers were selected from FPO membership rosters through simple random sampling, yielding 300 FPO members across the two districts. A comparison group of 100 non-member farmers was constructed by simple random sampling from village-level household lists in the same villages as the sampled FPOs, subject to the restriction that no selected household had any member currently enrolled in an FPO. A 3:1 member-to-non-member ratio was chosen to maximise the precision of the propensity-score-matched estimates while remaining feasible within survey resources, giving a combined analytical sample of 400 farmers.

For all 400 respondents, eight socio-personal characteristics were recorded for later use as propensity-score covariates: age (years), education (years of formal schooling), landholding size (hectares), annual household income (INR), family size, number of earning members, family type (nuclear or joint) and occupation type (agriculture-only versus mixed). These variables were selected because prior literature identifies them as the primary determinants of selection into FPO membership (Arvind et al., 2025; Singh et al., 2022).

3 Objective 1: Construction, standardisation and validation of the BCI

3.1 Identification of behavioural dimensions and item generation

The BCI was constructed and standardised following a standard, sequential scale-development procedure, beginning with the identification of behavioural dimensions and the generation of items and progressing through expert relevancy evaluation, dimension weighting, reliability testing and full psychometric validation, with each stage described in turn below. Twelve behavioural dimensions of farmer behavioural status were identified through a systematic review of the agricultural extension and producer-organisation literature (Jose et al., 2023; Gn et al., 2023; Singh et al., 2022; Naik et al., 2025; Rogers, 2003; Dessart et al., 2019). These dimensions are consistent with contemporary behavioural and agricultural innovation research, which views farmer decision-making as multidimensional and influenced by cognitive, motivational, informational, social and market-related factors The dimensions were: (1) Marketing Behaviour, (2) Decision-making Behaviour, (3) Risk-bearing Ability, (4) Innovativeness, (5) Achievement Motivation, (6) Communication Behaviour, (7) Economic Motivation, (8) Self-confidence, (9) Information-seeking Behaviour, (10) Business Skills, (11) Feedback Behaviour and (12) Credit-seeking Behaviour.

An initial pool of 142 statements was generated from existing behavioural indices, inputs from KVK scientists and the Palladium group, qualitative interviews with FPO field coordinators, and investigator observation. Items were drafted as first-person declarative statements scored on a five-point Likert agreement scale (1 = Strongly Disagree to 5 = Strongly Agree) (Likert, 1932; Edwards and Kilpatrick, 1948). The pool was then screened against Edwards’ (1957) fourteen-point editing criteria, covering clarity, single-idea expression, brevity, absence of double negatives and item-total relevance, reducing it to 120 statements suitable for expert rating.

3.2 Expert evaluation and item retention

The 120 statements were submitted to the 80-member expert panel for a relevancy test on a three-point continuum: Most Relevant (3), Relevant (2) and Least Relevant (1). For each statement, three indices were computed from the panel responses. The Relevancy Percentage (RP) was the proportion of judges rating the item Most Relevant or Relevant, expressed as a percentage Equation (1).

Relevancy percentage:

( Most relevant response × 3 + relevant response × 2 + least relevant response × 1 ) × 100 Maximum Possible score ( 80 × 3 = 240 ) (1)

Relevancy weightage:

( Most relevant response × 3 + relevant response × 2 + least relevant response × 1 ) Maximum Possible score ( 80 × 3 = 240 ) (2)

Mean relevancy score:

( Most relevant response × 3 + relevant response × 2 + least relevant response × 1 ) Number of Judges ( 80 ) (3)

The mean relevancy score (MRS) was the weighted mean rating [MRS = (3 M + 2R + 1 L)/J]. The Relevancy Weightage (RW) was the normalised weight [RW = MRS/3], where M, R and L are the numbers of judges rating an item Most Relevant, Relevant and Least Relevant respectively, and J is the total number of responding judges (80).

Items were retained only if they simultaneously satisfied RP ≥ = 75%, MRS ≥ = 2.25 and RW ≥ = 0.75. Applying these three criteria to the 120 statements yielded a final pool of 89 items distributed across all twelve dimensions.

3.3 Dimension weighting

Because the twelve dimensions were not expected to contribute equally to overall behavioural difference, expert-derived weights were assigned using the Direct Rating Method (Arbel, 1989), a multi-criteria decision-making approach that allows judges to rate each criterion on a continuous scale without forcing trade-offs across criteria. Each expert rated the twelve dimensions on a ten-point importance scale (1 = least important, 10 = most important). The mean importance rating for each dimension was computed across the 80 raters, and final weights (Wj) were obtained by normalising these means so that they summed to unity [W_j = (mean rating of dimension j)/(sum of mean ratings across all dimensions)]. The resulting weights ranged narrowly from 0.0808 to 0.0888, reflecting near-equal but differentiated expert importance across dimensions.

3.4 Internal consistency reliability

The 89-item scale was administered to the 40 non-sampled FPO members. Cronbach’s alpha was computed for the full scale and separately for each dimension

α = K K − 1 ( S T 2 − ∑ i S i 2 S T 2 ) (4)

Acceptable reliability was defined as alpha ≥0.70 (Cronbach, 1951; Nunnally and Bernstein, 1994; Tavakol and Dennick, 2011), with values of 0.60–0.69 treated as borderline acceptable and values below 0.60 as poor. Split-half reliability was assessed by randomly dividing the items of each dimension into equal halves, correlating the half-scale scores (r) and applying the Spearman-Brown correction. Corrected item–total correlations and alpha-if-item-deleted statistics were additionally examined for dimensions showing relatively lower internal consistency. Item retention was considered in conjunction with the statistical evidence, theoretical relevance and content coverage of the construct.

3.5 Psychometric validation of the BCI

The full 89-item scale was then administered to the 300 FPO members, and multiple complementary psychometric procedures were applied to establish its reliability and validity, consistent with contemporary recommendations for multidimensional scale development (DeVellis, 2017; Boateng et al., 2018).

3.5.1 Structural validity exploratory factor analysis

The 300 respondents provided a respondent-to-item ratio of approximately 3.4:1 for the 89-item instrument. Although conventional rules of thumb have suggested higher ratios, factor-analytic adequacy is not determined by the respondent-to-item ratio alone and is influenced by sample size and the strength and structure of the underlying factors (Comrey and Lee, 1992; MacCallum et al., 1999; Hair et al., 2022). Thus, the adequacy of the present validation sample was considered in conjunction with the empirical characteristics of the data and subsequent factor-analytic diagnostics.

Factorability of the 89 × 89 inter-item correlation matrix was supported by a significant Bartlett’s test of sphericity (χ² = 53,613.1, df = 3,916, p < 0.001) and a KMO value of 0.76, indicating satisfactory sampling adequacy (Kaiser, 1974). Exploratory factor analysis was then conducted using maximum-likelihood extraction with direct oblimin (oblique) rotation, oblique rotation being preferred because the behavioural dimensions were expected to correlate (Fabrigar et al., 1999; Costello and Osborne, 2005). The number of factors to retain was determined by parallel analysis (Horn, 1965; Hayton et al., 2004), which compares observed eigenvalues against those from 1,000 randomly simulated datasets of the same dimensions; factors whose observed eigenvalues exceeded the 95th percentile of the simulated distribution were retained. The 12-factor solution explained 68.3% of the total variance, providing additional empirical support for the adequacy and interpretability of the factor structure. The rotated (oblimin) factor loading matrix for all 89 items across the twelve retained factors is provided in Supplementary file.

3.5.2 Confirmatory factor analysis (CFA), composite reliability (CR), average variance extracted (AVE)

Confirmatory factor analysis (CFA) was subsequently conducted separately for each of the twelve theoretically specified behavioural dimensions to examine the extent to which the observed items represented their respective latent constructs. Model adequacy was evaluated using the comparative fit index (CFI), Tucker–Lewis index (TLI), root mean square error of approximation (RMSEA) and standardised root mean square residual (SRMR). Composite reliability (CR) and average variance extracted (AVE) were calculated to assess construct reliability and convergent validity, respectively. CR values ≥0.70 were considered indicative of satisfactory composite reliability, while AVE ≥ 0.50 was used as the conventional criterion for convergent validity (Fornell and Larcker, 1981).

Measurement invariance was examined across FPO members and non-members using sequential configural, metric and scalar models. Changes in CFI (ΔCFI) were used to evaluate invariance, with a change of ≤0.01 considered supportive of invariance (Cheung and Rensvold, 2002; Chen, 2007).

3.5.3 Discriminant validity HTMT ratios

Discriminant validity across the twelve dimensions was assessed using Heterotrait-Monotrait (HTMT) ratios for all 66 pairwise dimension combinations. Values below the conservative threshold of 0.90 were taken as evidence of acceptable discriminant validity, as the HTMT criterion has been shown to detect construct overlap more reliably than traditional approaches (Fornell and Larcker, 1981; Henseler et al., 2015).

3.6 Common method bias

To assess whether common method variance threatened the construct validity of the BCI, Harman’s single-factor test was applied (Podsakoff et al., 2003). An unrotated principal-components analysis was performed on all 89 items; common method bias is indicated when a single factor accounts for the majority (≥50%) of total variance. The proportion of variance explained by the first unrotated factor was reported as the test statistic.

3.7 Composite BCI score

For each respondent, the BCI score was computed as a weighted sum of the dimension-level mean scores, rescaled to a 20–100 percentage scale:

BCI = ∑ D 1 × W 1 + D 2 × W 2 + D 3 × W 3 + … D 12 × W 12 max 0 × 100 (5)

D₁, D₂, D₃,…,D₁₂ = Dimensions

W₁, W₂, W₃,…,W₁₂ = weighted mean of dimension.

∑ = Summation of all Dimensions.

Where W is the normalised weight of dimension, D is respondent mean score on the items of dimension (each scored 1–5), and Max₀ is the maximum attainable weighted score (equal to 5, since the weights sum to unity). Higher scores indicate more favorable behavioural orientations and higher level of the behavioural attributes captured by BCI. [The complete construction of BCI instrument has been added as Supplementary material].

4 Objective 2: Examining the association between FPO membership and behavioural status

The second objective was to examining the association of FPO membership on BCI scores in the presence of non-random selection into membership. Three methods were applied in sequence, an unadjusted comparison, propensity-score matching, and doubly-robust estimation, followed by a formal sensitivity analysis for potential unmeasured confounding.

4.1 Unadjusted member versus non-member comparison

As a baseline, the overall BCI and all twelve-dimension scores were compared between members (n = 300) and non-members (n = 100) without covariate adjustment. The overall comparison used Welch’s independent-samples t-test, which does not assume equal variances, supplemented by the Mann–Whitney U test as a non-parametric confirmation. Effect size was quantified using Cohen’s d based on the Hedges pooled standard deviation. For the twelve dimension-wise comparisons, Holm’s sequential correction was applied to control the family-wise error rate, and effect sizes were reported with 95% bootstrap confidence intervals (5,000 resamples). Values of d of 0.20, 0.50 and 0.80 denote small, medium and large effects, respectively.

4.2 Propensity score estimation and matching

Because members and non-members were not randomly assigned, observable pre-existing differences could confound the comparison. Because the study was observational and cross-sectional, PSM was used to reduce imbalance in measured pre-membership characteristics rather than to establish causality. Propensity-score matching (PSM) was therefore used to construct a balanced comparison group (Rosenbaum and Rubin, 1983; Dehejia and Wahba, 2002). The propensity score, the conditional probability of FPO membership given the eight covariates, was estimated by logistic regression, and model discrimination was assessed by the area under the receiver-operating-characteristic curve (AUC). A 1:1 nearest-neighbour matching algorithm was applied on the logit of the propensity score with a caliper width of 0.2 standard deviations of the logit, without replacement (Austin, 2011; see also Caliendo and Kopeinig, 2008). Common support was verified by inspecting the propensity-score distributions, and post-matching covariate balance was assessed using absolute standardised mean differences (SMDs), with values below 0.20 (ideally below 0.10) taken as evidence of adequate balance (Stuart, 2010). The matched BCI difference was estimated by a paired t-test on the successfully matched pairs.

4.3 Doubly-robust ATT estimation

To protect the membership-effect estimate against misspecification of either the propensity or the outcome model, a doubly-robust (DR) estimator was applied to the matched sample. The DR estimator combines inverse-probability weighting from the propensity model with an outcome-regression model, and yields a consistent estimate of the average treatment effect on the treated (ATT) if either, though not necessarily both, of the two models is correctly specified (Bang and Robins, 2005; Funk et al., 2011; Imbens and Wooldridge, 2009). The outcome regression included the same eight covariates as the propensity model, specified as a linear additive function of the BCI on the treatment indicator and covariates. The DR-ATT was reported with its standard error, t-statistic and p-value.

4.4 Robustness checks

As robustness checks, the adjusted association was additionally estimated using inverse-probability weighting (IPW), generalised full matching and kernel-based propensity-score matching. These alternative approaches were used to assess whether the estimated member–non-member difference was sensitive to the choice of matching or weighting procedure. Consistency in the direction and approximate magnitude of the estimates across approaches was considered evidence of robustness to the analytical specification.

4.5 Sensitivity analysis - E-value

Because matching can adjust only for measured confounders, the robustness of the DR-ATT to potential unmeasured confounding was quantified using the E-value framework (VanderWeele and Ding, 2017). The E-value is the minimum strength of association, on the risk-ratio scale, that an unmeasured confounder would need to have with both FPO membership and the BCI, over and above the measured covariates, to fully explain away the observed ATT. Larger E-values indicate greater robustness. The E-value was computed at both the point estimate and the lower bound of the 95% confidence interval of the ATT.

4.6 Statistical software

All analyses were conducted in R (version 4.3.x). The psych package was used for Cronbach’s alpha, split-half reliability, EFA and parallel analysis; MatchIt for propensity-score estimation, 1:1 nearest-neighbour matching and SMD balance diagnostics; boot for bootstrap confidence intervals; the E-value package for the sensitivity analysis and visualization of the data done by ggplot2 and ggpubr for all figures. The significance level was set at alpha = 0.05 for all inferential tests, with Holm’s correction applied to the twelve dimension-wise comparisons.

5 Results

5.1 Behavioural change index (BCI) - construction and validation

5.1.1 Dimension-wise internal consistency (Cronbach’s alpha)

The Behavioural Change Index (BCI) was constructed from 89 Likert-scale items spanning 12 dimensions. Table 2 presents Cronbach’s alpha for each dimension. The full-scale alpha was 0.95. Split-half reliability was r = 0.928, yielding a Spearman–Brown corrected coefficient of 0.963. Figure 1 displays the reliability profile visually.

Table 2. Presents Cronbach’s alpha for each dimension.

Although the Credit-Seeking Behaviour dimension exhibited relatively lower internal consistency (α = 0.556), it was retained because it comprised five theoretically relevant items and represented an important component of the multidimensional Behavioural Change Index. Item-level analysis was therefore conducted to further examine its internal consistency. The corrected item–total correlations (Table 3) for CRED_1 to CRED_5 were 0.735, 0.704, 0.556, 0.236 and 0.057, respectively. Deletion of CRED_4 or CRED_5 would increase the alpha to 0.588 and 0.632, respectively. However, the items were retained based on their conceptual relevance and contribution to the content coverage of the Credit-Seeking Behaviour construct. This decision is consistent with scale-development literature indicating that Cronbach’s alpha can be influenced by the number of items and their inter-item relatedness, and that theoretically important items should not be removed solely to increase internal consistency (Cortina, 1993; Tavakol and Dennick, 2011; Boateng et al., 2018). The relatively modest reliability of this dimension is acknowledged as a limitation and its findings are interpreted cautiously.

Table 3. Item delination analsysis for credit seeking behaviour.

5.1.2 Factorability of the item matrix

Bartlett’s test of sphericity was significant (χ² = 53,613.1, df = 3,916, p < 0.001), confirming that the correlation matrix was factorable. The Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy was 0.76, consistent with the intentional multidimensionality of the instrument, which spans 12 theoretically distinct constructs.

5.2 Exploratory factor analysis variance explained

An oblique-rotation EFA (maximum likelihood, oblimin rotation) was conducted. Parallel analysis retained 12 factors, consistent with the a priori 12-dimension BCI structure (Figure 2). The 12-factor solution explained 68.3% of total variance (Table 4).

Table 4. Presents eigenvalues and cumulative variance.

5.2.1 Confirmatory factor analysis (CFA), composite reliability (CR) and average variance explained(AVE)

CFA was undertaken as a complementary assessment of the theoretically specified twelve-dimensional structure. The fit indices varied across dimensions, with Credit-Seeking Behaviour showing close model fit, whereas several other dimensions showed less satisfactory conventional fit indices (Table 5). Such variation was considered alongside the reliability and validity evidence rather than interpreted in isolation, consistent with recommendations for evaluating multidimensional measurement models (Hu and Bentler, 1999; Kline, 2016). It should be noted that several dimensions (e.g., Marketing Behaviour, Innovativeness and Information-Seeking Behaviour) comprise a small number of items with correspondingly low degrees of freedom, a condition under which RMSEA and CFI are known to behave conservatively even when the underlying factor structure is substantively reasonable (Kenny et al., 2015); accordingly, model fit for these dimensions was interpreted jointly with their EFA loadings, composite reliability and item-content coherence, rather than against absolute fit thresholds in isolation.

Table 5. CFA/CR/AVE.

CR ≥0.70 and AVE ≥0.50 are commonly used reference criteria for composite reliability and convergent validity, respectively (Fornell and Larcker, 1981; Hair et al., 2022).

Composite reliability was satisfactory across all twelve dimensions (CR = 0.713–0.910), supporting the internal reliability of the latent constructs. AVE exceeded the conventional 0.50 criterion for Achievement Motivation, Communication Behaviour and Self-Confidence, whereas the remaining dimensions showed lower AVE values, ranging from 0.316 to 0.468 (Table 5). Thus, the evidence for convergent validity was stronger for some dimensions than others However, the CFA fit and AVE across dimensions also indicates that some dimensions may benefit from further item refinement and validation in independent samples given the acceptable composite reliability and the multidimensional nature of the instrument, the AVE results were interpreted alongside the factor loadings, CR and other validity evidence rather than as a stand-alone criterion (Fornell and Larcker, 1981; Hair et al., 2022).

Measurement-invariance analysis (Table 6) showed a ΔCFI of 0.007 between the configural and metric models, meeting the commonly used ≤0.01 criterion and supporting metric invariance across FPO members and non-members (Cheung and Rensvold, 2002; Chen, 2007). The scalar model produced a ΔCFI of 0.012 and slightly above the criterion 0.01 threshold and partially meet the conventional criterion for full scalar invariance. Accordingly, the results support comparability of the factor loadings across the two groups, while stronger assumptions concerning equality of item intercepts should be interpreted cautiously.

Table 6. Measurement invariance assessment.

5.3 Discriminant validity HTMT ratios

Heterotrait–Monotrait (HTMT) ratios were computed to assess discriminant validity across the 12 BCI dimensions (Table 7). The ratios ranged from 0.432 to 0.900, with all 66 construct pairs at or below the recommended 0.90 threshold; only two pairs (Decision-making-Economic Motivation and Communication-Business Skills) reached the threshold value of 0.900, while the remainder fell clearly below it, supporting adequate discriminant validity across the dimensions.

Table 7. Discriminant validity across the 12 BCI dimensions.

Values represent Heterotrait-Monotrait (HTMT) ratios of correlations between each pair of the twelve BCI dimensions, assessing discriminant validity. Lower values indicate stronger discriminant validity. Values at or below the conservative threshold of 0.90 (Henseler et al., 2015) support acceptable discriminant validity.

5.4 Common method bias Harman’s single-factor test

The first unrotated factor accounted for 25.9% of total variance, well below the 50% critical threshold, indicating that common method bias does not substantially threaten the validity of the BCI data.

5.4.1 Judges direct rating method on 10 scale continuum

Table 8 presents the expert-derived weights for the twelve BCI dimensions (Direct Rating Method; Arbel, 1989). Mean importance ratings ranged from 7.70 to 8.46 out of 10, giving normalised weights of 0.0808–0.0888 (summing to 1.000). Marketing Behaviour ranked highest (0.0888), followed by Decision-Making Behaviour (0.0857) and Business Skills (0.0851), while Innovativeness (0.0812) and Achievement Motivation (0.0808) ranked lowest. The narrow-spread shows experts rated all twelve dimensions as broadly comparable in importance, with market-facing and managerial competencies weighted marginally above the dispositional traits.

Table 8. Expert-derived weights for the twelve BCI dimensions.

Bold values indicate the highest-weighted (top three) behavioural dimensions: Marketing Behaviour (Rank 1), Decision-making Behaviour (Rank 2), and Business Skills (Rank 3).

5.5 Association between FPO membership and Behavioural status

5.6 Demographic profile

A total of 400 farmers were surveyed: 300 FPO members and 100 non-members. Table 9 presents socio-personal comparisons. Members had significantly higher education, landholding, annual income and family size, and a greater proportion of joint families (p < 0.05); they also reported significantly lower farming experience (p = 0.007). Age and the number of earning members did not differ significantly between the two groups.

Table 9. Demographic profile of members and non-members.

Education was coded on a six-point ordinal scale: 1 = Illiterate, 2 = Primary education, 3 = High school, 4 = Higher secondary education, 5 = Graduation, and 6 = Post-graduation and above. Landholding was coded according to the classification of the Ministry of Rural Development, Government of India (2019): 1 = Marginal (<1 ha), 2 = Small (1–2 ha), 3 = Semi-medium (2–4 ha), 4 = Medium (4–10 ha), and 5 = Large (>10 ha). Family type was coded as 1 = Nuclear and 2 = Joint. Age, farming experience, annual income, family size, and number of earning members were retained in their original units. The reported means for education and landholding represent the mean of the assigned category codes, while the mean for family type represents the mean of the binary coding.

5.7 Behavioural dimension wise member vs. non-member comparison

Table 10 presents mean scores, mean differences, t-statistics, Cohen’s d, and Holms-adjusted p-values for each of the 12 BCI dimensions. Members scored significantly higher on all dimensions (all p_adj < 0.001). Effect sizes were consistently large (d = 1.29 to 2.21). Economic Motivation (d = 2.21) and Feedback Behaviour (d = 2.14) showed the largest differences. Figure 3 presents effect sizes with 95% confidence intervals.

Table 10. Behavioural dimension wise member vs. non-member comparison.

5.8 Overall behavioural status: member vs. non-member comparison (unadjusted)

FPO members recorded a mean BCI of 84.53 (SD = 7.94) compared to 62.41 (SD = 10.36) for non-members a difference of 22.12 points (Welch t = −19.52, df = 139.8, p < 0.001; Mann–Whitney p < 0.001; Cohen’s d = 2.57). The distributions are plotted in (Figure 4).

5.9 Propensity score matching

To address selection bias, 1:1 nearest-neighbour PSM was performed using a logit model on eight covariates with a caliper of 0.2 SD. The propensity score model yielded AUC = 0.818. The PSM flow is shown in (Figure 5).

Propensity-score diagnostics were conducted to assess model performance, common support, overlap, positivity and covariate balance. The propensity-score model showed good discriminatory ability, with an area under the receiver operating characteristic curve (AUC) of 0.818. Visual inspection of the propensity-score density distributions (Figure 6) showed substantial overlap between FPO members and non-members across the lower and middle propensity-score ranges, although overlap was limited at the upper end of the distribution. A decile-based positivity assessment further indicated limited overlap in the highest propensity-score decile. Therefore, the matched estimates were interpreted within the region of common support and not extrapolated to areas with limited positivity. After matching, standardised mean differences fell below 0.20 for all covariates, confirming adequate balance (Figure 7). The matched BCI difference was 19.34 points (t = 12.06, p < 0.001; Cohen’s d = 1.92).

5.10 Doubly-robust ATT estimate

Doubly-robust estimation (Table 11) yielded an adjusted member–non-member difference of 19.71 BCI points (SE = 1.51, t = 13.04, p < 0.001), indicating a robust positive association between FPO membership and BCI scores after adjustment for the observed covariates.

Table 11. Doubly-robust ATT estimate.

5.10.1 Robustness checks using alternative matching/weighting approaches

As robustness checks, the adjusted association between FPO membership and BCI scores was re-estimated using three alternative propensity-score approaches: inverse-probability weighting (IPW), generalized full matching, and kernel-based matching (Table 12). The estimated ATT was 21.25 BCI points under IPW, 23.80 points under full matching, and 23.68 points under kernel matching, compared with 19.71 points from the primary doubly-robust estimator. All approaches yielded positive and statistically significant estimates of similar overall magnitude, indicating that the observed membership-associated difference was not dependent on the choice of the primary matching specification.

Table 12. Robustness checks using alternative propensity-score approaches.

5.11 E-value sensitivity analysis

The E-value at the point estimate was 6.685 and at the lower confidence bound was 5.408. An unmeasured confounder would need to be associated with both membership and BCI by a risk-ratio of at least 6.69 on top of all measured covariates to fully explain away the observed ATT. The matched estimate is therefore highly robust to unmeasured confounding.

6 Discussion

This study pursued two objectives: developing and validating a multidimensional Behavioural Change Index (BCI) for measuring farmers behavioural status following a developmental intervention, validated here on Farmer Producer Organisation (FPO) members; and examining the association between FPO membership and Behavioural status through a member–non-member comparison adjusted for self-selection. The findings for both objectives are discussed below in relation to the existing literature.

6.1 Development and validation of the behavioural change index

The BCI retained 89 of the original 120 statements on the basis of expert relevancy ratings, with every retained item meeting three simultaneous criteria: a relevancy percentage >75 per cent, a relevancy weightage above >0.75, and a mean relevancy score above >2.25. This rigorous retention process ensured that only items with strong content validity entered the final scale. The Edwards-style editing and Likert summated-rating procedure used here is well established in agricultural extension research for constructing composite behavioural indices (Jose et al., 2023; Gn et al., 2023).

Internal consistency was excellent at the full-scale level (Cronbach’s α = 0.95; split-half r = 0.928; Spearman–Brown corrected coefficient = 0.963), comfortably exceeding the conventional 0.70 threshold. These values are notably higher than those reported for comparable instruments in the FPO literature. Jose et al. (2023) reported adequate but modest reliability for their Entrepreneurial Behaviour Index among Kerala FPO members, and Mohammed et al. (2025) confirmed α > 0.70 for entrepreneurial behaviour dimensions among Tamil Nadu coconut farmers. The present BCI surpasses these benchmarks while covering a substantially broader domain of twelve weighted dimensions, strengthening its suitability for standardised assessment across farmer populations.

At the dimension level, reliabilities ranged from 0.556 (Credit-Seeking behaviour) to 0.879 (Communication behaviour), with seven of the twelve dimensions exceeding 0.70. The lower alpha for Credit-Seeking reflects a known measurement challenge: credit-seeking behaviour in rural India is shaped by informal norms, asymmetric access and household-specific factors that reduce item homogeneity (Malaisamy et al., 2026). The dimension was retained on theoretical grounds, since credit-seeking is a recognised pillar of market-oriented farmer behaviour (Singh et al., 2022; Naik et al., 2025) and its removal would have narrowed the construct coverage of the index. Item-level analysis further showed corrected item–total correlations ranging from 0.057 to 0.735; deletion of CRED_4 or CRED_5 increased alpha to 0.588 and 0.632, respectively, and the items were retained based on their conceptual relevance and contribution to construct coverage.

Structural validity was confirmed through exploratory factor analysis (maximum-likelihood extraction, direct oblimin rotation), in which parallel analysis retained exactly twelve factors explaining 68.3% of total variance a near-perfect correspondence with the a priori design of the scale. Discriminant validity was satisfactory across all 66 construct pairs, with HTMT ratios ranging from 0.432 to 0.900 and all pairs lying at or below the 0.90 threshold; only two pairs (Decision-making–Economic Motivation and Communication–Business Skills) reached the threshold value of 0.900, while the remainder fell clearly below it, confirming that the twelve dimensions measure empirically distinct aspects of farmer behaviour rather than a single undifferentiated trait. Common method bias was negligible, with the first unrotated factor accounting for 25.9% of variance, well below the 50% critical level. Taken together, these psychometric properties establish the BCI as a reliable, content-valid and multi dimension structure instrument with clear advantages over existing single-domain indices (Jose et al., 2023; Gn et al., 2023; Arvind et al., 2025), which capture only a subset of entrepreneurial traits without incorporating marketing, communication, feedback or credit-seeking behaviour.

The additional psychometric assessment provided complementary evidence supporting the proposed BCI. Composite reliability was satisfactory across all twelve dimensions, whereas AVE and CFA fit indices varied between dimensions, indicating that some dimensions may benefit from further item refinement. Metric invariance was supported across FPO members and non-members, whereas full scalar invariance was partially established. These findings suggest that the BCI demonstrates promising reliability and structural evidence while also highlighting the need for independent replication and longitudinal assessment of temporal stability. This balanced interpretation is consistent with contemporary recommendations that newly developed multidimensional instruments should be evaluated using multiple complementary sources of psychometric evidence rather than relying on a single criterion (DeVellis, 2017; Boateng et al., 2018).

Dimension weights assigned by the Direct Rating Method ranged narrowly from 0.0808 to 0.0888, with Marketing Behaviour ranked highest and Achievement Motivation lowest. This near-equal but differentiated weighting reflects expert consensus that all twelve dimensions contribute meaningfully to the multidimensional assessment of farmer behavioural status, while acknowledging their relative importance. The elevated weight of marketing behaviour aligns with Ajith and Binoo’s (2022) finding that marketing gains and entrepreneurial orientation are closely co-determined among FPC shareholders, and underscores the centrality of market engagement. The BCI’s multidimensional, weighted architecture makes it directly applicable as a monitoring and evaluation tool for intervention programmes at scale, addressing a gap noted in recent FPO policy reviews (Patil et al., 2025).

6.2 Association between FPO membership and behavioural status

FPO members recorded a mean BCI of 84.53 compared with 62.41 for non-members, a raw gap of 22.12 points (Cohen’s d = 2.57). After adjusting for self-selection on eight socio-personal covariates through 1:1 propensity-score matching, the adjusted member, non-member difference attenuated modestly to 19.34 points (d = 1.92), and the doubly-robust ATT estimate of 19.71 points (SE = 1.51) further confirmed this stability. Critically, the E-value at the point estimate was 6.69, meaning that an unmeasured confounder would need a risk-ratio association of at least 6.69 with both membership and the BCI, over and above all measured covariates, to nullify the estimated association. Such a confounding strength appears substantial in this setting, providing evidence that the membership-associated behavioural difference is robust to observed selection and sensitivity to unmeasured confounding. Although propensity-score matching was used to reduce observable selection bias based on baseline socio-economic characteristics, the possibility of residual confounding due to unmeasured baseline variables cannot be completely excluded. Variables such as extension contact, training exposure, institutional participation, market access and digital literacy were not included in the propensity-score model because they may represent post-membership factors rather than pre-treatment confounders. Nevertheless, the possibility of residual confounding due to unmeasured or inadequately measured pre-treatment characteristics cannot be completely excluded. Future longitudinal studies incorporating a broader set of pre-treatment characteristics may further strengthen adjustment for potential confounding and causal inference.

These results extend a growing body of evidence on FPO impact, most of which has focused on economic outcomes. Yazhini et al. (2025) found income gains for FPO members in Tamil Nadu through input-cost reduction and collective bargaining, while Malaisamy et al. (2025b) documented higher livelihood-security indices across education, food, health and economic domains, and Dudekula et al. (2025) reported significant improvements in income, savings and credit access among millet FPO members. The present study complements this economic evidence with the first selection-bias-adjusted to estimate the association between FPO membership on a comprehensive behavioural outcome a dimension increasingly recognised as foundational to FPO sustainability (Patil et al., 2025; Naik et al., 2025; Suresh et al., 2024). Methodologically, it advances the field beyond Malaisamy et al.'s (2025a) Heckman correction by adding doubly-robust estimation and E-value sensitivity testing and Veesam et al.’s (2025) PLS path-modeling approach to FPO performance assessment, offering a more transparent and rigorous observational framework for future behavioural impact studies. Similar conclusions have been reported internationally, where participation in producer organisations has been associated with improved market engagement and enhanced welfare outcomes among smallholders in Kenya and Rwanda (Fischer and Qaim, 2012; Verhofstadt and Maertens, 2015).

At the dimension level, members outperformed non-members on all twelve BCI components, with large and consistent effects (d = 1.29–2.21). Economic Motivation (d = 2.21), Feedback Behaviour (d = 2.14), Marketing Behaviour (d = 1.98) and Business Skills (d = 1.96) showed the greatest differentiation. This gradient is theoretically coherent: FPO participation may expose members to market transactions, input procurement, collective pricing and feedback-rich business environments that could be associated with stronger economic and marketing behaviours. Singh et al. (2022) made a parallel observation in Punjab, finding that functional FPO members excelled in business skills and economic motivation relative to non-functional counterparts; the present study extends this by quantifying effect sizes against a non-member baseline, clarifying what membership specifically adds beyond pre-existing differences. This gradient is theoretically informative rather than incidental: the dimensions showing the largest differences (Economic Motivation, Feedback Behaviour, Marketing Behaviour, Business Skills) are competence-based behaviours acquired through repeated practice, whereas FPO membership functions here less as a single undifferentiated treatment and more as a bundle of structured behavioural exposures to markets, feedback loops and collective pricing whose effects concentrate in the domains that membership actually rehearses. Arvind et al. (2025) corroborate the pattern from North Karnataka, where education, extension contact, social participation and income were all significantly related to entrepreneurial behaviour among FPO members the same profile variables that differentiated members from non-members in the present pre-matching sample.

Achievement Motivation recorded the smallest effect (d = 1.29), though it remains large by conventional benchmarks. This is consistent with Social Learning Theory (Bandura, 1977): dispositional motivational traits, rooted in early socialisation and personality formation, are less responsive to institutional exposure than skill-based or knowledge-mediated behaviours. Two complementary mechanisms plausibly account for this: institutional rehearsal, whereby recurring market transactions and feedback build competence-based behaviours directly, and differential malleability, whereby dispositional traits are comparatively stable once formed. Diffusion of Innovation theory (Rogers, 2003) offers a related reading, distinguishing the communication- and adoption-related stages of behaviour change, which institutional exposure accelerates directly, from deeper motivational reorientation, which typically requires more individually targeted intervention. The practical implication for extension programming is that routine FPO participation alone may be insufficient to shift achievement motivation, and that targeted interventions entrepreneurship development camps, peer mentorship by high-performing members, or aspirational exposure visits may be needed. Ofosu-Ampong et al. (2025), in a systematic review of behaviour-change studies in sustainable agriculture, found that empowerment-based and structural approaches produce more durable change than informational interventions alone, reinforcing the value of FPO membership as a structural platform while highlighting the need for targeted motivational inputs for harder-to-shift dimensions.

These findings also resonate with contemporary international research on farmer behaviour and producer organisation participation, which offers further theoretical purchase on the observed behavioural gradient. Erekalo et al. (2025) found that non-adopters’ future adoption intentions are driven mainly by social norms, whereas continuing adopters are driven by perceived compatibility and ease-of-use, a staged pattern broadly consistent with the present finding that skill- and practice-based behaviours, economic motivation, feedback and marketing, differentiated FPO members from non-members more strongly than the more dispositional achievement-motivation dimension. Tabe-Ojong et al. (2024) similarly demonstrated, using data from Cameroon and Kenya, that farmers’ aspirations, rather than resource access alone, predict the adoption of sustainable practices, lending cross-regional support to the interpretation that the FPO-associated behavioural differences documented here reflect psychological and institutional mechanisms rather than resource endowment alone.

In a closely comparable application, Mina et al. (2026) found that a multidimensional, behaviour-based Commitment Index for Philippine rice cooperatives was shaped more strongly by farming dependency and training exposure than by static demographic traits, echoing the present study’s finding that institutionally rehearsed behaviours respond more readily to membership than dispositional traits do. More broadly, Mamabolo et al. (2026) systematic review of socially inclusive approaches to sustainable agriculture found that participatory and empowerment-based interventions, of which FPOs are a prominent example, are more consistently associated with durable practice adoption than purely informational interventions; the present dimension-level results, showing the largest differences in market-facing and communicative behaviours, are consistent with this broader pattern and suggest that the behavioural returns to institutional membership are concentrated in domains that are directly rehearsed through participation rather than distributed evenly across all behavioural domains.

Taken together, the two objectives yield a mutually reinforcing set of findings. The BCI provides a validated, reproducible instrument for measuring the full behavioural scope of intervention impact, while the propensity-score analysis demonstrates that the adjusted membership–non-membership difference of approximately 19–20 BCI points remains substantial after accounting for the measured pre-existing farmer characteristics. Pallavi et al. (2024) argued that FPOs are underutilised as intermediaries of sustainable agricultural transitions; the higher behavioural scores observed among FPO members, spanning innovativeness, feedback responsiveness and market orientation, suggest this potential operates at the individual-farmer level. For national FPO policy, these findings support sustained investment in FPO promotion in rain-fed, smallholder-dominated contexts such as Odisha, and point to the BCI as a practical monitoring tool for agencies such as NABARD and SFAC to track behavioural outcomes alongside the economic metrics currently in use.

7 Conclusion

This study was motivated by a simple but important observation: development interventions are commonly evaluated in terms of economic outcomes, such as income, productivity, and market access, while their influence on farmers’ behaviour receives far less attention. However, lasting development depends not only on what farmers gain but also on how they think, make decisions, and respond to new opportunities. These behavioural attribute are crucial because they may influence whether the benefits of development interventions are sustained beyond the period of external support. Understanding such changes is therefore essential for assessing progress towards key Sustainable Development Goals, including poverty reduction (SDG 1), sustainable agriculture (SDG 2), and reduced inequalities (SDG 10).

To address this, we set out to develop a reliable way to assess multidimensional farmer behavioural status, and to examine whether behavioural differences were associated with FPO membership. The Behavioural Change Index that emerged showed pshycometric evidence and practical potential. Built from 89 expert-validated items across twelve weighted dimensions, it showed excellent reliability (α = 0.95), a clean twelve-factor structure explaining 68.3% of variance, and sound discriminant validity. Composite reliability provided satisfactory evidence of internal consistency across all dimensions, while variation in CFA fit and AVE across dimensions indicates that some dimensions may benefit from further refinement and independent validation. The instrument captures a far broader picture of farmer behaviour than the single-trait indices used until now, taking in market engagement, communication, feedback and credit-seeking alongside the usual dispositional traits. When applied to members and non-members, the index told a clear story. FPO members showed higher behavioural across all twelve dimensions, and their advantage of roughly 19 to 20 points held firm even after careful adjustment for the ways the two groups differ from the outset. The high E-value indicates that the observed association is relatively robust to potential unmeasured confounding, although it does not establish causal relationship. Together, these findings provide a psychometrically supported measurement tool and credible evidence that FPO membership is associated with more favourable farmer behavioural status. This has implications for sustainability because farmer behavioural attributes may influence the capacity to sustain development outcomes. By making behavioural dimensions more visible, the BCI may provide government agencies, development organisations and NGOs with complementary information alongside conventional economic indicators. The complete BCI instrument, including item wording, response scale, dimension-wise scoring procedure and the index-calculation formula, is provided in Supplementary file to support independent application and replication. It may help to identify behavioural areas that could benefit from further capacity-building and extension support and move small holders towards sustainable, equitable and resilient livelihoods and advancing the wider SDG agenda they serve.

8 Limitations and future studies

The findings suggest that the Behavioural Change Index (BCI) may serve as a useful complementary tool for assessing behavioural status alongside conventional economic indicators in FPO programmes. By providing a structured measure of multiple behavioural dimensions, the BCI may help implementing agencies and development organisations identify areas requiring capacity-building and extension support. However, the findings should be interpreted in light of certain limitations. First, the study employed a cross-sectional, ex post facto design; therefore, the observed differences between FPO members and non-members represent differences in behavioural status at the time of the study and cannot establish behavioural change over time or definitive causal effects of FPO membership. Although propensity score matching, doubly robust estimation, and sensitivity analysis using the E-value were employed to reduce the influence of observable selection bias, these approaches cannot completely eliminate the possibility of residual confounding due to unmeasured baseline characteristics. In addition, variables such as training exposure, extension contact, institutional participation, market access and digital literacy were not included in the propensity score model because they may represent post-membership factors rather than pre-treatment confounders. The study was conducted within a limited geographical setting representing contrasting FPO-density contexts in Odisha; therefore, the findings may not be directly generalisable to other agro-climatic regions, farming systems or institutional environments. The geographical scope of the study also limits the generalizability of the findings. Further, the BCI was developed and subsequently applied within the same study population, and test–retest reliability could not be assessed because data were collected at a single time point.

Future studies should therefore undertake longitudinal or panel-based research to examine behavioural differences over time and provide stronger evidence regarding temporal relationships and causal mechanisms. Independent validation of the BCI using separate samples from different geographical regions, states, agro-climatic zones, farming systems and institutional settings is also recommended to establish its stability, external validity and wider applicability. Future research should additionally assess test–retest reliability and examine the performance of the BCI across different agricultural and rural development interventions. Studies incorporating a broader set of pre-treatment characteristics may further strengthen adjustment for potential confounding. While the behavioural differences observed between FPO members and non-members are encouraging, these findings should be considered as evidence of association rather than definitive causal effects, and broader policy recommendations should be developed only after further longitudinal and multi-regional validation. Overall, the present study provides an initial empirical framework for assessing multidimensional behavioural status and demonstrates the potential value of behavioural indicators as a complement to conventional economic measures for evaluating FPOs and similar agricultural development interventions.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the authors used Claude (Anthropic, claude-sonnet-4-6, Anthropic PBC, San Francisco, CA, United States) to assist with language editing, sentence restructuring, and formatting. The authors reviewed and verified all content produced with this assistance, including the accuracy of all citations, data, and factual claims. All figures were prepared independently by the authors. The use of this tool did not involve any generation of original scientific content, data, or analysis.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fsufs.2026.1915367/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Raji Pushpalatha, Amrita Vishwa Vidyapeetham (Amritapuri Campus), India

Reviewed by: S. R. Bishnoi, Indian Agricultural Research Institute (ICAR), India

Sagar Wadkar, National Cooperative Union of India, India