Spatial dependence refers to the level of similarity or dissimilarity of observed data in space, and it is sometimes referred to as spatial autocorrelation (Anselin, 1988, 1990; Anselin et al., 2004; Haining, 2004; Postiglione et al., 2022). Spatial dependence analysis is crucial for comprehending interdependencies, spillover effects, and understanding the effects of misspecification in the presence of a model (Postiglione et al., 2022). Spatial dependence normally involves including in the model spatially lagged variables that are weighted averages of the observations collected for neighboring units of a given location (Anselin, 2010; Postiglione et al., 2022).

Spatial dependence in agricultural variables can be explained using Tobler’s first law of geography, which states that everything is related to something else, with near things being more related than distant things (Tobler, 1970). Crop diversification, like most agricultural variables, can exhibit similar values in adjacent areas, resulting in spatial clusters. Geographically close farm households can exchange crop production information, knowledge and transfer farming skills more easily, which can improve their crop diversification practices (Ambali and Begho, 2022; Liu et al., 2024). Differences in socio-economic, farm and farmer characteristics, agroecological, institutional, and transaction cost factors may drive spatial dependence in agricultural production. Consequently, decisions made by one farm household are likely to influence neighboring households due to differences or shared environmental conditions, market access, and social networks (Wollni and Andersson, 2014).

A growing number of studies have focused on the spatial patterns of agricultural variables. For instance, spatial patterns have been extensively explored in several agricultural and climate change studies (Bolea et al., 2024; Carletto et al., 2017; Chu et al., 2020; Feng et al., 2017; Jiang et al., 2023; Lin et al., 2023; Tanoh and Hashemi-Beni, 2023; Ward et al., 2011; Wardhana et al., 2017) Spatially explicit models have also been used to analyze drivers of agricultural technology adoption (e.g., Anselin et al., 2004; Case, 1992; Fang and Richards, 2018; Krishna and Qaim, 2012; Läpple et al., 2017). Other spatial analysis studies have also considered spatial dependence and spatial heterogeneity in organic farming, site-specific nitrogen management in corn production, and explained variation in agricultural variables like farm land values (Anselin et al., 2004; Lapple and Kelley, 2015; Maddison, 2009; Schmidtner et al., 2012; Wang, 2018; Yang et al., 2019).

Spatial econometrics approach has also been widely used in fields and topics such as regional economics (Anselin et al., 1996; Belotti et al., 2017; Fischer and Getis, 2010); food deserts (Tanoh and Hashemi-Beni, 2023); demand analysis (Case, 1991); international economics (Aten, 1997); public economics and local public finance (Murdoch et al., 1993). The approach has also been popularized in environmental economics to examine pesticide use in agriculture (Aida, 2018; Grogan and Goodhue, 2012; Pinkse and Slade, 1998; Wang et al., 2023; Wollni and Andersson, 2014); spatial connectivity between regions has also been well exploited (Bell and Bockstael, 2000; Magaudda et al., 2020; Putra et al., 2020). Putra et al. (2020) explored the spatial variation in food expenditure using a spatial regression technique. They reported disparities in economic growth, Gross Domestic Regional Product (GDRP) per capita, urbanization, poverty, and unemployment rate between provinces are factors that drive spatial variation in food expenditure.

Case (1992) examined neighborhood influence on agricultural technology adoption and argued that not controlling for spatial effects would result in biased estimates. Holloway et al. (2002), studied the adoption of high-yielding rice varieties, and Holloway and Lapar (2007) examined spatial dependence in market participation among Filipino smallholders. Using data from the 2014–15 waves of the Gallup World Poll, Smith and Floro (2021) found that spatial dependence in international and domestic remittances received by households decreases food insecurity in lower-middle-income countries. Case (1992) and Lapple and Kelley (2015) both examined the concept of spatial dependence in agriculture, but within the context of technology adoption. Lapple and Kelley (2015) employed the Bayesian Spatial Durbin Models (Bayesian SDMs) to analyze the spatial dependence in the adoption of organic drystock farming using cross-sectional data. They reported that farmers located in close proximity exhibit similar choice behavior with spatial dependence in organic dry stock farming influenced by social norms and farmer attitudes.

Wollni and Andersson (2014) also employed the Bayesian spatial autoregressive probit model to analyze the spatial patterns in the factors influencing farmers’ decisions to convert to organic agriculture. The results reveal that farmers who act in accordance with their neighbors’ expectations and with greater availability of extension information are more likely to adopt organic agriculture. Liu et al. (2024) examined the spatial effects of new farmers’ adoption of sustainable agricultural practices (SAPs). They identified farm and household characteristics such as gender, educational level, and farm size as exhibiting significant direct, spillover and total effects of farmers’ adoption of SAPs. For farm households, spatial dependence is important because crop diversification decisions are often influenced by neighbors’ practices through the pathway of information sharing, imitation, social networking, shared agroecological and market conditions.

This study is conceptualized along the framework that farm households’ crop diversification decisions are inherently spatial and interdependent. Drawing on Tobler’s First Law of Geography and spatial interaction theory (Anselin and Rey, 1991; Colombo, 2020; Tobler, 1970). The framework posits that geographically proximate farm households are more likely to exhibit similar crop diversification due to shared information, social learning and networking, imitation, labor arrangements, and exposure to common agroecological and institutional conditions. Consequently, crop diversification is expected to display spatial dependence, whereby the diversification behavior of one farm household is systematically influenced by that of neighboring farm households. However, spatial spillover effects in agricultural production systems, including crop diversification, can broadly be explained by farm characteristics, household characteristics, information characteristics, attitudinal characteristics, economic, policy, and institutional characteristics, and environmental characteristics (Donfouet et al., 2017; Lapple and Kelley, 2015; Liu et al., 2024; Wang et al., 2023; Wollni and Andersson, 2014).

Social networking membership and the duration of participation in social networking groups strengthen trust through information sharing among farm households, thereby driving positive spatial externalities. Land tenure security enhances incentives for long-term investment in diversified production and may indirectly influence neighboring households by shaping local norms and expectations regarding land use. In terms of labor arrangements, own, family, communal, and hired labor determine the feasibility of managing labor-intensive crop portfolios. In particular, communal labor systems facilitate coordinated farming activities and the diffusion of cropping practices across neighboring farms, reinforcing spatial interdependence (Vroege et al., 2020).

Agroecological heterogeneities, which capture location-specific characteristics that define variation in rainfall, temperature, soil fertility and vegetation that shape the cultivation of multiple crops, could influence the crop diversification practices of farm households. Without controlling for agroecological heterogeneities, observed spatial dependence in crop diversification may be confounded by the underlying ecological landscape similarities, leading to biased estimates of spatial parameters (Arndt and Helming, 2025; Cho et al., 2007). The agroecological variables may serve as ecological and environmental fixed effects, thus ensuring that spatial dependence reflects genuine behavioral and structural interdependencies rather than biophysical coincidence (Aida, 2018; Anselin, 1988; Cho et al., 2007; Grogan and Goodhue, 2012; LeSage and Pace, 2009).

Few studies focus on examining spatial dependence in crop diversification. For example, Donfouet et al. (2017) employed a spatial two-stage least squares (Spatial 2-SLS) estimation approach to examine the effect of crop diversity on crop production and identified labor use as exhibiting a significant positive spillover effect on crop diversity. Kumar et al. (2025) explored the spatial drivers of crop diversification using the spatial lag and spatial error models, and reported that location-specific factors influence crop diversification behavior of farm households. Vroege et al. (2020) studied the influence of neighboring farms on crop diversification decisions. They identified a positive spatial dependence in crop diversification, with spillover effects emanating from farm and farmer characteristics, socioeconomic and physical environment factors.

Put together, what therefore is of interest from the literature is that farm households do not make crop diversification decisions in isolation; they learn from, imitate and respond to neighboring farm households based on their peculiar characteristics. These local interactions generate spatial spillovers that influence their crop diversification outcomes, thus making farm household-specific and agroecologically targeted interventions more plausible than uniform one-size-fits-all interventions.

Despite widespread promotion and adoption of crop diversification practices among farm households in developing countries, little is known about how farm households’ crop diversification decisions are spatially interdependent, leading to underexploited spatial spillovers. Furthermore, studies on crop diversification that consider spatial dependence and spillover effects—especially spillover effects emanating from socioeconomic, farm and farmer characteristics, agroecological, transaction costs and institutional factors at the farm household level are also limited. Hence, empirical evidence on spatial dependence and spatial spillovers in crop diversification at the farm household level in sub-Saharan Africa, particularly in Ghana, needs thorough investigation.

The present study contributes to the literature by employing nationally representative survey data to account for spatial dependence and spillover effects in crop diversification practices among farm households in Ghana. Our results suggest that farm and household characteristics, agroecological, labor arrangements, and institutional factors significantly influence neighboring or across-household spillover effects in crop diversification behaviors. Notably, variables such as irrigation access, farm size, extension access, access to tractor and animal plow, social networking, years of social networking, tenure security, family and communal labor, and agroecological heterogeneities were significant variables that exhibited positive spillover or neighborhood effects.

The Ghana Socioeconomic Panel Survey (GSPS) dataset was used for the study. The GSPS is a nationally representative administrative micro panel dataset collected by the Economic Growth Centre (EGC) at Yale University, the Institute of Statistical, Social and Economic Research (ISSER) at the University of Ghana, and the Global Poverty Research Lab at Northwestern University, United States of America. The GSPS dataset is a detailed, multi-level, long-term scientific dataset of all individuals or a random subset over time. The baseline data was collected in 2009/2010 and intended to include five rounds over 15–21 years.

The wave three (3) cross-sectional dataset collected in 2017/2018 was employed for the analysis. The data collection method employed for the survey involved a two-stage stratified sampling design. The first stage involved selecting geographical precincts or clusters from an updated master sampling frame developed for the 2010 Population and Housing Census. 334 Enumeration Areas (EAs) from the regional clusters were selected for the interview. In the second stage of the sampling procedure, a simple random sample of 15 households was drawn from each EA cluster. The GSPS wave 3 cross-sectional dataset was appropriate for the study due to its large sample size and the well-defined geocodes needed for the spatial analysis.

Out of the 5,669 households in the GSPS wave 3 cross-sectional dataset, data management resulted in 4,292 farm households after missing observations were removed. Duplicate coordinates identified were jittered to allow for the full observations of 4,292 spatial units (farm households) to be used for the analysis. It must be noted that spatial units could be firms, regions, cities, municipalities, districts, households, and so forth, depending on the nature and study objectives being undertaken. Unlike studies relying on aggregated district/regional level data, the GSPS allows spatial dependence to be examined at the micro (farm household) level, where diversification decisions are made.

Spatial weighting matrix can be determined through exploratory indicators developed by Geary (1954) and Moran (1950). Both Moran (1950) and Geary (1954) developed a binary weights matrix W , where each element w ij was assigned a value of 1 if two observational units were neighbors and assumed to exert an influence on each other, and 0 otherwise, in a contiguity-based spatial weighting matrix approach. Spatial weighting matrix can also be defined based on threshold distance, which is often arbitrarily chosen since information on the exact size of the neighborhood does not exist (Mueller and Weiler, 2023). In this study, the approach by Lapple and Kelley (2015) and Roe et al. (2002) is followed, which assumes a certain threshold distance beyond which the spatial effect does not exert any influence or decays. The rationale is that using the inverse distance spatial weighting matrix allows for capturing spatial dependence where proximity influences crop diversification practices. For example, neighboring farm households are expected to adopt similar crop production strategies due to shared environmental conditions or knowledge diffusion, and thus spatial interaction among farm households is likely to weaken or diminish with distance (Ambali and Begho, 2022; Lapple and Kelley, 2015).

To accommodate multiple neighbors per farm household, a 2 km radius distance is chosen as the minimum distance, and models are also estimated for a 10 km radius and a 50 km radius to explain spatial dependence and as robustness checks. The inverse distance-based spatial weighting matrix ( w ij = 1 d ij ) is applied, where d ij is the absolute distance between spatial units—farm household i and j (Ambali and Begho, 2022; Lapple and Kelley, 2015). The spectral-normalized form of the spatial weighting matrix normalization was adopted. The spectral-normalized spatial weighting matrix normalizes the matrix such that its largest eigenvalue equals 1 (Haining, 2004; StataCorp, 2023).

The Moran’s I test statistic is computed to test the hypothesis H 0 : E ( u u ′ ) = σ 2 I from the linear expression y = X β + u where y is n × 1 dependent-variable vector, X is n × K matrix of covariates, β is K × 1 vector of regression parameters, and u is n × 1 vector of disturbances. It is assumed that u i is identically distributed with E ( u i ) = 0 and E ( u i 2 ) = σ 2 . The spatial weighting matrix W 1 gives a proper representation of spatial links for the error term u .

Formally, from Moran (1950), Moran’s I is computed from Equation 1 as follows:

Where u ∧ = y − X β ̂ are the estimated OLS residuals, and σ ̂ 2 = u ∧ u ∧ / n is the corresponding estimator for σ 2 . It follows from Kelejian and Prucha (2001, 2004) that I ~ N ( 0 , 1 ) and I 2 ~ χ 2 ( 1 ) . To address spatial dependence, the analytical framework comprised three specifications of spatial econometric models to correct for the spatial processes inherent in the dataset. These include the Spatial Lag Model (SLM), Spatial Error Model (SEM) and Spatial Durbin Model (SDM) (Benassi et al., 2023; Sansuk and Sornlorm, 2024; Taecharungroj, 2024; Tang et al., 2023).

The descriptive statistics of the variables used for the spatial analysis are presented in Table 1. A verification of multicollinearity in the independent variables was conducted using the variance inflation factor test. Multicollinearity often occurs when one covariate is a linear combination of some subset of other covariates. A perfect correlation in the covariates may render the least squares estimator biased. The variance inflation factor (VIF) test produced a value of 1.49 (VIF < 10), indicating no collinearity in the variables used for the estimation. After adding jitters to duplicate coordinates and transforming all coordinates into kilometers, the full observation of 4,292 spatial units was used for the analysis.

The maximum and minimum latitude values were 11.129 degrees and 4.798 degrees, respectively, with a mean value of 7.475 degrees, while the maximum and minimum longitude values were 1.129 degrees and −2.912 degrees, respectively, with a mean value of −1.128 degrees. Since the coordinates in the dataset were measured in degrees latitude and longitude, they were modified or scaled into kilometers so that distances between spatial units can be recorded as kilometer (km) radius following (Lapple and Kelley, 2015). We addressed duplicate coordinates in the data due to a house having two or more households and thus sharing the same coordinates before the data was declared in spatial settings for the analysis.

The descriptive statistics reveal that approximately 11% of the farm households had access to irrigation, while 14% had access to extension services. About 35% of the farm households had access to tractor and animal plow services. This result indicates that the provision and adoption of mechanization services are steadily gaining momentum among farm households in Ghana. Approximately 63% of farm households belong to and participate in a social networking group which facilitates peer-to-peer learning and information sharing on crop production. The average number of years a farm household participates in a social networking group is approximately 25 years. The descriptive statistics further reveal that in terms of agroecological distribution, 10% of farm households are in the Coastal savannah agroecological zone, 47% in the Seni-deciduous rainforest agroecological zone, 20% in the Guinea savannah agroecological zone, 10% in the Sudan savannah agroecological zone, and 12% in the Transitional agroecological zone.

To proceed with the spatial analysis, inverse distance spatial weighting matrices using the spectral-normalized form of the matrix were adopted. Spectral normalization ensures that the matrix is normalized so that its largest eigenvalue equals 1 (StataCorp, 2023). Other forms of matrix normalizations, like the minmax normalization and row normalization, also exist. However, the matrix can also be left unnormalized (StataCorp, 2023). From the GSPS data coordinates, we construct inverse distance spatial weighting matrices for distances of 2 km (WIdist2), 10 km (WIdist10), and 50 km (WIdist50) radii apart. These allow spatial interactions to be measured at different neighborhood distances, and the summary for each matrix is presented in Table 2. WIdist2 represents only neighbors within 2 km, WIdist10 represents only neighbors within 10 km, and WIdist50 represents only neighbors within 50 km. The minimum threshold distance cutoff of 2 km radius apart also prevented the occurrence of islands or missing neighbors in the spatial weighting matrix. If spatial units are islands or contain missing neighbors, the missing neighbors do not influence or are not influenced by other spatial units, which can affect model performance or interpretation (Anselin, 1988; Drukker et al., 2013).

Moran’s I test for spatial dependence in crop diversification (dependent variable) and the independent variables rejected the null hypothesis that the error terms are independently identically distributed (i.i.d)—indicating zero spatial autocorrelation. Table 3 presents Moran’s I test for spatial dependence in the dependent and independent variables. The test confirmed the presence of spatial autocorrelation since the OLS residuals in both the dependent variable and the independent variables are correlated with nearby OLS residuals under each of the inverse distance spatial weighting matrices. It is therefore important to examine the spatial dependence in crop diversification using rigorous spatial econometric specifications.

Figure 1 displays the spatial distribution of crop diversification among farm households across Ghana. Each point represents a geocoded farm household plotted using latitude and longitude coordinates. The color gradient reflects the number of crops cultivated, that is, the count of diversified crops (CD_count), with lighter colors indicating lower diversification and darker colors indicating higher diversification. The figure reveals clear spatial clustering in crop diversification patterns in Ghana. Higher levels of diversification are concentrated mainly in parts of the Semi-deciduous rainforest, Coastal savannah and Transitional zones, while lower diversification is more prevalent in the Guinea and Sudan savannah regions in the north. This spatial pattern suggests that crop diversification is not randomly distributed across space but is shaped by shared agroecological conditions, institutional environments, and social interactions among nearby households.

Table 4 presents the maximum likelihood estimates of the raw coefficients from the three spatial econometric model specifications with robust standard errors. The Spatial Autoregressive model (SAR) and Spatial Error Model (SEM) estimations reported only the parameter estimates of the lagged dependent variable and lagged error terms, respectively, but they did not report the parameter estimates for the lags of the independent variables, as shown in the SDM estimation. The SDM reported both the lags of the dependent variable and the independent variables. Overall, the direction and magnitude of the raw SDM coefficients in Table 4 indicate that spatial effects enter the crop diversification process through socioeconomic characteristics, farm and farmer characteristics, institutional/resource availability, social interactions, labor arrangements and agroecological specific factors. Together, these factors are the most economically important drivers of crop diversification, with spatial dependence playing a central role in shaping these outcomes. However, because raw SDM coefficients do not directly represent marginal effects, the substantive interpretation of these results is best supported by the direct, indirect, and total effects reported subsequently in Table 5.

The results for the spatial dependence parameter ( Rho − ρ ) , presented in Table 4 for WIdist2 (2 km inverse distance) and in Supplementary Appendix 1 for WIdist10 (10 km inverse distance) and WIdist50 (50 km inverse distance) from the maximum likelihood estimations, show a systematic decrease in the spatial dependence parameter as the distance between spatial units (farm households) increases. The spatial correlation in crop diversification, as defined by the spatial dependence parameters (Rho−𝜌) at the different spatial weighting matrices for the SDM estimations are all positive and significant, with values of 2.818540 at a 2 km inverse distance, 2.818562 at a 10 km inverse distance, and 2.818575 at a 50 km inverse distance. Intuitively, the findings imply that crop diversification decisions by one farm household influence the decisions of neighboring farm households. The spatial decay reflects Tobler’s law, whereby information flows, access to extension, and irrigation use are geographically clustered, with the effect being strongest at an inverse distance of 2 km and declining at the inverse distance of 10 km and beyond.

Furthermore, the spatial decay effect also reflects the localized nature of agricultural information flows in Ghana, where extension services, farmer groups, and informal networks tend to operate at the community or village level rather than across wider geographic areas. Moreover, physical distance increases transaction costs associated with learning and coordination, further confining diversification practices to immediate neighborhoods. While the spatial dependence parameter (Rho−𝜌) does not directly measure distance decay, the spatial weighting matrix used in generating Rho−𝜌 was constructed with the inverse distance, which implies that the weights assigned to neighbors decrease as the distance to them increases, ensuring that closer neighbors have a stronger influence on the location being modeled.

The Wald test of spatial terms is statistically significant in all three spatial econometric model specifications. The Wald test is a statistical test that assesses whether the factors in the models are jointly equal to zero (Anselin, 1988; Elhorst, 2010; LeSage and Pace, 2009). That is, it checks the null hypothesis that the spatial parameters ρ = 0 for the Spatial lag model specification, λ = 0 for the spatial error model specification and ρ = 0 for the Spatial Durbin Model specification. The results indicate that the p-value in each model is at least statistically significant at the 1% level of significance ( p < 0.01 ) , thus suggesting that spatial terms are important, and the models specified accounted for spatial effects in crop diversification.

The data employed for the study allows the identification of concrete mechanisms through which spatial spillover effects occur. As earlier discussed, the results presented in Table 4 are not interpreted as the effects of the independent variables on the dependent variable, but rather serve as components in a recursive computation of their marginal or incremental effects on the dependent variable (Elhorst, 2010). The coefficients from the SDM using the 2 km (WIdist2) inverse distance spatial weighting matrix (SWM) are interpreted. The average impacts (average marginal effects), which report the direct, indirect, and total effects obtained through a partial differential approach (Elhorst, 2010; Golgher and Voss, 2016; Liu et al., 2024), are shown in Table 5. The direct effect indicates how households’ internal or own characteristics influence spatial dependence in the dependent variable. It is represented by the mean of the diagonal terms of the partial derivatives matrix (Golgher and Voss, 2016; LeSage and Pace, 2009). The direct effect indicates how households’ internal or own characteristics influence spatial dependence in the dependent variable.

The indirect effect is represented as the mean of the off-diagonal elements in each row of the partial derivatives matrix (Golgher and Voss, 2016; LeSage and Pace, 2009). The indirect effects represent the cumulative effect of a change in the independent variable of neighboring households on the crop diversification of household i . The extent to which changes in neighboring households affect the crop diversification effort of household i depends on the distance between these households, which is defined by the spatial weighting matrix. The total effects of an independent variable are therefore the sum of the spatial direct effects and the spatial indirect spillover effects (LeSage and Pace, 2009). Total effect is represented as the sum of the direct and indirect effects of the partial derivatives matrix (Golgher and Voss, 2016; LeSage and Pace, 2009). However, the total effect’s standard error is not just the sum of the standard errors of the direct and indirect effects. It is computed from the full variance–covariance matrix, implying that the significance of the coefficients of both the direct and indirect effects does not necessarily translate into the significance of the coefficients of the total effects.

The SDM results for the 10 km (WIdist10) and 50 km (WIdist50) inverse distance weighting matrices are computed as robustness checks, and the outputs are presented in Supplementary Appendices 2, 3. The models are quite robust when estimates are compared across the different inverse distance SWM specifications. This finding is consistent with the conclusions of LeSage and Pace (2009) and Liu et al. (2024), who suggest that fine-tuning the spatial weighting matrix has little impact on the estimation. In other words, estimates from spatial analysis do not vary much regardless of the inverse distance weights used in generating the spatial weighting matrix.

The results indicate that, holding other covariates constant and accounting for spatial interdependencies, the own-household direct effect of whether a farm household head obtained higher education leads to a 15% increase in crop diversification relative to those without higher education. This result reinforces the idea that educational attainment enhances own-household’s ability to utilize resources effectively by increasing the number of crops cultivated. This finding is consistent with the findings by Liu et al., 2024 and Wang et al. (2023), who reported a significant and positive own-household effect of education. Similarly, the findings further indicate that household and farm sizes are associated with a 5 and 20% increase in crop diversification from the own-household direct effect perspective, respectively. Larger household sizes provide more available family labor, which is critical for managing multiple and different plots of crops in labor-intensive diversification strategies (Laurent et al., 2022). The findings suggest that the availability of labor within the household reduces the reliance on costly external labor, facilitating diversified crop production. Additionally, larger farm sizes offer greater land resources that can be allocated to the growing of multiple crops and maintaining different strands of crops, thus enabling the adoption of crop diversification as a risk-mitigating management strategy and to maximize land productivity.

The findings also indicate that the own-household direct effect of irrigation access and tractor and animal plow use lead to approximately a 36 and 2% increase in crop diversification, respectively. This result indicates that the own-household direct effect of irrigation access and use of tractors and animal plowing for farming operations resulted in the growing and maintaining different plots of crops by adopters relative to households that did not access irrigation and tractor and animal plowing services.

We also identify years of social networking and tenure security as significant correlates of crop diversification from the own-household direct effect perspective. With regards to years of social networking, the findings indicate that the own-household direct effect of an additional year of social networking increases crop diversification by about 0.6%. Intuitively, longer years of social networking engagement build and strengthen trust and cooperation, thereby enhancing the sharing of resources and information among farm households, which is critical for adopting diversified farming practices. This finding is consistent with the findings by Wang et al. (2023), who identified social networking as a positively significant variable in pesticide-free wheat production. The results further indicate that the own-household direct effect of cultivating crops on own farmland secured through purchase and inheritance is likely to increase crop diversification by approximately 13% relative to households that do not own their farming land. These findings demonstrate the importance of farmers networking with one another and cultivating crops on their own land.

Farm households in the Guinea savannah agroecological zone are less diversified compared to those in the coastal savannah agroecological zone. This finding could be a result of unfavorable climatic conditions in the Guinea savannah agroecological zone, which hinder the cultivation of diversified crops. Additionally, market access limitations, farming traditions and crop suitability could limit the crop diversification potential of farm households in the Guinea savannah zone from the own-household direct effect perspective.

The indirect effect or spillover effect analysis suggests that socioeconomic, farm and farmer characteristics, agroecological and institutional factors significantly influence neighboring or across-household spillover effect crop diversification behaviors. Notably, variables such as household size, irrigation use, farm size, extension access, tractor and animal plow use, social networking and years of social networking, tenure security and family and communal labor were the significant variables that exhibited spillover or neighborhood effects. The results indicate that the across-household spillover effect or neighborhood effect of irrigation use leads to a 12% increase in the level of crop diversification. This finding is consistent with the findings by Aida (2018), who reported a positive and significant spillover effect (neighborhood effect) of irrigation access on pesticide use. This indirect effect suggests that shared or visible irrigation systems promote local adoption of diversified farming. Irrigation infrastructure often serves multiple farms, such that nearby farming households observing the effect of irrigation use on multiple crop cultivation can also be encouraged to adopt diversified farming. The effect of irrigation on crop production has been well documented in the literature (Del Carpio et al., 2011; John, 2024; Mengistu et al., 2021; Ngango and Hong, 2021; Sibhatu et al., 2022).

Farm size was identified to have a significant and positive spillover effect on crop diversification. Intuitively, larger farms influence local norms or market conditions that affect neighbors, thus larger farms may be more diversified. The across-household effect suggests that neighbors might imitate the scale-dependent strategies or be affected by the local market power and land-use changes of households with larger farms. This finding is consistent with the findings by Wang et al. (2023), who reported a positive and significant indirect effect of farm size on pesticide-free wheat production. However, this finding contradicts the finding by Lapple and Kelley (2015), who did not report a significant effect of farm size on adoption of organic crop production practices.

The SDM estimation reveals that access to extension services was observed to exert a significant and positive effect on crop diversification level through spatial spillovers. The findings indicate that households located near others with access to extension services experienced, on average, a 1% increase in their crop diversification levels. This finding suggests that knowledge and practices gained through extension access via the own-household direct effect can spillover to neighboring farms through observation or social learning. The finding highlighted the importance of spatial externalities in agricultural development and support services that strengthen the reach and connectivity of extension systems delivery across rural farming households. This result also confirmed the long-held argument that extension information often spreads informally through observation and discussion among farm households, as outlined in the theory of technology diffusion (Dissanayake et al., 2022; Liu et al., 2024; Tey et al., 2014).

The spillover effect regarding the adoption of tractor and animal plowing indicates that the across-household spillover effect resulted in a 10% increase in crop diversification level for neighboring farm households. Tractor plow or mechanized services in Ghana are commonly shared, thus allowing nearby farmers to observe the impact on diversification and to also adopt by generating positive externalities. This indirect effect suggests that a higher level of crop diversification is triggered when farm households observe the benefits of tractor or animal plow use on neighbors’ farms.

The estimated spatial spillover effects for social networking, years of social networking, and communal labor all had a significant effect on crop diversification. The across-household effect for social networking and years of social networking is to increase neighboring households’ crop diversification by approximately 2 and 1%, respectively. This finding suggests that social group engagement promotes peer-to-peer knowledge and information dissemination, thereby facilitating crop diversification (Fang and Richards, 2018; van den Berg et al., 2020; Varshney et al., 2022). The across-household effect for tenure security, that is, neighboring households cultivating crops on their own land, was found to be positive but insignificant. The positive effect implies that secure landholding may inspire investment behavior among neighboring farm households.

The results further indicate that the across-household spillover effect or neighborhood effect of communal labor is to increase crop diversification by 2%. Communal labor, also described as rotational or shared labor (for example, nnoboa in the local language), promoted spatial spillovers in crop diversification. This finding implied that engagement of communal labor may likely create natural spillovers due to cropping techniques, timing and crop planning, as labor can be rotated among different fields of crops. The agroecological specific effect indicates that farm households in the Sudan savannah agroecological zone benefit more from crop diversification due to neighborhood influence.

The SDM results for the 10 km (WIdist10) and 50 km (WIdist50) inverse distance weighting matrices were computed as robustness checks following a similar approach by Lapple and Kelley (2015), and the outputs are presented in Supplementary Appendices 2, 3. The alternative estimates are quite robust when compared across the different inverse distance SWM specifications. This finding is consistent with the conclusions of LeSage and Pace (2009) and Liu et al. (2024), who suggested that fine-tuning the spatial weighting matrix has little impact on the estimation.

The robustness estimates indicate that, for direct effects, variables such as education, household size, irrigation use, farm size, extension access, tractor and animal plow use, years of social networking, and tenure security all show significant own-household direct effects. These findings suggest that these variables have direct impacts on a farm household’s crop diversification level. For indirect or spillover effects, household size, irrigation use, farm size, access to extension services, tractor and animal plow use, social networking, years of social networking, tenure security, family and hired labor exhibited significant spillover or neighborhood effects. The implication is that these variables generate positive externalities and influence the crop diversification levels of neighboring farm households.

The study has certain limitations. These limitations include our inability to explore interaction effects between crop diversification and household characteristics (e.g., age, gender of household head), farm size and agroecological factors, as well as the use of an alternative crop diversification measure to further test for robustness of the estimates. A future study should consider these limitations and analyze spatial dynamics in crop diversification from a panel data perspective.