The desert locust invasion between 2019 and 2020 affected some counties in Kenya, classified into three clusters (FAO, 2020d). Cluster 1 consisted of Baringo, Elgeyo Marakwet, Isiolo, Laikipia, Samburu, and Turkana. Cluster 2 consisted of Mandera, Wajir, and Marsabit, while Cluster 3 included Embu, Garissa, Kitui, Machakos, Meru, Tana River, and Tharaka Nithi (FAO, 2020d).

A quasi-experimental quantitative research design was used in this study. To get the sample for the study, a three-stage sampling procedure was applied. In the first stage, Isiolo and Meru counties were purposely selected because they experienced multiple desert locust invasions between 2019 and 2020. In addition, both areas have medium- to high-population settlements, which limits chemical application (GCRF African Swift, 2020). Isiolo County is characterized by semi-arid and arid zones where pastoralism and agro-pastoralism are practiced (County government of Isiolo, 2023). Meru County is a high agricultural potential area characterized by various levels ranging from the upper highlands to the lower midlands, where mixed farming, rain-fed cropping, and marginal mixed farming are practiced (County Government of Meru, 2023). In the second stage, Imenti North (Meru County) and Garbatulla sub-county (Isiolo County) and selected. Lastly, with the help of sub-county agricultural officers, one administrative ward was selected from each sub-county: Nyaki East ward (Miriga Mieru east) in Imenti North (Figure 1) and Kinna ward in Garbatulla sub-county (Figure 2). Simple random sampling method was applied to select the respondents.

Insights from key informant interviews (county directors of agriculture and officers involved in the desert locust control program), focus group discussions conducted with farmers and key informants, and information from previous desert locust studies (Emana, 2002; Showler, 2019; Pandey et al., 2021) were used to refine and validate the questionnaire tool. The final questionnaire covered various sections that included household characteristics, effects of desert locust invasion, control methods used, and 15 statements assessing perceptions of desert locust management. Five statements were related to the formal control measures; four statements were related to desert locust information; three statements were related to IDLM; and three statements were related to indigenous control methods. The statements are described in Equation 1.

The sampling formula of Cochran (Bartlett and Kotrlik, 2001) was used to calculate the sample size as described below:

where n is the sample size, Z is the desired confidence level, p is the estimated proportion with the attribute being studied, q is 1-p, and e is the margin of error. This study used a confidence level of 95% and a 5% margin of error. The p was assumed to take a value of 0.5 since the exact number of farmers in both study sites was unknown. The formula gave a sample size of 385. To account for non-response, the sample size was adjusted upwards by 25% following the 10–25% range suggested by Lindner et al. (2001), resulting in a total of 480 respondents, 240 for each study site. The sample size was close to other studies focusing on pest management, for instance, Tambo et al. (2020) and Weyori (2021), who used a sample size of 465 and 482 households, respectively. Due to incomplete data in some questionnaires, 7 households were dropped, leaving a valid sample size of 473 households for the analysis. The data were analyzed using Stata version 15.

Farmers can either adopt single or multiple technologies; thus, there is a need to account for this situation. The multivariate probit model (MVP) accounts for the simultaneous adoption of different technologies by considering the correlation of disturbance terms that may arise from the relationship between different agricultural technologies. The model has been applied in assessing factors influencing the adoption of agricultural technologies (Musungu et al., 2021; Bwiza et al., 2024; Geda et al., 2024; Tambo et al., 2024). In this case, unobserved factors simultaneously influence the adoption of each desert locust control method, and the MVP model would be appropriate (Tham et al., 2023).

In a multivariate model, different control methods, in this case, hitting objects (HO), lighting fire (UF), burying with soil (BT), avoiding invaded areas (AI), and on-farm spraying (OS), which are binary variables, can be modeled as:

where Z j k ∗ are the unobservable latent variables in the choice of desert locust control methods selected by a farmer. X is a vector of explanatory variables that influence the choice of desert locust control method, β represents a vector of parameters to be estimated, and ε j k is a set of multivariate error terms that are normally distributed with a zero conditional mean and variance normalized to unity ( U H O , U U F , U B T , U A I , U O S ) and a covariance matrix (Ω). This can be expressed as Equation 3:

where ρ (rho) is the pairwise correlation coefficient of error terms matching any two desert locust control methods. A significant ρ (rho) value indicates interdependence between the error terms (non-zero) result, and then Equation 2 becomes an MVP model.

Farmer awareness of desert locust control and adoption of control methods follow a two-step process. In the first stage, the farmer is aware or not, while in the second stage, a farmer adopts a control method. The second stage of adoption is a sub-sample of the first (awareness of desert locust control). Therefore, there is an increased likelihood that the sub-sample used in the second stage is non-random and different from the first stage (which included farmers who were not aware), which creates sample-selection bias (Rezaee et al., 2022). The Heckman two-stage procedure was used to correct for potential selection bias in the adoption decision (Heckman, 1979). In the first stage, an assessment of the factors influencing awareness of desert locust control methods was done. In the second stage, an MVP model corrected for selection bias was fitted. Both stages were fitted simultaneously using the CMP (conditional mixed processes) command in Stata to correct for self-selection bias following Bwiza et al. (2024). In the first stage, the following equation (Equation 4) was fitted as follows:

where A w a r e n e s s j k represents the jth farmer’s awareness about the kth desert locust control method, A w a r e n e s s j k = 1 if the farmer was aware and 0 otherwise, φ k are unknown parameters to be estimated, and ε j k is the error term assumed to be normally distributed. In the second stage, the following MVP (Equation 5)was fitted:

where Z j k represents the jth farmer’s decision to adopt the kth desert locust control method, 1 = hitting objects; 2 = lighting fire; 3 = burying with soil; 4 = Avoiding invaded areas; 5 = on-farm spraying.

Table 1 describes the explanatory variables used in MVP analysis and their expected signs. The choice of explanatory variables was informed by previous studies. Gender has been positively associated with the adoption of pest or disease control technologies (Nzuma and Mzera, 2023). Age is positively associated with the adoption of pest or disease control technologies; studies have found adoption to increase with age (Anang et al., 2021; Ateka et al., 2021). Households with more farming experience are more likely to adopt pest or disease control technologies (Miyittah et al., 2022). Households with more years of education have been found to likely adopt pest or disease control technologies (Owusu and Skevas, 2024). Households with large herd sizes, expressed in terms of tropical livestock units (TLUs), are more likely to adopt pest or disease control technologies (Musungu et al., 2021; Weyori, 2021). Household location reflects the geographical context in terms of pest management. Considering that desert locusts affect different locations; it was hypothesized that location would either positively or negatively influence the adoption of desert locust control methods.

Tambo et al. (2020) reported that previous experience with pest and disease shocks improved adoption of FAW management practices. It was therefore expected to be positively related to the adoption of desert locust control methods. Groups as key social networks provide an avenue for learning and increase the likelihood of uptake of control technologies (Ma et al., 2023; Mwenda et al., 2023; Tham et al., 2023).

Pests and diseases pose a major threat to livelihood incomes, and households engaged in commercial crop or livestock production are more likely to adopt alternatives to chemical control (Chelanga et al., 2023). It was hypothesized that households engaged in commercial agricultural production would adopt desert locust control methods.

Access to information from formal sources such as extension services has been positively linked with the adoption of pest control technologies (Ipara et al., 2021; Tambo et al., 2023; Adjei et al., 2024). Tham et al. (2023) highlighted the importance of including modern and indigenous communication channels in increasing the adoption of pest control technologies. It was therefore expected that informal sources of information (information sharing among community members) would positively influence the adoption of desert locust control methods.

Both structural equation modeling (SEM) and principal component analysis (PCA) are techniques for dimensionality reduction, but they serve distinct purposes. The PCA is primarily a descriptive method that simplifies data by identifying principal components, linear combinations of variables that capture the maximum variance within a dataset (Jolliffe and Cadima, 2016). In contrast, SEM goes beyond data reduction by simultaneously modeling relationships between observed and latent variables and examining causal pathways (Mishra et al., 2024). While PCA focuses on summarizing data patterns, SEM incorporates theoretical frameworks to test both direct and indirect relationships between variables, making it a more complex and theory-driven approach.

The SEM is widely used in studies examining behavioral patterns and the adoption of new technologies, often guided by frameworks like the technology acceptance model (TAM) (Ejigu and Yeshitela, 2024; Forouzani et al., 2024). While SEM and path analysis are advanced techniques for assessing factors influencing farmer behavior, they require data on sequential constructs such as awareness, interest, desire and action (AIDA) that predict human behavior over time; such data were not available in our case. In contrast, the current study examined how perception, as a single construct (Filmer and Pritchett, 2001; Olana et al., 2023), affects the adoption of integrated desert locust management methods. To do this, PCA is used to create a perception index, which is then applied in subsequent analyses as a simplified representation of perception. Following Okello et al. (2021), the PCA method was used to reduce the 15 statements (Supplementary Table 1) to derive a perception index. This was specified as Equation 6:

where P m is the perception index for the mth farmer, c n is the weight or factor loading of the nth perception statement, d m n is the response of the mth farmer for the nth perception statement, d n , and s n are the mean and standard deviation of the nth perception statement, respectively. The derived perception index was categorized as a binary variable, where a household either has positive perceptions or otherwise. It has been shown that households with favorable or positive perceptions toward pest or disease control methods are more likely to adopt them (Misango et al., 2022; Bryant et al., 2024). It was therefore expected that positive perceptions toward desert locust management would increase the adoption of control methods.

Prior to fitting the empirical models, multicollinearity was tested using the variance inflation factor (VIF) approach. The results revealed no multicollinearity, as all values were below 10 (Gujarati and Porter, 2009). The “hetprobit” command was used to test for heteroskedasticity. The null hypothesis of homoskedasticity was not rejected [chi2(6) = 0.17, p = 0.99], implying that the error terms had constant variance. Potential selection bias was controlled using awareness as a dependent variable in the selection Equation 4 and incorporated into Equation 5 using the CMP command in Stata version 15.

Table 3 presents the results on the substitutability and complementarities of desert locust control methods. The correlation coefficients between desert locust control methods were positive and significant, except for the correlation between avoiding invaded areas and lighting fire, leading to the conclusion that the control methods used by farmers are complementary and not substitutes.

Table 4 presents the probit regression results of the selection equation (awareness) and the MVP regression (outcome equation) results of the factors influencing the adoption of desert locust control methods. The Wald Chi-square ( χ 2 = 219.25, ρ = 0.000) for the goodness of fit model is significant at the 1% level, justifying the use of a multivariate probit model for analysis (Table 4). The values of the “atanhrhos” (test of correlation among residuals) were significant (Supplementary Table 2), implying the presence of selection bias in the adoption of desert locust control methods. This necessitated the need for corrective measures by applying a corrected MVP model to account for selection bias (Bwiza et al., 2024).

The socio-economic characteristics of households significantly influence their adoption of desert locust control methods. Households with larger livestock holdings, for instance, were more inclined to adopt control strategies. Desert locust invasions severely impact pasture lands, diminishing feed availability for livestock and making control measures critical to reducing these losses (Landmann et al., 2023). Notably, a slightly higher percentage of households in Isiolo adopted control methods compared to those in Meru, likely due to Isiolo County experiencing more severe locust invasions (FAO, 2020e). Over 60% of households previously exposed to various pests and diseases were more likely to implement control measures, a rate surpassing the 45% adoption reported by Tambo et al. (2020). The low dissemination of desert locust information through both formal and informal channels likely reflects the inadequate locust management framework in place at the onset of the invasion (FAO, 2023). The most common desert locust control method was making noise by hitting objects, a finding that aligns with observations reported by FAO (2020e).