This study was conducted in Mecha District, Northern Ethiopia. Mecha District is one of the 13 Districts of the West Gojjam Administrative Zone and is located 30 km southwest of Bahir Dar and 525 km northwest of Addis Ababa (Alebachew et al., 2015). The District is geographically situated between 11^°10′ N and 11^°30′ N latitudes and between 37^°00′ E and 37^°20′ E longitudes (Figure 1). According to the District Agricultural Office report (2018), the district has a total population of 301,188 of which 151,480 are men and 149,708 are women, and ranges from 1,500 to 2,500 m.a.s.l. with a mean annual rainfall of 1,560 mm and a mean daily temperature ranging from 16 to 20^°C. The topography of the District is predominantly flat (75%), as only about a quarter of its area is characterized by jagged topography, valleys, and mountains (with a total coverage of 156,027 ha) (Tefera and Kassa, 2017).

To choose sample household heads, a multistage sampling method was used. Multistage sampling is a method in which sampling is done in stages with smaller units being defined and selected at each stage within the units selected at the prior stage (Shimizu, 1998), aimed to select samples that are concentrated in a few geographical regions (Hamed, 2016). The method has been used by Asfaw et al. (2022) to select sample households among Date Producers in Afar Region, Ethiopia, Abebaw and Dea (2016) to determine factors affecting Eucalyptus woodlot planting in the Meket district. Similarly, Wu et al. (2023) adopted a multistage sampling procedure for estimating the forest volume. Thus, in this study, first, among the Districts in the Amhara region, Mecha District was purposefully chosen for its potential and extensive Eucalyptus plantations. As the District agriculture and rural development experts indicated, all the villages in the District are potential areas for Eucalyptus production and marketing. In the second stage, by considering farmers’ tree planting experience, production potentiality, and current expansion rate on productive and irrigation areas, out of the total 39 villages, three villages, namely Enashenifalen, Addisameba, and Rim were purposefully selected. Then, 186 sample respondents were selected using a random sampling technique from the three selected villages and allocated proportionally to the population size of the selected villages. The list of households was acquired from the village agricultural office.

As a reliable estimate of the population variance is often unavailable, selecting a sample size using proportional estimates of variability is frequently favored (Cochran, 1977). The method has been applied in many research studies. For example, Abebaw and Dea (2016) used it to determine the sample size of farmers’ decision-making for planting Eucalyptus trees in Meket District, North Wollow, Ethiopia, and Zeenatul et al. (2022) used it in their study of integrated environment-smart agricultural practices. Similarly, Tamirat (2022) used the Cochran (1977) sample size determination method to investigate factors in chemical fertilizer adoption decisions and their impact on food security in southern Ethiopia. Thus, for this study, a statistical formula developed by Cochran (1977) was used to determine the sample size at a 95% confidence level with a degree of variability of 0.5 and a level of precision of 7% (0.07) and distributed proportionally (Supplementary Appendix Table 1).

Where: n = sample size of household heads; P = 0.5, the maximum level of variability taken when previous population variability is unknown; q = 1-0.5, i.e., 0.5; e is the precision level (i.e., 0.07); and N = total population size of the selected villages (3,727), obtained from the administrative office of the selected villages.

For this study, a cross-sectional mix of primary and secondary data was used. Various tools were used to gather the primary data. This includes in-depth interviews, key informant interviews, direct field observation, and focused group discussions. Both quantitative and qualitative data were collected. Semi-structured questionnaires (open and closed-ended) were used to conduct an in-depth interview. The household surveys were carried out to collect quantitative data that was used to investigate the factors influencing farmers’ adoption and intensity of adoption of Eucalyptus woodlots by farmers. A key informant (KI) who had a broad knowledge of the farming practices of Eucalyptus plantations in the area was given priority for selection. A total of 13 KI interviews were conducted with nine village agricultural experts (three from each village) and one District expert, and three farmers who have long-standing know-how in farming, production, and marketing of Eucalyptus were contacted. Three focus group discussions, one at each village level, with a group of 12 farmers using men’s, women’s, and youth groups, were conducted. The discussion was focused on the status of Eucalyptus woodlot management, reasons for the expansion of Eucalyptus woodlots by converting cropland, and opportunities and challenges for Eucalyptus production and management.

Secondary data were collected from literature and the work of other researchers. Mecha District Agricultural Office (MDAO), Amhara Region Environment, Forest and Wildlife Protection and Development Authority (AREFWPDA) were contacted to gather information about Eucalyptus planting experiences and their determinates in the district.

In this study, an econometric model was used to identify the factors that affect the adoption decision and adoption intensity of Eucalyptus woodlots by farmers. Different methods can be employed to analyze factors influencing the adoption and adoption intensity of Eucalyptus woodlots. One approach to analyzing the issue is to use the well-known Tobit model. The standard Tobit model, originally formulated by James Tobin in 1958, was the first model to attempt a censored dependent variable. However, the Tobit model has a number of potential shortcomings due to the restrictive assumptions it makes. First, it assumes that both the decision to participate in an activity and the level of participation are determined by the same variables with the same sign (Wooldridge, 2002), which means both the value of continuous observations on the dependent variable and the discrete variables were determined by the same stochastic process. Hence, according to the Tobit model, the decision to participate in Eucalyptus woodlot adoption and the intensity of the adoption are jointly determined and influenced by the same parameters. Such an assumption may not be appropriate in this study because the factors that affect whether or not a household adopts Eucalyptus woodlot may be significantly different from the factors that affect the adoption intensity of Eucalyptus woodlot. Secondly, the Tobit model assumes that all the zero observations are, in fact, standard corner solutions (Garcia, 2013) and those households that do not adopt are a result of their economic circumstances. This means, the zero values on the dependent variable, intensity of woodlot adoption in this case, are attributable to economic factors alone. However, the Tobit model currently has been generalized by the use of the double-hurdle model that the possibility of zeros is due to non-participation decision, which arises from factors such as economic, institutional and demographic characteristics (Martínez-Espiñeira, 2006). A test for the double-hurdle model against the Tobit model was made by estimating three regression models (Tobit, truncated and probit regression models) separately the result indicated the rejection of the Tobit model (Supplementary Appendix Table 5).

Cragg’s (1971) double-hurdle model, relaxes the restrictive assumption of the Tobit model by specifying determinants of adoption and the level of intensity of adoption into two separate steps used to determine the adoption decision and level of adoption of Eucalyptus woodlots. The pattern suggests that in every decision-making process, farmers face two obstacles: deciding whether to adopt and determining the level of adoption (Asfaw et al., 2010, Kiyingi et al., 2016). The model has been widely used in the literature on decision-making behavior across different research fields, such as consumer purchasing behavior (Asfaw et al., 2022), consumption analysis (Wei et al., 2023) and agricultural and forest technology adoption (Mal et al., 2012; Kiyingi et al., 2016; Jara-Rojas et al., 2020; Gebru et al., 2021). The essential premise of this study is that a farmer makes two decisions: whether to allocate land to establish a Eucalyptus woodlot and the intensity of the area assigned for a Eucalyptus woodlot. In the first hurdle, a probit model was employed to determine the probability that the households adopt Eucalyptus woodlot. In the second hurdle, a truncated regression model was used to determine the intensity of the adoption. The general form of Cragg’s double-hurdle model employed in this study is as follows:

Where, D_i* is the latent variable describing the farmer’s decision of whether or not to adopt a Eucalyptus woodlot and takes the value of 1 if the farmer adopted and 0 otherwise. D_i is the observed variable that represents the farmer’s adoption decision; Z_i’ is a vector of explanatory variables influencing the farmer’s initial adoption decision; α is a vector of parameters to be estimated; and μ_i is the error term.

Where Y_i* is the latent variable describing its decision on the proportion of land to allocate to the Eucalyptus woodlot. Y_i is the area allocated for the Eucalyptus woodlot in hectares, suggesting the intensity of the adoption. X_i is the vector of explanatory variables influencing how much the farmer adopts Eucalyptus woodlot; β is a vector of parameters to be estimated; and ε_i is the error term.

If both decisions made by the individual farmers are independent, the error term is assumed to be independently and normally distributed as:

The log-likelihood function, which is the sum of the log-likelihood from a Probit and a truncated regression (Cragg, 1971), is estimated using the maximum likelihood estimation (MLE) technique following Cragg (1971).

Variance-inflating factor (VIF) was used to detect multicollinearity problems among continuous and discrete variables (Gujarati, 2004) and a VIF value of more than 10 indicates a high correlation while less than 10 indicates a weak association among explanatory variables. Contingent coefficient (C.C) techniques were used to detect multicollinearity problems among discrete variables and a value of 0.75 or more indicates stronger associations while a value less than 0.75 indicates weak associations among explanatory variables.

Additionally, a hypothesis test for the double-hurdle model against the Tobit model was made by estimating three regression models (Tobit, truncated and Probit regression models) separately and then conducting a likelihood ratio test that compares the Tobit with the sum of the log-likelihood functions of the Probit and truncated regression models. The LR statistic was computed using the formula developed by Greene (2003) as Γ = −2[lnLT − (lnLP + lnLTR)] ∼ χ²k.

Where LT = likelihood for the Tobit model; LP = likelihood for the Probit model; LTR = likelihood for the truncated regression model and k is the number of independent variables in the equations. The test hypothesis is written as H₀: λ = βσ and H₁: λ ≠ βσ; the null hypothesis (H0) that the Tobit model is the best-fit model will be rejected on a pre-specified significance level if Γ > χ² k.

Grounded on the existing knowledge from literature review including Ewnetu (2008), Duguma and Hager (2010), Gebreegziabher et al. (2010), Rafael (2011), Jenbere et al. (2012), Anik and Salam (2015), Abebaw and Dea (2016), Kiyingi et al. (2016), Zerga and Woldetsadik (2016), Gizachew (2017), Sete (2017), and Derbe et al. (2018) the possible explanatory variables hypothesized to affect adoption and level of adoption of Eucalyptus woodlot were specified (Supplementary Appendix Table 2).

Descriptive statistics such as mean, maximum, minimum, standard deviation, and frequency distribution were employed to get a clear picture of the socio-economic, institutional, and demographic characteristics of sample households. A chi-square test and an independent sample t-test were used to test the significance of variables associated with the adoption and intensity of Eucalyptus woodlots adoption and to identify variables that vary significantly between adopters and non-adopters of Eucalyptus woodlots. MS Excel, the Statistical Package for Social Sciences (SPSS) version 19.0, and Stata version 12.0 software were used for entering and analyzing the data.

As indicated by Derbe et al. (2018), a Eucalyptus woodlot is a parcel of land devoted to Eucalyptus plantations and has a size of 0.1 hectares, or what is conventionally called a “block,” which is 40 m by 25 m and above. In the study area, almost all smallholder farmers own Eucalyptus trees. However, for the interest of this study, farmers who had a Eucalyptus woodlot size of 0.1 ha were considered as Eucalyptus woodlot adopters. Therefore, of the total 186 sample respondents interviewed, 157 (84.41%) were found to be adopters of Eucalyptus woodlot while 29 (15.59%) were non-adopters. Adopters of Eucalyptus have allocated the high coverage of their land 65.38% (0.84 ha with a standard deviation of 0.24 ha on average) to Eucalyptus woodlot.

The sample respondent’s ages ranged from 25 to 77 with an average age of 45.04 years. The minimum and maximum years of schooling for adopters were 0 and 10 years, while the non-adopters have a minimum of 0 and a maximum of 6 years of schooling. Concerning family size, woodlot adopters were found to be 5.64 per family while non-adopters were found to be 4.36 per family on average (Supplementary Appendix Table 6). Among the total respondents, 80.60% of them are male-headed while 19.40% of them are female-headed households. Out of 157 Eucalyptus adopter participant households, 17.80% are female-headed and the remaining 82.20% are male-headed. The mean land holding for Eucalyptus woodlot adopters was 1.30 ha with maximum and minimum of 4.0 and 0.25 ha, respectively. The corresponding figure for the non-adopter households was 0.62 ha with maximum and minimum of 1.75 and 0.10 ha, respectively. Thus, the non-adopters had a limited land area. It indicates that farmers with large farm sizes are more likely to participate in the adoption and intensity of adoption of Eucalyptus woodlot than those farmers who have smaller farm sizes.

Adopters were found to own more livestock than their counterparts. It was observed that adopters of Eucalyptus woodlot had livestock holding 4.88 Tropical Livestock Units (TLU) on average, while non-adopters had 3.72 TLU. It indicates as farmers own large livestock units, the chance of participating in Eucalyptus woodlot adoption become increases. It is because livestock ownership provides the draught power (oxen) for the establishment of Eucalyptus woodlots. In the study area, Eucalyptus is commonly planted after plowing the land with oxen for crop production. Moreover, as farmers own large livestock populations, the availability of financial income in the household from the sale of livestock, especially bulls for meat, small ruminants, and chickens, increases, which in turn leads to investment decisions in Eucalyptus woodlot production. Of the total 186 sample households, only 9.14% (9.55% of adopters and 6.90% of non-adopters) of the sample respondents had got training related to Eucalyptus planting by the village representatives, the other (90.86%) of the respondents did not have access to training (Supplementary Appendix Table 7). It shows farmers are planting Eucalyptus simply because of its economic impacts without or with less government support. On average, Eucalyptus woodlot adopters earn 18,349.36 Ethiopian Birr (ETB), while non-adopters had 12,253.45 ETB annually from crop and livestock production. Of all respondents, who had market access 66.24% were Eucalyptus woodlot adopters and the other 17.24% were non-adopters. The result indicates that Eucalyptus woodlot adopters had better market access than non-adopters did. It means that farmers with market access are more likely to adopt and intensify their adoption of Eucalyptus woodlots than farmers without market access.

Farmers’ perception of the relative advantage of Eucalyptus woodlot production is one of the factors that can facilitate or undermine their participation in Eucalyptus woodlot adoption. Farmers were asked to respond as to how they perceived the importance of Eucalyptus woodlot production. From the total respondents, about 77.70% of Eucalyptus woodlot adopters and 31.00% of non-adopters were perceived positively on Eucalyptus woodlot production. It means farmers who had a positive perception of Eucalyptus woodlot production adopted and allocated more land for Eucalyptus woodlot production than those who had a negative perception.

The chi-square test result indicates that farmers’ perception towards Eucalyptus woodlot production (χ² = 25.603, df = 1, p = 0.000), off-farm work engagement (χ² = 4.737, df = 1, p = 0.012), adjacent farm (χ² = 29.799, df = 1, p = 0000) and market access (χ² = 24.227, df = 1, p = 0.000) were showed a variation between adopter and non-adopters. For the t-test, family size (t = 2.928, df = 190, p = 0.003), land holding size (t = 4.891, df = 190, p = 0.000), number of parcels of land (t = 5.57, df = 190, p = 0.000), livestock holding (t = 2.742, df = 190, p = 0.006) and income of the household head (t = 3.822, df = 190, p = 0.000) were compared and showed variation between the adopter and non-adopter groups.

Before running regression analysis, both the continuous and discrete explanatory variables were checked for the existence of multicollinearity using the Variance Inflating Factor (VIF) and the Contingency Coefficient (CC) methods, respectively. Thus, the VIF values displayed in Supplementary Appendix Table 3 show that all the continuous explanatory variables had no serious multicollinearity problem, as the VIF is less than 10. Likewise, the values of the contingency coefficient are less than the rule of thumb of 0.75, (Supplementary Appendix Table 4), implying a weak degree of association among the variables considered. The likelihood ratio test that compares the Tobit with the sum of the log-likelihood functions of Probit and Truncated regression indicated the rejection of the Tobit model and acceptance of the Double-hurdle model. The calculated statistical value of the likelihood ratio was 50, which was greater than the tabulated or critical value of χ² (15) = 32 at a 1% level of significance, which shows the existence of two separate decision-making stages during the adoption process (Supplementary Appendix Table 5). This proves the farmers’ independent decision-making possibilities regarding the adoption and intensity of adoption of Eucalyptus woodlot. Similarly, Heckman’s two-step model was regressed and the independence between the participation and intensity decision was tested and the corresponding Wald test of independence suggests the independence assumption cannot be rejected (p-value = 0.2), as Mills ratio is insignificant, which strengthens the use of double hurdle-approach. In other words, the sample-selection bias is not an issue in this study.

The likelihood ratio of the chi-square result revealed that the overall fitness of the model was found to be significant at a 1% probability level, indicating the partial effects of each explanatory variable on the response of the dependent variable and similar to the intensity of adoption. Out of the fifteen explanatory variables, four variables from the probit model and six variables from the truncated model were found to be significantly determining adoption and intensity (area of Eucalyptus planted in hectare) of adoption of Eucalyptus woodlot at 5% and 1% level of significance, respectively Supplementary Appendix Tables 8, 9).

The number of parcels of land was found to positively and significantly influence the probability of adoption and intensity of adoption of Eucalyptus woodlot at less than 5 and 1% level of significance, respectively. It indicates that if the other factors are kept constant, a unit increase in the number of plots of land increases the probability of adoption and intensity of adoption of Eucalyptus woodlot by 8.80, and 10.60%, respectively. Similarly, the positive perception of farmers towards Eucalyptus woodlot production increased the adoption and its level by 14.03 and 7.74%, (1 and 5% of significance level), respectively.

Off-farm work engagement and credit availability were found to positively and significantly influence the probability of adoption of Eucalyptus woodlot at a 5% level of significance. The marginal effect estimate indicates that if the other factors are kept constant, participation in off-farm work increases the adoption of Eucalyptus woodlot by 10.23% and a 1% increase in credit availability for households increases the adoption of Eucalyptus woodlot by 8.8%. While land holding size, adjacent farms (the farm of another framer which is adjacent or close to the owner of Eucalyptus woodlot) and access to market had a positive and significant influence on farmers’ decision to allocate land for Eucalyptus woodlot at less than 1% significant level, which increases the land allocated for Eucalyptus woodlot by 47.27%. Likewise, keeping all other variables constant, the intensity of adopting Eucalyptus woodlot was increased by 18.72 and 19.78% for the households who have more land near to Eucalyptus woodlot and have access to market, respectively. While, a one-person increase in the family size decreases the land allocated for Eucalyptus woodlot by 3.34% at a 1% significance level, ceteris paribus.

Moreover, lack of support (59%), lack of training (34%), lack of segregation of land (22.14) and disease (18%) were the major constraints on tree growers (Supplementary Appendix Figure A1). Moreover, policies that encourage Eucalyptus planting activities have not existed and rules and regulations on production and marketing of Eucalyptus are not functional as evidenced from the key informant interviews and focus group discussions.