A Typology-Based Approach for Assessing Qualities and Determinants of Adoption of Sustainable Water Use Technologies in Coping with Context Diversity: The Case of Mechanized Raised-Bed Technology in Egypt

: Mechanized raised-bed technology (MRBT) is recognized as an important measure to achieve higher crop productivity and water-use e ﬃ ciency in intensive irrigated systems. Development e ﬀ orts on spreading this technology require adequate understanding of the qualities and drivers of farmers’ adoption of MRBT. Research in agricultural innovation adoption has identiﬁed the importance of the socio-ecological context (SEC) that inﬂuences the livelihood of farmers adopting new technologies. This study introduces an agricultural livelihood systems (ALS) typology-based approach for guiding concrete analytical steps and statistical methods in evaluating the e ﬀ ects of system SEC diversity in two Egyptian governorates. We objectively classify a population of sampled farming households into a limited number of ALS types and use inferential statistics for the whole sampled population and individual ALS types to discover adoption drivers. Values added by the ALS approach conﬁrm the widespread role of common determinants of MRBT adoption across ALS types, household groups subject to the e ﬀ ects MRBT, and show new causal e ﬀ ects. The presented advanced approach and empirical ﬁndings will be useful for enhancing targeting and out-scaling of MRBT practices toward achieving sustainable agricultural water uses at scale.


Introduction
Water scarcity for agriculture in Egypt has been, and will continue to be, a profound problem. Average water availability per person has declined below the threshold value of 1000 m 3 /capita/year, and is estimated to fall to 500 m 3 /capita/year in 2025 if there is no significant improvement in management [1,2]. Moreover, negative effects of climate change on agricultural production introduce further problems for water allocation to agriculture. Recent projection suggests that Egyptian agricultural production could decrease by 8-47% by 2060, with employment losses of up to 39% [3]. Thus, the current and future challenge in Egypt is producing more food with less water resources. The benefits of each drop applied could be maximized by adopting appropriate and suitable irrigation scheduling and adapted irrigation practices [2,4].
household farm lists of several districts in the two governorates: 180 household farms practicing traditional farming methods and the remaining 180 were adopters of MRBT.
In the surveyed household farms, the main crops grown in the summer season are rice (57% of the total summer cropping area, found only in Sharkia), maize (49%), cotton (26%, only in Sharkia), and sorghum (51%, only in Assiut). The main winter crops are wheat (91%) and clover (9%). Surveyed traditional household farms in Sharkia Governorate comprise 14 from Zagazig, 30 from Awlad Saqr, five from Menia Al-Qamh, 15 from Hehia, 21 from Abo-Ahmed, and five from the Faqos district. The surveyed adopters of MRBT in Sharkia include 14 from Zagazig, 29 from Awlad Saqr, five from Menia Al-Qamh, 15 from Hehia, 20 from Abo-Ahmed, and seven from the Faqos district. In Assiut Governorate, the surveyed traditional farmers include 45 from Manfalot and 45 from the Al-Fat'h district, all of which are small farmers (100%). The surveyed MRBT adopters in Assiut comprise 60 from Al-Fat'h district and 30 from Abnob district. In addition, 19%, 64%, and 17% of them own land on the head, middle, and tail of Mesqa, respectively [23]. In the surveyed household farms, the main crops grown in the summer season are rice (57% of the total summer cropping area, found only in Sharkia), maize (49%), cotton (26%, only in Sharkia), and sorghum (51%, only in Assiut). The main winter crops are wheat (91%) and clover (9%).

Analytical Framework
Surveyed traditional household farms in Sharkia Governorate comprise 14 from Zagazig, 30 from Awlad Saqr, five from Menia Al-Qamh, 15 from Hehia, 21 from Abo-Ahmed, and five from the Faqos district. The surveyed adopters of MRBT in Sharkia include 14 from Zagazig, 29 from Awlad Saqr, five from Menia Al-Qamh, 15 from Hehia, 20 from Abo-Ahmed, and seven from the Faqos district. In Assiut Governorate, the surveyed traditional farmers include 45 from Manfalot and 45 from the Al-Fat'h district, all of which are small farmers (100%). The surveyed MRBT adopters in Assiut comprise 60 from Al-Fat'h district and 30 from Abnob district. In addition, 19%, 64%, and 17% of them own land on the head, middle, and tail of Mesqa, respectively [23].

Analytical Framework
The analytical framework of agricultural livelihood typology-based approach proposed by [24] is applied. This framework includes sequential steps of empirical analyses toward achieving the stated objectives ( Figure 2). The analytical framework of agricultural livelihood typology-based approach proposed by [24] is applied. This framework includes sequential steps of empirical analyses toward achieving the stated objectives ( Figure 2).

Methods for Identifying Livelihood Typology of Smallholder Farming Systems
Livelihood typology of smallholder farming systems is built on the concept of household/farm livelihood sustainability, including its adaptability and resilience in the vulnerability context. The sustainable livelihoods framework describes the essential resources and capabilities at household/farm disposal and livelihood strategies built from these resources in coping with the vulnerability context [25]. These resources comprise five types of livelihood assets that are used to achieve the households' or community's livelihood outcomes. Human assets include variables of labor, health, education, and capabilities. Natural assets comprise attributes of land (amount and quality), livestock, and water resources. Financial assets include incomes, savings, and loans from different sources. Physical assets consist of variables for housing conditions, access to infrastructure, and equipment for agricultural production. Social assets include supports and advantages from social networks (both in vertical and horizontal connections), engagement with rural development institutions, positions, and projects/programs. In addition, from the resilience approach, the five livelihood assets interactively determine the buffering capacity of livelihood systems. The adaptability and transformability of a household's livelihood strategies will also be determined by its community's institution for self-organizing and learning capacities [26]. This livelihood framework was used to guide the development of the content of questionnaires for livelihood surveys

Methods for Identifying Livelihood Typology of Smallholder Farming Systems
Livelihood typology of smallholder farming systems is built on the concept of household/farm livelihood sustainability, including its adaptability and resilience in the vulnerability context. The sustainable livelihoods framework describes the essential resources and capabilities at household/farm disposal and livelihood strategies built from these resources in coping with the vulnerability context [25]. These resources comprise five types of livelihood assets that are used to achieve the households' or community's livelihood outcomes. Human assets include variables of labor, health, education, and capabilities. Natural assets comprise attributes of land (amount and quality), livestock, and water resources. Financial assets include incomes, savings, and loans from different sources. Physical assets consist of variables for housing conditions, access to infrastructure, and equipment for agricultural production. Social assets include supports and advantages from social networks (both in vertical and horizontal connections), engagement with rural development institutions, positions, and projects/programs. In addition, from the resilience approach, the five livelihood assets interactively determine the buffering capacity of livelihood systems. The adaptability and transformability of a household's livelihood strategies will also be determined by its community's institution for self-organizing and learning capacities [26]. This livelihood framework was used to guide the development of the content of questionnaires for livelihood surveys and indicators for analyses and assessments. Table 1 defines quantitative variables, which were specified using the sustainable livelihoods framework, for statistical analyses toward identifying livelihood types. First, the principal component analysis (PCA) procedure wasused for discovering key factors explaining the majority of variation in the multivariate livelihood data, as well as reducing dimensionality of the data [27]. The meaning of each principal component (PC) is interpreted in terms of the original variables with higher weights/loadings. Because the extracted PCs are independent of each other, the use of component scores for subsequent analysis avoids the multi-collinearity problem that some explanatory variables are correlated and can cause problems for the interpretation of a regression model. There are 30 variables entered in PCA in this study (Table 1). Next, K-mean cluster analysis (K-CA) was used for deriving typical household/farm groups defined by livelihood criteria. K-CA is a multivariate statistic method that groups similar objects into groups called clusters. The endpoint of this cluster analysis is a set of clusters, where each cluster is distinct from each other cluster, and the objects within each cluster are broadly similar to each other. Unlike hierarchical clustering methods, the K-CA method avoids problems of chaining and artificial boundaries and works on the original input data rather than on a similarity matrix. For a large dataset (e.g., hundreds of cases), K-CA should be chosen because it would be difficult to interpret grouping results using hierarchical cluster analysis given that the considered sample is large [28]. Data entered into K-CA can be the scores of PCs extracted by the earlier PCA, or original livelihood variables that are highly correlative with the extracted PC.
To determine the number of clusters, a previously described procedure was used [29]. The optimal cluster number is defined as the minimal cluster number with the highest cluster homogeneity. First, K-CAs are run with the number of clusters set to all values K = 2, 3 . . . , 10. For each K-CA (with a concrete K-value), the mean distance of cases to their assigned cluster centers was calculated. These mean distance values were then plotted against the increasing cluster number (K = 2, 3 . . . , 10). The optimal cluster number was chosen by examining the "knee" of the curve: The point from which the overall cluster quality, i.e., the reduction of the mean distance from cases to their cluster centers, or the overall cluster homogeneity is not substantially improved when K increases [30]. The livelihood groups of households/farms defined at this stage are just potentially functional livelihood types. Livelihood types will be functional regarding MRBT adoption outcome if MRBT adoption rates and drivers differ among the livelihood types. The functional livelihood types are tested in adoption assessment in the next step.

Dependent Variables
The first dependent variable used for inferential statistics (binary logistic regression) is the existence of MRBT practice in household-farms (MRB_PRACTICE). This is a straightforward (yes/no) adoption that is often seen in many publications.
The second dependent variable used for inferential statistics [multiple linear regression (MLR)] is the composite index of adoption quality that is conceptualized in the following formula: where AQ is the adoption quality index, P i is an on/off controlling factor of for AQ, and Q j is quality attributes for the performance and/or impact of the practice of considered technology. In this study, P i is the existence of a MRBT practice: If there is no MRBT practice (P i = 0), then AQ = 0 regardless of Q j value. We collected Q j on farmers' reflections on different benefits of MRBT, including improvements of household's machinery ability (MRB_MA), technical knowledge and skills (MRB_KT), cost of adoption (MRB_AC), crop yield (MRB_YD), water saving (MRB_WS), marketability (MRB_MKA), and market price received (MRB_MKP). The short definitions of these adoption quality components are displayed in Table 2. Therefore, in this study, the equation for AQ takes the following form:   (1) and (2)) Score between 0 and 14

Inferential Statistical Models
For the MRB_PRACTICE, because it is in binary scale (1 if the household practices MRBT and 0 otherwise, i.e., traditional farm), binary logistic regression (bi-logit) is used to identify factors determining MRBT adoption. As site-specific constraints and potentials influence MRBT outcomes, the unit of MRBT adoption analysis is recommended to be a field rather than a household. The effect of the hypothesized socio-ecological variables on the adoption of MRBT by a household can be modeled as: where P(MRBT) is the probability of MRBT adoption. X i and β i (i = 1, 2, 3, . . . , n) are explanatory variables and their weight coefficients (or affecting coefficients), respectively, and µ is a random error term. Performance evaluation of binary logistic regressions includes: • Chi-squared test for overall statistical significance of the regression model (Hosmer-Lemeshow test); • Probability of correct prediction, and; • Receiver Operating Characteristic (ROC) statistics.
Although the pseudo-coefficient of determination (pseudo-R 2 ) in bi-logit mimics the widely used R 2 in linear regression, there are no agreed benchmark values of pseudo-R 2 for determining whether model performance is acceptable. As an alternative, the goodness-of-fit of the model uses ROC statistics, as recommended by several experts in binary logistic regressions [31][32][33]. The ROC curve depicts the model sensitivity (true positive fraction) and model specificity (true negative fraction) over all possible cut-off points. The area under the ROC curve (theoretically ranging within 0.5-1) is used as the basis for evaluating model performance. If the area value is significantly (p < 0.05) higher than 0.5, then the model predicts the output better than chance. Area values of 0.7-0.8 show acceptable model performance, values of 0.8-0.9 demonstrate excellent performance, and values >0.9 indicate outstanding performance [31]. From another side, MLR is used for inferring determinants of AQ. Performance of MLR models is measured by F-test and R 2 .

Explanatory Variables
The vector of explanatory variables [X i ] (i = 1, 2, 3, . . . , n) is from the indicators of livelihood assets of the household that owns or operates the land. Vector [X i ] can have some overlap with the variables in the earlier PCA, but not necessarily. In general, inclusion of livelihood variables in [X i ] should be based on an understanding (through either literature or common sense) of the rationales of their effects on adoption of MRBT. The explanatory variables for regression analyses in this study are shown in Table 3. Both regression models require the checking of multi-collinearity problems that the explanatory variables can cause for the regression models. High multi-collinearity (i.e., strong linear relationship among two or more explanatory variables) is a problem because it undermines the statistical significance of an independent variable. Before conducting regression analyses, we tested for multi-collinearity of all explanatory variables using Variance Inflation Factors (VIF). Ideally, there will be serious collinearity problems for VIF > 5 [34]. We remove explanatory variables of VIF > 10 to reduce serious multi-collinearity [35]. All statistical analyses were done using SPSS 16.

Factors Explaining the Differences in ALS Types
The PCA run extracted nine principal components with total eigenvalues greater than 1.0, explaining 74% of the total variance of the 30 original livelihood variables. The rotated component matrix then helped to determine what the components represent ( Table 4). The PC1 was strongly related to variables approximating access/connection to market, both physical market access (H_DISTANCE_TOWN) and institutional market connection (H_AMA_EFFECTIVE and H_AMA_MEMBER). Thus, this component was named "Market factor", and explained 13.4% of total variance of the original dataset. The PC2 was most weighted by income variables, especially the share of income from agriculture (H_AGR_INCOME). Therefore, we labeled this "Agricultural income factor" and it accounted for 11.7% of total variance of the original dataset. The PC3 accounted for 9.5% of total variation and was most highly correlated with three livestock variables-H_LIVESTOCK, H_LIVESTOCK_PERS, and H_CATTLE-and, thus, was called the "Livestock factor". Pairwise correlations of the three variables showed that they were strongly correlated. Because of the high loading value, H_LIVESTOCK was selected to represent the Livestock factor. The PC4 was highly correlated with the demographic variables of surveyed household (H_SIZE and H_LABOR) and demographic pressure on land (H_AREA_PERS), and was called the "Labor factor".
The PC5 was explained mainly by the variable of family mean age (H_AGE_MEAN) and dependency ratio (H_DEPENRATIO) was called the "Age factor". The PC6 was mainly related to the share of on-agricultural income (H_NAGR_INCOME), thus named "Non-agricultural income factor". The PC7 was strongly associated with variables indicating house quality (H_NFLOORS and H_NROOMS), thus referred to as "House quality factor". The PC8 was mainly explained by distance from household house to farm, reflecting the household's physical transaction cost in daily farming, thus named "Access-to-farm factor". The PC9 was inversely related to household access to loans (H_LOAN_ACCESS), thus called "Less access-to-loans factor". However, as PC9 accounted for only 4% of the total data variation, and the loading of H_LOAN_ACCESS was not high, the role of this variable for distinguishing ALS types may not be important.

ALS Types
Twenty-two variables with loading >0.6 in Table 4 were used for subsequent cluster analyses (K-CA). We ran eight K-CA with K = 2, 3, . . . , 10 that used these 22 variables, and calculated the sum of squared errors (SSE) for every K-CA run. Figure 3 depicts the distribution of SSE versus the running number of clusters (K), which shows a "knee" point at K = 3, suggesting an optimal cluster number of K = 3 for the final K-CA. Increasing the cluster number further from this point will not effectively increase the average clumsiness of each cluster (clustering quality inversely approximated by SSE) [36].
The PC7 was strongly associated with variables indicating house quality (H_NFLOORS and H_NROOMS), thus referred to as "House quality factor". The PC8 was mainly explained by distance from household house to farm, reflecting the household's physical transaction cost in daily farming, thus named "Access-to-farm factor". The PC9 was inversely related to household access to loans (H_LOAN_ACCESS), thus called "Less access-to-loans factor". However, as PC9 accounted for only 4% of the total data variation, and the loading of H_LOAN_ACCESS was not high, the role of this variable for distinguishing ALS types may not be important.

ALS Types
Twenty-two variables with loading >0.6 in Table 4 were used for subsequent cluster analyses (K-CA). We ran eight K-CA with K = 2, 3, …, 10 that used these 22 variables, and calculated the sum of squared errors (SSE) for every K-CA run. Figure 3 depicts the distribution of SSE versus the running number of clusters (K), which shows a "knee" point at K = 3, suggesting an optimal cluster number of K = 3 for the final K-CA. Increasing the cluster number further from this point will not effectively increase the average clumsiness of each cluster (clustering quality inversely approximated by SSE) [36]. The final K-CA with K = 3 resulted in three clusters of farm-households. An ANOVA was used to test the differences among the mean values of the 22 used variables (data not shown). The final selection of a limited set of variables for characterizing the clusters was based on the statistical differences via ANOVA tests. In many variables, although there were significant differences among their mean values across ALS types, the variables were not picked to characterize the clusters because the mean differences were either less or equal to only one measuring unit of the variables (e.g., the case of H_SIZE, H_LABOR, H_NFLOORS, and H_NROOMS). Some variables had significantly high The final K-CA with K = 3 resulted in three clusters of farm-households. An ANOVA was used to test the differences among the mean values of the 22 used variables (data not shown). The final selection of a limited set of variables for characterizing the clusters was based on the statistical differences via ANOVA tests. In many variables, although there were significant differences among their mean values across ALS types, the variables were not picked to characterize the clusters because the mean differences were either less or equal to only one measuring unit of the variables (e.g., the case of H_SIZE, H_LABOR, H_NFLOORS, and H_NROOMS). Some variables had significantly high correlations with each other. In these cases, the most meaningful variable of them was selected. The key variables for characterizing ALS types are shown in Table 5.
ALS type 1 (n = 196, 55%): Poor household-farms (1900 EGP (EGP = 0.059 US$ as an average exchange rate for 2019)/person/year) with main income based on non-agricultural activities at low cost (agricultural income being 3% of the total income), less access to local market institutions such as agricultural market associations (AMA). It is likely these household members have frequent low-cost non-farm activities in town as they live closer to urban centers (3.7 km) compared to the two other ALS types (5.5-5.8 km).

Adoption Responses versus ALS Types
The MRBT adoption practice, adoption quality attributes, and AQ versus ALS types are shown in Table 6. The adoption quality attributes responsive to ALS types include farmers' reflection on MRBT benefits on technical knowledge and skills (MRB_KT), marketability (MRB_MKA), and market price received (MRB_MKP). The other adoption attributes did not respond differently to diverse ALS types, in which household-farms of ALS type 2 were most positive in MRBT adoption for these three quality attributes.

Determinants of MRBT Adoption (Yes/No Adoption)
The results of binary logistic regression for identifying determinants of MRBT practice adoption for the whole sample population and three ALS types are presented in Table 7. For whole sampled population and sub-groups of ALS types 1 and 2, the Hosmer and Lemeshow tests for the bi-logit models showed acceptable results: p > 0.05 showed no significant difference between predicted data for MRBT adoption and observed data, meaning that there was a good fit of the model to the data used. The percentage of overall correct predictions of the models were relatively high (76-80%). The calculated area under the ROC curve ranged within 0.69-0.85, indicating that performance of the models varied from acceptable to excellent for identifying the determinants of MRBT adoption. Although there was no sign of failure in the parameters of model performance tests, the bi-logit model for ALS type 3 had no explanatory power, with no variable demonstrating a significant effect. The poor statistical performance of ALS type 3 may be due to the small size of this farm cluster (n = 61) compared to a long list of used explanatory variables (30 variables). Notes: Symbols *, **, and *** indicate a statistical significance at 90% (p < 0.1), 95% (p < 0.05), and 99% (p < 0.01), respectively. Table 7 shows 14 variables with a significant effect on MRBT adoption. These 14 determinants of MRBT adoption can be of four categories: Common determinants for MRBT adoption (two variables): Common determinants of adoption were the explanatory variables found significant for the whole population and individual ALS types, and sharing the same direction of effect. If all determinants were common, then the treatment of adoption analyses for individual ALS types would have no benefit. In Table 7, among the 14 variables with significant effects, only two variables within the physical asset category were common determinants: H_NROOMs (supporting MRBT adoption) and H_EQUIPTMENT (discouraging MRBT adoption).
ALS type-specific determinants for MRBT adoption-type 1 (four variables): A determinant for adoption in this type was an explanatory variable found significant for the whole population, and for one of a few individual ALS type(s) rather than all individual ALS. In this case, the added value given by the treatment of individual ALS was to further narrow the specific zone(s)/condition(s) where the effect actually took place. Determinants of this type included average age of the family (H_AGE_MEAN) and income (H_INCOME) which effects are happen in ALS type 2. Membership of agricultural market association (H_AMA_MEMBER) encouraged farmers to adopt MRBT in general, but the effect was further narrowed in ALS type 3 (the medium and livestock-based farmers).
ALS type-specific determinants for MRBT adoption-type 2 (three variables): Determinants for adoption in this type were explanatory variables found for one of a few individual ALS type(s) but not realized through analysis of the whole population, reflecting the added value given by the use of individual ALS types. There were four determinants for MRBT adoption of this type, including farm size (H_FARM_SIZE) and income per capita (H_INCOME_PERS), for which effects are found only in ALS type 1 (poor and low-cost non-agriculture based). With larger farm size there was more adoption of MRBT in this group of farmers. This finding suggests that large farms should be targeted to widely spread adoption of MRBT. The negative effect of income per capita is understandable because the income of this group is largely based on non-agricultural activities. Here, the result of group characterization, as a step of the approach, eased the interpretation of the effect. A negative effect of the age of household head (H_AGE_HEAD) on MRBT adoption was found only in ALS type 3 (medium and livestock-based), but was statistically weak.
Determinants for MRBT adoption found only in whole-population analysis (five variables): The determinants of this category included five variables: Education of household head (H_EDU_HEAD) and sheep (H_SHEEP) with positive effects; and severity of water table raising (H_WATER_TABLE), share of agricultural income (H_INCOME_AGR), and distance from house to farm (H_DISTANCE_FARM) with negative effects. Except for H_INCOME_PERS, the effect directions of determinants agreed with common sense. The existence of determinants in this category demonstrates the complementary role of whole-population adaption analysis except for the ALS type specific ones.

Determinants of MRBT AQ
The empirical findings of the MLR analysis, with MRBT AQ as the dependent variable is shown in Table 8. The list of explanatory variables used is shorter than those shown in Table 3 (Section 3.2) because variables with high VIF were excluded with the purpose to minimize the problem of multi-collinearity for MLR models. The F-tests indicated that all MLR models for explaining MRBT AQ significantly differed at confidence levels of 95% (p < 0.05) or 99% (p < 0.01). Notes: Symbols *, **, and *** indicate a statistical significance at 90% (p < 0.1), 95% (p < 0.05), and 99% (p < 0.01), respectively. Table 8 shows nine variables with a significant effect on AQ, which can be grouped into three main categories: (1) Common determinants for MRBT adoption (one variable): The common determinant of MRBT adoption quality was household equipment (H_EQUIPMENT)(significantly negative). (2) ALS type-specific determinants for MRBT adoption, with two types in this category: -Type 1 (three variables): Determinants of this type included the share of non-agricultural income (H_INCOME_NAGR) and number of rooms in house (H_NROOMS) that were further narrowed for ALS type 1. The effect of household sheep (H_SHEEP) (significantly positive) was common for the whole population and ALS types 2 and 3. -Type 2 (two variables): The strong positive effects of effectiveness of Water Use Association (WUA) (H_WUA_EFFECTIVE) on MRBT AQ were found only in ALS type 1. Without the ALS type-specific adoption analyses, this determinant-being meaningful for policy and management practice-would not have been realized. The negative effect of agricultural market association effectiveness (H_AMA_EFFECTIVE) on MRBT adoption quality was found only in ALS type 2, which is quite unusual and needs further interpretation with additional information.
(3) Determinants for MRBT adoption found only in the whole-population analysis (three variables): The determinants of this category included distance from house to farm (H_DISTANCE_FARM), number of house floors (H_NFLOORS) (both significantly negative), and membership in agricultural market association (H_AMA_MEMBER) (significantly positive). Table 9 shows evaluation remarks on the added values and limitations of the ALS typology-based method compared to the traditional approach. The ALS typology-based approach, presented in this study, is the complementary use of both ALS type-specific and whole-sample analyses. The traditional approach, often found in current literature of adoption analysis, is the use of whole-sample analysis only. The evaluation remarks result from the findings of presented adoption analyses.

Evaluation of Added Values of the ALS Typology-Based Method Compared to Traditional Approach
Overall, the ALS typology-based approach captured more quantities and comprehension of causalities in MRBT adoption. The ALS typology-based approach found 14 and nine determinants of MRBT adoption and AQ, respectively; compared to 10 and seven corresponding determinants for the traditional approach. The ALS typology-based approach added the following values: • Confirmed widespread role of common determinants of MRBT adoption across ALS types; • Household groups subjected to the effects of MRBT adoption that were found in whole-sample analysis; • Discovered new causal effects that the traditional approach could not. The effects of effectiveness of agricultural institutions, such as WUA and AMA, can only be realized with ALS type-specific analyses rather than the whole-sample analysis; • By complementary use of whole-sample analysis, the ALS typology-based approach utilized the large size of the whole sample to increase statistical power, thus off-setting the problem of weak statistical power for small ALS groups. Capture more quantity and comprehension of causalities in MRBT adoption

Conclusions and Recommendations
This study empirically investigated the issues using a system-based option-by-context approach, or ALS typology-based approach, for guiding concrete analytical steps and statistical methods in coping with the challenges of system complexity and contextual diversity in two governorates of Egypt: Sharkia and Assiut. A sustainable livelihood framework was used to define candidate variables entered in the sequential multivariate statistical analyses. First-subsequent to PCA-K-mean cluster analysis and ANOVA statistical tests were used to objectively define three distinct ALS types. Consequently, we conceptualized the formula of AQ for statistical analysis in a quality-focused way, in addition to consideration of yes/no adoption. Relevant inferential statistical methods (bi-logistic and linear regressions) were applied to the whole/merged sample, and sub-samples corresponding to ALS types, to identify determinants of farmers' adoption of MRBT.
The main finding of the study is that classifying the considered agrarian population into a limited number of ALS types and conducting multivariate inferential statistics for both (1) whole sample population and (2) each specific ALS types helped discover hidden causal relationships shaping MRBT adoption which could not be identified by considering the whole sample only. For instance, roles of effectiveness in agricultural institutions, such as WUA and AMA, in MRBT adoption were found in specific ALS types rather than looking at the merged population. This finding highlighted the need to strengthen the capacity of these existing formal organizations and to establish new ones. Some causal relationships were found significant via inferential statistics for the whole sample, but actually had an effect in a specific ALS group (e.g., the effect of farm size on MRBT adoption of poor and non-farm-based income group). This result suggests that farm size play a critical role in the adoption process of MRBT, and this factor was one important determinant of technology adoption. Indeed, this technology requires a significant investment of capital and time but may offer cost savings and higher yields through more precise management of inputs. This validates the previous findings.
From the technical and political points of view, understanding the factors that influence or hinder adoption of MRBT using an ALS typology-based approach is essential in planning and executing technology related programs for meeting the challenges of food production in Egypt. Therefore, to enhance technology adoption by farmers, it is important for policy makers and developers of new technology to understand the farmers' needs as well as their ability to adopt technology.
From the methodological point of view, the added values of the ALS typology-based approach that were indicated in this study were: (1) Confirmed widespread role of common determinants of MRBT adoption across ALS types, (2) household groups subjected to the effects of MRBT, (3) discovered new causal effects that could not be found by traditional approach, and (4) utilized the large size of the whole sample to increase statistical power, thus off-setting the problem for the small ALS group.
Finally, the ALS typology-based approach has a particular limitation regarding poor performance of inferential statistical models for small ALS groups, and consequently some limitations should be noted when using this approach. These limitations are probably due to two reasons: (1) Small sample size reduces the statistical power of parametric methods; and (2) ALS classification tends to reduce within-cluster variation, causing poor performance of statistical models for MRBT adoption analysis if the same data used for ALS grouping are also used for adoption analysis. A possible solution for the former issue would be testing the use of additional non-parametric methods for small ALS groups. An alternative for the latter problem would be the use of additional data sources for adoption analysis and assessment.