2.1 | Study area The study area is located in the sub-district of Birganj, part of the Dinajpur district in North-Western Bangladesh (25°44'N and 88°40'E). The district is located in the Old Himalayan Piedmont plan agro-ecological zone (AEZ-1), and the land types are highland (HL; 5%), medium highlands (MHL; 37%) and medium lowlands (MLL; 5%), respectively. Under the Köppen climate classification, the climatic condition of the North-Western part of Bangladesh including Dinajpur district is ‘Cfa’. (Humid Subtropical Climate). These areas also experience high temperatures and limited soil moisture as well as low and erratic rainfall. The annual average rainfall in the Dinajpur district is 1710 mm which mainly occurs during the monsoon and varies widely both by season and year. For example, rainfall recorded in 1982 was 1,342 mm, while in 2015 it was 1,965 mm. The average annual maximum and minimum temperature in the region is 35.11°C and 20.28°C, respectively. These conditions make the region drought-prone, leading to poor crop productivity. Thus, the livelihood of people in the area is threatened by climate extremes, particularly drought, in the late winter season. This region is also geographically vulnerable to natural hazards such as flash floods, heat waves and cold spells, which have resulted in increased food shortage (Barma et al., 2019; Hossain & Teixeira da Silva, 2013; Paul et al., 2013). Historically, the regional economy has depended on the agricultural sector, with predominance of cereal crops, especially rice, as well as other major crops like maize, wheat, potato and pulses (Mainuddin et al., 2020). In addition to this, aquaculture, the rearing of livestock, poultry and off-farm activities provide additional income to the farm households in this area.
2.2 | Data collection A sampling framework was constructed in consultation with the relevant local extension personnel, in particular, agricultural officers, before the final sampling. A multistage sampling procedure was used for this study to select a research area and sample farmers. A total of 92 farms were randomly selected to address the research objective as well as to identify the different farm household types. Afterward, a draft semi-structured interview schedule was used on 10 non-sampled respondents for necessary modification. Finally, data were collected from the selected households from February to March 2019 by face-to-face interviews. This month was selected to minimize the possible recall bias relating to the quantities of inputs used and output (grain and residues) obtained. The primary data collection was carried out with the support of three agricultural graduates, and the interviews were conducted either at the respondent's house or in one of the farmer service centres where farmers regularly meet. Seven sections were included in the interview schedule, the first focusing on information about household characteristics and the second including questions on access to food and other major assets. In the third section, questions were targeted at farm activities including herd size, land use, cropping patterns and field management practices etc. The fourth section focussed on crop residue management. The fifth section concentrated on livestock, such as breed type, herd structure, dynamics and feeding strategies. The sixth section was aimed at collecting information regarding the adoption of new technology, ease of access to a market and key constraints to farming. Finally, the last section recorded information about household income and expenditure.
2.3 | Typology construction Farm household data were analysed by using a multivariate statistical approach comprising principal component analysis (PCA) and cluster analysis (CA). PCA was used to reduce the data set and create a smaller set of independent components. The new set of independent components was used as an input for cluster analysis and, later on, identifying the farm household in the research area. The technique has been widely used in many studies to classify farm households (e.g. Bidogeza et al., 2009; Goswami et al., 2014; Kuswardhani et al., 2014). All analyses were done using the ade4 package (available online http://pbil.univ-lyon1.fr/ade-4 from R 3.6.0. software R Core Team (2019).
2.3.1 | Principal component analysis The first steps for the PCA was data quality control, including identifying missing values and variables with strong correlations. The data set based on the 27 variables was carefully examined and missing data was identified. Based on Kaiser's criterion, all PC having an eigenvalue of one was retained for further analysis (Field, 2005; Herve, 2001). If the number of variables is less than 30, Kaiser's criterion is considered to be accurate (Field, 2005). In our study, there were 27 variables, thus making it appropriate for this research. The number of the axis for principal component analysis can be determined based on the minimum cumulative percentage of variance, 60% or higher is usually best for PCA (Hair et al., 2010). In our case, it was about 69% which was suitable for our research. In addition, loading of less than 0.40 was not considered for the interpretation of our objectives.
2.3.2 | Cluster analysis The nine components from the PCA were used to develop hierarchical clustering following Ward's method (Reynolds et al., 2006). Although there is no single procedure to determine the appropriate number of clusters, a two-step approach (i.e. the hierarchical method and the partitioning method) was used (Hair et al., 2006). The k-cluster solution was created by connecting with two clusters from the k+1 cluster solution, whereas the partitioning method was employed to isolate the farm household into a given number of clusters (Lattin et al., 2005). Ward's hierarchical method was used to define the number of groups as it was widely used to minimize the variation within the cluster and successively join with equal clusters (Kuivanen et al., 2016). A key point in this procedure is where to cut the tree to identify an appropriate number of clusters that is realistic for the study area. Shifting the cutting line from A to B reduces the number of clusters to four; hence, line C denotes only two farm types. The number of clusters should reflect the real situation in the study area. By using cutting line C, two clusters based on the partitioning method were appropriate, but it did not represent the real situation in the area. Finally, using information from the dendrogram and taking into account expert knowledge in the study area, the number of clusters was chosen which was meaningful and realistic. To identify the variance between clusters, one-way analysis of variance was carried out and this was largely used to analyse the clusters (Field, 2005).