A cross-sectional study was conducted in Madagascar from December 2005 to April 2006 in three geographical areas. These areas were selected purposively based on the following criteria. First, the study areas were stratified according to the three main climatic zones of Madagascar (Table 1). Then, they are important regions for pig production where outbreaks of swine diseases are regularly reported by pig farmers. Finally, the selected areas were relatively easy to access, in order to limit logistic constraints for the data collection. As shown in Fig. 1, the study areas were named after their largest towns: Ambatondrazaka, Arivonimamo and Marovoay.

The unit of interest in this study was the individual pig farm, defined as any premise where at least one pig was reared. Our target population comprised of all such pig farms in the three study areas. A sample size of 300 pig farms per study area was specified as a target which would allow estimating husbandry practice frequencies at a 95% confidence level with a precision of 5%, assuming frequencies of 50% for dichotomous factors. Within each study area, since no register of pig farms was available, pig farms were selected using a multi-stage sampling approach. First, as primary sampling units a purposive sample of fokontanies ? the smallest administrative unit in Madagascar ? was selected. In each study area, the sample of fokontanies covered rural and urban areas, with settlements connected by commercial exchanges. Because of the limited road network, fokontanies with few human settlements or which were in remote locations were excluded. The secondary sampling units were the pig farms within the selected fokontanies of the three study areas. The number of pig farms to select in each fokontany was calculated assuming random proportional sampling, based on a constant sampling fraction for each study area. For this purpose, the total number of pig farms in each selected fokontany had been estimated with the assistance of veterinarians, local associations supporting pig farmers and local administrators. The sampling fraction was 20% in Ambatondrazaka, 30% in Arivonimamo and 17.5% in Marovoay. Since no sampling frame was available for the secondary sampling units, the selection of pig farms in each fokontany was done along line transects. Each line transect started in the middle of the fokontany, its direction was chosen randomly, and all pig farms of the fokontany situated along the transect were selected. The same approach was repeated until the sample size was reached for each fokontany.

A questionnaire was developed to collect data on husbandry practices. Using 49 closed and semi-closed questions, the following aspects of pig farms were investigated: demographics, housing, commercial exchanges, method(s) of reproduction, contacts with other animals, contacts with people and vehicles, feeding, animal health care, waste handling and biosecurity measures.

The questionnaire was developed in French, and administered in Malagasy to ensure that farmers would understand all the questions (the questionnaire in French is available upon request to the first author). It was pre-tested with 10 pig farmers and 3 interviewers outside our study areas and questions were refined according to feedback from both farmers and interviewers. The final questionnaire was administered by 9 interviewers ? 3 per study area ? trained before the start of the study. For biosecurity reasons, interviewers did not enter the premises and the collected information relied solely on farmers? responses. It was emphasised that the answers would be processed anonymously, and that correctness of the answers was necessary for the study results to be sufficiently meaningful so that they are suitable for informing the development of tailored recommendations for farmers.

MFA (^{Escofier and Pages, 1994}) examines the relationships existing between variables separated into different groups. It can be considered as a factor analysis (principal component analysis for quantitative variables, multiple correspondence analysis for qualitative variables) applied to the whole set of variables within which each group is weighted. All elements of the dataset (individuals, variables, groups of variables) are represented graphically in a Euclidean space. The principle of factor analysis is to define projections ? or factors ? representing an optimised quantitative summary of the relationships between variable categories. A factor is therefore a linear combination of the variables and is characterised by its eigenvalue, which indicates the variance ? or inertia ? of the data it represents. The first factor is the projection which represents the highest amount of variance, and each other factor is defined so that it captures the variance not explained by the previous factor. The cumulative percentage of inertia of a given number of factors indicates the percentage of variance of the dataset they explain. Factor analysis can include both active variables, used to calculate factors, and supplementary variables, not used to define factors but projected on these factors. In a MFA, factor analyses are first performed independently for each group of variables. The results are then normalised by dividing individual scores by the square root of the first eigenvalue, in order to make the different groups of variables comparable in a global analysis. A factor analysis is then performed for the merged dataset (obtained by juxtaposing the individual normalised datasets), where each group of variables has an equal a priori influence. By generating common factors for both variables and groups of variables, MFA takes into account the heterogeneity of groups of variables in terms of biological meaning, and allows the identification of the main variables and groups of variables that differentiate between the individuals.

Data entry and data coding were performed using Epi Info 3.3.2, data manipulation using Microsoft Access 2003, and descriptive statistics using STATA 9.2 (Statistical Software: Release 9.2., Stata Corp., College Station, TX, USA). Multivariate analyses were conducted with the statistical software R 2.5.1 (^{R Development Core Team, 2007}), using the package ade4 for MFA (^{Chessel et al., 2004; Dray et al., 2007}).

A total of 42 variables describing farm management and biosecurity practices were included in the dataset. They were grouped according to different aspects of husbandry practices assumed to have the same influence on the risk of introduction of contagious diseases: 5 variables described the structure of the farm, 8 referred to contacts with other animals and pigs from other farms, 7 considered person- and vehicle-contacts, 7 referred to feeds, and 7 focused on sanitary aspects. Eight variables illustrating the demographics of the farm were also included. Binary variables for which less than 5% of farmers gave a positive answer were excluded from the analysis. Otherwise, these variables with few observations may be outliers and have a dominant influence on the definition of factors. Farms with missing values were also excluded from the analysis, as factor analyses such as MFA do not allow missing values.

MFA was conducted on data from all pig farms as well as separately for each study area, in order to identify the main aspects of management and biosecurity practices discriminating farms and investigate whether these differed between study areas. The number of factors selected for interpretation was determined using the scree plot, which indicates the eigenvalue of each consecutive factor. A break on the scree plot separates factors with large and small eigenvalues. Only factors with large eigenvalues were retained for interpretation. Hierarchical cluster analysis (HCA) (^{Everitt, 1974}) was then used to differentiate groups of farms with similar management and biosecurity practices. HCA were conducted on pig farms? MFA scores, using Ward's criteria for linkage. The principle of Ward's aggregation method for the HCA is to group individuals ? pig farms ? in a way that minimises intra-cluster variance and maximise inter-cluster variance. The characterisation of the farming practices significantly associated with each group of farms was then done by calculating test-values (^{Morineau, 1984}). Test-values measure the distance between the within-group value and the overall value for each variable category ? or practice.

Data on management and biosecurity practices were obtained from a total of 853 pig farms, 709 of which had no missing values and were included in the analysis: 272 farms from Ambatondrazaka (38.3%), 233 from Arivonimamo (32.9%) and 204 from Marovoay (28.8%).

The 42 categorical variables describing management and biosecurity practices were separated into 5 active groups and a group of supplementary variables. Tables 2?6 list the 34 variables separated into the following groups: structure of the farm, animal-contacts, person- and vehicle-contacts, feeding, and sanitary aspects. The 8 variables introduced as supplementary variables are presented in Table 7. Six variables were excluded from the analysis: use of proper farm clothing (0.6%), use of disinfection baths or sprays at the entrance of the premises (3.3%), quarantine for pigs entering the farm (3.6%), presence of a sickbay (0.9%), equipment shared with other farm (s) (4.0%), systematic working route progressing from clean to dirty zones of the farm (0.7%).

The global displays of groups (Fig. 2) present the groups of variables in a two-dimensional space defined by the 2 first factors for each of the MFAs performed. The global display indicates the importance of the groups of variables for differentiating between pig farms: the larger their inertia on the factors 1 and 2, the more they differentiate between pig farms. For the MFA conducted on the 709 farms, the cumulative percentage of inertia for the 2 first factors was 19.2%. The husbandry practices differentiating between farms were, in decreasing order of importance: structure of the farm, sanitary aspects, feeding, animal-contacts and, to a lesser extent, person- and vehicle-contacts (Fig. 2a). For MFAs performed separately on data from Ambatondrazaka, Arivonimamo and Marovoay, we selected factors explaining 22.5, 36.5 and 19.3% of the variation, respectively. For Ambatondrazaka, Fig. 2b showed that the inertia of the five groups of variables in relation to factors 1 and 2 were low, and therefore pig farms were poorly differentiated. The plot for Arivonimamo (Fig. 2c) indicated that pig farms were differentiated by the five groups of variables on factor 1. In Marovoay, aspects of husbandry and biosecurity practices discriminating farms were, in decreasing order of importance: feeding, sanitary aspects, animal-contacts and, to a lesser extent, person-contacts (Fig. 2d). The results of the MFAs performed on data from the 3 different study areas thus showed that the management and biosecurity practices discriminating between pig farms differed greatly between regions.

In Fig. 3, the representation of all pig farms and study areas in relation to the first two factors of the MFA showed that groups of farms with similar practices existed, and this grouping seemed related to the study areas.

The results from the MFAs therefore suggested that the main husbandry practices differed between regions.

The HCAs carried out on each study area's MFA scores resulted in the identification of four clusters of farms in each of two study areas (Arivonimamo and Marovoay), while no clustering of farms was found in Ambatondrazaka. Although this latter area was not homogeneous in terms of husbandry practices, no clear groupings of farms could be identified. The main characteristics defining the clusters of farms are presented in Table 8, and a description of the clusters is provided below.

It was not possible to select a random sample of pig farms in Madagascar, as there is no central or even village-level register of pig farms. The three study areas were a purposive sample of Malagasy regions where pig production is important. Within these areas, a multi-stage sampling approach using line transects was applied to limit selection bias. The selection of the primary sampling units, the fokontanies, could not be performed randomly due to the absence of sampling frame, the limited road network and difficult access to remote rural settlements. Therefore, the study sample may have been biased towards pig farms better connected to transport networks. Because of the non-random sampling approach applied in this study, results are relevant to the study sample but one should be cautious in making inference about the study population.

The effective sample size achieved in the study was sufficient to estimate frequencies of management and biosecurity practices at a 90% confidence level and with a precision of 5%. The data collected should thus provide acceptable estimates of the farming practices in the study sample.

Our study relied on data about farm management practices as reported by farmers, and not on direct observations. It was decided not to enter premises in order to not compromise farm biosecurity. During the pre-testing of the questionnaire, farm visits were done after completing the questionnaire, to check the validity of the responses given. It appeared that the good level of accuracy of information reported by farmers could be further improved by asking for a description of their protocols, specifically in relation to animal health care and biosecurity measures. Semi-open questions were subsequently included in the final version of the questionnaire. Positive answers were only considered validated when an appropriate description of the products and protocols used was provided for the following practices: animal health care, control of rodents and insects, cleaning and disinfection, use of proper farm clothing, use of disinfection baths or sprays at the entrance of the premises, quarantine for pigs entering the farm, presence of a sickbay, equipment shared with other farm(s), and systematic working route progressing from clean to dirty zones of the farm. Some farmers probably still gave answers reflecting what they thought they were supposed to do rather than what they actually did, and others refused to answer some questions. In order to reduce this information bias, interviewers explained the aim of the study to farmers, and emphasized that we were interested in practices actually applied on the farm, with the aim of providing adapted recommendations for limiting the risk of introduction of contagious disease into their pig herds.

Another potential source of bias in this study was the administration of questionnaires by 9 interviewers. In accordance with recommendations made by ^{Smith and Morrow (1996)} to reduce such bias, the interviewers took part in the testing of the questionnaire and their feedback was taken into account when refining questions. In addition, detailed instructions for the interviewers were provided in the questionnaire and interviewers were trained for the interview process.

Multiple factor analysis and hierarchical cluster analysis are particularly well suited for this study, as they are adapted to questionnaire data with correlated variables. MFA allows the analysis of variables separated into groups according to distinct biological meanings, which enables an intuitive interpretation of the results. The use of multivariate analyses for describing management and biosecurity practices has an advantage over the scoring system used in the study by ^{Julio Pinto and Santiago Urcelay (2003)}, as it can give an indication not only of the level of biosecurity in farms, but also of the main differences of practices existing between groups of farms. Other studies have used multivariate analyses to investigate swine management and biosecurity practices (^{Hurnik et al., 1994; Rose and Madec, 2002; Boklund et al., 2004; Ribbens et al., 2008}), using either factor analysis or multiple correspondence analysis and cluster analysis. In these studies, although practices were initially presented according to different aspects of practices (e.g. management and personnel, transport to slaughter, biosecurity status), all variables were then pooled for the multivariate analysis. Only in the study by ^{Ribbens et al. (2008)} were management practices and biosecurity status considered separately, in two independent analyses. In this context, the advantage of MFA over multiple correspondence analysis is that it allows the simultaneous analysis of different aspects of husbandry and biosecurity practices.

The results showed that the biosecurity is poor in Malagasy swine farms, despite a campaign implemented in 2000 to train pig owners on swine diseases ? and ASF in particular ? and measures to prevent them (Malagasy des Professionels de l?Elevage, unpublished data). Most pig farms were small farming systems where almost no sanitary measures were applied. In addition, many opportunities for contacts existed, such as with pigs from other farms for commercial exchanges or natural services, with other animals present on the premises and with various stakeholders within the pig production sector.

The results of the MFA showed heterogeneity between regions in terms of management and biosecurity practices. This heterogeneity may be due, partially at least, to differences in ethnicity and culture, climate and agro-systems between the three areas selected for this study.

Additional information was collected during informal interviews with veterinarians, representatives of farmer associations and other key stakeholders from the pig production sector. This information suggested that training and technical support to farmers vary between regions, and this may be another reason for the presence of distinct profiles of husbandry practices in the three regions. In both Ambatondrazaka and Arivonimamo, veterinarians and animal health workers are present in cities and work with pig farmers. A dynamic farmers? association is also present in Arivonimamo city. In the region of Marovoay, veterinarians work mainly with zebu and beef cattle producers. A larger number of farmers reported little training and limited access to technical advice in this area, although farmers? associations were set up in some localities. The results obtained at the regional level reflected these differences: pig farms in Marovoay had lower biosecurity standards than in Ambatondrazaka, where the pig farming systems had in turn lower biosecurity standards than in Arivonimamo.

The results of the HCA showed that within-area differences existed and different types of farms were differentiated in Arivonimamo and Marovoay areas. In Ambatondrazaka, the absence of grouping in terms of swine management and biosecurity practices suggested that pig farmers have individually adapted their practices in accordance with their own perceptions of practices influencing the introduction of contagious diseases. In Marovoay and Arivonimamo, within-area differences might be partially explained by variation in: access to professional expertise on swine health, technical support and training on farm management and diseases. Moreover, the variation observed probably reflected differences in household wealth and whether owners reared pigs for subsistence or commercial purpose. Unfortunately, this cannot be confirmed as it was not investigated by the questionnaire. Finally, in Arivonimamo and Marovoay, farmers associations aiming to improve production are present. This might be an indication of some farmers? willingness to work together and share expertise in order to reduce disease risk and improve productivity.

The most important result of this study is the identification and characterisation of the different clusters of farms existing in the investigated study areas of Madagascar, as it will allow the development of tailored recommendations for improving productivity and reducing the risk of introduction of contagious diseases. This is essential in Madagascar, where pig owners have limited financial resources and can only make limited changes to their practices. Support to smallholder farmers should help them in making informed changes according to their management profiles and their respective influence on the risk of disease, rather than providing farmers with a long list of management practices and biosecurity measures they would never be able to all implement. Studies as the one presented here represent the first step in this direction.

In conclusion, this study provided a description of management and biosecurity practices in smallholder pig farms in three regions of Madagascar, and the identification of farm groups based on different patterns of husbandry practices. The practices investigated in this study were assumed to have a potential influence on the introduction of contagious diseases. The results presented here will be used in a subsequent study to investigate the association between the main practices implemented in Malagasy pig farms and clinical signs suggestive of ASF, CSF and Teschen disease. Results of both studies will then be used to develop tailored recommendations, in order to reduce the risk of introduction of diseases amongst the different types of smallholder pig farms identified in Madagascar.