Both ‘household’ and ‘slum’ have multiple technical definitions. Different agencies define ‘slum’ in different ways. For example, UN Habitat specifies that a slum area is “any specific place, whether a whole city, or a neighbourhood, [...] if half or more of all households lack improved water, improved sanitation, sufficient living area, durable housing, secure tenure, or combinations thereof .” While UNESCO use: “A contiguous settlement where the inhabitants are characterised as having inadequate housing and basic services .” In our context, the study sites are well-known ‘slum’ areas whose boundaries are defined by the communities themselves or agreed among relevant stakeholders. Regardless of the definition, though, our method was designed to be applicable to any complex, irregular, or unplanned urban environment. For this article a necessary condition to be defined as a “household” was to be living in the same housing unit or connected premises. Other criteria such as having common cooking arrangements are frequently used, but are not required by this method.
The aims of the method described in this paper are to produce a valid spatially-regulated sample and to conduct a survey using this sample. Our sampling frame within each study-site is therefore the complete set of geo-located household locations. We describe firstly how access to slum sites might be obtained, secondly how a sampling frame is generated, thirdly how to sample from this sampling frame, and fourthly the process of collecting, managing and storing data from slum sites.
Access to slum sites
Negotiating and obtaining access to conduct research in slum sites, which are often physically challenging and socially complex informal environments, requires an in-depth understanding of the political and social structures at local, community, and national levels. Slums are heterogeneous and the political and social structures within them vary between countries, cities, and even within the same city where more than one slum is present (e.g. [7, 8]).
The first step is a stakeholder mapping and engagement phase for each slum site. This is required in order to negotiate and obtain access, both from local authorities and local community leaders. The identification of all the relevant stakeholders (i.e. the “mapping”) should be undertaken by a local research team with knowledge of the national and local policy-drivers/makers as well as the political and social structures of the slum communities and local governments. There is not necessarily an optimal method to conduct the stakeholder mapping and engagement. A focus-group approach to stakeholder mapping can be cost-effective, rapid, and easily adaptable to a wide range of contexts. However, it may not always be suitable in practice as it is not always easy to bring together busy people. Similarly, political sensitivities or social hierarchies may affect group dynamics and ‘who’ speaks. Therefore, a combination of focus groups and one-on-one engagements may be more practical and appropriate.
The mapping and engagement exercise must consider each stakeholder’s interests and influences based upon their agenda, power-base, credibility, and consequences of the research for them, to ensure successful engagement. Understanding the sociology and political economy in slums is a key objective of stakeholder engagement exercises; however that aspect of the research is beyond the scope of this article, and here we focus on the issue of engaging stakeholders to negotiate access and site entry.
Table 1 provides some examples of access negotiations in the Slum Health Project. In all sites, we met with local community leaders and government officials, as well as NGOs and different community-based groups. The relevant government authority was the first point of contact followed by local community leaders. A “snowball” approach was taken, so that any additional stakeholders identified in the meetings were also engaged. Once access was negotiated, key stakeholders were kept appraised of all research activities, including times and dates when field workers would be present.
Constructing a sampling frame
In order to generate a spatially-regulated sample, the sampling frame must list all households in the area of interest and their precise locations. In high-income country settings, listings of households and their addresses are well-maintained, along with accurate detailed maps permitting the enumeration and geo-location of each household. Frequently, neither maps nor household listings exist for slums as the population is often not legally resident on the land, the structures that would be shown on maps are temporary and changeable, and there is little incentive for the state or private enterprise to produce maps. As a result there is a lack of information, formal or otherwise, about the location and function of structures and where households reside. Therefore, both a detailed map of all structures and a listing of households linked to locations on that map is required. Each country in the project formed a local mapping team, which included research staff and local community members and who were trained locally on each of the tasks described below. Figure 1 shows a simplified flow diagram of the processes to generate the sampling frame.
Generation of digital map data from satellite imagery
In this project, the slum boundaries were defined in collaboration with the local research team and slum community leaders. Official administrative or electoral boundaries can be used but can often be considered incorrect or out of date, particularly given the dynamic nature of the slum (e.g. ). Optical satellite images covering the study sites were procured from Airbus Intelligence at a resolution of ~ 30 cm – note that the resolution of freely available satellite imagery such as LandSat (~ 15 m) is insufficient for identifying the relevant features.
An online mapping platform was set up using the Humanitarian OpenStreetMap Team (HOT) Tasking Manager, which is a free online Geoweb infrastructure for coordinating remote participatory mapping, i.e. generation of map data from satellite imagery by a varied team in multiple locations (e.g. ). The HOT Tasking Manager subdivides the area of interest into smaller grids, each referred to as a “Task” that can be selected by a participant and mapped (Fig. 1, Step 2). Local project teams were first trained before recruiting additional participants including slum residents, OpenStreetMap communities (local and non-local), and other project team members. Once the digital maps are completed and validated against the satellite imagery, they need to be validated through “ground-truthing”, i.e. comparing the mapped features with observations on the ground. Each task is validated by an experienced mapper. The generated data are uploaded onto the OpenStreetMap online database (Fig. 1, Step 3) .
Onsite participatory mapping
The onsite participatory mapping is the “ground-truthing” stage: the accuracy of the digital map produced from the online mapping is checked in the field. This stage involves a number of steps. First, roads and footpaths are tracked in the study sites with handheld GPS devices to confirm their locations as mapped from the satellite imagery and produced in the digital map. Second, each structure is verified in two ways: if its geometry is incorrect, any changes are drawn on printed versions of the digital maps in the field, which are scanned and overlaid with the digital maps to make any corrections using the FieldPapers.org service; and third, each structure is surveyed using the digital data collection tools to generate unique identifiers for each structure, which are also marked indelibly on the structure for future identification, and to determine its function (e.g. residence or shop).
Where structures are identified as dwellings, each household as defined is recorded and identified by the name of the head of household or family name. High population turnover and lack of tenure or rental contract can result in households departing without notice. For the “Slum Health Project”, any household that had not been observed by neighbours for a period of 3 months or more is no longer considered ‘resident’. The spatial locations of the identified households are then linked to the relevant structure by the structure’s unique identifier. The location of the household is specified as the structure’s centroid. Once all structures are surveyed the sampling frame is complete.
As slums typically exhibit very substantial spatial heterogeneity, it is desirable that the sampled locations span the whole of the site. A geometrically simple way to achieve this is to sample the households at, or as close as possible to, the points of a regular lattice overlaid on the mapped site. However, this has the disadvantage that it is biased in favour of sampling relatively isolated households. A completely random sample removes the bias but also results in uneven spatial coverage of the site. These considerations led Chipeta et al.  to suggest using an inhibitory sampling design, in which sampled locations are chosen at random subject to the spatially regulating constraint that no two sampled locations can be less than a specified distance d apart. The packing density of an inhibitory design is the fraction of the site area occupied by discs of diameter d centred on each sampled location. The maximum achievable packing density depends on the spatial arrangement of the available locations, here households, but in high-density settings a value of around 0.4 produces a highly regulated sample. Inhibitory designs are generally efficient for capturing spatial variation on the scale of the whole site, but cannot neither capture small-scale spatial variation nor distinguish it from non-spatial variation amongst the individuals who populate the sampled households. For this reason, Chipeta et al.  recommended tempering an inhibitory design by including a number of close pairs, i.e. augmenting an inhibitory design with a number of sampled locations, each one of which is located less than a specified distance e from the closest point of the inhibitory design, with e < d. In this context “close pairs” is taken to be households residing in the same structure (e = 0). The number of close pairs and value of d will determined based on data from pilots conducted at each study site of between 20 and 30 households, which are purposively selected to maximise spatial variation.
Sample size in this context is often based on pragmatic considerations including resourcing and time. The “effective sample size” of a spatially correlated sample, i.e. the equivalently sized uncorrelated sample that provides the same information on a statistic of interest, is dependent on the degree of spatial correlation and the location of samples, among other things. Griffith  for example, provides a conceptual frame work for consider the effective sample size of spatially correlated samples.
Data collection methods
Data collection methods in a slum context are dictated by similar considerations as any other context: data quality, data security (including relevant legislation such as the General Data Protection Regulation (GDPR) that came into force in the European Union in May 2018 ), ease of use, and costs. Based on the above criteria we recommend the use of digital tablet devices over paper-based forms as they reduce the risk of transcribing error, protect data security through encryption and not requiring the transport and storage of multiple paper forms, and reduces costs by not requiring extensive data entry. For the Slum Health Project digital tablet devices were purchased for all field workers and locked with a password. Field workers were required to sign agreements to use the tablets responsibility and all tablets are signed in and out.
In terms of software, a number of both proprietary and open-source options are available, including Open Data Kit, RedCap, and Survey CTO. We opted for the open-source Open Data Kit suite of software, which will improve sustainability . Importantly this software permits offline data collection, automatic encryption, and uploads all submissions when the device is connected to the internet. Form programming was completed using xlsform . These tools permit complex survey design and skip structures, can restrict responses to reduce errors, and collects locations, signatures, and images as required. A data aggregation server was set up using a (GDPR-compliant) cloud server provider that permits full control of the server and location of data storage to which access was strictly limited. Access to the server was secured using password-protected 256-bit SSH keys. A second data storage server was set up at the University of Warwick, to permit access to the data for project members and constitute a backup. The data collection process is as follows (Fig. 2):
Field workers conduct the interviews, completing the forms on the tablet devices. The responses are checked by a field supervisor for any potential errors. Additional quality control steps include spot checks by field supervisors, i.e. returning to a sampled house and re-asking a subset of questions, and sit-ins on interviews by supervisors. If any potential errors are identified, the field worker returns to confirm responses, otherwise the form is finalised. The software encrypts finalised forms using AES-256 encryption; following which forms are no longer accessible and are submitted automatically to the server when online and are deleted from the tablet (Fig. 2(1)).
Encrypted data are stored on the data aggregation server. When requested, the data are decrypted, the unique identifiers of submissions extracted and checked against a list of previous submissions, and if there are new submissions the data are processed (i.e. redundant columns removed, wide-form data converted to long-form, and separate data sets combined if they are from the same form, to facilitate use), and then re-encrypted using AES-256 encryption with a separate set of country-specific SSH keys. The data are then submitted to the data storage server via SFTP (Fig. 2(2)). Any unencrypted data are deleted.
Data are stored encrypted on the data storage server until required (Fig. 2(3) and (5)). Quality checks are conducted (Fig. 2(3) and (4)). On completion of data collection, the data will be ‘cleaned’ and merged into one final data set.