Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

Satellite-aided survey sampling and implementation in low- and middle-income contexts: a low-cost/low-tech alternative

Satellite-aided survey sampling and implementation in low- and middle-income contexts: a... Background: The increasing availability of online maps, satellite imagery, and digital technology can ease common constraints of survey sampling in low- and middle-income countries. However, existing approaches require special- ised software and user skills, professional GPS equipment, and/or commercial data sources; they tend to neglect spatial sampling considerations when using satellite maps; and they continue to face implementation challenges analogous to conventional survey implementation methods. This paper presents an alternative way of utilising satel- lite maps and digital aides that aims to address these challenges. Results: The case studies of two rural household surveys in Rajasthan (India) and Gansu (China) compare conven- tional survey sampling and implementation techniques with the use of online map services such as Google, Bing, and HERE maps. Modern yet basic digital technology can be integrated into the processes of preparing, implementing, and monitoring a rural household survey. Satellite-aided systematic random sampling enhanced the spatial repre- sentativeness of the village samples and entailed savings of approximately £4000 compared to conventional house- hold listing, while reducing the duration of the main survey by at least 25 %. Conclusion: This low-cost/low-tech satellite-aided survey sampling approach can be useful for student researchers and resource-constrained research projects operating in low- and middle-income contexts with high survey imple- mentation costs. While achieving transparent and efficient survey implementation at low costs, researchers aiming to adopt a similar process should be aware of the locational, technical, and logistical requirements as well as the meth- odological challenges of this strategy. Keywords: Survey, Sampling, Implementation, Rural, India, China, Google Maps, Bing Maps Survey researchers conventionally rely on manual Background household listing and mapping approaches to address the Survey research in low- and middle-income countries is problem of missing sampling frames, but this approach often subject to stifling resource, time, and administra - comes with high resource requirements. In order to man- tive constraints—especially in the case of small-scale and age costs and workload, often only segments of larger Ph.D. student research projects. A typical challenge in clusters are listed, or researchers rely on random walk this respect is the absence of sampling frames for rural approaches without constructing a sampling frame [3]. and urban household surveys in low- and middle-income Such economising solutions can result in a clustering of countries, for example in cluster random sampling responses if nearby dwelling units share similar charac- designs [1, 2]. teristics (e.g., because they are located in a slum area). Because it is difficult to locate clearly the boundaries of a cluster (or village) in these approaches, it is also possi- *Correspondence: marco.haenssgen@ndm.ox.ac.uk Centre for Tropical Medicine and Global Health, Nuffield Department ble that marginalised households at the village fringes are of Medicine, University of Oxford, Oxford, UK Full list of author information is available at the end of the article © 2015 Haenssgen. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/ publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 2 of 10 more likely to be omitted from the sample than centrally skills, equipment, or data are required is Flynn et  al. located dwelling units. [7], where, however, the sampling strategy is limited to The increasing availability of online maps, satellite a random walk. Furthermore, despite utilising satellite imagery, and digital technology can facilitate the sam- imagery and aerial maps, spatial considerations in the pling process in low- and middle-income contexts. For sampling strategy mostly focus on defining the catch - example, Wampler et  al. [4] carried out cluster random ment area. The aforementioned solutions are useful sampling in Haiti and Escamilla et al. [5] implemented a to ensure the inclusion of marginalised houses in the simple random sampling design in Malawi. In both cases, sampling frame, but they do not automatically lead to the authors use Google Earth to map the sites and mark improved spatial representation. The sub-cluster weight - eligible structures for sampling, geographical informa- ing approach of Shannon et al. [3] is a notable exception, tion system (GIS) software to extract the coordinates yet also this solution does not guarantee that the sample of the identified structures, additional software tools to is spatially stratified. An additional problem is to locate select the sample based on the list of coordinates, and selected houses using GPS units, which can make it dif- global positioning system (GPS) handhelds to upload the ficult to distinguish neighbouring dwelling units. In some sample coordinates and to locate the houses in the field. situations, the conspicuous use of GPS handhelds can Besides, Escamilla et al. [5] use GPS handhelds to define also pose a security issue. the survey site from which they draw the sample. The objective of this paper is to discuss the advantages A slightly different solution is proposed by Shan - and limitations of an alternative approach to satellite- non et  al. [3]. Although the authors use Google Earth aided survey sampling that (a) can make survey sam- as well, they also utilise commercially procured aerial pling and implementation feasible where conventional maps. From these geo-coded sources, Shannon et  al. approaches would be logistically challenging and pro- [3] sampled coordinates randomly and selected hous- hibitively expensive; that (b) does not require specialised ing structures within a 20  m radius of each location. As software skills, equipment, or data sources; and that (c) the number of structures in each location was known enables better spatial representation of rural communi- through manual enumeration, the authors applied sample ties in low- and middle-income contexts. To this end, I weights to each selected building to improve the socio- describe two household surveys in rural Rajasthan (India) geographic representativeness of the sample. In addi- and rural Gansu (China) and compare the sampling accu- tion, rather than uploading the sample coordinates to racy and costs between conventional household listing GPS handhelds, survey fieldworkers were equipped with and my satellite-aided approach. printed maps to locate the households. In contrast to other strategies utilising satellite imagery, Satellite-aided approaches are also employed to facili- the procedure presented in this paper does not require tate random walk strategies. Galway et  al. [6] describe a any specialised software knowledge nor professional technique in which they export geospatial cluster data equipment to locate households. Latest yet basic off-the- from GIS software to Google Earth, subsequently super- shelf laptop computers, low-cost smartphones, printed impose a grid and randomly choose cells therein as start- maps, and browser-based map services such as Google, ing points for random walks in their study site. Like Bing, and HERE maps are sufficient to select the sample Shannon et  al. [3], Galway et  al. [6] face a security-sen- and to facilitate implementation [11–13]. In addition, by sitive research environment and use printed maps rather ensuring better spatial representation of the survey areas, than handheld GPS units to locate the starting household the procedure presented here is likely more efficient than in the field. Not relying on any specialised software or conventional cluster sampling approaches with random data source, Flynn et al. [7] use the browser-based Google walk procedures [6, 14, 15]. The approach proposed in Maps service to randomly select the starting point for this paper is therefore intended as an addition to survey random walks in urban environments in Canada. Moreo- researchers’ methodological toolbox. ver, satellite-aided approaches can also be found in pro- gramme monitoring and disease surveillance, involving Methods specialised GIS software in nearly all cases [8–10]. I illustrate my low-cost/low-tech approach through a case These new satellite-aided survey sampling and imple - study analysis of two surveys in rural Rajasthan (India) mentation approaches face a number of difficulties. One and rural Gansu (China). In Rajasthan, administrative, problem is the reliance on specialised software and user skills, professional GPS equipment, and/or commercial The research was approved by the Oxford Department of International data sources. This can make the strategies inaccessible Development’s Departmental Research Ethics Committee in accord‑ ance with the procedures laid down by the University of Oxford for ethics or unattractive to researchers from a methodological and approval of all research involving human participants (CUREC1A/ODID financial standpoint. The only study where no specialised C1A 14‑031). Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 3 of 10 logistical, and economic conditions permitted conven- by country. Resource constraints prevented such a sur- tional survey sampling and implementation. The sur - vey design, instead requiring cluster random sampling vey context was more constrained in Gansu, where high in order to stay within the available budget by reducing labour costs, logistical challenges, and administrative survey fixed costs [17]. At the same time, spatial sam - constraints rendered the conventional approach econom- ple stratification helped to improve the effective sample ically infeasible. size, which would otherwise be diminished by cluster I describe the survey case studies to highlight that the random sampling designs [18]. The multi-stage stratified tools and techniques involved in my new approach can cluster random sample results directly from these con- replace conventional survey sampling approaches, and siderations and constraints. The sample size was limited that they require a lower level of technical sophistication to 400 respondents per site because of budget limitations than existing satellite-aided approaches. In addition, the (approx. £11,000 per country). Gansu case study outlines opportunities to appreciate The survey design therefore followed a four-step pro - spatial considerations in survey sampling that go beyond cess in order to survey 400 rural dwellers in each country: conventional and other satellite-aided approaches. In order to assess the suitability of my approach, I 1. purposive selection of representative sub‑districts in describe the survey results from Rajasthan and Gansu each study area; in terms of identification accuracy and refusal numbers, 2. random selection of 16 villages across these sub‑dis considering that satellite maps allow the researcher to tricts per country; identify houses but not necessarily households. I also 3. random selection of 25 households in each village compare the financial costs and benefits of my survey through interval sampling; and sampling approach with the expected costs of conven- 4. random selection of one respondent in each house tional household sampling across various scenarios. hold through the use of age‑order tables [19]. Results Rajasthan: village and household sampling Survey case studies The first survey was implemented in Rajasthan, with The surveys described in this section are part of a mixed assistance from the Indian Institute of Health Manage methods study on mobile phone diffusion and health - ment Research (IIHMR), Jaipur. I keep the description of care access in India and China. This research was car - village and household sampling short because very few ried out among the adult village population in selected difficulties were encountered and the survey therefore districts of Rajasthan (Udaipur, Rajsamand) and Gansu adhered to standard practice. (Baiyin, Dingxi, Lanzhou). Following a preceding qualita- Village selection in Rajasthan took place without nota tive research stage, the field sites were revisited in order ble complications. Village-level census data was available to test hypotheses derived from a qualitatively grounded and Valid International provided geographical location theoretical framework. data for all registered villages in Rajasthan. I sampled the No prior study had been conducted on the studied villages proportionally to population size and stratified phenomena (i.e., the effects of phone diffusion on health - by their distance to the nearest sub-district town (above care behaviour). Sample size decisions in this kind of or below sub-district average distance). I used the geo exploratory research are difficult because commonly graphic coordinates to calculate these distances. Two vil- used statistical power calculations are not applicable for lages of the original sample had to be replaced because phenomena whose effect size is unknown a priori [16]. heavy rainfall made them inaccessible. In addition, the survey had an explicit spatial component Following the village sampling, a team of six field inves because my analysis assumed that mobile phones influ - tigators and two supervisors relocated to the survey sites. ence people’s health behaviour differently depending on Half a team day each was required to map the village and their location. list all households in the selected areas (i.e., two segments Given that this is the first study attempting to meas - of less than 250 households in villages larger than 500 ure the effects of mobile phones on healthcare access on households). Village segmentation was required in order the micro level, sample size considerations focused on to maintain a manageable workload. This was the case in assumptions of the prevalence about phone-aided health 6 of the 16 villages, in each of which the number of seg action in the population based on the preceding qualita- ments ranged from three to six. These segments covered tive research. Ideally, this would have involved a simple random sample across my field sites with sample sizes in I defined a household as a group of people who normally live in the same excess of 500 per country in order to capture rare behav- dwelling, eat together, who may or may not be related, and who had resided iours and to permit logistic regression analyses stratified in the village for at least 6 months prior to the survey. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 4 of 10 on average 55  % of the total number of households per duplicates, or because they could not be located in Google village. Overall, the estimated population in the total or Maps or elsewhere. partial villages ranged from 161 to 1272. When the data extraction was complete, I calculated After the household listing, one team-day per village the distance between each village and their associated was required to approach, interview, and, if necessary, townships. While the main purpose of the distance cal- revisit 25 households in each village. An interval sample culation was village sample stratification, it also helped was drawn from this household list based on the total to detect outlier villages with extreme distances to the number of households in the segment and a randomly township, which led these villages to be re-entered, re- selected starting point, accounting for up to ten replace- classified to another township or district, or dropped. ment households (i.e., 40 % oversampling). After validating the geographical data, 1553 villages remained in the sampling frame. Gansu I selected 16 villages plus 32 replacement villages in Contrary to expectations, the sampling strategy could Gansu, which tended to be larger than their Rajasthan not be replicated in Gansu. Neither village-level census counterparts, ranging from estimated 481 to 3141 inhab- information nor geographical data was available—merely itants (estimates based on survey data). Two villages had a register containing village and township names could to be replaced because weather conditions rendered be accessed [21]. In addition, financial resources were them inaccessible; one village was replaced because it did insufficient to list households in the villages through not have high-resolution satellite imagery available. the local survey team. Despite the administrative and Following the village selection, the online map and resource challenges, the Gansu survey would have to be satellite image providers Google Maps and Bing Maps implemented in a comparable four-stage design as in facilitated the selection of survey households (in many Rajasthan. I describe in the following how satellite map- instances, Bing Maps provided superior-quality images). ping solutions and other digital aides compensated for Given that resource limitations were an obstacle for the lack of administrative data and facilitated sample household listing through the survey teams, I relied on selection and survey implementation. satellite imagery to replace this process. All selected vil- lages or their first replacement had high-resolution satel - Village and household sampling The village register was lite images available on Google Maps, Bing Maps, or both the only available administrative data source for the vil- (the highest available resolution was 1:670). lage selection, but it did not contain population or geo- In order to extract the satellite maps, a catchment area graphical data. While population data on the registered within a 1 km radius from the selected village centre was villages remained inaccessible, a data set containing geo- screen-capped at the highest resolution (uninhabited graphical village coordinates could be constructed using areas omitted), pasted into Microsoft PowerPoint, and Google Maps (this could have been done alternatively re-assembled to yield one high-resolution map of the with the free software Google Earth; Bing Maps and HERE catchment area (see Fig. 1 for an extract; the detail shows Maps proved inferior to Google Maps for locating Chi- the numbered houses, the selected structures, and the nese villages [11–13, 22]). I located the registered villages assigned field investigators as explained in the following through Google Maps manually, using their Chinese paragraphs). This map was further complemented with names. The process was repeated for 1736 registered lower-resolution overview maps that indicated approach townships and villages in the eight selected sub-districts roads and nearby towns for navigation. A complete area in Gansu. This procedure did not only help to extract the map of a village could thus comprise between 5 and 40 village coordinates, but it also served to validate the infor- individual high-resolution maps, containing up to 950 mation in the village register. A number of entries had to residential buildings or 380 on average per village. be removed from the list because they were urban areas, Once the complete location map was assembled, I divided the village (or settlements) into natural segments to ensure that both central and marginal areas were represented in my survey. For example, if a village com- The survey team administered paper questionnaires, which were subse‑ prised one core segment of 400 houses and an adjacent quently entered into the CSPro software [20]. yet detached group of ten residential buildings belong- In contrast to Google Maps, Google Earth would enable exporting the coordinates in the metric Universal Transverse Mercator coordinate system, ing to the village, the latter group would form a separate which does not require the researcher to apply trigonometric formulae in order to calculate distances. Whether it is more time‑ saving to manually Part of the reason for the larger village size (apart from socioeconomic dif‑ extract the coordinates from a browser‑based interface or to automati‑ ferences between the sites) is the 1  km catchment area in Gansu, meaning cally export the coordinates of pinned locations in Google Earth ultimately that surrounding settlements could be included in the sample. A primary depends on the researcher’s skills and preferences. I thank the anonymous sampling unit therefore does not necessarily correspond to one village. referee for pointing this out. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 5 of 10 Fig. 1 Excerpt of segmented and numbered village map with detail. Source: Survey material for survey fieldworkers, 2014 map data from Microsoft Corporation, DigitalGlobe, HERE segment with at least one sampled household. Segmenta- Interval sampling was chosen to ensure a better spatial tion was necessary for this survey because random sam- representation of the segment than a simple random pling would not guarantee that households at the village draw. margins are included in the sample. I selected 75 houses per village through this method, The next step in the process consisted of identifying which includes 50 replacements or two for each selected and listing all residential properties in the village. Hous- house because it is not clear beforehand whether these ing structures in the field site were homogeneous, con - buildings are locked or abandoned. The replacements— sisting of yards surrounded by two or three buildings and marked in yellow in Fig.  1—were located within the encircled with a wall and a gate. None of the villages con- intervals (i.e., an interval was selected for 75 houses, tained apartment buildings. The detailed aerial images and every third house is the first choice, followed by two therefore enabled reliable identification and consecutive replacements). The sample size per segment was assigned numbering of houses using the annotation features of according to the number of residential structures. Microsoft PowerPoint and a touchscreen laptop (i.e., This process allowed the mapping and listing of all resi - each residential structure was labelled with a handwrit- dential structures in one village by a single person within ten number; see detail in Fig.  1). Difficulties only arose 3–5  h [compare this to the approx. 32 labour hours with respect to other village structures that were later required to list all households in a village in the Rajasthan identified as schools, village councils, or factory build - survey (including investigators, supervisor, and driver)]. ings. These structures received placeholders to be As this sampling strategy could be executed without included in the sampling frame should they prove to be actually visiting the villages, and because neither sal- residential during the village visit. This was not the case, ary nor additional equipment were required for the lead however. researcher (a Ph.D. student), the process took place with- In accordance with the Rajasthan leg of the study, I out notable costs. selected the household sample through systematic ran- While I chose interval sampling stratified by village dom sampling based on the household number per seg- segments, other survey designs can superimpose a grid ment, a random starting point, and a fixed interval [23]. structure on the villages to select households [6, 24]. Where spatial considerations are not relevant for sample As an alternative to manual labelling, researchers using Google Earth selection, researchers can consider random draws from could mark and export all residential structures (this would forgo the option the sampling frame, which can be created through map of using alternative satellite map services such as Bing Maps). Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 6 of 10 annotations or through Google Earth as in Wampler et al. because the maps make clear from the very beginning [4] and Escamilla et al. [5]. which segments are relatively difficult to access, more athletic investigators could be assigned to more demand- Survey implementation Digital technology did not only ing areas. All these planning activities helped to reduce facilitate village and household selection, but it also sim- slack time when approaching and navigating the villages, plified the field investigators’ and supervisors’ work, and making the village visit as efficient as possible. it helped to locate and approach the villages and house- In order to facilitate the work of the field investiga - holds. tors, each investigator received a complete set of A4 print The detailed satellite maps enabled up-front plan - copies of the village maps. Reducing an entire area map ning of the route to each village because—unlike road to the size of A4 for the investigators is impractical for maps—major and minor roads can be easily discerned household selection because maps details become indis- even if they are not officially mapped. When approaching tinguishable. Handling maps larger than A4 is similarly the village, a low-budget smartphone with satellite map cumbersome for the field investigators (e.g., poster-sized applications (Lumia 638 with HERE Maps) enabled the at highest resolution). I instead extracted and magnified team to precisely track the current position of the vehicle the map segments relevant to each investigator, and pro- and to select the correct approach road (see Fig. 2 for an vided lower-resolution overview maps to facilitate village example of unchartered approach roads and their repre- navigation. While the resolution of the detailed village sentation on a satellite image). This saved valuable time, maps varied, these variations were of no consequence for which the team could then spend more usefully on carry- the survey implementation because the field investigators ing out the survey. received the maps in advance and could raise concerns if I used the numbered village maps to assign each of the they were unable to identify their assigned houses. six field investigators to their survey households, to plan The survey supervisors briefed and accompanied the the drop-off of the investigators in different segments of investigators where they were unsure about the exact the village, and to coordinate the final meeting and pick- location of a house (e.g., in the centre of a densely pop- up of the investigators prior to departure. This level of ulated village). A compass, their own smartphones with planning was necessary in order to complete 25 1-h sur- satellite navigation applications, and local residents fur- vey questionnaires during 1 day, considering travel times ther helped the investigators to navigate the village and of up to 3 h to the village. approach the correct household according to their maps. The detail in Fig.  1 illustrates the investigator assign- ment through numbered boxes drawn around the Sampling accuracy selected houses and their replacements (in this case, Pow- The final samples in Rajasthan and Gansu comprised 400 erPoint’s annotation features are superior to markers in adults each. In Rajasthan, 33 replacements (or up to eight Google Earth). The selected houses were assigned to the in one village) were necessary because the selected investigators in such a way that relocation time and dis- respondents were unavailable or refused participation. In tance between the interviews was minimal. In addition, Gansu, the satellite-based approach did not permit Fig. 2 Example of unchartered road and corresponding satellite map. Source: own photograph from field survey; 2014 map data from Google Inc., CNES/Astrium. Dotted red lines in b indicating perspective of photographer Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 7 of 10 judgements as to whether the selected houses were Each of these scenarios depicts the estimated costs inhabited or vacant. A total of 223 houses or up to 29 of the household listing process, including daily allow- houses per village had to be replaced because they were ances, transportation, insurance, and accommodation. vacant or locked. Manual household listing and mapping Also an international travel allowance of £50 per day is through the survey team—had it been possible—would calculated, given that the conventional strategy would have filtered out a large portion of this number by design. effectively extend the total duration of the survey. The In addition, refusal was slightly higher in Gansu as well, estimated household listing costs thus range from with a total number of 38 households, or up to four in a £2860 to £5710, depending on the difficulty of village single village. The surveys therefore required 33 replace - access. ments in Rajasthan and 261 in Gansu. In contrast, the actual expenditures in this survey Whereas houses in Gansu were often found vacant, it pertain to mobile phone equipment and operating costs was rarely the case that houses were omitted from the to implement the strategy. These expenses amounted to sampling frame or inhabited by more than one house- approximately £170. In other circumstances, research- hold. Only one house was shared by two households, and ers may have to budget their own work time spent on initially unidentified structures turned out to be schools, mapping the households and acquire an adequate lap- factories, and village councils. A notable mismatch top computer and GPS units to cross-check the field between the satellite maps and the village realities arose investigators’ household selection. These optional only in one village where an entire segment of buildings expenditures are indicated in parentheses and amount was still under construction and thus uninhabited. “On- to £1840. the-fly” updates of the village samples using the printed Compared to the assumed costs of conventional sur- maps and a recalculated sampling interval helped to solve vey sampling in Gansu, my satellite-aided approach saved this problem prior to commencing the village survey. between £850 and £5540. In any of these scenarios, satel- lite-aided household selection proves cost efficient. Given Financial cost‑benefit scenario analysis total actual survey costs in Gansu of £11,470, the cost Using satellite imagery through online map services and savings correspond to a 26 % reduction of the anticipated other digital aides enabled low-cost survey sampling and expenses of the Gansu leg of this study (given actual streamlined implementation logistics. This relaxed other - implementation costs versus the “realistic” conventional wise binding constraints of available team time and had sampling scenario; if we include the “optional” costs, this direct financial implications. But not all of the cost sav - would be a 14 % reduction). In addition, labour time sav- ings and efficiency improvements are quantifiable. For ings for household sampling correspond to a reduction of example, navigating unchartered rural roads with satel- the main survey time by at least 25 %. lite maps saved travel time (compared to using road maps The net benefit of this strategy depends on local cost and asking for directions). While the time savings were conditions and the difficulty of creating the village sam - beneficial for team performance and morale, their finan - pling frame using conventional methods. In Rajasthan, cial impact was limited because team members were paid the actual costs of conventional household listing were on a per diem basis. approximately £1250 (including staff, transport, and In order to provide an indication of the relative financial overheads), which is far below the potential expendi- costs and benefits involved in my satellite-aided sampling tures for the satellite-aided approach (i.e., up to £2010). strategy, I compare in Table  1 the savings and expendi- A satellite-aided survey sampling approach might not be tures involved my chosen strategy in Gansu versus three required on financial grounds in such situations. scenarios involving conventional household listing. The three scenarios are based on actual field experiences: Discussion Firstly, the optimistic scenario assumes that the team can The survey sampling and implementation process out - completely list two villages in 1  day, with limited over- lined above had considerable advantages over con- night stays required. The second, more realistic, scenario ventional approaches, especially in financial terms. appreciates the dispersion of the villages in rural Gansu, However, my proposed strategy is not universally appli- making it more difficult to access and map two villages cable. Researchers intending to adopt a similar approach in 1 day. In a third and yet more conservative scenario, it should be aware of the preconditions for utilising satellite would only be possible to map and list one village per day, maps and digital aides successfully, and of the remaining requiring up to eight overnight stays for the team. logistical and methodological challenges. After the data were entered, first ‑ stage error rates were 0.074  % per data field in Rajasthan and 0.035  % in Gansu. All these issues were clarified or rectified after revision. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 8 of 10 Table 1 Cost comparison of conventional versus actual sampling strategy in Gansu Assumptions Scenarios Actual Conventional household listing satellite‑aided household listing Optimistic Realistic Conservative (easy access) (medium access) (difficult access) Number of sites for listing 16 16 16 16 Number of team days per site 0.5 0.75 1 0 Number of required overnight stays 4 6 8 0 Household listing costs in planned design Optimistic Realistic Conservative Actual Item Unit Costs/unit Total costs Total costs Total costs Total costs Field investigator (×6) Daily rate p.p. £18.42 £884.21 £1326.32 £1768.42 Supervisor (×2) Daily rate p.p. £23.68 £378.95 £568.42 £757.89 Vehicle (incl. driver, fuel) Daily rate £76.84 £614.74 £922.11 £1229.47 Insurance (×8) Daily rate p.p. £0.17 £11.02 £16.53 £22.05 Team accommodation (×9) Night p.p. £15.79 £568.42 £852.63 £1136.84 Overseas travel allowance Daily rate £50.00 £400.00 £600.00 £800.00 Smartphone Unit £100.00 £100.00 Printing Total charges £40.00 £40.00 Mobile data charges Total charges £30.00 £30.00 Labour costs for satellite sampling Hourly rate £15.00 (£960.00) Touchscreen laptop Unit £700.00 (£700.00) GPS units Unit £18.00 (£180.00) Total £2857.34 £4286.01 £5714.68 £170–2010 Source: own elaboration Data based on actual expenditures. Assumed exchange rate: GBP 1.00 = CNY 9.50. Expenditures in parentheses are likely to arise in other research but did not accrue in the present study Locational, technical, and logistical prerequisites mainly in relation to missing household registers. Satel- Survey researchers aiming to adopt the sampling proce- lite-aided sampling approaches become less attractive if dure outlined in this paper have to be aware of locational, detailed administrative household lists are available (pro- technical, and logistical factors that influence the viability vided these lists are not politically influenced); where vil - and success of using satellite imagery and digital aides in lages and households are easily accessible; where labour, their work. transport, and accommodation costs are comparatively Locational factors influence the feasibility and viability low; and if the survey focuses on a very small geographi- of a satellite-aided sampling strategy. First, if up-to-date cal area (e.g., one sub-district). The strategy is better high-resolution satellite images are not available via any suited to rural household surveys that extend over a large provider (e.g., Google Maps, Bing Maps, HERE Maps), area and that involve dispersed settlements. their use is evidently ineffective. The same would apply if In terms of technical requirements, neither specialist housing structures are highly irregular, if they are indis- equipment nor dedicated software packages are needed to tinguishable from non-residential buildings, if the popu- implement this strategy (unlike e.g., [4–6]). I used off-the- lation is very mobile (e.g., nomads), or if rural dwellings shelf equipment and software (Lenovo Yoga, Microsoft generally accommodate more than one household, for Lumia 638, Microsoft Office 2013) to stratify the villages example in the case of apartment buildings. Such condi- according to their distance to the nearest town, to label tions would hamper the correct identification of house - village households on the extracted satellite maps, to cal- holds, leading to the omission of parts of the population culate the interval for household selection, and to locate from the sampling frame. Knowledge of the local living and navigate within the selected villages. Laptop models conditions prior to the household sampling, for example with the required specifications (touchscreen, 8 GB RAM, through a qualitative pre-study, can help to detect such 256  GB hard drive) currently trade at less than £500, situations. Internet-enabled smartphones with an adequate satellite Locational factors also affect the economic viability mapping application are available for less than £100. In of this sampling strategy. The strategy offers its benefits general, it is possible to implement this strategy with basic Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 9 of 10 and comparatively low-cost equipment, and without hav- villages. In order to correct for respondents’ higher prob- ing to acquire specialised software skills. ability of selection in smaller villages, the researcher can Logistically, it is important that the lead researcher has estimate the village population ex post based on the vil- at least basic training in survey sampling and is comfort- lage household count (derived from the map-based able working with maps. In order to ensure the quality of household list) and the average number of household the household sample, it is necessary to train the super- members in the surveyed dwellings. Sample weights can visors and field investigators intensively in using maps then correct for the higher chances of residents in small and to carry out briefing sessions in each village. Super - villages to be included in the sample [18]. visors should be instructed in detail about the journey Household sampling through satellite maps has at least to the village and the deployment of the investigators on two further methodological implications. Firstly, expe- the day before the village is surveyed. In order to identify riences from the Rajasthan leg of the study suggest that yet unknown structures in the village, it is also useful to manual household listing through the survey team can discuss the locational specifics with village leaders before be an opportunity to build trust with the residents before deploying the team. they are being surveyed. It is possible that this can reduce refusal rates compared to research teams who spend only Challenges 1 day in each survey village. Satellite-aided household sampling approaches come Secondly, even where housing units are homogeneous, with an idiosyncratic set of advantages and challenges. it is difficult to identify shared and abandoned houses Methodologically, map-based household selection makes through aerial images. This is a disadvantage compared to it easier to list all households in a large village rather manual household listing and mapping, which identifies than only segments thereof, which can improve the rep- households rather than houses and filters out uninhab - resentativeness of the sample [5]. In addition, the use of ited dwellings when establishing the sampling frame. If satellite maps enables spatial village stratification in order segments of the village contain a disproportionate share to ensure that marginalised dwellers are included in the of vacant houses, then the inhabitants of this segment sample as well (while I chose interval sampling stratified would be overrepresented in the village. It is therefore by village segments, other survey designs can superim- advisable to discuss the village maps with local leaders pose a grid structure on the villages to select households and update the sampling frame “on the fly” if necessary. [6, 24]). As a special form of stratification, this approach is (in Conclusions theory) at least as efficient as simple random sampling and This paper illustrated and discussed the merits, require - superior to incompletely generated household lists in dis- ments, and challenges of using map services, satellite persed villages or approaches using a random walk (pro- imagery, and basic IT equipment to facilitate sample vided that observed effects are correlated across proximate selection and survey implementation in a rural low- households [6, 14, 15]). In other words, spatial sampling income context. I argued that, despite its challenges, this approaches can help to reduce the extent of clustering in strategy can be a cost-efficient and transparent alterna - a village, which can increase the effective sample size in tive to conventional village and household listing meth- complex multi-stage sampling designs [15, 18]. ods. The cost-benefit scenario analysis underlined the Despite these advantages, satellite-aided village and usefulness of my approach in the Gansu context with household sampling also raises methodological ques- high economic and logistical constraints, whereas a con- tions. One challenge arises from the use of Google Maps ventional approach can be preferable in the Rajasthan or Google Earth for listing villages and recording their setting with low survey implementation costs and few coordinates. It is conceivable that road and satellite map administrative constraints. information does not fully correspond to official village Provided minimal investments in basic equipment, registers or that village registers are politically influenced, this approach can be replicated in other contexts where both of which can lead to the systematic exclusion of par- resources for household listing are limited, where sam- ticularly small and remote communities. Though more pling frames cannot be produced from administrative time-consuming, an alternative to using village lists is to data, and where residential structures are homogenous inspect the satellite imagery in the selected regions and and distinctive. Satellite-aided survey sampling and record the location of all identifiable settlements. implementation helped to reduce main survey time by In either case, compared to sampling through cen- one quarter and saved approximately £4100 in the pre- sus data, the methods are insensitive to population size. sent study. This and similar approaches can therefore Large and small villages are equally likely to enter the potentially improve the affordability of surveys especially sample [2], which can bias the sample towards smaller for student researchers and resource-constrained studies. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 10 of 10 Abbreviations 8. Chang A, Parrales M, Jimenez J, Sobieszczyk M, Hammer S, Copenhaver 2G: second generation; CNY: Chinese Renminbi Yuan; GB: gigabyte; GBP: D, et al. Combining Google Earth and GIS mapping technologies in a pound sterling; GIS: geographical information system; GPS: global positioning dengue surveillance system for developing countries. Int J Health Geogr. system; IIHMR: Indian Institute of Health Management Research; IT: informa- 2009;8:49. doi:10.1186/1476-072X-8-49. tion technology; RAM: random access memory; VPN: virtual private network. 9. Gammino VM, Nuhu A, Chenoweth P, Manneh F, Young RR, Sugerman DE, et al. Using geographic information systems to track polio vaccina- Author details tion team performance: pilot project report. J Infect Dis. 2014;210(suppl Centre for Tropical Medicine and Global Health, Nuffield Department of Med- 1):S98–101. doi:10.1093/infdis/jit285. icine, University of Oxford, Oxford, UK. CABDyN Complexity Centre, Saïd 10. Morland KB, Evenson KR. Obesity prevalence and the local food Business School, University of Oxford, Oxford, UK. Green Templeton College, environment. Health Place. 2009;15(2):491–5. doi:10.1016/j. University of Oxford, Oxford, UK. healthplace.2008.09.004. 11. Microsoft Corporation. Bing Maps. 2015. http://www.bing.com/maps. Acknowledgements Accessed 07 Jan 2015. This paper arises from research funded by the John Fell Oxford University 12. Google Inc. Google Maps. 2015. http://maps.google.com. Accessed 07 Press (OUP) Research Fund. Further support from the UK Economic and Social Jan 2015. Research Council, the Scatcherd European Scholarship, the Oxford Depart- 13. HERE Maps BV. HERE Maps. 2015. http://www.here.com. Accessed 07 Jan ment of International Development, the University of Oxford Vice-Chancellors’ 2015. Fund and Hertford College is gratefully acknowledged. I thank Proochista 14. Delmelle E. Spatial sampling. In: Fotheringham AS, Rogerson PA, editors. Ariana, Ernest Guevarra, Felix Reed-Tsochas, and Xiaolan Fu and participants at The SAGE handbook of spatial analysis. London: Sage; 2009. p. 182–207. the 6th Conference of the European Survey Research Association for helpful 15. Working group for mortality estimation in emergencies. Wanted: stud- discussions in relation to survey design and implementation. I further received ies on mortality estimation methods for humanitarian emergencies, excellent research assistance from IIHMR in Rajasthan, especially SD Gupta, suggestions for future research. Emerg Themes Epidemiol. 2007;4:9. Nutan Jain, Arindam Das, Jagjeet Prasad Singh, Vidya Bhushan Tripathi, and doi:10.1186/1742-7622-4-9. Matadin Sharma; and from Liu Xingrong and the School of Public Health at 16. Fowler FJ. Survey research methods. 4th ed. Thousand Oaks: Sage Publi- Lanzhou University, Li Jian, and Wang Wei in Gansu. cations; 2009. 17. Groves RM, Fowler FJ, Couper MP, Lepowski JM, Singer E, Tourangeau R. Competing interests Survey methodology. In: Wiley series in survey methodology. 2nd ed. The author declares that he has no competing interests. Chichester: Wiley; 2009. 18. Heeringa S, West BT, Berglund PA. Applied survey data analysis. In: Chap- Received: 24 June 2015 Accepted: 4 December 2015 man & Hall/CRC statistics in the social and behavioral sciences series. Boca Raton: Chapman & Hall/CRC; 2010. 19. Gaziano C. Comparative analysis of within-household respondent selec- tion techniques. Public Opinion Q. 2005;69(1):124–57. doi:10.1093/poq/ nfi006. 20. United States Census Bureau, ICF Macro. Census and survey processing References system (CSPro) 5.0. Washington: United States Census Bureau; 2014. 1. Overton J, van Diermen P. Using quantitative techniques. In: Scheyvens 21. China Standard Press. 中华人民共和国行政区划代码资料手册 [Peo- R, Storey D, editors. Development fieldwork: a practical guide. London: ple’s Republic of China administrative division code book]. Beijing: China Sage; 2003. p. 37–56. Standard Press; 2010. 2. Zarkovich SS. Some problems of sampling work in underdeveloped 22. Google Inc. Google Earth. Mountain View: Google Inc; 2015. countries. In: Bulmer M, Warwick DP, editors. Social research in developing 23. Daniel J. Sampling essentials: practical guidelines for making sampling countries: surveys and censuses in the third world. London: UCL Press; choices. Thousand Oaks: Sage; 2012. 1993. p. 103–10. 24. Grais R, Rose A, Guthmann J-P. Don’t spin the pen: two alternative meth- 3. Shannon H, Hutson R, Kolbe A, Stringer B, Haines T. Choosing a survey ods for second-stage sampling in urban cluster surveys. Emerg Themes sample when data on the population are limited: a method using global Epidemiol. 2007;4:8. doi:10.1186/1742-7622-4-8. positioning systems and aerial and satellite photographs. Emerg Themes Epidemiol. 2012;9:5. doi:10.1186/1742-7622-9-5. 4. Wampler P, Rediske R, Molla A. Using ArcMap, Google Earth, and global positioning systems to select and locate random households in rural Haiti. Int J Health Geogr. 2013;12:3. doi:10.1186/1476-072X-12-3. 5. Escamilla V, Emch M, Dandalo L, Miller WC, Martinson F, Hoffman I. Sam- pling at community level using satellite imagery and geographical analy- sis. Bull World Health Organ. 2014;92(9):690–4. doi:10.2471/BLT.14.140756. 6. Galway L, Bell N, Sae A, Hagopian A, Burnham G, Flaxman A, et al. A two- stage cluster sampling method using gridded population data, a GIS, and Google Earth ( TM) imagery in a population-based mortality survey in Submit your next manuscript to BioMed Central Iraq. Int J Health Geogr. 2012;11:12. doi:10.1186/1476-072X-11-12. and we will help you at every step: 7. Flynn A, Tremblay PF, Rehm J, Wells S. A modified random walk door-to- door recruitment strategy for collecting social and biological data relat- • We accept pre-submission inquiries ing to mental health, substance use/addictions and violence problems in • Our selector tool helps you to find the most relevant journal a Canadian community. Int J Alcohol Drug Res. 2013;2(2):7–16. • We provide round the clock customer support • Convenient online submission • Thorough peer review • Inclusion in PubMed and all major indexing services • Maximum visibility for your research Submit your manuscript at www.biomedcentral.com/submit http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Emerging Themes in Epidemiology Springer Journals

Satellite-aided survey sampling and implementation in low- and middle-income contexts: a low-cost/low-tech alternative

Emerging Themes in Epidemiology , Volume 12 (1) – Dec 23, 2015

Loading next page...
 
/lp/springer-journals/satellite-aided-survey-sampling-and-implementation-in-low-and-middle-xZL1GAuFSY

References (34)

Publisher
Springer Journals
Copyright
Copyright © 2015 by Haenssgen.
Subject
Medicine & Public Health; Epidemiology
eISSN
1742-7622
DOI
10.1186/s12982-015-0041-8
pmid
26705402
Publisher site
See Article on Publisher Site

Abstract

Background: The increasing availability of online maps, satellite imagery, and digital technology can ease common constraints of survey sampling in low- and middle-income countries. However, existing approaches require special- ised software and user skills, professional GPS equipment, and/or commercial data sources; they tend to neglect spatial sampling considerations when using satellite maps; and they continue to face implementation challenges analogous to conventional survey implementation methods. This paper presents an alternative way of utilising satel- lite maps and digital aides that aims to address these challenges. Results: The case studies of two rural household surveys in Rajasthan (India) and Gansu (China) compare conven- tional survey sampling and implementation techniques with the use of online map services such as Google, Bing, and HERE maps. Modern yet basic digital technology can be integrated into the processes of preparing, implementing, and monitoring a rural household survey. Satellite-aided systematic random sampling enhanced the spatial repre- sentativeness of the village samples and entailed savings of approximately £4000 compared to conventional house- hold listing, while reducing the duration of the main survey by at least 25 %. Conclusion: This low-cost/low-tech satellite-aided survey sampling approach can be useful for student researchers and resource-constrained research projects operating in low- and middle-income contexts with high survey imple- mentation costs. While achieving transparent and efficient survey implementation at low costs, researchers aiming to adopt a similar process should be aware of the locational, technical, and logistical requirements as well as the meth- odological challenges of this strategy. Keywords: Survey, Sampling, Implementation, Rural, India, China, Google Maps, Bing Maps Survey researchers conventionally rely on manual Background household listing and mapping approaches to address the Survey research in low- and middle-income countries is problem of missing sampling frames, but this approach often subject to stifling resource, time, and administra - comes with high resource requirements. In order to man- tive constraints—especially in the case of small-scale and age costs and workload, often only segments of larger Ph.D. student research projects. A typical challenge in clusters are listed, or researchers rely on random walk this respect is the absence of sampling frames for rural approaches without constructing a sampling frame [3]. and urban household surveys in low- and middle-income Such economising solutions can result in a clustering of countries, for example in cluster random sampling responses if nearby dwelling units share similar charac- designs [1, 2]. teristics (e.g., because they are located in a slum area). Because it is difficult to locate clearly the boundaries of a cluster (or village) in these approaches, it is also possi- *Correspondence: marco.haenssgen@ndm.ox.ac.uk Centre for Tropical Medicine and Global Health, Nuffield Department ble that marginalised households at the village fringes are of Medicine, University of Oxford, Oxford, UK Full list of author information is available at the end of the article © 2015 Haenssgen. This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/ publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 2 of 10 more likely to be omitted from the sample than centrally skills, equipment, or data are required is Flynn et  al. located dwelling units. [7], where, however, the sampling strategy is limited to The increasing availability of online maps, satellite a random walk. Furthermore, despite utilising satellite imagery, and digital technology can facilitate the sam- imagery and aerial maps, spatial considerations in the pling process in low- and middle-income contexts. For sampling strategy mostly focus on defining the catch - example, Wampler et  al. [4] carried out cluster random ment area. The aforementioned solutions are useful sampling in Haiti and Escamilla et al. [5] implemented a to ensure the inclusion of marginalised houses in the simple random sampling design in Malawi. In both cases, sampling frame, but they do not automatically lead to the authors use Google Earth to map the sites and mark improved spatial representation. The sub-cluster weight - eligible structures for sampling, geographical informa- ing approach of Shannon et al. [3] is a notable exception, tion system (GIS) software to extract the coordinates yet also this solution does not guarantee that the sample of the identified structures, additional software tools to is spatially stratified. An additional problem is to locate select the sample based on the list of coordinates, and selected houses using GPS units, which can make it dif- global positioning system (GPS) handhelds to upload the ficult to distinguish neighbouring dwelling units. In some sample coordinates and to locate the houses in the field. situations, the conspicuous use of GPS handhelds can Besides, Escamilla et al. [5] use GPS handhelds to define also pose a security issue. the survey site from which they draw the sample. The objective of this paper is to discuss the advantages A slightly different solution is proposed by Shan - and limitations of an alternative approach to satellite- non et  al. [3]. Although the authors use Google Earth aided survey sampling that (a) can make survey sam- as well, they also utilise commercially procured aerial pling and implementation feasible where conventional maps. From these geo-coded sources, Shannon et  al. approaches would be logistically challenging and pro- [3] sampled coordinates randomly and selected hous- hibitively expensive; that (b) does not require specialised ing structures within a 20  m radius of each location. As software skills, equipment, or data sources; and that (c) the number of structures in each location was known enables better spatial representation of rural communi- through manual enumeration, the authors applied sample ties in low- and middle-income contexts. To this end, I weights to each selected building to improve the socio- describe two household surveys in rural Rajasthan (India) geographic representativeness of the sample. In addi- and rural Gansu (China) and compare the sampling accu- tion, rather than uploading the sample coordinates to racy and costs between conventional household listing GPS handhelds, survey fieldworkers were equipped with and my satellite-aided approach. printed maps to locate the households. In contrast to other strategies utilising satellite imagery, Satellite-aided approaches are also employed to facili- the procedure presented in this paper does not require tate random walk strategies. Galway et  al. [6] describe a any specialised software knowledge nor professional technique in which they export geospatial cluster data equipment to locate households. Latest yet basic off-the- from GIS software to Google Earth, subsequently super- shelf laptop computers, low-cost smartphones, printed impose a grid and randomly choose cells therein as start- maps, and browser-based map services such as Google, ing points for random walks in their study site. Like Bing, and HERE maps are sufficient to select the sample Shannon et  al. [3], Galway et  al. [6] face a security-sen- and to facilitate implementation [11–13]. In addition, by sitive research environment and use printed maps rather ensuring better spatial representation of the survey areas, than handheld GPS units to locate the starting household the procedure presented here is likely more efficient than in the field. Not relying on any specialised software or conventional cluster sampling approaches with random data source, Flynn et al. [7] use the browser-based Google walk procedures [6, 14, 15]. The approach proposed in Maps service to randomly select the starting point for this paper is therefore intended as an addition to survey random walks in urban environments in Canada. Moreo- researchers’ methodological toolbox. ver, satellite-aided approaches can also be found in pro- gramme monitoring and disease surveillance, involving Methods specialised GIS software in nearly all cases [8–10]. I illustrate my low-cost/low-tech approach through a case These new satellite-aided survey sampling and imple - study analysis of two surveys in rural Rajasthan (India) mentation approaches face a number of difficulties. One and rural Gansu (China). In Rajasthan, administrative, problem is the reliance on specialised software and user skills, professional GPS equipment, and/or commercial The research was approved by the Oxford Department of International data sources. This can make the strategies inaccessible Development’s Departmental Research Ethics Committee in accord‑ ance with the procedures laid down by the University of Oxford for ethics or unattractive to researchers from a methodological and approval of all research involving human participants (CUREC1A/ODID financial standpoint. The only study where no specialised C1A 14‑031). Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 3 of 10 logistical, and economic conditions permitted conven- by country. Resource constraints prevented such a sur- tional survey sampling and implementation. The sur - vey design, instead requiring cluster random sampling vey context was more constrained in Gansu, where high in order to stay within the available budget by reducing labour costs, logistical challenges, and administrative survey fixed costs [17]. At the same time, spatial sam - constraints rendered the conventional approach econom- ple stratification helped to improve the effective sample ically infeasible. size, which would otherwise be diminished by cluster I describe the survey case studies to highlight that the random sampling designs [18]. The multi-stage stratified tools and techniques involved in my new approach can cluster random sample results directly from these con- replace conventional survey sampling approaches, and siderations and constraints. The sample size was limited that they require a lower level of technical sophistication to 400 respondents per site because of budget limitations than existing satellite-aided approaches. In addition, the (approx. £11,000 per country). Gansu case study outlines opportunities to appreciate The survey design therefore followed a four-step pro - spatial considerations in survey sampling that go beyond cess in order to survey 400 rural dwellers in each country: conventional and other satellite-aided approaches. In order to assess the suitability of my approach, I 1. purposive selection of representative sub‑districts in describe the survey results from Rajasthan and Gansu each study area; in terms of identification accuracy and refusal numbers, 2. random selection of 16 villages across these sub‑dis considering that satellite maps allow the researcher to tricts per country; identify houses but not necessarily households. I also 3. random selection of 25 households in each village compare the financial costs and benefits of my survey through interval sampling; and sampling approach with the expected costs of conven- 4. random selection of one respondent in each house tional household sampling across various scenarios. hold through the use of age‑order tables [19]. Results Rajasthan: village and household sampling Survey case studies The first survey was implemented in Rajasthan, with The surveys described in this section are part of a mixed assistance from the Indian Institute of Health Manage methods study on mobile phone diffusion and health - ment Research (IIHMR), Jaipur. I keep the description of care access in India and China. This research was car - village and household sampling short because very few ried out among the adult village population in selected difficulties were encountered and the survey therefore districts of Rajasthan (Udaipur, Rajsamand) and Gansu adhered to standard practice. (Baiyin, Dingxi, Lanzhou). Following a preceding qualita- Village selection in Rajasthan took place without nota tive research stage, the field sites were revisited in order ble complications. Village-level census data was available to test hypotheses derived from a qualitatively grounded and Valid International provided geographical location theoretical framework. data for all registered villages in Rajasthan. I sampled the No prior study had been conducted on the studied villages proportionally to population size and stratified phenomena (i.e., the effects of phone diffusion on health - by their distance to the nearest sub-district town (above care behaviour). Sample size decisions in this kind of or below sub-district average distance). I used the geo exploratory research are difficult because commonly graphic coordinates to calculate these distances. Two vil- used statistical power calculations are not applicable for lages of the original sample had to be replaced because phenomena whose effect size is unknown a priori [16]. heavy rainfall made them inaccessible. In addition, the survey had an explicit spatial component Following the village sampling, a team of six field inves because my analysis assumed that mobile phones influ - tigators and two supervisors relocated to the survey sites. ence people’s health behaviour differently depending on Half a team day each was required to map the village and their location. list all households in the selected areas (i.e., two segments Given that this is the first study attempting to meas - of less than 250 households in villages larger than 500 ure the effects of mobile phones on healthcare access on households). Village segmentation was required in order the micro level, sample size considerations focused on to maintain a manageable workload. This was the case in assumptions of the prevalence about phone-aided health 6 of the 16 villages, in each of which the number of seg action in the population based on the preceding qualita- ments ranged from three to six. These segments covered tive research. Ideally, this would have involved a simple random sample across my field sites with sample sizes in I defined a household as a group of people who normally live in the same excess of 500 per country in order to capture rare behav- dwelling, eat together, who may or may not be related, and who had resided iours and to permit logistic regression analyses stratified in the village for at least 6 months prior to the survey. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 4 of 10 on average 55  % of the total number of households per duplicates, or because they could not be located in Google village. Overall, the estimated population in the total or Maps or elsewhere. partial villages ranged from 161 to 1272. When the data extraction was complete, I calculated After the household listing, one team-day per village the distance between each village and their associated was required to approach, interview, and, if necessary, townships. While the main purpose of the distance cal- revisit 25 households in each village. An interval sample culation was village sample stratification, it also helped was drawn from this household list based on the total to detect outlier villages with extreme distances to the number of households in the segment and a randomly township, which led these villages to be re-entered, re- selected starting point, accounting for up to ten replace- classified to another township or district, or dropped. ment households (i.e., 40 % oversampling). After validating the geographical data, 1553 villages remained in the sampling frame. Gansu I selected 16 villages plus 32 replacement villages in Contrary to expectations, the sampling strategy could Gansu, which tended to be larger than their Rajasthan not be replicated in Gansu. Neither village-level census counterparts, ranging from estimated 481 to 3141 inhab- information nor geographical data was available—merely itants (estimates based on survey data). Two villages had a register containing village and township names could to be replaced because weather conditions rendered be accessed [21]. In addition, financial resources were them inaccessible; one village was replaced because it did insufficient to list households in the villages through not have high-resolution satellite imagery available. the local survey team. Despite the administrative and Following the village selection, the online map and resource challenges, the Gansu survey would have to be satellite image providers Google Maps and Bing Maps implemented in a comparable four-stage design as in facilitated the selection of survey households (in many Rajasthan. I describe in the following how satellite map- instances, Bing Maps provided superior-quality images). ping solutions and other digital aides compensated for Given that resource limitations were an obstacle for the lack of administrative data and facilitated sample household listing through the survey teams, I relied on selection and survey implementation. satellite imagery to replace this process. All selected vil- lages or their first replacement had high-resolution satel - Village and household sampling The village register was lite images available on Google Maps, Bing Maps, or both the only available administrative data source for the vil- (the highest available resolution was 1:670). lage selection, but it did not contain population or geo- In order to extract the satellite maps, a catchment area graphical data. While population data on the registered within a 1 km radius from the selected village centre was villages remained inaccessible, a data set containing geo- screen-capped at the highest resolution (uninhabited graphical village coordinates could be constructed using areas omitted), pasted into Microsoft PowerPoint, and Google Maps (this could have been done alternatively re-assembled to yield one high-resolution map of the with the free software Google Earth; Bing Maps and HERE catchment area (see Fig. 1 for an extract; the detail shows Maps proved inferior to Google Maps for locating Chi- the numbered houses, the selected structures, and the nese villages [11–13, 22]). I located the registered villages assigned field investigators as explained in the following through Google Maps manually, using their Chinese paragraphs). This map was further complemented with names. The process was repeated for 1736 registered lower-resolution overview maps that indicated approach townships and villages in the eight selected sub-districts roads and nearby towns for navigation. A complete area in Gansu. This procedure did not only help to extract the map of a village could thus comprise between 5 and 40 village coordinates, but it also served to validate the infor- individual high-resolution maps, containing up to 950 mation in the village register. A number of entries had to residential buildings or 380 on average per village. be removed from the list because they were urban areas, Once the complete location map was assembled, I divided the village (or settlements) into natural segments to ensure that both central and marginal areas were represented in my survey. For example, if a village com- The survey team administered paper questionnaires, which were subse‑ prised one core segment of 400 houses and an adjacent quently entered into the CSPro software [20]. yet detached group of ten residential buildings belong- In contrast to Google Maps, Google Earth would enable exporting the coordinates in the metric Universal Transverse Mercator coordinate system, ing to the village, the latter group would form a separate which does not require the researcher to apply trigonometric formulae in order to calculate distances. Whether it is more time‑ saving to manually Part of the reason for the larger village size (apart from socioeconomic dif‑ extract the coordinates from a browser‑based interface or to automati‑ ferences between the sites) is the 1  km catchment area in Gansu, meaning cally export the coordinates of pinned locations in Google Earth ultimately that surrounding settlements could be included in the sample. A primary depends on the researcher’s skills and preferences. I thank the anonymous sampling unit therefore does not necessarily correspond to one village. referee for pointing this out. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 5 of 10 Fig. 1 Excerpt of segmented and numbered village map with detail. Source: Survey material for survey fieldworkers, 2014 map data from Microsoft Corporation, DigitalGlobe, HERE segment with at least one sampled household. Segmenta- Interval sampling was chosen to ensure a better spatial tion was necessary for this survey because random sam- representation of the segment than a simple random pling would not guarantee that households at the village draw. margins are included in the sample. I selected 75 houses per village through this method, The next step in the process consisted of identifying which includes 50 replacements or two for each selected and listing all residential properties in the village. Hous- house because it is not clear beforehand whether these ing structures in the field site were homogeneous, con - buildings are locked or abandoned. The replacements— sisting of yards surrounded by two or three buildings and marked in yellow in Fig.  1—were located within the encircled with a wall and a gate. None of the villages con- intervals (i.e., an interval was selected for 75 houses, tained apartment buildings. The detailed aerial images and every third house is the first choice, followed by two therefore enabled reliable identification and consecutive replacements). The sample size per segment was assigned numbering of houses using the annotation features of according to the number of residential structures. Microsoft PowerPoint and a touchscreen laptop (i.e., This process allowed the mapping and listing of all resi - each residential structure was labelled with a handwrit- dential structures in one village by a single person within ten number; see detail in Fig.  1). Difficulties only arose 3–5  h [compare this to the approx. 32 labour hours with respect to other village structures that were later required to list all households in a village in the Rajasthan identified as schools, village councils, or factory build - survey (including investigators, supervisor, and driver)]. ings. These structures received placeholders to be As this sampling strategy could be executed without included in the sampling frame should they prove to be actually visiting the villages, and because neither sal- residential during the village visit. This was not the case, ary nor additional equipment were required for the lead however. researcher (a Ph.D. student), the process took place with- In accordance with the Rajasthan leg of the study, I out notable costs. selected the household sample through systematic ran- While I chose interval sampling stratified by village dom sampling based on the household number per seg- segments, other survey designs can superimpose a grid ment, a random starting point, and a fixed interval [23]. structure on the villages to select households [6, 24]. Where spatial considerations are not relevant for sample As an alternative to manual labelling, researchers using Google Earth selection, researchers can consider random draws from could mark and export all residential structures (this would forgo the option the sampling frame, which can be created through map of using alternative satellite map services such as Bing Maps). Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 6 of 10 annotations or through Google Earth as in Wampler et al. because the maps make clear from the very beginning [4] and Escamilla et al. [5]. which segments are relatively difficult to access, more athletic investigators could be assigned to more demand- Survey implementation Digital technology did not only ing areas. All these planning activities helped to reduce facilitate village and household selection, but it also sim- slack time when approaching and navigating the villages, plified the field investigators’ and supervisors’ work, and making the village visit as efficient as possible. it helped to locate and approach the villages and house- In order to facilitate the work of the field investiga - holds. tors, each investigator received a complete set of A4 print The detailed satellite maps enabled up-front plan - copies of the village maps. Reducing an entire area map ning of the route to each village because—unlike road to the size of A4 for the investigators is impractical for maps—major and minor roads can be easily discerned household selection because maps details become indis- even if they are not officially mapped. When approaching tinguishable. Handling maps larger than A4 is similarly the village, a low-budget smartphone with satellite map cumbersome for the field investigators (e.g., poster-sized applications (Lumia 638 with HERE Maps) enabled the at highest resolution). I instead extracted and magnified team to precisely track the current position of the vehicle the map segments relevant to each investigator, and pro- and to select the correct approach road (see Fig. 2 for an vided lower-resolution overview maps to facilitate village example of unchartered approach roads and their repre- navigation. While the resolution of the detailed village sentation on a satellite image). This saved valuable time, maps varied, these variations were of no consequence for which the team could then spend more usefully on carry- the survey implementation because the field investigators ing out the survey. received the maps in advance and could raise concerns if I used the numbered village maps to assign each of the they were unable to identify their assigned houses. six field investigators to their survey households, to plan The survey supervisors briefed and accompanied the the drop-off of the investigators in different segments of investigators where they were unsure about the exact the village, and to coordinate the final meeting and pick- location of a house (e.g., in the centre of a densely pop- up of the investigators prior to departure. This level of ulated village). A compass, their own smartphones with planning was necessary in order to complete 25 1-h sur- satellite navigation applications, and local residents fur- vey questionnaires during 1 day, considering travel times ther helped the investigators to navigate the village and of up to 3 h to the village. approach the correct household according to their maps. The detail in Fig.  1 illustrates the investigator assign- ment through numbered boxes drawn around the Sampling accuracy selected houses and their replacements (in this case, Pow- The final samples in Rajasthan and Gansu comprised 400 erPoint’s annotation features are superior to markers in adults each. In Rajasthan, 33 replacements (or up to eight Google Earth). The selected houses were assigned to the in one village) were necessary because the selected investigators in such a way that relocation time and dis- respondents were unavailable or refused participation. In tance between the interviews was minimal. In addition, Gansu, the satellite-based approach did not permit Fig. 2 Example of unchartered road and corresponding satellite map. Source: own photograph from field survey; 2014 map data from Google Inc., CNES/Astrium. Dotted red lines in b indicating perspective of photographer Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 7 of 10 judgements as to whether the selected houses were Each of these scenarios depicts the estimated costs inhabited or vacant. A total of 223 houses or up to 29 of the household listing process, including daily allow- houses per village had to be replaced because they were ances, transportation, insurance, and accommodation. vacant or locked. Manual household listing and mapping Also an international travel allowance of £50 per day is through the survey team—had it been possible—would calculated, given that the conventional strategy would have filtered out a large portion of this number by design. effectively extend the total duration of the survey. The In addition, refusal was slightly higher in Gansu as well, estimated household listing costs thus range from with a total number of 38 households, or up to four in a £2860 to £5710, depending on the difficulty of village single village. The surveys therefore required 33 replace - access. ments in Rajasthan and 261 in Gansu. In contrast, the actual expenditures in this survey Whereas houses in Gansu were often found vacant, it pertain to mobile phone equipment and operating costs was rarely the case that houses were omitted from the to implement the strategy. These expenses amounted to sampling frame or inhabited by more than one house- approximately £170. In other circumstances, research- hold. Only one house was shared by two households, and ers may have to budget their own work time spent on initially unidentified structures turned out to be schools, mapping the households and acquire an adequate lap- factories, and village councils. A notable mismatch top computer and GPS units to cross-check the field between the satellite maps and the village realities arose investigators’ household selection. These optional only in one village where an entire segment of buildings expenditures are indicated in parentheses and amount was still under construction and thus uninhabited. “On- to £1840. the-fly” updates of the village samples using the printed Compared to the assumed costs of conventional sur- maps and a recalculated sampling interval helped to solve vey sampling in Gansu, my satellite-aided approach saved this problem prior to commencing the village survey. between £850 and £5540. In any of these scenarios, satel- lite-aided household selection proves cost efficient. Given Financial cost‑benefit scenario analysis total actual survey costs in Gansu of £11,470, the cost Using satellite imagery through online map services and savings correspond to a 26 % reduction of the anticipated other digital aides enabled low-cost survey sampling and expenses of the Gansu leg of this study (given actual streamlined implementation logistics. This relaxed other - implementation costs versus the “realistic” conventional wise binding constraints of available team time and had sampling scenario; if we include the “optional” costs, this direct financial implications. But not all of the cost sav - would be a 14 % reduction). In addition, labour time sav- ings and efficiency improvements are quantifiable. For ings for household sampling correspond to a reduction of example, navigating unchartered rural roads with satel- the main survey time by at least 25 %. lite maps saved travel time (compared to using road maps The net benefit of this strategy depends on local cost and asking for directions). While the time savings were conditions and the difficulty of creating the village sam - beneficial for team performance and morale, their finan - pling frame using conventional methods. In Rajasthan, cial impact was limited because team members were paid the actual costs of conventional household listing were on a per diem basis. approximately £1250 (including staff, transport, and In order to provide an indication of the relative financial overheads), which is far below the potential expendi- costs and benefits involved in my satellite-aided sampling tures for the satellite-aided approach (i.e., up to £2010). strategy, I compare in Table  1 the savings and expendi- A satellite-aided survey sampling approach might not be tures involved my chosen strategy in Gansu versus three required on financial grounds in such situations. scenarios involving conventional household listing. The three scenarios are based on actual field experiences: Discussion Firstly, the optimistic scenario assumes that the team can The survey sampling and implementation process out - completely list two villages in 1  day, with limited over- lined above had considerable advantages over con- night stays required. The second, more realistic, scenario ventional approaches, especially in financial terms. appreciates the dispersion of the villages in rural Gansu, However, my proposed strategy is not universally appli- making it more difficult to access and map two villages cable. Researchers intending to adopt a similar approach in 1 day. In a third and yet more conservative scenario, it should be aware of the preconditions for utilising satellite would only be possible to map and list one village per day, maps and digital aides successfully, and of the remaining requiring up to eight overnight stays for the team. logistical and methodological challenges. After the data were entered, first ‑ stage error rates were 0.074  % per data field in Rajasthan and 0.035  % in Gansu. All these issues were clarified or rectified after revision. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 8 of 10 Table 1 Cost comparison of conventional versus actual sampling strategy in Gansu Assumptions Scenarios Actual Conventional household listing satellite‑aided household listing Optimistic Realistic Conservative (easy access) (medium access) (difficult access) Number of sites for listing 16 16 16 16 Number of team days per site 0.5 0.75 1 0 Number of required overnight stays 4 6 8 0 Household listing costs in planned design Optimistic Realistic Conservative Actual Item Unit Costs/unit Total costs Total costs Total costs Total costs Field investigator (×6) Daily rate p.p. £18.42 £884.21 £1326.32 £1768.42 Supervisor (×2) Daily rate p.p. £23.68 £378.95 £568.42 £757.89 Vehicle (incl. driver, fuel) Daily rate £76.84 £614.74 £922.11 £1229.47 Insurance (×8) Daily rate p.p. £0.17 £11.02 £16.53 £22.05 Team accommodation (×9) Night p.p. £15.79 £568.42 £852.63 £1136.84 Overseas travel allowance Daily rate £50.00 £400.00 £600.00 £800.00 Smartphone Unit £100.00 £100.00 Printing Total charges £40.00 £40.00 Mobile data charges Total charges £30.00 £30.00 Labour costs for satellite sampling Hourly rate £15.00 (£960.00) Touchscreen laptop Unit £700.00 (£700.00) GPS units Unit £18.00 (£180.00) Total £2857.34 £4286.01 £5714.68 £170–2010 Source: own elaboration Data based on actual expenditures. Assumed exchange rate: GBP 1.00 = CNY 9.50. Expenditures in parentheses are likely to arise in other research but did not accrue in the present study Locational, technical, and logistical prerequisites mainly in relation to missing household registers. Satel- Survey researchers aiming to adopt the sampling proce- lite-aided sampling approaches become less attractive if dure outlined in this paper have to be aware of locational, detailed administrative household lists are available (pro- technical, and logistical factors that influence the viability vided these lists are not politically influenced); where vil - and success of using satellite imagery and digital aides in lages and households are easily accessible; where labour, their work. transport, and accommodation costs are comparatively Locational factors influence the feasibility and viability low; and if the survey focuses on a very small geographi- of a satellite-aided sampling strategy. First, if up-to-date cal area (e.g., one sub-district). The strategy is better high-resolution satellite images are not available via any suited to rural household surveys that extend over a large provider (e.g., Google Maps, Bing Maps, HERE Maps), area and that involve dispersed settlements. their use is evidently ineffective. The same would apply if In terms of technical requirements, neither specialist housing structures are highly irregular, if they are indis- equipment nor dedicated software packages are needed to tinguishable from non-residential buildings, if the popu- implement this strategy (unlike e.g., [4–6]). I used off-the- lation is very mobile (e.g., nomads), or if rural dwellings shelf equipment and software (Lenovo Yoga, Microsoft generally accommodate more than one household, for Lumia 638, Microsoft Office 2013) to stratify the villages example in the case of apartment buildings. Such condi- according to their distance to the nearest town, to label tions would hamper the correct identification of house - village households on the extracted satellite maps, to cal- holds, leading to the omission of parts of the population culate the interval for household selection, and to locate from the sampling frame. Knowledge of the local living and navigate within the selected villages. Laptop models conditions prior to the household sampling, for example with the required specifications (touchscreen, 8 GB RAM, through a qualitative pre-study, can help to detect such 256  GB hard drive) currently trade at less than £500, situations. Internet-enabled smartphones with an adequate satellite Locational factors also affect the economic viability mapping application are available for less than £100. In of this sampling strategy. The strategy offers its benefits general, it is possible to implement this strategy with basic Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 9 of 10 and comparatively low-cost equipment, and without hav- villages. In order to correct for respondents’ higher prob- ing to acquire specialised software skills. ability of selection in smaller villages, the researcher can Logistically, it is important that the lead researcher has estimate the village population ex post based on the vil- at least basic training in survey sampling and is comfort- lage household count (derived from the map-based able working with maps. In order to ensure the quality of household list) and the average number of household the household sample, it is necessary to train the super- members in the surveyed dwellings. Sample weights can visors and field investigators intensively in using maps then correct for the higher chances of residents in small and to carry out briefing sessions in each village. Super - villages to be included in the sample [18]. visors should be instructed in detail about the journey Household sampling through satellite maps has at least to the village and the deployment of the investigators on two further methodological implications. Firstly, expe- the day before the village is surveyed. In order to identify riences from the Rajasthan leg of the study suggest that yet unknown structures in the village, it is also useful to manual household listing through the survey team can discuss the locational specifics with village leaders before be an opportunity to build trust with the residents before deploying the team. they are being surveyed. It is possible that this can reduce refusal rates compared to research teams who spend only Challenges 1 day in each survey village. Satellite-aided household sampling approaches come Secondly, even where housing units are homogeneous, with an idiosyncratic set of advantages and challenges. it is difficult to identify shared and abandoned houses Methodologically, map-based household selection makes through aerial images. This is a disadvantage compared to it easier to list all households in a large village rather manual household listing and mapping, which identifies than only segments thereof, which can improve the rep- households rather than houses and filters out uninhab - resentativeness of the sample [5]. In addition, the use of ited dwellings when establishing the sampling frame. If satellite maps enables spatial village stratification in order segments of the village contain a disproportionate share to ensure that marginalised dwellers are included in the of vacant houses, then the inhabitants of this segment sample as well (while I chose interval sampling stratified would be overrepresented in the village. It is therefore by village segments, other survey designs can superim- advisable to discuss the village maps with local leaders pose a grid structure on the villages to select households and update the sampling frame “on the fly” if necessary. [6, 24]). As a special form of stratification, this approach is (in Conclusions theory) at least as efficient as simple random sampling and This paper illustrated and discussed the merits, require - superior to incompletely generated household lists in dis- ments, and challenges of using map services, satellite persed villages or approaches using a random walk (pro- imagery, and basic IT equipment to facilitate sample vided that observed effects are correlated across proximate selection and survey implementation in a rural low- households [6, 14, 15]). In other words, spatial sampling income context. I argued that, despite its challenges, this approaches can help to reduce the extent of clustering in strategy can be a cost-efficient and transparent alterna - a village, which can increase the effective sample size in tive to conventional village and household listing meth- complex multi-stage sampling designs [15, 18]. ods. The cost-benefit scenario analysis underlined the Despite these advantages, satellite-aided village and usefulness of my approach in the Gansu context with household sampling also raises methodological ques- high economic and logistical constraints, whereas a con- tions. One challenge arises from the use of Google Maps ventional approach can be preferable in the Rajasthan or Google Earth for listing villages and recording their setting with low survey implementation costs and few coordinates. It is conceivable that road and satellite map administrative constraints. information does not fully correspond to official village Provided minimal investments in basic equipment, registers or that village registers are politically influenced, this approach can be replicated in other contexts where both of which can lead to the systematic exclusion of par- resources for household listing are limited, where sam- ticularly small and remote communities. Though more pling frames cannot be produced from administrative time-consuming, an alternative to using village lists is to data, and where residential structures are homogenous inspect the satellite imagery in the selected regions and and distinctive. Satellite-aided survey sampling and record the location of all identifiable settlements. implementation helped to reduce main survey time by In either case, compared to sampling through cen- one quarter and saved approximately £4100 in the pre- sus data, the methods are insensitive to population size. sent study. This and similar approaches can therefore Large and small villages are equally likely to enter the potentially improve the affordability of surveys especially sample [2], which can bias the sample towards smaller for student researchers and resource-constrained studies. Haenssgen Emerg Themes Epidemiol (2015) 12:20 Page 10 of 10 Abbreviations 8. Chang A, Parrales M, Jimenez J, Sobieszczyk M, Hammer S, Copenhaver 2G: second generation; CNY: Chinese Renminbi Yuan; GB: gigabyte; GBP: D, et al. Combining Google Earth and GIS mapping technologies in a pound sterling; GIS: geographical information system; GPS: global positioning dengue surveillance system for developing countries. Int J Health Geogr. system; IIHMR: Indian Institute of Health Management Research; IT: informa- 2009;8:49. doi:10.1186/1476-072X-8-49. tion technology; RAM: random access memory; VPN: virtual private network. 9. Gammino VM, Nuhu A, Chenoweth P, Manneh F, Young RR, Sugerman DE, et al. Using geographic information systems to track polio vaccina- Author details tion team performance: pilot project report. J Infect Dis. 2014;210(suppl Centre for Tropical Medicine and Global Health, Nuffield Department of Med- 1):S98–101. doi:10.1093/infdis/jit285. icine, University of Oxford, Oxford, UK. CABDyN Complexity Centre, Saïd 10. Morland KB, Evenson KR. Obesity prevalence and the local food Business School, University of Oxford, Oxford, UK. Green Templeton College, environment. Health Place. 2009;15(2):491–5. doi:10.1016/j. University of Oxford, Oxford, UK. healthplace.2008.09.004. 11. Microsoft Corporation. Bing Maps. 2015. http://www.bing.com/maps. Acknowledgements Accessed 07 Jan 2015. This paper arises from research funded by the John Fell Oxford University 12. Google Inc. Google Maps. 2015. http://maps.google.com. Accessed 07 Press (OUP) Research Fund. Further support from the UK Economic and Social Jan 2015. Research Council, the Scatcherd European Scholarship, the Oxford Depart- 13. HERE Maps BV. HERE Maps. 2015. http://www.here.com. Accessed 07 Jan ment of International Development, the University of Oxford Vice-Chancellors’ 2015. Fund and Hertford College is gratefully acknowledged. I thank Proochista 14. Delmelle E. Spatial sampling. In: Fotheringham AS, Rogerson PA, editors. Ariana, Ernest Guevarra, Felix Reed-Tsochas, and Xiaolan Fu and participants at The SAGE handbook of spatial analysis. London: Sage; 2009. p. 182–207. the 6th Conference of the European Survey Research Association for helpful 15. Working group for mortality estimation in emergencies. Wanted: stud- discussions in relation to survey design and implementation. I further received ies on mortality estimation methods for humanitarian emergencies, excellent research assistance from IIHMR in Rajasthan, especially SD Gupta, suggestions for future research. Emerg Themes Epidemiol. 2007;4:9. Nutan Jain, Arindam Das, Jagjeet Prasad Singh, Vidya Bhushan Tripathi, and doi:10.1186/1742-7622-4-9. Matadin Sharma; and from Liu Xingrong and the School of Public Health at 16. Fowler FJ. Survey research methods. 4th ed. Thousand Oaks: Sage Publi- Lanzhou University, Li Jian, and Wang Wei in Gansu. cations; 2009. 17. Groves RM, Fowler FJ, Couper MP, Lepowski JM, Singer E, Tourangeau R. Competing interests Survey methodology. In: Wiley series in survey methodology. 2nd ed. The author declares that he has no competing interests. Chichester: Wiley; 2009. 18. Heeringa S, West BT, Berglund PA. Applied survey data analysis. In: Chap- Received: 24 June 2015 Accepted: 4 December 2015 man & Hall/CRC statistics in the social and behavioral sciences series. Boca Raton: Chapman & Hall/CRC; 2010. 19. Gaziano C. Comparative analysis of within-household respondent selec- tion techniques. Public Opinion Q. 2005;69(1):124–57. doi:10.1093/poq/ nfi006. 20. United States Census Bureau, ICF Macro. Census and survey processing References system (CSPro) 5.0. Washington: United States Census Bureau; 2014. 1. Overton J, van Diermen P. Using quantitative techniques. In: Scheyvens 21. China Standard Press. 中华人民共和国行政区划代码资料手册 [Peo- R, Storey D, editors. Development fieldwork: a practical guide. London: ple’s Republic of China administrative division code book]. Beijing: China Sage; 2003. p. 37–56. Standard Press; 2010. 2. Zarkovich SS. Some problems of sampling work in underdeveloped 22. Google Inc. Google Earth. Mountain View: Google Inc; 2015. countries. In: Bulmer M, Warwick DP, editors. Social research in developing 23. Daniel J. Sampling essentials: practical guidelines for making sampling countries: surveys and censuses in the third world. London: UCL Press; choices. Thousand Oaks: Sage; 2012. 1993. p. 103–10. 24. Grais R, Rose A, Guthmann J-P. Don’t spin the pen: two alternative meth- 3. Shannon H, Hutson R, Kolbe A, Stringer B, Haines T. Choosing a survey ods for second-stage sampling in urban cluster surveys. Emerg Themes sample when data on the population are limited: a method using global Epidemiol. 2007;4:8. doi:10.1186/1742-7622-4-8. positioning systems and aerial and satellite photographs. Emerg Themes Epidemiol. 2012;9:5. doi:10.1186/1742-7622-9-5. 4. Wampler P, Rediske R, Molla A. Using ArcMap, Google Earth, and global positioning systems to select and locate random households in rural Haiti. Int J Health Geogr. 2013;12:3. doi:10.1186/1476-072X-12-3. 5. Escamilla V, Emch M, Dandalo L, Miller WC, Martinson F, Hoffman I. Sam- pling at community level using satellite imagery and geographical analy- sis. Bull World Health Organ. 2014;92(9):690–4. doi:10.2471/BLT.14.140756. 6. Galway L, Bell N, Sae A, Hagopian A, Burnham G, Flaxman A, et al. A two- stage cluster sampling method using gridded population data, a GIS, and Google Earth ( TM) imagery in a population-based mortality survey in Submit your next manuscript to BioMed Central Iraq. Int J Health Geogr. 2012;11:12. doi:10.1186/1476-072X-11-12. and we will help you at every step: 7. Flynn A, Tremblay PF, Rehm J, Wells S. A modified random walk door-to- door recruitment strategy for collecting social and biological data relat- • We accept pre-submission inquiries ing to mental health, substance use/addictions and violence problems in • Our selector tool helps you to find the most relevant journal a Canadian community. Int J Alcohol Drug Res. 2013;2(2):7–16. • We provide round the clock customer support • Convenient online submission • Thorough peer review • Inclusion in PubMed and all major indexing services • Maximum visibility for your research Submit your manuscript at www.biomedcentral.com/submit

Journal

Emerging Themes in EpidemiologySpringer Journals

Published: Dec 23, 2015

There are no references for this article.