over time. Godoy et al. (2009) used panel data collected annually (2002–2006) from the Tsimane’ Amazonian Panel Study (TAPS) to fill this gap. Results suggest a general improvement in well-being over time, with the highest rate of change in villages closest to the market town. This longitudinal study provides a better test of theory than would a cross-sectional design.
Panel studies need not be multiyear designs. Wutich (2009) combined participant observation with a five-wave panel survey in a study of common pool water institutions in Cochabamba, Bolivia. The panel design consisted of a household survey (N = 72) in which Wutich and her assistants interviewed households every two months for 10 months. This design – spanning two wet seasons and one dry – made it possible to examine seasonal variability in climate and water security in a way that would not have been possible with a cross-sectional study.
Case-control studies (Figure 4.4) are relatively rare in anthropology, although they are among the most common designs in epidemiology. The classic example in medical anthropology is Rubel and colleagues’ (1984) study of susto, a folk illness reported in many parts of Latin America. Based on ethnographic accounts, Rubel et al. developed specific hypotheses about the sociocultural factors that shape susceptibility to susto. To test these hypotheses, they compared a sample of people who suffered from susto (cases) with people who did not (controls) in three communities with different histories, language, and cultures in the Oaxaca Valley of Mexico (Zapotec, Chinantec, and mestizo).
Cases and controls were matched to form pairs of people who differed in whether they reported susto but were similar in other respects: age, gender, community, and complaints of being sick. Rubel and colleagues then tested cases and controls for differences in social stress, psychiatric symptoms, and physical health problems. There were no significant associations between psychiatric symptoms and susto, but people suffering from susto did experience more social stress, including perceived difficulty in performing important social roles – a pattern Rubel et al. expected from ethnography. And this difference really made a difference: Seven years later, 17% of people who complained of susto had died, but all the controls were still alive.
Rubel and colleagues’ study also teaches a valuable lesson about research design in general. It is a model of good design, but it wasn’t perfect – things happen in the field. Rubel et al. are honest about this fact and acknowledge their uncertainty. They realized, for example, that the gendered stigma of susto may have resulted in fewer Chinantec being willing to label themselves with the illness. This pattern would have biased the researchers’ conclusions about gender differences in the experience of susto. One of hallmarks of well-designed research design is that it makes such problems public.
SAMPLING
Some researchers collect data on an entire population, but most medical anthropologists work with samples, or subsets of the population in which they are interested. Thus, one of the key tasks of research design is to select an appropriate sampling strategy.
There are three steps in developing a sampling strategy: (1) defining the population, (2) identifying the unit of analysis (e.g., individual, household, clinic), and (3) selecting units of analysis for inclusion in the sample. The goal is to be able to say something about the units of analysis that were not selected for the sample. To meet this goal, it helps to be clear about what you’d like to say about your units of analysis. Do you want to estimate the average age in a population? Or, do you want to describe what it means to get older in that population? Medical anthropologists are likely to ask both types of questions, and they imply different types of sampling strategies.
We can describe this distinction in terms of “individual attribute” versus “cultural” data (Bernard 2018, p. 114). Attribute data refer to the characteristics of people (or other units of analysis) in a sample (e.g., age, income, blood pressure). Researchers in most health-related social sciences collect primarily attribute data of this type. Medical anthropologists collect attribute data, too, but many are also interested in questions such as “What does it mean to get older around here?” or “How important is income to a person’s status in this community?” or “How do you know if you have high blood pressure, and how is it distinct from other illnesses?” These questions elicit cultural data, because they capture shared and socially transmitted systems of meaning that organize how people make sense of the world.
Attribute data generally require probability samples; cultural data do not, because the shared and socially constructed nature of cultural phenomena violates the assumption of case independence in classical sampling theory (Gravlee 2005, p. 953). This assumption is often warranted with attribute data – your age is unrelated to mine. But if we participate in the same culture, your understanding of what it means to get older and mine are bound together, because people acquire cultural knowledge through social interaction. Thus, efficient ethnographic samples should select units of analysis to represent the range of variability in life experiences and social contexts related to the transmission of culture (Guest 2015; Johnson 1990). Probability samples are not necessary for achieving this aim, as probability and nonprobability samples have been shown to yield identical conclusions about cultural data (Handwerker and Wozniak 1997).
Probability and Nonprobability Sampling
There are many options for probability and nonprobability sampling designs. Miles and Huberman (1994, p. 28) list 16 types of nonprobability sampling. Onwuegbuzie and Leech (2007) identify 24 sampling designs, and Teddlie and Tashakkori (2009, p. 170) delineate 26, including a mix of probability and nonprobability methods. These complex typologies build on a small set of basic sampling designs summarized in Table 4.1. For details about probability and nonprobability sampling in anthropology, see Bernard (2018), Guest (2015), Johnson (1990), and Schensul and LeCompte (2013, pp. 280–318).
Table 4.1 Basic probability and nonprobability sampling designs
| Method | Purposes | Procedures |
|---|---|---|
| Probability methods | ||
| Simple random sampling | Generate representative sample when adequate sampling frame is available | List all members of population (sampling frame); select subset at random |
| Systematic random sampling | Generate representative sample; may not need to enumerate entire sampling frame | Select random starting point in sampling frame; select every Nth case |
| Stratified random sampling | Ensure that key subpopulations are represented; maximize between-group variation to increase precision | Divide population into subgroups; select random sample from each subgroup |
| Cluster sampling | Generate representative sample when no convenient sampling frame exists; sample dispersed populations efficiently | Divide population into clusters (e.g., neighborhoods, clinics); select random sample of clusters; sample within clusters |
| Nonprobability methods | ||
| Purposive sampling | Sample theoretically important dimensions of variation | Identify important theoretical criteria; select cases to satisfy criteria; multiple criteria-based methods are available |
| Quota sampling | Generate sample with fixed proportions of key subpopulations |
Divide population into subgroups; purposively select cases to fill quotas
|