The Discipline II: Fieldwork and the Fieldwork Tradition, the Methods from Ethnography and Observation to Grounded Theory, Excavation and GIS, Statistics from Variables and Sampling to Regression, and Content, Discourse and Narrative Analysis

Fieldwork is the prolonged, first-hand study of a people in their own setting, in their own language, through daily participation in their life. The armchair evolutionists of the 1870s (Tylor, Frazer, Morgan, who did visit the Iroquois) worked from correspondents' reports; the shift to the field came in stages. Boas wintered on Baffin Island among the Inuit in 1883–84. The Cambridge expedition to the Torres Straits in 1898 (Haddon, Rivers, Seligman, Myers, McDougall) took a team of scientists to collect systematically, and Rivers developed the genealogical method there. Radcliffe-Brown worked in the Andamans in 1906–08 and Seligman among the Vedda. Then Malinowski, marooned by the First World War, spent about two years in the Trobriand Islands (1915–16, 1917–18), living in a tent in the village, learning Kiriwina and writing Argonauts of the Western Pacific (1922), whose introduction laid down the method: pitch the tent among the natives, record the "imponderabilia of actual life", and grasp "the native's point of view". Participant observation, a year or more in one place, a field diary and a monograph became the standard. In India the tradition began with S.C. Roy's Chotanagpur monographs, continued with Majumdar among the Ho and Elwin among the Baiga and Muria, and matured in the village studies of the 1950s (Srinivas at Rampura, Dube at Shamirpet). Its ethical side — informed consent, confidentiality, no harm, reciprocity to the community — is now codified by the professional associations, and reflexivity (attention to the fieldworker's own position) is expected of every ethnography.

A variable is any characteristic that takes different values across cases: the independent variable is the presumed cause, the dependent the effect, and an intervening or confounding variable can sit between or behind them. Stevens's four levels of measurement (1946) decide which statistic is allowed: nominal (categories — caste, blood group; mode, chi-square), ordinal (ranked — wealth rank, Likert scale; median, Spearman's rho), interval (equal units, no true zero — temperature in °C; mean, SD, t-test) and ratio (true zero — height, income; all of these plus ratios). Sampling draws a manageable part of the universe. Probability samples give every unit a known chance of selection: simple random (lottery or random numbers), systematic (every kth house from a random start), stratified (divide by caste or age, then sample each stratum; proportionate or not), cluster and multi-stage (sample villages, then households). Non-probability samples do not: purposive or judgement (the elders who know the ritual), quota, convenience, and snowball (each informant names the next — the method for hidden populations). Sampling error shrinks with sample size; non-sampling error (bad questions, non-response) does not.

The measures of central tendency are the mean (sum divided by n), the median (the middle value when ranked; unaffected by extreme values, so preferred for income) and the mode (the most frequent value; the only one usable for nominal data). In a symmetrical distribution they coincide; in a positively skewed one (a long right tail, as with income) mean > median > mode. The measures of dispersion are the range, the quartile deviation, the mean deviation, the variance (the mean of squared deviations from the mean) and its square root, the standard deviation; the coefficient of variation, CV = (SD ÷ mean) × 100, compares the spread of series in different units. Worked: the heights of five adult men are 160, 165, 170, 175 and 180 cm. Mean = 850 ÷ 5 = 170. Deviations −10, −5, 0, 5, 10; squares 100, 25, 0, 25, 100; sum 250. Population variance = 250 ÷ 5 = 50 and SD = √50 ≈ 7.07 cm; the sample (n − 1) variance is 250 ÷ 4 = 62.5 and SD ≈ 7.91 cm. CV = 7.07 ÷ 170 × 100 ≈ 4.2%. When a question describes the values as the whole group, use the population form (divide by n); when it calls them a sample, divide by n − 1.

Parametric tests assume interval or ratio data drawn from a normally distributed population with similar variances: the t-test (Gosset, "Student", 1908; one-sample, independent-samples, paired) compares means, the F-test and analysis of variance (Fisher) compare more than two means, and Pearson's product-moment correlation r measures a linear relation between −1 and +1. Non-parametric ("distribution-free") tests need only nominal or ordinal data: the chi-square test (Pearson, 1900) for association between categorical variables, the Mann–Whitney U and Wilcoxon tests in place of the t-test, the Kruskal–Wallis test in place of ANOVA, the sign and run tests, and Spearman's rank correlation rho. Bivariate analysis relates two variables; multivariate analysis relates three or more at once — multiple regression, factor analysis, cluster analysis, discriminant analysis. Simple linear regression fits y = a + bx, with b the slope (change in y per unit x) and a the intercept, and reads a continuous dependent variable — stature from femur length is the forensic classic. Logistic regression is for a binary dependent variable (stunted or not, migrated or not): it models the log of the odds, ln[p/(1 − p)] = a + bx, and each coefficient is read as an odds ratio, e^b, the factor by which the odds of the outcome change per unit of the predictor.

Worked chi-square. In a village, 120 households are classified by whether the head is literate and whether the household uses the public health centre. Observed: literate and uses, 45; literate and does not, 15; non-literate and uses, 30; non-literate and does not, 30. Row totals 60 and 60; column totals 75 and 45. Expected count = (row total × column total) ÷ N: literate-uses 60 × 75 ÷ 120 = 37.5; literate-not 22.5; non-literate-uses 37.5; non-literate-not 22.5. χ² = Σ(O − E)²/E = (7.5²/37.5) + (7.5²/22.5) + (7.5²/37.5) + (7.5²/22.5) = 1.5 + 2.5 + 1.5 + 2.5 = 8.0. Degrees of freedom = (rows − 1)(columns − 1) = 1; the critical value at 0.05 is 3.84 and at 0.01 is 6.63, so the association between literacy and use of the centre is significant at both levels. A goodness-of-fit chi-square works the same way against theoretical proportions: 120 offspring expected in a 3:1 ratio (90 : 30) but observed 75 : 45 give χ² = (15²/90) + (15²/30) = 2.5 + 7.5 = 10 on 1 df, and the 3:1 hypothesis is rejected.

Content analysis, defined by Berelson (1952) as the objective, systematic and quantitative description of the manifest content of communication, counts: how often a newspaper names a tribe, which adjectives a textbook attaches to "primitive", how many folk tales feature a trickster. Its steps are to define the universe of texts, draw a sample, fix the unit (word, theme, item), build a coding frame, code with two coders for inter-coder reliability, and tabulate; Krippendorff's alpha measures that reliability, and qualitative content analysis codes latent as well as manifest meaning. Discourse analysis treats language as social action: it asks how a way of talking (about "tribal backwardness", about "development") constructs its object, who may speak, and what is made unsayable. Its lineage runs from Foucault's discourses as regimes of knowledge and power, through Fairclough's critical discourse analysis of text, discursive practice and social practice, to conversation analysis (Sacks, Schegloff) of turn-taking in talk. Narrative analysis takes the story itself as the unit: Labov's structure of oral narrative (abstract, orientation, complicating action, evaluation, result, coda), Riessman's thematic, structural, dialogic and visual approaches, and the anthropological use of life stories, myths and illness narratives (Kleinman's illness narratives) to see how people make experience meaningful and how a community's stories order its past.