Triangulation

Triangulation is a research strategy in which one brings multiple forms of evidence to bear on a single research question. The term triangulation is often modified by an adjective (e.g., within-method triangulation) to reflect how the evidentiary variation is introduced. In the years since the term was introduced into the research methods lexicon, two primary accounts of the value of triangulation have emerged. One account views triangulation as valuable for what it reveals about the validity of a descriptive or causal inference. The second views triangulation as valuable for how it enriches the perspective one gains on the question under investigation. The endorsement of triangulation for one or both of these reasons is widespread within the social sciences, but skeptical and cautious voices are heard as well. This entry discusses varying perspectives of triangulation and points to new developments.

Early references to triangulation focused on measuring concepts with data gathered from multiple methods. The classic reference is to the 1966 volume Unobtrusive Measures by Eugene J. Webb and colleagues. The advice was threefold: (1) supplement measures of social and political behavior drawn from interviews and questionnaires with (unobtrusive) measures drawn from physical trace evidence, documents, and observation; (2) test any given hypothesis repeatedly using the different measures of key concepts; and (3) treat the hypothesis as more credible if all the tests converge in support. The reasoning is that measures built from different data sources will each be flawed in ways that the researcher should be able to anticipate, if not correct, in advance. But the particular flaws will vary across the measures. A survey-based measure of alcohol consumption, for example, would likely suffer from social desirability bias, which would not threaten the validity of a measure based on (covert) observation of how many liquor bottles were found in a person's trash, even though the latter will have its own limitations. If, therefore, one finds similar results when testing a hypothesis with each measure, one can draw conclusions with greater confidence. It becomes implausible to view the findings as an artifact driven by biases unique to each measure.

Note that the major concern here is systematic error in the measures (a question of measurement validity) not random error (a question of measurement reliability). Hence the advice is to repeat one's analysis using each measure, not to form indices from the set as if they were simply parallel measures. This strategy loses value, however, if it is plausible to view the bias in each measure as having concordant effects; even if the hypothesis were false, flaws in the measures could produce converging results. Hence, it is important to select measures where the biases are expected to produce discordant effects on the results.

Influenced by Norman Denzin's more expansive treatment of the topic in 1970, social science understanding of triangulation quickly extended beyond the measurement context. To represent the varying ways in which different forms of evidence can be combined, triangulation is now routinely differentiated into types. Between-method triangulation involves multiple methods of data collection (focus groups, sample surveys, participant observation in field settings, content analysis of documents, laboratory experiments, and so on). In within-method triangulation, the data collection method is held constant but the design (e.g., analyze a panel and a cross-sectional survey) or measurement technique (e.g., work with different survey questions or question types) may vary. Data triangulation refers to the use of data on multiple samples, obtained at different times and/or in different contexts. With investigator triangulation, multiple investigators work at least semi-independently on a joint venture. In analysis triangulation, the same data are analyzed using multiple techniques (e.g., apply narrative and conversational analysis to in-depth interview texts or analyze quantitative data with both Bayesian and non-Bayesian statistical techniques).

...