1142 Triangulation

Hendry, D. F. (1995). Dynamic econometrics. Oxford, UK:Oxford University Press.

TRIANGULATION

Triangulation refers to the use of more than oneapproach to the investigation of a research questionin order to enhance confidence in the ensuing find-ings. Because much social research is founded on theuse of a single research method, and as such maysuffer from limitations associated with that methodor from the specific application of it, triangulationoffers the prospect of enhanced confidence. Triangu-lation is one of the several rationales for multimethodresearch. The term derives from surveying, where itrefers to the use of a series of triangles to map outan area.

TRIANGULATION AND MEASUREMENT

The idea of triangulation is very much associatedwith measurement practices in social and behavioralresearch. An early reference to triangulation was inrelation to the idea of unobtrusive method pro-posed by Webb, Campbell, Schwartz, and Sechrest(1966), who suggested, “Once a proposition has beenconfirmed by two or more independent measurementprocesses, the uncertainty of its interpretation is greatlyreduced. The most persuasive evidence comes througha triangulation of measurement processes” (p. 3). Thus,if we devise a new survey-based measure of a con-cept such as emotional labor, our confidence in thatmeasure will be greater if we can confirm the distri-bution and correlates of emotional labor through theuse of another method, such as structured observa-tion. Of course, the prospect is raised that the twosets of findings may be inconsistent, but as Webbet al. observed, such an occurrence underlines theproblem of relying on just one measure or method.Equally, the failure for two sets of results to convergemay prompt new lines of inquiry relating to either themethods concerned or the substantive area involved.A related point is that even though a triangulationexercise may yield convergent findings, we should bewary of concluding that this means that the findingsare unquestionable. It may be that both sets of data areflawed.

TYPES OF TRIANGULATION

Denzin (1970) extended the idea of triangulationbeyond its conventional association with researchmethods and designs. He distinguished four forms oftriangulation:

1. Data triangulation, which entails gatheringdata through several sampling strategies so thatslices of data at different times and in differentsocial situations, as well as on a variety ofpeople, are gathered

2. Investigator triangulation, which refers to theuse of more than one researcher in the field togather and interpret data

3. Theoretical triangulation, which refers to theuse of more than one theoretical position ininterpreting data

4. Methodological triangulation, which refers tothe use of more than one method for gatheringdata

The fourth of these, as the preceding discus-sion implies, is the most common of the mean-ings of the term. Denzin drew a distinction betweenwithin-method and between-method triangulation.The former involves the use of varieties of the samemethod to investigate a research issue; for example, aself-completion questionnaire might contain two con-trasting scales to measure emotional labor. Between-method triangulation involved contrasting researchmethods, such as a questionnaire and observation.Sometimes, this meaning of triangulation is taken toinclude the combined use of quantitative research

and qualitative research to determine how farthey arrive at convergent findings. For example, astudy in the United Kingdom by Hughes et al. (1997)of the consumption of “designer drinks” by youngpeople employed both structured interviews and focusgroups. The two sets of data were mutually con-firming in that they showed a clear pattern of agedifferences in attitudes toward these types of alcoholicdrinks.

Triangulation is sometimes used to refer to allinstances in which two or more research methodsare employed. Thus, it might be used to refer tomultimethod research, in which a quantitative and aqualitative research method are combined to provide amore complete set of findings than could be arrived atthrough the administration of one of the methods alone.However, it can be argued that there are good reasons

Trimming 1143

for reserving the term for those specific occasions inwhich researchers seek to check the validity of theirfindings by cross-checking them with another method.

EVALUATION OF TRIANGULATION

The idea of triangulation has been criticized onseveral grounds. First, it is sometimes accused of sub-scribing to a naive realism that implies that there canbe a single definitive account of the social world. Suchrealist positions have come under attack from writersaligned with constructionism and who argue thatresearch findings should be seen as just one amongmany possible renditions of social life. On the otherhand, writers working within a constructionist frame-work do not deny the potential of triangulation; instead,they depict its utility in terms of adding a sense ofrichness and complexity to an inquiry. As such, trian-gulation becomes a device for enhancing the credibilityand persuasiveness of a research account. A secondcriticism is that triangulation assumes that sets of dataderiving from different research methods can be unam-biguously compared and regarded as equivalent interms of their capacity to address a research question.Such a view fails to take account of the different socialcircumstances associated with the administration ofdifferent research methods, especially those associatedwith a between-methods approach (following Den-zin’s, 1970, distinction). For example, the apparentfailure of findings deriving from the administrationof a structured interview to converge with focus

group data may have more to do with the possibilitythat the former taps private views as opposed to themore general ones that might be voiced in the morepublic arena of the focus group.

Triangulation has come to assume a variety of mean-ings, although the association with the combined useof two or more research methods within a strategyof convergent validity is the most common. In recentyears, it has attracted some criticism for its apparentsubscription to a naively realist position.

—Alan Bryman

REFERENCESDenzin, N. K. (1970). The research act in sociology. Chicago:

Aldine.Hughes, K., MacKintosh, A. M., Hastings, G., Wheeler, C.,

Watson, J., & Inglis, J. (1997). Young people, alcohol, and

designer drinks: A quantitative and qualitative study. BritishMedical Journal, 314, 414–418.

Webb, E. J., Campbell, D. T., Schwartz, R. D., & Sechrest, L.(1966). Unobtrusive measures: Nonreactive measures in thesocial sciences. Chicago: Rand McNally.

TRIMMING

A fundamental problem with the sample mean is thatslight departures from normality can result in relativelypoor power (e.g., Staudte & Sheather, 1990; Wilcox,2003), a result that follows from a seminal paper byTukey (1960). Roughly, the sample variance can begreatly inflated by even one unusually large or smallvalue (called an outlier), which in turn can result inlow power when using means versus other measures ofcentral tendency that might be used.

There are two general strategies for dealing with thisproblem, and each has advantages and disadvantages.The first strategy is to check for outliers, discard anythat are found, and use the remaining data to compute ameasure of location. The second strategy is to discarda fixed proportion of the largest and smallest obser-vations and then average the rest. This latter strategyis called a trimmed mean, and, unlike the first strat-egy, empirical checks on whether there are outliers arenot made.

A 10% trimmed mean is computed by removing thesmallest 10% of the observations, as well as the largest10%, and averaging the values that remain. To computea 20% trimmed mean, remove the smallest 20%, thelargest 20%, and compute the usual sample mean usingthe data that remain. The sample mean and the usualmedian are special cases that represent two extremes:no trimming and the maximum amount.

A basic issue is deciding how much to trim. Interms of achieving a low standard error relative toother measures of central tendency that might be used,Rosenberger and Gasko (1983) recommend 20% trim-ming, except when sample sizes are small, in whichcase they recommend 25% instead. This recommenda-tion stems in part from the goal of having a relativelysmall standard error when sampling from a normaldistribution. The median, for example, has a rela-tively high standard error under normality, which inturn can mean relatively low power when testing hypo-theses. This recommendation is counterintuitive basedon standard training, but a simple explanation can be