Family Practice© Oxford University Press 1996

Vol. 13, No. 6Printed in Great Britain

Sampling for qualitative researchMartin N Marshall

Marshall, MN. Sampling for qualitative research. Family Practice 1996; 13: 522-525.The probability sampling techniques used for quantitative studies are rarely appropriatewhen conducting qualitative research. This article considers and explains the differencesbetween the two approaches and describes three broad categories of naturalistic sampling:convenience, judgement and theoretical models. The principles are illustrated with practicalexamples from the author's own research.Keyword. Qualitative sampling.

IntroductionThe benefits of a qualitative approach to health careresearch are becoming increasingly recognized by bothacademics and clinicians, but misunderstandings aboutthe philosophical basis and the methodological approachremain. The impression is sometimes given thatqualitative research differs from the hypothetico-deductive model simply in terms of the way that datais collected. The process of sampling is one of the prin-cipal areas of confusion, a problem not helped by theinadequate way that it is covered in the literature, wherethere is little agreement on definitions and authors fre-quently invent new and complex terms which cloudsimple fundamental issues. In this article I will describeboth quantitative and qualitative methods of samplingand consider the basic differences between the twoapproaches in order to explain why the sampling tech-niques used are not transferable. I will consider issuesrelating to sample size and selection in qualitativeresearch and illustrate the principles with practicalexamples.

Quantitative samplingChoosing a study sample is an important step in anyresearch project since it is rarely practical, efficient orethical to study whole populations. The aim of all quan-titative sampling approaches is to draw a representativesample from the population, so that the results ofstudying the sample can then be generalized back tothe population. The selection of an appropriate methoddepends upon the aim of the study. Sometimes lessrigorous methods may be acceptable, such as incidentalor quota samples, but these methods do not guarantee

Received 30 May 1996; Accepted 15 July 1996.Institute of General Practice, University of Exeter, PostgraduateMedical School, Barrack Road, Exeter EX2 5DW, UK.

a representative sample. The most common approachis to use random, or probability samples. In a randomsample the nature of the population is defined and allmembers have an equal chance of selection. Stratifiedrandom sampling and area sampling are variants ofrandom sampling, which allow subgroups to be studiedin greater detail.

The size of the sample is determined by the optimumnumber necessary to enable valid inferences to be madeabout the population. The larger the sample size, thesmaller the chance of a random sampling error, but sincethe sampling error is inversely proportional to the squareroot of the sample size, there is usually little to be gainedfrom studying very large samples. The optimum samplesize depends upon the parameters of the phenomenonunder study, for example the rarity of the event or theexpected size of differences in outcome between the in-tervention and control groups.

Comparing the quantitative and qualitative approachesThe choice between quantitative and qualitative researchmethods should be determined by the research ques-tion, not by the preference of the researcher. It wouldbe just as inappropriate to use a clinical trial to examinebehavioural differences in the implementation of clinicalguidelines as it would be to use participant observationto determine the efficacy of antibiotics for upper respira-tory tract infections. The aim of the quantitative ap-proach is to test ore-determined hypotheses and producegeneralizable results. Such studies are useful for answer-ing more mechanistic 'what?' questions. Qualitativestudies aim to provide illumination and understandingof complex psychosocial issues and are most usefulfor answering humanistic 'why?' and 'how?' questions.The principal fundamental differences in both thephilosophical foundation of and the methodologicalapproach to the two disciplines are summarized inTable 1.

522

by guest on February 16, 2011

fampra.oxfordjournals.org

Dow

nloaded from

Sampling for qualitative research 523

Why is random sampling inappropriate forqualitative studies?The process of selecting a random sample is well definedand rigorous, so why can the same technique not beused for naturalistic studies? The answer lies in the aimof the study; studying a random sample provides thebest opportunity to generalize the results to the popula-tion but is not the most effective way of developing anunderstanding of complex issues relating to humanbehaviour. There are both theoretical and practicalreasons for this.

First, samples for qualitative investigations tend tobe small, for reasons explained later in this article. Evenif a representative sample was desirable, the samplingerror of such a small sample is likely to be so large thatbiases are inevitable. Secondly, for a true random sam-ple to be selected, the characteristics under study of thewhole population should be known; this is rarely pos-sible in a complex qualitative study. Thirdly, randomsampling of a population is likely to produce a represen-tative sample only if the research characteristics are nor-mally distributed within the population. There is noevidence that the values, beliefs and attitudes that formthe core of qualitative investigation are normallydistributed, making the probability approach inappro-priate. Fourthly, it is well recognized by sociologists1

that people are not equally good at observing, understan-ding and interpreting their own and other people'sbehaviour. Qualitative researchers recognize that someinformants are 'richer' than others and that these peopleare more likely to provide insight and understandingfor the researcher. Choosing someone at random toanswer a qualitative question would be analogous to ran-domly asking a passer-by how to repair a broken downcar, rather than asking a garage mechanic—the formermight have a good stab, but asking the latter is likelyto be more productive.

Sample sizeQuantitative researchers often fail to understand theusefulness of studying small samples. This is relatedto the misapprehension that generalizability is theultimate goal of all good research and is the principalreason for some otherwise sound published qualitativestudies containing inappropriate sampling techniques.2

An appropriate sample size for a qualitative study isone that adequately answers the research question. Forsimple questions or very detailed studies, this might bein single figures; for complex questions large samplesand a variety of sampling techniques might benecessary. In practice, the number of required subjectsusually becomes obvious as the study progresses, asnew categories, themes or explanations stop emergingfrom the data (data saturation). Clearly this requiresa flexible research design and an iterative, cyclical

approach to sampling, data collection, analysis and in-terpretation. This contrasts with the stepwise design ofquantitative studies and makes accurate prediction ofsample size difficult when submitting protocols tofunding bodies.

Sample strategiesThere are three broad approaches to selecting a samplefor a qualitative study.

Convenience sampleThis is the least rigorous technique, involving the selec-tion of the most accessible subjects. It is the least costlyto the researcher, in terms of time, effort and money,but may result in poor quality data and lacks intellec-tual credibility. There is an element of conveniencesampling in many qualitative studies, but a morethoughtful approach to selection of a sample is usuallyjustified.

Judgement sampleAlso known as purposeful sample, this is the most com-mon sampling technique. The researcher actively selectsthe most productive sample to answer the research ques-tion. This can involve developing a framework of thevariables that might influence an individual's contribu-tion and will be based on the researcher's practicalknowledge of the research area, the available literatureand evidence from the study itself. This is a more in-tellectual strategy than the simple demographicstratification of epidemiological studies, though age,gender and social class might be important variables.If the subjects are known to the researcher, they maybe stratified according to known public attitudes orbeliefs. It may be advantageous to study a broad rangeof subjects (maximum variation sample), outliers(deviant sample), subjects who have specific experiences(critical case sample6) or subjects with special exper-tise (key informant sample). Subjects may be able torecommend useful potential candidates for study(snowball sample). During interpretation of the data itis important to consider subjects who support emerg-ing explanations and, perhaps more importantly, sub-jects who disagree (confirming and disconfirmingsamples).

Theoretical sampleThe iterative process of qualitative study design meansthat samples are usually theory driven to a greater orlesser extent. Theoretical sampling necessitates buildinginterpretative theories from the emerging data andselecting a new sample to examine and elaborate on thistheory. It is the principal strategy for the groundedtheoretical approach3 but will be used in some formin most qualitative investigations necessitatinginterpretation.

by guest on February 16, 2011

fampra.oxfordjournals.org

Dow

nloaded from

524 Family Practice—an international journal

TABLE 1 Comparison of quantitative and qualitative methods

PhilosophicalfoundationAim

Study plan

Position ofresearcherAssessing qualityof outcomes

Measures ofutility of results

Quantitative

Deductive,reducdonalistTo test pre-sethypothesisStep-wise,predeterminedAims to be detachedand objectiveDirect tests ofvalidity and reliabi-lity using statisticsGeneralizability

Qualitative

Inductive, holistic

To explore complexhuman issuesIterative, flexible

Integral part ofresearch processIndirect qualityassurance methods oftrustworthinessTransferability

It is apparent from the above description that thereis considerable overlap even between these three broadcategories. The relative balance will depend upon theresearch question and the chosen style of data analysisand interpretation. It is important to recognize thatthe essence of the qualitative approach is that it isnaturalistic—studying real people in natural settingsrather than in artificial isolation. Sampling thereforehas to take account not only of the individual'scharacteristics but also temporal, spatial and situationalinfluences, that is, the context of the study. The re-searcher should consider the broader picture: would thisindividual express a different opinion if they were in-terviewed next week or next month? Would they feeldifferently if they were interviewed at home or at work?Should I study mem when they are under stress or re-laxed? There is no correct answer to these questions,just as there is no perfect way to sample, but the in-fluence that these factors might have on the trustwor-thiness of the results should be acknowledged.

A practical example of sampling strategyIn practice, qualitative sampling usually requires a flex-ible, pragmatic approach. This may be illustrated bymy own study of the professional relationship betweenGPs and specialists (in progress).

The way that the two branches of the medicalprofession work together is a key component of theprimary-secondary care interface, which, in terms ofsociological interaction, is largely unresearched. Thestudy aims to describe the current relationship, com-pare this with the historical literature, and elucidate theprincipal factors causing a change in the interaction be-tween the two main branches of the medical profession.Four methods of data collection have been used, eachof which view the interaction from differing perspectivesand each of which have required different samplingstrategies.

The first stage involved the use of key informant

interviews,4 an anthropological technique utilizing richinformation sources, which has defined sample selec-tion criteria.5 A sample of 10 national figures in posi-tions of leadership and responsibility within theprofession were chosen. Since the total population ofpossible key informants is small, this was necessarilya convenience sample, though there was an element ofa judgement approach, since efforts were made to ensurethat participants came from a range of clinical,academic, managerial and political backgrounds. Theadvantage of this approach lies in its simplicity but itwas difficult to determine at the sampling stage whetherthe informants fulfilled the published selection criteria.

The second stage involved in-depth interviews withpractising clinicians throughout the South and WestRegion. The aim was to develop an understanding andan interpretative framework of the process of interactionbetween specialists and GPs. I started with a judgementsample framework including variables such as time sincequalification, gender, geographical location, rurality,fundholding status and teaching hospital status. As thedata was collected and analysed, an interpretativeframework was constructed, so the sampling strategychanged from largely judgement to largely theoretical,in order to build on the developing theory. New themesstopped emerging after about 15 interviews and an ac-ceptable interpretative framework was constructed after24 interviews—the stage of thematic and theoreticalsaturation.

The third stage of the study brought GPs andspecialists together in focus groups to collect the dif-ferent level of data produced by personal interaction.For pragmatic reasons, this had to be conducted in myown locality, and it was important for the study thatthe participants were able to interact in a productive,rather than dysfunctional way. I was able to use mylocal knowledge to satisfy these sampling requirementsusing a combination of convenience and purposivetechniques.

The three qualitative stages of the study will formthe basis of a Likert survey to test out emergent themesand which will be distributed to a stratified randomsample of the whole population of clinicians work-ing in the Region. This will represent a different, notnecessarily a stronger, perspective of the professionalrelationship.

ConclusionSampling for qualitative research is an area of con-siderable confusion for researchers experienced in thehypothetico-deductive model. This largely relates tomisunderstanding about the aims of the qualitativeapproach, where improved understanding of complexhuman issues is more important than generalizabilityof results. This basic issue explains why probabilis-tic sampling is neither productive nor efficient for

by guest on February 16, 2011

fampra.oxfordjournals.org

Dow

nloaded from

Sampling for qualitative research 525

qualitative studies and why alternative strategies areused. Three broad categories of naturalistic samplingtechniques have been described—convenience, judge-ment and theoretical sampling—though in practice thereis often considerable overlap between these approaches.

AcknowledgementsI would like to thank Ms Nicky Britten for her com-ments on this paper. The study of the professional rela-tionship between specialists and GPs is funded by aResearch and Training Fellowship from the Researchand Development Directorate of the South and WestRegional Health Authority.

References1 Jackson JA. Professions and professionalisation. Cambridge:

Cambridge University Press, 1970.2 Pound P, Bury M, Gompertz P, Ebrahim S. Stroke patients'

views on their admission to hospital. Br MedJ 1995; 311:1&-22.

3 Glaser BG, Strauss AL. The discovery of grounded theory:Strategies for qualitative research. London: Weidenfield andNicholson, 1968.

4 Marshall MN. The key informant technique. Fam Pract 1996;13: 92-97.

5 Burgess RG (ed.). Field research: a sourcebook and manual.London: Routledge, 1989.

6 Bradley C. Turning anecdotes into data—the critical incidenttechnique. Fam Pract 1992; 9: 98-103.

by guest on February 16, 2011

fampra.oxfordjournals.org

Dow

nloaded from