Download Thematic Analysis and more Study notes Qualitative research in PDF only on Docsity!
[citations: alpha order] [spelling UK ize] [recto running head: Thematic analysis]
Thematic Analysis
Gareth Terry, Nikki Hayfield, Victoria Clarke and Virginia Braun
Introduction
This chapter introduces thematic analysis (TA), a method that has become a widely-used tool for analysing qualitative data, both in psychology and beyond. We first outline the history and context of TA, and identify key issues that need to be considered when conducting TA. We discuss the flexibility TA can offer, and highlight the need for deliberate and careful research. This flexibility can apply to theoretical assumptions, research questions, data collection and analysis. We include a detailed worked demonstration of the processes and procedures of undertaking a TA, illustrated with examples from Nikki Hayfield, Victoria Clarke, Sonja Ellis and Gareth Terry’s research on the lived experiences of childfree women (see Box 2 .1). Our discussion of how to complete a TA is based on a widely used version of TA – the approach developed by Virginia Braun and Victoria Clarke ( 2006 ). We conclude by considering the limitations and applications of TA, as well as future directions. [TS: Insert Box 2 .1 about here] Box 2.1 Introducing the lived experience of childfree women (child - freedom) study
Thematic analysis: History and context
What is thematic analysis (TA)? This question invites many different answers. TA practitioner Joffe (2012) credits philosopher of science Gerald Holton with founding TA in his work on ‘themata’ in scientific thought (Holton, 1975), but the term does seem to pre-date Holton’s use of it. Since the early part of the twentieth century, if not earlier, the term ‘thematic analysis’ has been used to refer to a number of different things, including, but not limited to, data analysis techniques in the social sciences. Some earlier instances of the use of TA are similar to contemporary use – a method for identifying themes in qualitative data (e.g. Dapkus, 1985). It has also been used interchangeably with content analysis to refer to both qualitative (Baxter, 1991) and quantitative (Christ, 1970) content analysis, and some have claimed that TA developed from content analysis (Joffe, 2012). Procedures for using TA as a qualitative technique only began to be published in the 1990s (e.g. Aronson, 1994), but qualitative researchers have described their approach to analysis as ‘thematic’, without an explicit reference to a developed method, both pre- and post-specific procedural advice being published. This complexity is why, in 2006, Virginia Braun and Victoria Clarke described TA as ‘a poorly demarcated and rarely acknowledged, yet widely used qualitative analytic method’ (Braun and Clarke, 2006: 77). Since the publication of what became a landmark paper, TA as a ‘named and claimed’ method has gained hugely in popularity and has entered the qualitative canon as a recognisable and reputable method of analysis. Other notable accounts of TA procedures published prior to Braun and Clarke’s have also grown in popularity (e.g. Boyatzis, 1998). However, some confusion remains about what TA is, and indeed whether it is anything in particular. Our task in the remainder of this section is to map the terrain of TA, and identify some of the similarities and differences between various approaches to TA. This provides context for our subsequent discussion and demonstration of what has become the
there are a few useful guides [on how to carry out TA], including Boyatzis (1998), Braun and Clarke (2006) and Joffe and Yardley (2004). This chapter moves to laying out the set of key steps involved in a TA. (Joffe, 2012: 215, our emphasis) Hence authors of methodological texts thus often fail to acknowledge diversity within TA. This is potentially confusing for qualitative beginners seeking clear guidance, but much more importantly, it obscures important theoretical and conceptual differences between different TA approaches. This diversity covers the overall conceptualisation of what TA is or offers, where it sits theoretically, and processes and procedures for (best practice) analysis. For a first broad categorisation, we find a distinction between ‘experiential’ and ‘critical’ orientations to qualitative research useful (Braun and Clarke, 2013 ; Reicher, 2000 ). Experiential orientations focus on what participants think, feel and do, and are underpinned by the theoretical assumption that language reflects reality (either a singular universal reality, or the perspectival reality of a particular participant). Critical orientations seek to interrogate dominant patterns of meaning and theoretically understand language as creating, rather than reflecting, reality. Some writers situate TA as only and always an experiential approach. Others describe TA as a theoretically independent – and thus flexible – approach, but still see it as particularly compatible with certain theoretical orientations, such a particular kind of phenomenology, or phenomenology in general (Guest et al., 2012; Joffe, 20 12 ). It is rarely explained why TA is seen as particularly compatible with these approaches – and the claimed compatibility seems to rely on the assumption that TA is an experiential orientation. Moreover, any claimed theoretical independence is often circumscribed in two (related) ways. First, TA is often described as an approach that bridges a quantitative (positivist) and qualitative (interpretative) divide (Boyatzis, 1998). The idea that TA can bridge a divide between quantitative and qualitative research depends on a particular definition of qualitative research as offering techniques or tools for collecting (and analysing) qualitative data. With a
conceptualisation of qualitative research as (only) about techniques and tools, TA is understood as offering a bridge over a divide, because it either provides qualitative techniques for use within a (post-)positivist paradigm, and/or allows for (post-)positivist standards like reliability to be utilised. However, this idea of what qualitative research offers is remarkably limited, and dominated by (post-)positivism, a framework that many qualitative researchers reject. An understanding of qualitative research as a paradigm (or multiple paradigms, Grant and Giddings, 2002) characterised by values and standards quite different from those espoused within (post-)positivist empiricist traditions dominates much qualitative scholarship. Therefore, any attempt to bridge qualitative and quantitative through TA therefore relies on limited conceptualisation of what qualitative research is (and can be). The second way the flexibility of TA is circumscribed stems from this point: critical orientations within qualitative research are rarely acknowledged. This absence results in a very limited account of what TA can offer. The approach to TA we have developed and that we expand on in this chapter offers full theoretical flexibility, potential for an experiential or critical orientation, and locates TA fully within a qualitative paradigm (e.g Braun and Clarke, 2006). The importance of these broader tensions is revealed through looking at the different procedures for conducting TA that are described. Despite variations across different versions, there seem to be two basic approaches: (1) an approach defined by an emphasis on coding reliability; (2) a more qualitative approach that advocates for a flexible approach to coding and theme development. Coding reliability approaches are often deductive, and echo the scientific method – moving from theory (deduction) to hypothesis/prediction (identifying themes), to evidence gathering/testing hypotheses (coding). This means analysis moves from familiarisation to some form of theme development then to coding. Themes are often at least partly determined in advance of full analysis, guided by existing theory and reflected in interview questions (in some instances, it is recommended that interview questions form the
rater reliability can only show that two coders have been trained to code the data in the same way, not that the coding is somehow ‘accurate’ (Braun and Clarke, 2013). In contrast, in more qualitative versions of TA such as our own (e.g. Braun and Clarke, 2006, 2013, Clarke, Braun and Hayfield, 2015a, Braun, Clarke and Terry, 2015), the subjectivity of the researcher is seen as integral to the process of analysis. Within such approaches, an inductive approach to coding and theme development is more common. Analysis once again starts with familiarisation, but close similarities with ‘coding reliability’ approaches to TA end there. Coding is treated as an organic and flexible process, where good coding requires a detailed engagement with the data. The assumption is that coding ‘gets better’ (i.e. develops depth and moves beyond the obvious surface level) through immersion in, or repeated engagement with, the data – something unlikely to be achieved with a code- book approach. Themes are developed from coding and working with the data and codes, rather than pre-existing the coding process. They are the outcome of the analytic process, rather than a starting point. They are not imagined or anticipated early on, and do not drive analytic direction. Coding and theme development are assumed to be subjective and interpretative processes. This means the outcomes of these processes can be stronger or weaker, but they cannot be right or wrong in any objective sense. The analysis is seen as something created by the researcher, at the intersection of the data, their theoretical and conceptual frameworks, disciplinary knowledge, and research skills and experience; it is not seen as something waiting ‘in’ the data to be found. Quality remains a vital concern, but quality-assurance strategies, such as a review of candidate themes (Braun and Clarke, 2 006 ), are focused on encouraging reflection, rigour, a systematic and thorough approach, and even greater depth of engagement, rather than focusing on coding ‘accuracy’.
Considering the differences among published existing versions of TA, we think they can be divided into two broad ‘schools’: (1) ‘Small q’ TA that retains a foothold in positivist research (e.g. Boyatzis, 1998; Guest et al., 2012, Joffe, 2012) and is concerned with establishing coding reliability; (2) a ‘Big Q’ approach to TA, that operates within a qualitative paradigm and is characterised by (genuine) theoretical independence and flexibility, and organic processes of coding and theme development (e.g. Braun and Clarke, 2006 ; Langridge, 2004). For readers unfamiliar with the small q/Big Q distinction, small q qualitative research describes the use of qualitative tools and techniques, particularly around data generation, within a positivist framework; Big Q refers to the use of these tools and techniques within the qualitative paradigm (Kidder and Fine, 1987). As our discussion above has illustrated, this distinction is important with regard to TA, because small q and Big Q approaches are underpinned by very different conceptualisations of knowledge, research, and the researcher. In small q TA, the researcher is like an archaeologist sifting through soil to discover buried treasures. Analysis is a process of discovering themes that already exist within a dataset, or finding evidence for themes that pre-exist the data. In Big Q TA, the researcher is more like a sculptor, chipping away at a block of marble. The sculpture is the product of an interaction between the sculptor, their skills and the raw materials. Analysis becomes a creative rather than technical process, a result of the researcher’s engagement with the dataset and the application of their analytic skills and experiences, and personal and conceptual standpoints. This section has highlighted that TA is far from the singular, homogeneous approach it is often treated as being, and the diversity within TA is consequential for research. Researchers need to both understand, and then locate their use of TA in relation to, this diversity – we often see authors stating they are doing TA, then referencing two different and contradictory approaches. We advocate an approach that is theoretically independent and flexible but
Table 2 .1 offers definitions of three broad ontological orientations typical within TA, and what the research is then assumed to capture, as well as offering some example questions. [TS: Insert Table 2.1 about here] Table 2.1 Ontologies and research questions
Methods of data collection and sample size
The flexibility of TA means it is suitable to analyse a wide range of data types: TA can be used to analyse data from ‘traditional’ face-to-face data collection methods such as interviews (e.g. Niland et al., 2014) and focus groups (e.g. Neville et al., 2015). It can also be used with textual data from qualitative surveys (e.g. Hayfield, 2013; Terry and Braun, 2016), diaries (e.g. Leeming et al., 2013), story based methods such as vignettes and story completion tasks (e.g. Clarke et al., 2015b), as well as online discussion forums (e.g. Bennett and Gough, 2013), and other media sources (e.g. Frith, 2015). The most important aspect of data type or mode of collection is quality of the data. Rich and complex data on a given topic are the crown jewels of qualitative research, allowing us deep and nuanced insights. Quantity (e.g. sample size) is also a consideration, but should not be conflated with quality. Key in thinking about sample size in TA is to recognise that it produces accounts of patterns across the dataset (this is not intended as a case-study approach, although some researchers are using TA in case studies, see Cedervall and Åberg, 2010). Sample size is a fraught, contentious, and debated topic in qualitative research. We offer some broad indicative size recommendations across TA projects of different scale for reference in Table 2.2 – linked to student projects. However, what is deemed ‘publishable’ is an entirely separate, and also fraught, issue, often linked to an editor’s view, but not necessarily shared by all qualitative scholars. [TS: Insert Table 2.2 about here]
Table 2.2 Project sample size recommendations (adapted from Braun and Clarke, 2013)
Possibilities for coding
The flexibility of TA applies also to the analysis, where the researcher again needs to make some deliberate choices about their approach to data and analysis. One consideration is theoretical stance (as outlined above). Another is whether to approach the data inductively or deductively – either exclusively, or as a primary mode of engagement. Inductive coding and theme development involves working ‘bottom up’ from the data, and developing codes (and ultimately themes) using what is in the data as the starting point; the data provide the bedrock for identifying meaning and interpreting data. Of course, there is some fallacy in this idea, as the researcher is never a blank slate, and inevitably brings their own social position and theoretical lens to the analysis, but an inductive orientation signals a data-led analysis. In a deductive approach, the analytic starting point is more ‘top down’ – the researcher brings in existing theoretical concepts or theories that provide a foundation for ‘seeing’ the data, for what ‘meanings’ are coded, and for how codes are clustered to develop themes; it also provides the basis for interpretation of the data (Braun et al., 2015). A deductive orientation is less bound by the semantic meaning in the data than an inductive orientation. Whether to focus semantically or latently (again, exclusively, or primarily) is a second analytic choice. In semantic coding, codes capture explicit meaning; they are identified at the surface level of the data. In latent coding, the codes capture implicit meaning, such as ideas, meanings, concepts, assumptions which are not explicitly stated; a ‘deeper’ level of analysis is required to code in this way (see Box 2 .2). In the early stages of coding, particularly for those new to qualitative analysis, the analysis is often more semantic/surface. However, with ‘immersing yourself in’ the data – and/or becoming more experienced – analysis can develop towards a more latent orientation. The appropriateness of each approach needs to fit with research question, and overall theoretical framework too: on the whole, more experiential and
Familiarisation and coding (phases 1–2)
Familiarisation , a process common across many qualitative analytic approaches, is the bedrock for doing good TA. Familiarisation provides the researcher with an entry point into analysis – it’s a way of engaging with, and gaining insight into, what can sometimes appear to be an overwhelming mass of data. When done poorly, or not at all, the rest of the analysis often suffers. So as tempting may it be, skipping over familiarisation, or only doing it once over lightly, does not provide the best launching pad for a high quality TA. Familiarisation is the researcher’s first opportunity for what’s referred to as immersion in the dataset. While the term evokes a very passive, and possibly terrifying-sounding, process, like floating in a tank of water, it is nothing like that. Familiarisation is about intimately knowing the dataset – this facilitates a deep engagement with the data. It requires the researcher to get into a mode of reading that actively engages with the data as data – this means being observant, noticing patterns or quirks, starting to ask questions, and so on, rather than just absorbing the information therein, as when reading a good crime novel. In practice, this means reading and re-reading all textual data, making casual observational notes. It might involve (re)listening or (re)watching, if the dataset is audio or video. This first phase is about generating very early and provisional analytic ideas, and this requires being curious, and asking questions of the data. The sorts of questions vary by form of TA, combined with the research question: they could be about the way participants orient themselves to questions; about assumptions they make; about worldviews they drawing from; about the implications of their accounts for themselves and those around them; about (more semantically) the different emotional responses to the research topic; and so on (but keeping the general research question in mind). Familiarisation involves moving through the entire dataset. Keeping notes (e.g. in transcript margins; in a separate notebook) ensures these early
analytic observations are remembered and can be referred back to. To make the most of this process, the researcher can synthesise observations and notes into ideas or insights related to the dataset as a whole, related to the research focus. Box 2 .3 provides examples of familiarisation notes from the child-freedom study – related to one single participant and then across all transcripts. [TS: Insert Box 2.3 about here. TS Note that in the Box text file the author requests that the ‘handwritten’ typeface is retained] Box 2.3 Familiarisation from one interview and the entire dataset In the child-freedom study, the four researchers – Nikki Hayfield, Victoria Clarke, Sonja Ellis and Gareth Terry – each independently familiarised themselves with some or all the transcripts, and a research team meeting was held to discuss the insights generated. As Box 2.3 shows, one of the notes from across the dataset was about the precariousness of many of the women’s accounts. When reading the transcripts, Gareth noted early on that many of the women spoke about the various points in their lives at which they might have had children, or stages where, if they had made different choices – such as when partners put them under pressure – it may have resulted in rethinking their identity as childfree. This stood out to him, as in his previous research with vasectomised childfree men (see Terry and Braun, 2012), the participants instead emphasised a distinct, lifelong, unyielding resistance to children, a much more fixed identity. This demonstrates the way research is a subjective process (this is illustrated further in Box 2.4, where Nikki and Gareth reflect on what they brought to the project, and how it impacted on their analytic process). An overenthusiastic researcher might take a familiarisation noticing like this, and attempt turn it into a theme early on, before it has been identified across the data. There are two risks of a ‘fast and loose’ approach like that, which impacts quality: (1) the risk of ‘cherry picking’, or selectively choosing data to suit an
(label) ideally contains enough information about the content of that data extract, and sometimes analytic interpretation, that it is meaningful without needing to refer back to the data. We’ve called this the ‘take away the data’ test (Braun and Clarke, 2013). This is might seem annoyingly pedantic, but it becomes particularly important later in the process, when developing themes from codes. [TS: Insert Table 2.3 about here.] [TS: Note that entries in column 2 must keep position relative to text in column 1, as shown in text file] Table 2.3 Example of coding in P17 (‘Millers’) Table 2 .3 gives an example of a coded data extract for the child-freedom project. The broadly semantic codes (e.g. quality of life would be impacted) reflect what the participant (‘Millers’) explicitly said about, and the meanings she ascribed to, being childfree. More latent codes capture ideas or concepts embedded within, or underpinning, the explicit content (e.g, resistance to engaging in superwoman/supermum position). Millers did not talk explicitly about a supermum/superwoman discourse – the notion that women should be able to be primary caregivers of their children, hold down a full time job, and still do both with high levels of competence (e.g. Sasaki and Hazen, 2010). However, this concept was useful for making sense of her logic when she talked about ‘spreading herself thin’ – an idea that is part of the discourse. The coding process is iterative and flexible, and code revision and development is part of this. Codes developed later in the process might capture a particular concept more clearly than earlier ones, and researchers tend to refine and revise codes throughout the process – it pays not to get too attached too early on (this is a bit of a mantra for doing our version of TA). The researcher often circles back through data items to clarify, or modify, earlier
coding, which also helps with coding consistency – avoiding having hundreds or even thousands of unique codes with lots of overlap. Coding is there to help the analyst make sense of the data, develop insight, and provide a rigorous and thorough foundation for the analysis (it can also help to tighten or modify a research question). In terms of the practicalities of coding, we recommend researchers use whatever method works best for them: write codes in the margins of hard copies of the data items; use Microsoft Word’s comment function; use computer software designed for qualitative coding (see Chapter 23 to tag and collate data. People also use file cards, or cut and paste (either physically or digitally) data segments into new files or onto clean pages. Recently, we have seen people start to claim that computer programs provide the best way to code. We definitely do not agree with this sentiment as a generic position. Any researcher needs to identify the right tools for them, in the context of their particular project. For instance, software might facilitate code sharing and development in a large team project; a low-tech researcher working on a small individual project may find file cards work best for them. Coding is a process not a technology, and the same quality can be achieved through various means. Poor quality coding is thin, with limited interpretative work, and/or sloppy – inconsistent and partial; good quality coding is the opposite, deep, consistent and thorough. After coding all data items thoroughly, this phase ends with the production of a compiled list of codes that adequately identify both patterning and diversity of relevant meaning within the dataset. Collating associated tagged data segments is the last task before moving on to theme development.
Theme development (phase 3)
Establishing a deep understanding of the dataset through familiarisation and coding sets up the researcher well to begin constructing themes. Rather than describing themes as
At this point in the analysis, it is really easy to get attached. But it is extremely rare that first attempts at theme development will produce a final thematic mapping. If themes emerged preformed, this might be an understandable way of looking at the process. However, as we view themes as constructed or generated through a productive, iterative, reflective process of data-engagement, it makes more sense to treat each clustering of codes as possibilities. At this stage, they are provisional or ‘candidate’ themes – imagining them as candidate themes gives the researcher the opportunity to discard them, to explore other possibilities, before eventually settling on a final set of themes. In order to facilitate this process of shifting mapping of various patterns, we encourage researchers to make use of visual aids, such as thematic maps (see Figure 2 .1) (see also, Braun and Clarke, 2013; Braun et al., 2015) or tables (see Table 2 .4). As with coding, such (visual) mapping aids are tools that enhance the researcher’s ability to identify and understand potential themes in relation to each other, and the overall dataset. Such tools provide a way of identifying what the boundaries of, and the relationships between, each theme might be, as well as how different themes work together to tell an overall story about the data. Good quality themes should be distinctive, with little ‘bleeding’ of codes between themes; themes should also be linked to, and work alongside, the other themes in the analysis
- and each needs to have its own distinct central organising concept. Which mapping tools the researcher uses to construct themes, how they work with those tools, and whether they ever present the maps or tables in a research write-up, is dependent on what works best for them – and the context of the study (we, for instance, would be unlikely to report thematic maps, unless we were writing a methodological piece, like this one, but encourage undergraduate and postgraduate students to do so in the appendices of their projects). [TS: Insert Figure 2.1 about here]
Figure 2.1 An early thematic map [TS: Insert Table 2. 4 about here] Table 2.4 Four candidate themes from the child-freedom study, with example codes In the child-freedom study, the idea of a ‘precarious identity’ noted earlier continued to appear across the dataset. Women spoke of the decision to be childfree as one that wasn’t always straightforward, linear, or even at times, coherent. One of the women, Mary, spoke of this ‘precariousness’ in terms of percentages: ‘We might have a really nice interaction with a child and you’re like “ooh I’m seventy-five percent today” or “sixty percent today”’ (see Box 2 .6 below). Thus, following a thorough coding, we did have evidence of a prevalent pattern across the data (a large number of data extracts across different interviews), and therefore a potential candidate theme – this addressed an important aspect of the lived experiences of childfree women. Two other candidate themes we generated also related to the lived experiences of childfree women: first, that children would interfere with the freedoms and quality of life the women enjoyed; second, that living in a pronatalist society meant they experienced marginalisation, but simultaneously, they would often deny any explicit stigma. All three themes spoke, centrally, to our research question, which was to ‘explore the lived experience of voluntary childlessness across the life course for a diverse group of women’. A fourth candidate theme ‘not a maternal bone in my body” (see Table 2 .4) captured a prevalent meaning, but proved not to be of value to the research question, because it related more to women’s initial reasons for not having children, rather than to their current lived experiences of being childfree. This discrepancy was addressed as we reviewed and defined our (candidate) themes.
Reviewing and defining themes (phases 4–5)
