Article

Sociological Methods & Research

2021, Vol. 50(2) 708-739
Flexible Coding of ª The Author(s) 2018
Article reuse guidelines:
In-depth Interviews: sagepub.com/journals-permissions
DOI: 10.1177/0049124118799377
A Twenty-first-century journals.sagepub.com/home/smr

Approach

Nicole M. Deterding1 and Mary C. Waters2

Abstract
Qualitative coding procedures emanating from grounded theory were lim-
ited by technologies of the 1960s: colored pens, scissors, and index cards.
Today, electronic documents can be flexibly stored, retrieved, and cross-
referenced using qualitative data analysis (QDA) software. We argue the oft-
cited grounded theory framework poorly fits many features of contemporary
sociological interview studies, including large samples, coding by teams, and
mixed-method analysis. The grounded theory approach also hampers
transparency and does not facilitate reanalysis or secondary analysis of
interview data. We begin by summarizing grounded theory’s assumptions
about coding and analysis. We then analyze published articles from American
Sociological Association flagship journals, demonstrating that current con-
ventions for semistructured interview studies depart from the grounded
theory framework. Based on experience analyzing interview data, we suggest
steps in data organization and analysis to better utilize QDA technology. Our
goal is to support rigorous, transparent, and flexible analysis of in-depth
interview data. We end by discussing strengths and limitations of our
twenty-first-century approach.

1
Institute for Research on Poverty, University of Wisconsin–Madison, Madison, WI, USA
2
Harvard University, Cambridge, MA, USA

Corresponding Author:
Nicole M. Deterding, Institute for Research on Poverty, University of Wisconsin–Madison, 1180
Observatory Drive, Madison, WI 53706. USA.
Email: [email protected]
Deterding and Waters 709

Keywords
in-depth interviews, qualitative coding, qualitative data analysis software,
grounded theory, sociological research methods

Well into the twenty-first century, most qualitative research training is still
either consciously or unconsciously based on grounded theory—a model
designed in the 1960s. Many books or articles based on in-depth qualitative
interviews begin the methodological section with a citation to this classic
approach. Grounded theory, developed by Glaser and Strauss (1967) and
elaborated by Glaser (1992), Strauss (1987), Strauss and Corbin (1990), and
Charmaz (2000), provides a set of steps for conducting and analyzing
qualitative research. Although often cited, a true grounded theory approach
is less common among sociologists doing interview studies today. Its cen-
tral prescriptions—theoretical sampling toward saturation, strongly induc-
tive analysis, and full immersion in the research field—bear little
resemblance to the actual methods used by many large-scale interview
researchers. And yet, grounded theory continues to have an enormous
influence on how qualitative research design and interview coding are
taught in graduate programs. How should twenty-first-century interview
researchers proceed?
In this article, we argue that the coding procedures emanating from
grounded theory were limited by the technology available to researchers at
the time: paper, scissors, index cards, and colored pens. In a path-dependent
process, open descriptive coding aggregating to conceptual abstraction
shaped the features of computer assisted qualitative data analysis (QDA)
software. Despite the fact that electronic documents can be flexibly stored
and snippets of text easily cross-referenced and retrieved, the result is that the
procedures outlined by Glaser and Strauss were turned into algorithms for
how to do qualitative research. Indeed, most sociologists reporting the use of
QDA software in their published work appear to do little more than apply
virtual sticky notes and sort piles of electronic note cards (White, Judd, and
Poliandri 2012).
Reliance on the twentieth-century approach poses a number of problems
for today’s researchers. First, most researchers have many theoretical ideas
and concepts they will apply to a single set of data. Codes that work well for
one chapter of a book or dissertation are not necessarily the ones a scholar
wants for a different substantive chapter or article, and grounded theory
710 Sociological Methods & Research 50(2)

coding does not easily allow for flexible reanalysis. Second, coding line by
line takes a great deal of time and effort before one establishes a set of codes
that can be applied to the entire set of data; establishing the reliability and
validity of coding is a different challenge when using software over hard-
copy coding. Third, the grounded theory approach was developed based on
projects with a relatively small number of interviews, generally conducted,
and analyzed by the researcher himself or herself. Increasingly, qualitative
studies involve large numbers of interviews, often numbering near 100.
While reliably and validly applying thematic codes across 80–100 interviews
is a logistical problem for even a single researcher, today’s large-N studies
often frequently involve teams of people who interview, code, and write
separate papers based on a shared pool of material.
Finally, structuring data using many small codes does not easily facilitate
analytic transparency or secondary analysis by researchers not involved in
the initial data collection. More and more, qualitative researchers are
exhorted to clearly communicate the logical steps in their data analysis, yet
conventions for doing so have yet to appear. Relatedly, norms of open sci-
ence—such as data archiving and secondary analysis—that are taking hold in
quantitative social science have largely yet to make it to qualitative research.
While secondary analysis of interview data is an opportunity that is rarely
taken advantage of today, we believe it should be encouraged. Other
advances in technology—such as secure servers—enable researchers with
institutional review board (IRB) permission to analyze existing qualitative
data, even if the researcher may not be at the same institution as the data.
However, to date, there is little methodological guidance on how to plan for
or undertake secondary analysis.
In this article, we briefly describe the assumptions involved in the
grounded theory approach to coding and analysis. Next, we analyze the
methodology sections of in-depth interview studies in American Sociolo-
gical Association (ASA) journals to highlight their mismatch with current
conventions for semistructured interview research in sociology. We then
propose a set of procedures—which we call flexible coding—that flips the
script to take advantage of modern QDA technology. Rather than limit
ourselves to how most qualitative interviewers are taught to code or to how
beginning qualitative researchers think they ought to approach the craft, our
goal is to better reflect how one can practically go about analyzing large-
scale interview data. We end by weighing the strengths and limitations of a
twenty-first-century approach to qualitative analysis compared to the pre-
vious approaches.
Deterding and Waters 711

Background
Grounded Theory: Origins and Approach
In 1967, two sociologists, Barney G. Glaser and Anselm L. Strauss, pub-
lished The Discovery of Grounded Theory: Strategies for Qualitative
Research. Building on Znaniecki’s (1934) “analytic induction,” their book
promoted a systematic, inductive approach to qualitative research, suggest-
ing qualitative researchers should abstract conceptual relationships from data
rather than deduce testable hypotheses from existing theory. After publica-
tion, Glaser and Strauss came to disagree about central features of the method
and parted ways. They each have been active since (Strauss until his death in
1994), publishing competing guides to grounded theory (Glaser 1998, 2005;
Strauss 1987; Strauss and Corbin 1990).
Much ink has been spilled explicating their variations on a theme, but the
primary divergence is that Glaser stresses induction as the core of the
approach, ultimately going so far as to advocate that researchers avoid any
literature relevant to the subject being studied until after the interviews are
complete and initial coding of the data is done (Glaser 1998). While main-
taining a focus on induction, Strauss became more concerned with validation
and systematic procedure. Most modern variants of grounded theory favor
Strauss over Glaser in their attention to the particular words of study respon-
dents. Charmaz (2006), Bryant (2002), and Bryant and Charmaz (2007) built
on Strauss, developing what they label “constructivist grounded theory.”
This approach stresses the coproduction and construction of concepts and
interpretations by researchers and participants, recognizing their different
positions, roles, backgrounds, and values.
There are most likely several reasons for the outsized influence of
“grounded theory” in qualitative interview training. Glaser and Strauss
offered an early attempt to systematically describe how qualitative research
ought to be conducted. Their grounded theory offered a “scientific” prescrip-
tion for what midcentury positivists had diagnosed as the biased, impressio-
nistic, and anecdotal condition of qualitative research. The inductive
approach begins with the data itself and exhorts the researcher to produce
concepts and theory from it, promising theory generation rather than mere
theory testing. Later, as broad questions of epistemological authority roiled
the social sciences, the method was easily adapted to emphasize the cate-
gories of meaning offered by respondents themselves (Charmaz 2006). This
combined practical and epistemological appeal remains even today. By pro-
viding step-by-step directions, the grounded theory approach can be taught to
students, including applied researchers, regardless of their disciplinary
712 Sociological Methods & Research 50(2)

background. The approach also likely remains cited because most qualitative
research has at least an element of induction. However, many scholars have
speculated that researchers cite Glaser and Strauss merely to imply that they
took an inductive approach.
Based on our informal discussions with contemporary researchers using
large-scale interview data, it appears that few actually implement the unfold-
ing data collection suggested by a grounded theory approach: theoretical
sampling toward conceptual saturation. Practical demands of modern aca-
demic life, including adhering to the requirements of IRB, grant writing to
secure research funding, balancing competing professional demands and
multiple projects, working with research teams, and coordinating transcrip-
tion of large samples all run contrary to the image of a researcher wholly
immersed and spontaneously pursuing their interview project. This is one
way that what interview researchers actually do departs from the methodo-
logical guidance.

Coding Interviews in Grounded Theory

Another departure between the methodological literature and everyday prac-
tice is the process of coding, which is our particular focus. Coding is the
actual work of analyzing the interview data that have been collected, and it is
rarely described in detail in publications. Often researchers will write that
they “coded the data using grounded theory techniques,” that “data were
transcribed and coded,” or that they “coded the data with ATLAS.ti”. Coding
is the most laborious and time-consuming part of interview research. Yet it is
granted little space in published academic articles and books. It is not sur-
prising, then, that coding is the most mysterious part of the research cycle for
many beginning researchers. Glaser and Strauss outlined a procedure for
coding that follows their inductive approach. Strauss and Corbin provide
an even more detailed set of steps involving three major stages. Charmaz
(2006, 2014) outlines the same three stages but uses different terminology. In
all three models, coding begins by describing small parts of the data: lines,
sentences, and even paragraphs or words.
In each of the approaches derived from grounded theory, the advice to
researchers is clear: begin by generating many codes that reflect only the
data. Later on, cull the list of codes by deleting some and combining others.
When at all, it is only in late stages that the analyst should connect the data to
previous research, including to the theories and concepts in the preexisting
literature on the subject.
Deterding and Waters 713

This approach to coding makes some sense when we imagine that the data
were printed in hard copy; the codes that were applied were written in the
margins of the transcripts; and the resulting categories were cut up into bits of
paper, taped onto index cards, and then sorted into thematic piles. However,
the widespread development of word processing in the early 1980s and the
development of qualitative data analysis software soon after have given
qualitative researchers new tools to do their research. It is unfortunate, how-
ever, that these tools have not led to a fundamental rethinking of how coding
is actually performed. This is what we propose.
Data reduction is a fundamental part of qualitative analysis (Miles and
Huberman 1994). When grounded theory was developed, chopping up the
data into small pieces was a very time-consuming, physical process. After
that came the stage of tossing aside material that was no longer needed and
then combining the small pieces into larger codes. An example would be
taking the pile of pieces of text labeled “attitudes about congressmen” and
putting it together with a pile of text labeled “attitudes about mayors” and
then labeling this new pile “attitudes about elected officials.” The initial level
of abstraction mattered deeply: If the researcher started by cutting up the
transcripts into piles with broader categories—attitudes toward elected offi-
cials, for instance—when she decided to analyze mayors differently than
congressmen, transcripts would need to be cut up again. It is easier when
dealing with paper to sort codes and chunks of text in different ways to see
how they come together, rather than to look at large chunks of text more
broadly to determine how the data should be split. In short, the idea of
starting with tiny bits of text and many codes and aggregating up made a lot
more sense than starting with big chunks of text when researchers were doing
it all by hand, with little piles of paper on their dining room tables. This is no
longer required.

Epistemology of Grounded Theory

The second reason that grounded theory advises starting with small codes
and aggregating is that Glaser and Strauss operated from an epistemology
that almost pathologically advocated inductive research. Wacquant (2002)
has colorfully characterized this approach as “an epistemological fairytale”
(p. 1481). Others have less-colorfully noted that this is not how actual
researchers work.
While other types of qualitative research—such as ethnography—may
begin data collection with a broad research site rather than a set research
question, the vast majority of interview researchers do not design their
714 Sociological Methods & Research 50(2)

interview protocols without connecting to previous literature on the topic. It

would be absurd to invest all of the time and money in conducting an
interview study without knowing whether it had already been done, what
were the main findings, and what remaining theoretical puzzles and empiri-
cal gaps needed to be addressed. As such, sociologists who do interview
studies typically know a great deal about their topics, even before entering
the field. And for the most part, their interview schedules are structured or
semistructured. The questions are open ended to be sure, but they generally
follow a logical order designed to create a conversation, put respondents at
ease, build trust, and importantly—focus the discussion on the researcher’s
questions, not just what the respondent feels like talking about. As a result,
the questions that are asked in an in-depth interview reflect the researchers’
hypotheses about the phenomena under study, the previous research that has
been done on the subject, and the hunches or findings the researcher has
learned as the research project has progressed.
While Glaser and Strauss (and Charmaz) stress the inductive nature of
grounded theory to emphasize that researchers should learn from their
respondents, allowing for surprising results and phenomena to emerge, con-
temporary qualitative sociology privileges cumulative theory building. Miles
and Huberman (1994) developed a sourcebook of qualitative methods that
allowed more theory-driven research, primarily stressing data display
matrices as a way of advancing coding and analysis of transcripts. Timmer-
mans and Tavory (2012) suggest that the best qualitative researchers are
sensitized by the previous research early on; their projects are not wholly
inductive. At the same time, one should remain very open to surprising or
unexpected findings; projects are also not solely deductive. They define this
approach as abductive, suggesting good qualitative researchers adopt:

[a] qualitative data analysis approach aimed at theory construction. This

approach rests on the cultivation of anomalous and surprising empirical find-
ings against a background of multiple existing sociological theories and
through systematic methodological analysis. As such, it requires a fundamental
rethinking of core ideas associated with grounded theory, specifically the role
of existing theories in qualitative data analysis and the relationship between
methodology and theory generation. (p. 169)

We believe this description more accurately reflects what many sociolo-

gists conducting in-depth interview studies aim to do. Yet, while Timmer-
mans and Tavory have proposed a useful epistemological substitute for
grounded theory, they have not provided advice about how the actual process
Deterding and Waters 715

of coding could support their recommended move away from grounded

theory suppositions.
Enter QDA Software
When QDA software was developed in the 1980s, software developers fol-
lowed the only accepted guidelines for qualitative analysis in existence—
grounded theory. Thus, the many different software packages reflect these
historical roots. The software allows researchers to code line by line, to write
memos while coding, and to combine codes into broader more abstract codes.
Most packages also produce visual models that relate the codes to build a
conceptual model. Weiss (1994) describes the low-tech ways in which qua-
litative researchers developed these visual models for data reduction before
QDA:

Many researchers find visual display of the elements of their story a valuable
means of achieving both local integration and inclusive integration. Miles and
Huberman recommend diagrams on paper, with lines linking related issues, to
display graphically the conceptual framework of the final report. Becker sug-
gests putting data and memos about the data on file cards that can then be
spread on a large flat surface and arranged and rearranged until they achieve a
logical sequence. Agar suggests finding an empty classroom full of black-
boards on which can be drawn maps of concepts and their interrelations.
(p. 162).

Slowly, however, QDA software packages have developed into more

powerful and versatile tools that reflect their early origins but also better
support researchers in what they actually do when they analyze data. We
argue below that the best use of these tools for coding data is exactly the
opposite from the steps outlined by the grounded theorists. Researchers
should get to know their data by indexing transcripts with broad codes that
reflect the questions they asked in their interviews and the concepts they
sought to examine as they went into the field. By writing analytic memos
during this process, they can generate conceptual themes that appear across
cases. Then, capitalizing on the existing index for data reduction, they can
apply more fine-grained codes to subsections of the interviews. Finally,
analytic coding can be examined across respondent attributes to document
the depth of textual evidence for the claims about key relationships the
analyst makes. We believe this method of coding can facilitate reliability,
validity, and transparency; allows the researcher to capitalize on features of
the software beyond the simple retrieval of quotes; and will provide better
716 Sociological Methods & Research 50(2)

opportunities for team work, complex projects, secondary analysis, and repli-
cation studies.

The State of the Field

We came to our recommendations out of necessity. As a graduate student,
Deterding joined Waters’s in-progress study of low-income parents’ recov-
ery following 2005’s Hurricane Katrina, the Resilience in Survivors of
Katrina (RISK) project (www.riskproject.org), leading a team of graduate
students in coding the interview data. RISK includes four waves of long-
itudinal survey data (N ¼ 780) and three rounds of life history interviews
with a subset of survey respondents (N ¼ 125 interviews; 30 respondents were
interviewed twice). The project was overseen by a multidisciplinary team of
principal investigators (sociology, psychology, and economics), and the data
needed to be accessible to researchers across three universities. Because of the
mandates of Federal funding sources—and the historical importance of the
topic—data archiving for secondary analysis was a concern from the start of
the project (Waters 2016). While the RISK project may be of unusual size and
scope, including the opportunity to match survey responses to interview tran-
scripts (Deterding 2015a; Morris and Deterding 2016), many of the challenges
we faced are common to large-scale interview studies.
Before the RISK study, Waters had conducted four other increasingly
large qualitative interview projects. Her first project, a study of ethnic iden-
tity, was based on 60 interviews and the analysis was conducted with typed
transcripts, scissors, and index cards (Waters 1990). A second project
included 120 in-depth interviews on the topic of West Indian immigrant
identity and was analyzed using word processing software (Waters 1999). In
the early 2000s, Waters codirected a large study of second-generation young
adults in New York City that included 333 in-depth interviews and was ana-
lyzed using an early version of ATLAS.ti (Kasinitz et al. 2008), then in 2011,
she codirected a study of young adulthood in New York, Iowa, San Diego, and
Minneapolis based on 437 life history interviews. This project was also ana-
lyzed using ATLAS.ti and involved a small army of researchers (Waters et al.
2011). In each case, the analysis of interviews followed a very different logic
and flow than that suggested by grounded theory.

Current Conventions in Sociology

It is not only our own work that produces large volumes of interview data. To
document broader disciplinary convention in U.S. sociology, we compiled a
data set of all interview-based articles published in the two flagship journals
Deterding and Waters 717

of the discipline (American Sociological Review and the American Journal of

Sociology) and three ASA-sponsored subfield journals (Journal of Health
and Social Behavior, Sociology of Education, and Social Psychology Quar-
terly) between January 2010 and December 2015. These pieces are listed in
Online Appendix (which can be found at http ://smr.sagepub.com/supple-
mental/). While noting that many interview-based studies are published as
books and that specialty qualitative journals and regional or subfield journals
regularly publish interview-based studies, we suggest these leading journals
reflect what is rewarded in the mainstream of the discipline, particularly by
general audiences.
The goal of this analysis was to assess the accuracy of our claims about
current disciplinary convention. For each article, we recorded the journal,
year, author, title, number of interviews, any additional modes of data col-
lection, and the authors’ textual descriptions of the procedure they followed
for data coding and analysis. In total, we identified 98 articles that included
semistructured interviews as a data source. Analysis of these articles con-
firms that twenty-first-century qualitative researchers generally conduct
studies with large numbers of interviews, are very likely to use mixed-
methods approaches combining interviews and other data sources, provide
very little information on how they coded their data, and nevertheless con-
tinue to cite or gesture toward grounded theory as their methodology.

Large Ns
The main consensus among interview methodologists regarding how many
interviews a researcher should plan is “it depends” (Baker and Edwards
2012). Yet we hypothesized that the typical number of interviews in socio-
logical studies has expanded far beyond what would be required for a
grounded theory threshold of conceptual “saturation” (Glaser and Strauss
1967; for an empirical test suggesting the number may be as low as 12, see
Guest, Bunce, and Johnson 2006). Exactly how much does practice diverge
from what is recommended by this benchmark of grounded theory?
Table 1 reports information on the number of studies including semistruc-
tured interviews by journal. The minimum, median, and maximum number
of interviews is reported. Overall, only 19 articles (19.4 percent) were based
solely on interview data. Consistent with the idea that that a single data
collection often yields multiple analyses, several studies appeared in more
than one publication. One study was represented three times (Calarco 2011,
2014a, 2014b), and five were represented twice (Ispa Landa, Rivera, Teeger,
Turco, and Wilkins). As far as could be ascertained from the methodology
718 Sociological Methods & Research 50(2)

Table 1. Characteristics of American Sociological Association–Sponsored Journal

Articles Reporting Semistructured Interviews (2010–2015).

Studies N Interviewsa

Interviews Min.
Journal All Only (interview only) Median Max.

American Journal of Sociology (AJS) 29 4 36 (50) 85 208

American Sociological Review (ASR) 28 5 12 (71) 55 170
Journal of Health and Social 7 1 18 (46) 46 137
Behavior (JHSB)
Social Psychology Quarterly (SPQ) 15 5 17 (25) 35 189
Sociology of Education (SoE) 19 4 12 (54) 53 170
Total 98 19 12 (25) 55 208
a
N interviews excludes three studies for which the number of interviews analyzed was unclear:
ASR (0), AJS (2), JHSB (0), SoE (0), and SPQ (1). These were historical case studies, ethnographic
fieldwork that included a semistructured interview component, or repeated interviews within a
stated number of respondents.

discussions in the articles, only three of the studies (Brown 2013; Collett,
Vercel, and Boykin 2015; Deterding 2015a) included secondary analysis of
semistructured interview data.
While qualitative methodologists will rarely state a hard cutoff for the
number of required interviews, these data demonstrate that large studies end
up published in the core disciplinary journals. The number of interviews
ranged from 12 to 208, with only 19 studies containing fewer than 30 inter-
views. The median number for the whole set is 55, and interview-only studies
generally had larger samples than studies that combined data sources. One in
four articles (24 total) reported on 100 or more interviews. Clearly, a sub-
stantial portion of contemporary qualitative researchers face a large amount
of interview data to handle during analysis.

Opaque Coding
Table 2 reports information coded from authors’ descriptions of their coding
process. By and large, the type of information included in descriptions of the
coding process is not standardized, either within or across journals. The most
commonly cited methodologists were the grounded theorists and their suc-
cessors: Charmaz, Glaser and Strauss, and Miles and Huberman. More than
half of the articles (55 percent) did not cite a specific methodological text,
instead using terms that are drawn from these authors: “inductive,”
Deterding and Waters 719

Table 2. Descriptions of Coding/Analysis in Studies Using Semistructured Interview

Data (2010–2015).

No Describes
Methodologist Coding Mentions
All Cited Procedure QDAS
Studies
Journal N N Percent N Percent N Percent
a
American Journal of Sociology 29 21 72.4 20 68.9 9 42.8
American Sociological Review 28 16 57.1 23 82.1 11 39.2
Journal of Health and Social 7 4 57.1 7 100.0 4 57.1
Behavior
Social Psychology Quarterly 15 7 46.7 13 86.7 4 26.7
Sociology of Education 19 7 36.8 18 94.7 11 57.9
Total 98 55 55.1 61 82.6 30 39.7
a
In American Journal of Sociology several studies used in-depth interviews as part of a historical
case method approach. None of these mentioned coding, but it’s a different method of analysis.

“emergent,” “open, “axial,” “focused,” or “thematic.” A handful of studies

adopted particular approaches such as “frame analysis,” “narrative analysis,”
or “semiotic techniques” or reported drawing their coding categories from
prior literature. Only a minority of articles (40 percent) explicitly mentioned
use of QDA software (typically NVivo or ATLAS.ti). Nearly, 20 percent of
articles reporting semistructured interview data did not discuss their coding
procedure at all.

Why So Large?
We’d briefly like to speculate on why sociological interview projects are
typically so large. Given that contemporary sociology privileges describing
mechanisms, many sociologists today design their work to highlight con-
trasts between groups of respondents or research sites. By designing a study
across hypothesized salient differences, scholars are able to construct argu-
ments that illuminate the relationship between context and concept. This
approach requires enough respondents in each “cell” to establish patterns
of difference by group or site. Studies designed this way produce large sets of
textual data, which presents logistical challenges for the researcher during
the analysis phase. We believe studies examining the contours of group
difference are particularly well suited to the coding procedure we outline
below.
720 Sociological Methods & Research 50(2)

Flexible Coding: Outlining Twenty-first-century

Analysis
We describe a method of “flexible coding” that is better suited than grounded
theory to the realities of analyzing in-depth interviews with QDA software.
While emerging methods push computerized analysis further, including
automated analysis of textual data (Bail 2016; Chakrabarti and Frye 2017),
we do not formally incorporate quantitative text analysis here.
Our techniques, we believe, may not be entirely new to many seasoned
researchers, as the process reflects both what we discovered as useful for
ourselves and, loosely, what many peers have told us that they do. Our goal in
outlining the method is to bring the craft of twenty-first-century interview
coding into the light, reflecting how sociologists with large interview sam-
ples can set themselves up for rigorous, flexible analysis.

Assumptions of Our Approach

While there is no shortage of how-to guides for popular QDA software
packages, they largely sidestep disciplinary epistemologies in favor of broad
applicability to applied research settings (e.g., Bazeley and Jackson 2013).
As demonstrated by the analysis of published articles above, this work has
not produced a common language or set of conventions for sociologists.
We take as our starting point that the arguments qualitative sociologists
make should be soundly supported by the data at their disposal and that the
published research product should transparently reflect that. Other assump-
tions derive from common features of empirical sociology based on in-depth
interviews:

 A semistructured interview protocol with a relatively large number of

interviews (N > 30).
 Some combination of induction and literature- or theory-based cod-
ing; research that is not completely inductive, even in the first steps of
analysis.
 Comfort with the analytic process of relating variables (independent/
dependent) to one another.

In other words, we start from the premise of an empirical qualitative

sociology that is in dialogue with existing theory and findings from previous
studies, including quantitative research. It may be the case that a priori
categories actually have no importance to respondents, are not relevant to
the research setting, or are not the main factor driving the phenomenon we
Deterding and Waters 721

are trying to explain. Evidence for such an argument is produced by system-

atically examining the text against these categorizations. Our approach may
not be well suited to scholars who completely reject variable-based empirical
work.
Other scholars have devoted substantial energy to establishing contempo-
rary definitions of “rigor” in the design and collection phases of qualitative
projects, including two National Science Foundation workshops (Lamont
and White 2009; Ragin, Nagel, and White 2004). There has been no similar
update for the craft of qualitative data analysis. We focus below on specific
steps to help researchers organize data and develop an argument once the
interviews have been conducted and transcribed. It is beyond the scope of
this piece to delve into ongoing conflicts among sociologists over the epis-
temological limits of interview research (see Sociological Methods &
Research Forum, Jerolmack and Khan 2014, and responses). Finally, analy-
sis based on a small number of interviews or using highly unstructured
protocols (such as phenomenological interviewing) are probably not well
suited to these methods. When there are very few interviews or researchers
have a narrow scope to their analysis, it might not make sense to make deep
use of QDA, as the investment of time and money will likely not yield much
beyond what one can do using word processing documents or paper and
pencil to find and retrieve data.
Given these parameters, what is a twenty-first-century approach to qua-
litative analysis? At minimum, we believe that researchers should be able to
communicate the logical steps underpinning their argument and report these
as they write up their study findings. How did they generate and operatio-
nalize their concepts? How did they verify the conceptual relationships they
propose? Did they attempt to rule out alternative explanations or explain
negative cases? The answers to these questions should be included in any
discussion of how a researcher reaches a conclusion in published work. It is
now rarely if ever discussed. The potential of QDA software to aid these
tasks remains largely untapped, at least based on what is typically reported in
the Methods sections of books or peer-reviewed articles.

A Note on Choosing QDA Software

QDA technology is rapidly evolving. The computer-assisted QDA network-
ing project in the United Kingdom (http://caqdas.soc.surrey.ac.uk) is an
excellent—and regularly updated—source for comparable reviews of and
guidance on choosing QDA software. In 1995, they listed 24 QDA packages
supporting a variety of specialized approaches to analyzing qualitative data
722 Sociological Methods & Research 50(2)

(Weitzman and Miles 1995). As of this writing, the site lists 10. Common
tools across the set of programs include the ability to write and track memos;
index or code data with thematic or conceptual labels; add demographic or
other categorical information to compare subgroups; run searches or queries;
develop visual models or charts; and generate reports or output from the data
(Lewins and Silver 2007). That major software options share these features is
evidence of the fact that, over time, QDA software has “simultaneously
become more comprehensive, more applicable to a diverse range of meth-
odologies, and more homogenous” (Bazeley and Jackson 2013:6). We ela-
borate our approach using NVivo Version 10, though we have substantial
experience with several iterations of ATLAS.ti and some experience with the
first version of Dedoose. We are confident that a similar approach can be
used with these programs. Any software package with the major capabilities
above should be able to facilitate our approach to organizing and analyzing
interview data. Based on our analysis of published in-depth interview studies
appearing in ASA journals above, ATLAS.ti and NVivo appear to be the
most commonly used among sociologists.

QDA for In-depth Interviews

Below, we discuss key considerations for designing a QDA database and a
three-step process for flexible coding and analysis. Codes are of three
types—attributes, index codes, and analytic codes. After applying these
codes, it is possible to use the software to test and document the key relation-
ships in the data.
We suggest that analysis begins with getting a handle on the big picture(s)
in the data. Generally, our first step can be thought of as data exploration and
preparation. It involves indexing the transcripts, anchoring content to the
interview protocol. During this phase, the researcher will also produce
respondent-level and cross-case memos, beginning to document the analytic
process by developing hypothesized relationships between concepts. In stage
2, the researcher will use the index for data reduction, applying analytic
codes to focused sections of the transcript, prioritizing reliability and validity
of the coding. Analytic codes represent the concepts to explore in a single
paper or book chapter and integrate emergent findings with what is known
from the literature. In stage 3, the researcher will use the software’s tools for
conceptual validation, model building, and the testing and refinement of the
data-based theory.
In our recent work, we have used this process for organizing qualitative
analysis with a research team, which included scholars from different
Deterding and Waters 723

disciplines writing about many separate topics. The RISK qualitative data
have been used to examine physical and mental health outcomes of Katrina
(Bosick 2015; Lowe, Rhodes, and Waters 2015; Morris and Deterding 2016);
experiences of racism during the hurricane and its aftermath (Lowe, Lustig,
and Marrow 2011); changes in marital and partner relationships (Lowe,
Rhodes, and Scoglio 2012); residential choices and social mobility (Asad
2015); posthurricane (im)migration and race relations (Tollette 2013); and
educational planning and return (Deterding 2015a). Other methodological
treatments of team-based coding focus on the intercoder reliability of analy-
tic codes in a single study, which is noted as a very time-intensive iterative
process (Campbell et al. 2013; Price and Smith 2017). We present an alter-
native model for teamwork in qualitative analysis. Because our initial
“indexing” step is broad rather than fine-grained, the task can be more easily
distributed among members of a research team than can line-by-line abstrac-
tion. Once the index is established, either an individual analyst or a research
team can proceed to the analytic stage, using the software’s capacity for data
reduction to enhance the reliability and validity of analytic codes. These
steps are further discussed below.

Basic Organization: Transcription and Database Design

To effectively use QDA technology, first collect all of the source materials
and design the database that will house them. Setting up for flexible analysis
requires importing transcripts and any related materials (audio recordings;
postinterview field notes or memos). It also includes connecting these doc-
uments to attributes or the conceptual categories that guided the research
design.
Attention to the particularities of QDA analysis before and during tran-
scription can save a lot of time. Some basic organization will make life easier
when dealing with a large number of interviews, whether or not transcription
is done by the researcher(s). For instance, it is a good idea to institute uniform
naming conventions of files. If documents are named with respondent ID
numbers (e.g., “001-Transcript” for the first respondent’s transcript), many
software programs will import the document name as an anchor for attaching
respondent-level attributes. This automation will save a lot of time and
clicking if respondents are the unit of analysis. Using respondent ID or a
pseudonym instead of “respondent” in the body of the transcript will make it
easier to identify excerpted text during data analysis. And if the project has
multiple interviewers, we recommend using the interviewer initials in the
transcript. This will also allow the researcher to know who is talking when
724 Sociological Methods & Research 50(2)

dealing with reduced data and facilitate an examination of interviewer effects

(see Lowe et al. 2011).
If the researcher(s) will transcribe the interviews, it is possible do so from
within the QDA program. In NVivo, there is a transcribing function that
allows audio replay, setting the speed, and pausing and rewinding, much
as one would use a transcribing machine. This is helpful because some of
the very best ideas about what the data mean can arise in the middle of the
drudgery of transcribing. Transcribing within the program enables synchro-
nizing audio or video with the transcription in order to easily revisit the audio
while reading the transcript. It may be helpful to ask questions such as: What
was the respondent’s tone of voice when they said this quote? Were they
laughing, being sarcastic, or crying? Sounding strong or weak? While a deep
discussion about the value of revisiting audio is beyond our scope here, the
actual voices of respondents can be helpful for the duration of the analysis if
the challenge of locating the appropriate section is overcome. If transcription
is outsourced, it may be worth paying to include periodic time stamps in the
transcripts.
Interviewers may have written field notes immediately after conducting
the interview. These documents should be imported into the database and
linked to the respondent so that they can easily be accessed; they represent
the beginning of respondent-level memos. If the researcher(s) do the tran-
scribing, it’s a good idea to add to these memos by inserting additional
thoughts about the interview case at the moment they occur and linking these
memo notes to the appropriate part of the transcript.
Finally, beginning to link documents to attributes is a fundamental part of
QDA database setup. Attributes are the salient personal characteristics of the
interviewees that played a role in the study design. These may be demo-
graphic variables such as age, gender, race, occupation, socioeconomic sta-
tus, or specific contextual data such as experimental group or control group,
state, or neighborhood of residence. To use the language of variables, these
are the demographic and structural independent variables that were posited to
shape the qualitative topic of the study. For example, in the New York
Second-generation Study, Waters and colleagues sampled second-
generation young adults whose parents had immigrated from five sending
areas—China, Dominican Republic, South America, West Indies, and the
former Soviet Union, along with young adults with native-born parents who
were white, black, and Puerto Rican. These ethnic and racial origins were
assigned as attribute codes to each respondent and later used in analyses that
examined different conceptions of success in life. In this particular case,
ethnic origin was the independent variable (and an attribute code attached
Deterding and Waters 725

to each individual) and attitudes toward success (an analytic code applied to
the text later) was the dependent variable (Kasinitz et al. 2008).
In mixed-methods data, attributes can include person-level data from
respondent screening sheets, demographic information sheets, or surveys.
For instance, using the RISK data, we connected longitudinal survey
responses to transcripts using NVivo Version 10 “classification sheets,”
allowing us to describe qualitative mechanisms for mental health differences
among survey respondents with proximal and distant social networks (Morris
and Deterding 2016). Attributes may not be limited to person-level charac-
teristics; other units of analysis are possible. For instance, if the project has
multiple research sites, “site” can be designated as an attribute, allowing the
analyst to query the transcripts and memos to examine thematic differences
between sites. If this information is not available in a demographic informa-
tion form, the researcher will need to record respondent attributes when
reading the interview text and applying index codes (discussed below).
Why is it important to connect transcripts with attributes? QDA’s query-
ing capabilities rely on the intersection of codes, and attributes are codes that
are applied to the entire transcript. Attributes are applied by linking docu-
ments to a classification sheet (NVivo version 10), descriptors (Dedoose), or
a primary document family (ATLAS.ti). In each of these software options,
documents can be assigned attributes manually from within the software.
However, with more than a handful of respondents, attributes are more
quickly and reliably imported via a spreadsheet that includes the respondent
ID. Because of this, we recommend recording attributes for all of the respon-
dents in a spreadsheet and then importing this to the database. The work of
identifying attributes can be done while sampling and interviewing is
ongoing or at any point afterward. Although it is likely that other important
conceptual categories warranting inclusion as an attribute will arise during
analysis, identifying and coding initial respondent attributes is an important
part of early data preparation. Querying the intersection of attributes and
analytic codes will reappear in the third stage of the data analysis, when
testing the robustness of the theoretical claims.

Stage 1—The Big Picture: Indexing and Memos

Our formal analytic procedure begins once the interview transcripts are
complete and the database is set up. How does one begin to apply codes to
what may be hundreds or even thousands of pages of text? We believe that
early reads of the transcripts play an important role in generating concepts
and theories—identifying the main “stories” in the data. From conducting the
726 Sociological Methods & Research 50(2)

interviews, researchers may have a provisional idea of emergent themes

across them. They may also want to explore relationships hypothesized by
prior literature. At the beginning of the reading, however, these concepts
likely will not be refined enough to apply analytic codes across a large
number of interviews with any degree of reliability. This will be especially
the case if one is part of a qualitative research team with multiple inter-
viewers and coders, though it is likely true even if a single researcher wrote
detailed memos in the field or during the transcription process. Coding
reliability is of central importance to responsible use of QDA software.
Textual coding should begin by applying index codes to the text, building
out respondent memos, and beginning cross-case conceptual memos. The
inverse of line-by-line coding, index codes represent large chunks of text,
enabling data reduction and retrieval as the analyst proceeds through con-
structing and documenting their argument. Setting the data up this way
allows subsequent rounds of reading to be more focused and analytic coding
to be more reliable. A logical starting place for designing the index is the
interview protocol. Include a code for each question on the structured or
semistructured interview schedule. If the interview protocol is organized
with questions inside of broader topics, it is possible to aggregate question
codes to the topic level as well. While most sociological interviews are
semistructured, if interviews are very structured, it is possible to take advan-
tage of the MS Word subheading structure to label each question as a new
subheading. Then, when the documents are imported into NVivo, the pro-
gram can autocode the text, applying the question label to the text that
immediately follows.
In addition to helping the analyst become familiar with the data through a
deep reading of the transcript, indexing takes advantage of the transcripts’
natural organization to enable easy text location and retrieval. In a life history
interview, for example, the protocol may begin by asking about childhood,
then neighborhood, and then schools; each section may have several ques-
tions or prompts. The index would have “childhood” as the major heading
and would include answers to several specific questions in the protocol, each
with their own code. The answers to the question “Tell me about the neigh-
borhood you grew up in?” would receive the index code “childhood-
neighborhood”; “Who did you live with when you were growing up?” would
receive the index code “childhood-household.” Combining both would cre-
ate a broad “childhood” code.
Given that most sociological interviews are only semistructured, the task
of indexing is not as straightforward as it may first seem. To preserve con-
versational flow, the interviewer likely did not ask again questions that were
Deterding and Waters 727

answered in the natural course of the conversation. Software features such as

NVivo’s “matrix coding query” or ATLAS.ti’s “code-by-document matrix”
are extremely helpful for making sure that material for index codes is cap-
tured completely. These tools create a chart of all codes and transcripts,
showing which transcripts appear to be missing information. If, after the first
pass at indexing, it appears that a respondent has not answered a question in
the protocol, it is time to reread to make sure the answer was not simply
overlooked when naturally intertwined in the conversation.
Because it is based on the interview protocol, indexing represents coding
for the broad topics that were pursued in the research. In our study of Hurri-
cane Katrina survivors, index codes included topics such as “evacuation
experiences” and “physical and mental health” (Asad 2015; Waters 2016).
A study of immigration topics might include “migration experiences,”
“naturalization,” and “American identity.” These large topics likely include
pages of text spanning several different questions or even different points in a
given interview. The goal is to locate and link together answers to broad
content areas wherever these topics appeared during the interview.
While identifying respondent attributes and reading the text for indexing,
the researcher will begin to develop ideas about the transcripts. We agree with
Small (Lamont and White 2009:169) that it helps to think of qualitative inter-
views as a series of case studies. In reading each case, the researcher will
develop an idea of the important concepts and their linkages in the data—
provisional answers to the “how” and “why” questions at the center of the
research. These can be documented in respondent-level memos, which we
recommend writing for each interview as it is read and indexed. Whether
respondent memos begin immediately after conducting the interview or while
one analyzes cases upon completion of data transcription, cross-case analysis
of respondent memos will produce several ideas about “the story” in the data.
When passing through multiple cases in the first round of reading and coding,
we recommend compiling a list of concepts and relationships between them
that appear to describe multiple cases and begin describing the contours of
relationships in thematic memos. This is the beginning of cross-case analysis,
when we recommend starting to flesh out the analytic codes to apply in stage 2.

Great quote and “aha”. Throughout all stages of the coding process, research-
ers should take care to note chunks of text where respondents are particularly
concise, articulate, or poignant. Include or make a separate code for snippets
of text that trigger “aha” moments in understanding the data so that these are
easily retrieved later. When writing, queries of the overlap of “great quote”
or “aha” with analytic codes will identify quotes to include in the paper.
728 Sociological Methods & Research 50(2)

Our recommended first stage is in direct contrast to the grounded theory

approach, which would begin with line-by-line coding. Based on our own
and others’ experiences, we have concluded that using the software to apply
conceptual abstraction to the text is ill-advised at such an early stage. The
benefit of QDA software is that it is easy to apply as many codes as one
would like; the challenge of software is that it allows far more codes than is
either useful or reliable. Line-by-line coding requires much time is later spent
on aggregation, and coding this way produces a single story in the data,
making reanalysis difficult. This is why we suggest that the first read of the
text be used for indexing the data and developing contours of the concepts in
thematic memos; the themes and theories are not yet ready to be applied to
text in a reliable or valid way, and we do not recommend limiting analytic
options too early.
To allow for maximum flexibility, it is important to begin a list of analytic
codes—and their definitions—as ideas and insights occur during indexing.
However, it is best to wait to apply them to the data. Valid and reliable
coding is necessary for taking full advantage of the data (and the power of
QDA software) in later stages of analysis. The second, more focused, reading
is when we recommend applying analytic codes one research question or
paper at a time.
The process of indexing transcripts and applying attributes may seem
time-consuming, but it is time well spent. In contrast to the unwieldy piles
of text snippets created by open coding, after this stage is complete, the data
are ready for use in several different thematic projects. The data set will also
be ready to share among research team members or with researchers who
may be interested in using the data for secondary analysis. If the research
funder requires that transcript data are available to others, sharing an indexed
data file with respondent attributes will meet this standard—and increase
accessibility—without requiring additional work. Team members and sec-
ondary analysts can apply their own analytic codes to suit their own research
questions and goals. In other words, we recommend producing a version of
the coded project containing only attribute and index codes. This version
should be archived for secondary users and distributed to the research team.
Then, each researcher can apply their own analytic codes as they analyze the
data.

Stage 2: Applying Analytic Codes

After a researcher or research team goes through the process of indexing
transcripts and writing respondent and cross-case memos, there will be
Deterding and Waters 729

several ideas that could be pursued in many papers or book chapters. The list
will include, but not be limited to, the original research questions that
informed the study; some themes will be truly “emergent.” The work done
during respondent and cross-case memoing will offer a variety of possible
directions for the first paper from this project. It is very common for large
qualitative projects to unfold over time and result in a range of products,
beginning with journal articles and culminating in a book. Being strategic
about the analytic process can help researchers meet very real publication
pressures. Also, to avoid being overwhelmed by possibilities, we suggest
approaching the application of analytic codes one research question or
paper at a time. It is not necessary to write all of the papers at once! The
familiarity with the transcripts built during the first reading means research-
ers will have a good idea where to find the relevant chunks of the transcript
for a single research question.
A major problem of analyzing a large interview data set is applying codes
in a reliable manner (Campbell et al. 2013). While sometimes this issue is
solved by coding in teams, it is possible to use the software to aid the process.
On the second reading, we suggest limiting reading to the relevant text only,
considerably reducing the task of applying analytic codes. We also recom-
mend ignoring respondent attributes while applying the analytic codes. Only
after coding thematically across all transcripts should the researcher examine
whether there are patterns of qualitative difference by attributes. This tech-
nique allows the analyst to avoid confirmation bias, keeping the final argu-
ment as close to the text as possible.
Use the index code to display the relevant sections of the transcripts and
apply only one or two analytic codes at a time to this text. By focusing the
analytic process in this way, it is possible to increase the reliability and
validity of coding. For example, Deterding (2015a) built concepts of
“instrumental education” and “expressive education” in conceptual memos
during her initial read of respondents’ discussion of their college plans. Using
the index, she was then able to apply instrumental/expressive codes to text in
127 transcripts in about 10 hours. By limiting herself to the code indexed at
“education history” and “successful adulthood,” her second read covered
approximately 20 percent of the full transcripts, a piece of information she
reported in her methodology section. The indexing reading of the 120 tran-
scripts took a group of graduate student coders about 250 hours plus time for
cleanup and matrix checking, and it would be nearly impossible to reliably
apply well-defined analytic codes over such a long period of time.
In addition to the a priori attributes discussed above, some analytic codes
should be applied to the entire interview, becoming attributes. This is
730 Sociological Methods & Research 50(2)

especially true if a typology is used to organize the data. A typology is a way

of generalizing from concrete cases by defining a common core within a set
of cases (Weiss 1994:173). This common core defines the specific way in
which the cases cluster together. A typology is not a theory, but rather a way
of organizing data as a step toward producing a theory.
For example, Deterding (2015b) read the reduced interview excerpts,
identifying instrumental and expressive logic toward education as they
appeared in the text. At the end of each case, she made a judgment about
whether the person should be typologized as instrumental, expressive, or
mixed based on the textual evidence and added this to the attribute file.
Together, the analytic and attribute codes marked the transcript with textual
evidence for her person-level typology. Similarly, Waters (1990) constructed
a typology of racial and ethnic identity among second-generation West
Indian teens. She coded each respondent based on their responses to many
different questions as (1) identifying as African American, (2) identifying as
ethnic American (Jamaican American or Trinidadian American), or (3)
immigrant identified (identifying as Jamaican or Trinidadian). She con-
ducted this study before she began to use QDA software, but if she had been
using the software, these designations would have been applied to the entire
transcript via Primary Document (PD) families in ATLAS.ti or a classifica-
tion sheet in NVivo to help her understand the relationship between ethnic
identity and lived experiences in the United States.
Analytic codes may also arise from the literature. Sometimes, qualitative
researchers seek to apply concepts from previous research to a new study
population. Here, reliability and validity are also very important, if qualita-
tive data are to be used to explicitly test theoretical concepts derived by
others. Reducing the transcripts to the appropriately indexed information,
this task becomes much less unwieldy.

Stage 3: Exploring Coding Validity, Testing, and Refining Theory

In the end, researchers using the flexible coding process aim to construct an
account of the data that meets the threshold of theoretical validity (Maxwell
1992). A theoretically valid explanation is an “abstract account that proposes
to explain what has been observed,” where concepts and their relationships
have a strong basis in the data (Silbey 2009:81). Without a clear procedure
for validating that the argued relationships are truly grounded in the data,
interview researchers risk accusations that their work is not rigorous. During
respondent-level and cross-case memoing, flexible analysts built hunches
about the important stories in the data. Next they applied analytic codes across
Deterding and Waters 731

the body of data. Now it is time to explore how deeply the story is grounded in
the entire body of text. Software can aid this process, by helping researchers
identify trends across cases, investigate alternative explanations, and quickly
locate negative cases that help refine or limit the theoretical explanation.
QDA software also makes it relatively simple to examine the cross-case
reliability of the thematic coding. While other authors have suggested
options for enhancing reliability that involve multiple coders (see Campbell
et al. 2013; Price and Smith 2017), an alternate option is to query the inter-
section of the typology (stored as an attribute) and analytic codes. To con-
tinue the example above, for each person-level categorization (instrumental/
expressive/mixed), Deterding output the analytic codes for instrumental and
expressive logic in order to make sure that the textual evidence identified
groups that were truly distinct. The process of reducing data down from full
transcripts, to indexed extracts, and finally to grouped analytic codes allowed
her to judge whether she had applied uniform qualitative criteria across the
sample, increasing reliability or construct validity. When looking at the data
in this organized, reduced form, some respondents seemed misclassified. She
then revised their classification, assuring the construct validity of her
typology.
Other features of the software can help researchers test and refine the
theoretical explanations they have developed. There is considerable debate
over methods for determining what counts as a robust finding in qualitative
research. We do not believe it is necessary for a phenomenon to apply to the
entire sample to be analytically important. However, a systematic treatment
of alternative explanations and negative cases is an important part of con-
textualizing findings and creating a convincing theory. The intuition is that it
is possible to learn about the scope of the theory and refine an understanding
of important relationships by examining (and interpreting) where it does not
appear to apply. Blee (2009:148) sets out the following criteria for a quali-
tative analytical plan, arguing that it should take into account “how will data
be assessed to ensure that (1) all data are considered (2) spectacular/extraor-
dinary events are not overly stressed (3) data that diverge from the pattern are
not discounted without a clear rationale to do so.” Querying the data with
QDA software can aid in this process.
While the easy production of frequency tables is a useful feature of QDA
software, taking advantage of software does not require a frequentist per-
spective. If one approached analysis from a case-based perspective, a single
disconfirming case or cluster of exceptions may crystalize the conceptual
limitations of the trend or help refine the working theory to account for the
exceptions. From this perspective, it is not the number of exceptions to the
732 Sociological Methods & Research 50(2)

theme that is analytically important but how the exceptions help to refine the
theory. The data querying capacity of QDA software allows one to easily
identify cases that are exceptions to trends and require further examination,
meeting a typical analytical requirement in the methods literature that negative
cases should be explicitly treated in the analysis (Katz 1982; Luker 2008).
If the index and analytic codes are applied reliably and analytic attribute-
level categorization is consistent, it is possible to run text queries to docu-
ment the robustness and the limitations of the findings (Maxwell 1992:48).
On the frequentist end of the spectrum, which the software easily facilitates,
the analyst may want to make statements such as “N respondents demon-
strated this logic.” As cautioned by scholars such as Small (2011), however,
it’s important to make sure that such statements are appropriate for the form
of the data. For instance, if the interview protocol evolved over the course of
the study, and the same questions were not asked of everyone, it may not be
appropriate to report coding counts. It may also be the case that one wants to
write about a topic that only applies to half of the interviewees. By querying
codes, it is possible identify in how many transcripts the topic appeared,
which might be considered a more accurate number of interviews to report
in the Methods section of the paper than the full interview sample.

Conclusion, Strengths, and Limitations

In this article, we summarized the basic features of a grounded theory
approach to qualitative interview analysis, assessed the current state of qua-
litative analysis of sociological interviews, and highlighted mismatches
between grounded theory’s prescriptions and current practice in the field.
Based on our own experience working with large-scale interview data, we
outlined a three-step process that flips traditional grounded theory on its head
to take advantage of modern QDA technology. Rather than begin with small
codes and aggregate up to a single set of concepts, we advocate for starting big
with index coding and capitalizing on the flexibility of electronic retrieval for
data reduction and conceptual refinement over time. We have seen many
benefits to our approach in our own work and believe that flexible coding can
help move the field of qualitative data analysis into the twenty-first century.
First, flexible coding begins with indexing hundreds of pages of data for
easy retrieval, setting interview researchers up to take full advantage of the
power of QDA—and of their large-scale data. The collection of in-depth
interview data is a time- and resource-intensive process, and the ability to
quickly enter and flexibly reanalyze data can help qualitative researchers to
get as large a return as possible from their effort. Since it is not fine-grained,
Deterding and Waters 733

index coding can be more easily dispersed among members of a research

team than reliable analytic coding. The indexed data can then be distributed
to members of the team so that researchers can apply their own thematic
codes, a different model for qualitative teamwork than teamwork for the
purpose of interrater reliability. Formatting data with an index and attributes
prepares qualitative data for electronic archiving and makes data sharing
more feasible, as recommended by the National Science Foundation Work-
shop on Interdisciplinary Standards for Systematic Qualitative Research
(Lamont & White 2009). A flexible index also could reduce the burden of
reanalyzing data in combination with data collected by other researchers or
over time.
Second, the process of indexing allows one to get to know their data.
While applying index codes, we read whole transcripts and begin writing
person-level and thematic memos. At this stage, the analyst can begin play-
ing with and refining concepts across interview cases. The initial thematic
memos become the basis for analytic codes that are then applied across the
full data set; it is not necessary to begin applying fine-grained coding to
meaningfully begin (and begin documenting) analytic work.
Third, the process enables researchers to (more) reliably and efficiently
apply analytic codes. After using memo writing for thematic development,
the index allows researchers to reduce the amount of data under analysis.
Compared to coding as suggested by grounded theory, our process can avoid
hours of time muddying the story by overcoding. Applying thematic codes to
reduced data during validation and refinement increases the likelihood that
the coding is reliable and valid across hundreds or thousands of pages of data.
In short, the process allows researchers to spend more time developing and
elaborating concepts—and meaningfully linking them to data—than on
cleaning up and combining line-by-line codes.
Fourth, by systematically attending to respondent attributes, researchers
are positioned to take full advantage of QDA’s data query capabilities.
Examining the intersection of analytic codes at the text and attribute (typol-
ogy) levels provides an opportunity to easily validate coding reliability and
examine patterns in themes. Querying also identifies meaningful exceptions
to themes. These negative cases are important for refining a data-based
theory or limiting its scope.
Finally, a clear and systematic analytic process can increase transparency
in qualitative research. To reduce the opacity of qualitative data analysis,
methodology sections must be more specific than “I analyzed the data using
qualitative software.” Reporting additional information, such as the portion
of a larger transcript that was relevant to the topic at hand, provides a clearer
734 Sociological Methods & Research 50(2)

idea of the breadth of information the author used to draw conclusions in a

piece of research reporting. This may be particularly important for secondary
analysis of existing data or qualitative work drawn from large, team-driven
projects. Regardless of whether the data collection was conducted by a single
researcher or across a research team, analytic transparency makes qualitative
analysis more easily replicable or reinterpretable, which we believe is impor-
tant for the rigor and credibility of qualitative research.

Limitations
While we use QDA software to flip the script on grounded theory, flexible
coding may not be appropriate for every interview project. It might not be
worth fully indexing transcripts for projects with a small number—fewer
than 30—of interviews. If the research question is tightly circumscribed and
the data are intended for a single article, it may not be important for the
researcher (or others) to revisit data in the future. Finally, many of
the published papers we examined included in-depth interviews as one of
the multiple data sources. Some of these pieces were historical case studies,
others were ethnographies, and some articles drew on a handful of interviews
to add color to primarily quantitative analyses. In these circumstances, less
intensive forms of interview data preparation and analysis may well be
appropriate. And, of course, if the researcher truly performed a grounded
theory study, with an unfolding data collection, where transcripts were ana-
lyzed as they were completed and the protocol modified toward a final
emergent concept, then our method is not for them.
Finally, we do not want to overemphasize the importance of easy quanti-
fication and tabulation of data, pressing too far beyond our primary concern
about the internal validity of qualitative explanations. Our goal has been to
suggest ways that qualitative researchers analyzing interview data can iden-
tify findings that are firmly grounded in the data they have collected, prop-
erly contextualize findings in the set of data as a whole, and easily identify
negative cases for the refinement of theory. In the end, we believe that the job
of generalizing theories generated by qualitative data falls to future research
using more representative methodologies.

Conclusion
If the methodology sections of recent sociology journal articles are to be
believed, qualitative data analysis appears stuck in the twentieth century.
Grounded theory certainly deserves credit for the role it played in
Deterding and Waters 735

systematizing analysis of in-depth interview data. But we suspect that—

cursory citations to the contrary—it does not match the practice of many
contemporary qualitative researchers. Today’s in-depth interview studies are
typically grounded in prior literature, comprised of large samples, matched
with other data sources, collected or analyzed by teams of researchers, and
drawn upon for multiple research products. We have argued that these
changes in convention and the introduction of QDA technology demand
renewed attention to the process of qualitative coding.
Our analysis of research articles published in ASA-sponsored journals
demonstrates that the work of coding remains relatively opaque to readers of
qualitative work. Our discussions with many other researchers who regularly
use interview methodology, however, suggest they cite methodological classics
in absence of better guidance, after reaching their own detailed process through
a series of trials and errors. While tricks of the trade are likely shared informally,
our goal here is to offer guidance for researchers facing hundreds of pages of text
imported to QDA software. As we prepare to analyze the large volume of data
resulting from today’s disciplinary conventions, thoughtful attention to the
structure and process of coding can produce an archive of information that may
be revisited time and again, including by secondary analysts, improving the
transparency and accountability of qualitative research.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research,
authorship, and/or publication of this article.

Funding
The author(s) disclosed receipt of the following financial support for the research, author-
ship, and/or publication of this article: This work was supported by National Institutes of
Health Grants R01HD046162, RO1HD057599, and P01HD082032. Waters was
supported by a Robert Wood Johnson Investigator Award in Health Care Policy.

ORCID iD
Nicole M. Deterding https://orcid.org/0000-0001-5819-8935

Supplemental Material
Supplemental material for this article is available online.

References
Asad, Asad L. 2015. “Contexts of Reception, Post-disaster Migration, and Socio-
economic Mobility.” Population and Environment 36(3):279-310.
736 Sociological Methods & Research 50(2)

Bail, Christopher A. 2016. Terrified: How Anti-Muslim Fringe Organizations

Became Mainstream. Princeton, NJ: Princeton University Press.
Baker, Sarah Elsie and Rosalind Edwards. 2012. “How Many Qualitative Interviews
Is Enough? Expert Voices and Early Career Reflections on Sampling and Cases in
Qualitative Research.” National Centre for Research Methods Review Paper,
Economic and Social Research Council, Southampton, England. Retrieved Sep-
tember 13, 2018 (http://eprints.brighton.ac.uk/11632/1/how_many_interviews.
pdf).
Bazeley, Patricia and Kristi Jackson. 2013. Qualitative Analysis with NVivo. London,
England: Sage.
Blee, Kathleen. 2009. “Defining and Communicating Standards for Qualitative
Research Rigor in Sociology.” Pp. 148-52 in Workshop on Interdisciplinary Stan-
dards for Systematic Qualitative Research, edited by M. Lamont and P. White.
Arlington, VA: National Science Foundation.
Bosick, Stacy. 2015. “Pushed Out on My Own: The Impact of Hurricane Katrina in
the Lives of Emerging Adults.” Sociological Perspectives 58(2):243-63.
Brown, Hana E. 2013. “Racialized Conflict and Policy Spillover Effects: The Role of
Race in the Contemporary U.S. Welfare State”. American Journal of Sociology
119(2):394-443.
Bryant, Antony. 2002. “Regrounding Grounded Theory.” Journal of Information
Technology Theory and Application 4(1):25-42.
Bryant, Antony and Kathy Charmaz. 2007. The SAGE Handbook of Grounded The-
ory. London, England: Sage.
Calarco, Jessica McCrory. 2011. “‘I Need Help!’ Social Class and Children’s Help-
seeking in Elementary School.” American Sociological Review 76(6):862-82.
Calarco, Jessica McCrory. 2014a. “Coached for the Classroom: Parents’ Cultural
Transmission and Children’s Reproduction of Educational Inequalities.” Ameri-
can Sociological Review 79(5):1015-37.
Calarco, Jessica McCrory. 2014b. “The Inconsistent Curriculum: Cultural Tool-kits
and Student Interpretations of Ambiguous Expectations.” Social Psychology
Quarterly 76(2):186-209.
Campbell, John L., Charles Quincy, Jordan Osserman, and Ove K. Pedersen. 2013.
“Coding In-depth Semistructured Interviews: Problems of Unitization and Interco-
der Reliability and Agreement.” Sociological Methods & Research 42(3):294-320.
Chakrabarti, Parijat and Margaret Frye. 2017. “A Mixed-methods Framework for
Analyzing Text Data: Integrating Computational Techniques with Qualitative
Methods in Demography.” Demographic Research 37:1351-82.
Charmaz, Kathy. 2000. “Grounded Theory: Objectivist and Constructivist Methods.”
Pp. 509-35 in Handbook of Qualitative Research, edited by N. K. Denzin and Y. S.
Lincoln, 2nd ed. Thousand Oaks, CA: Sage.
Deterding and Waters 737

Charmaz, Kathy. 2006. Constructing Grounded Theory. London, England: Sage.

Charmaz, Kathy. 2014. Constructing Grounded Theory. 2nd ed. London, England:
Sage.
Collett, Jessica L., Kelcie Vercel, and Olevia Boykin. 2015. “Using Identity Pro-
cesses to Understand Persistent Inequality in Parenting.” Social Psychology Quar-
terly 78(4):345-364.
Deterding, Nicole M. 2015a. “Instrumental and Expressive Education: College Plan-
ning in the Face of Poverty.” Sociology of Education 88(4):284-301.
Deterding, Nicole M. 2015b. “Start, Start Again: The College Pathways of
Economically-vulnerable Mothers.” Unpublished doctoral dissertation. https://
dash.harvard.edu/handle/1/17463981. Accessed 20 September 2018.
Glaser, Barney G. 1992. Emergence vs. Forcing: Basics of Grounded Theory Anal-
ysis. Mill Valley, CA: Sociology Press.
Glaser, Barney G. 1998. Doing Grounded Theory: Issues and Discussions. Mill
Valley, CA: Sociology Press.
Glaser, Barney G. 2005. The Grounded Theory Perspective II: Description’s Remo-
deling of Grounded Theory Methodology. Mill Valley, CA: Sociology Press.
Glaser, Barney G. and Anselm L. Strauss. 1967. The Discovery of Grounded Theory:
Strategies for Qualitative Research. Chicago, IL: Aldine.
Guest, Greg, Arwen Bunce, and Laura Johnson. 2006. “How Many Interviews Are
Enough? An Experiment with Data Saturation and Variability.” Field Methods
18(1):59-82.
Jerolmack, Colin and Shamus Khan. 2014. “Talk Is Cheap: Ethnography and the
Attitudinal Fallacy.” Sociological Methods & Research 43(2):178-209.
Kasinitz, Philip, John H. Mollenkopf, Jennifer Holdaway, and Mary C. Waters. 2008.
Inheriting the City: The Children of Immigrants Come of Age. Cambridge, MA:
Harvard University Press.
Katz, Jack. 1982. Poor People’s Lawyers in Transition. New Brunswick, NJ: Rutgers
University Press.
Lamont, Michele and Patricia White. 2009. Workshop on Interdisciplinary Standards
for Systematic Qualitative Research (Final Report). Arlington, VA: National Sci-
ence Foundation.
Lewins, Ann and Christina Silver. 2007. Using Software in Qualitative Research: A
Step-by-step Guide. Los Angeles, CA: Sage.
Lowe, Sarah R., Kara Lustig, and Helen B. Marrow. 2011. “African American
Women’s Reports of Racism during Hurricane Katrina: Variation by Interviewer
Race.” New School Psychology Bulletin 8(2):46-57.
Lowe, Sarah R., Jean E. Rhodes, and Arielle A. J. Scoglio. 2012. “Changes in Marital
and Partner Relationships in the Aftermath of Hurricane Katrina.” Psychology of
Women Quarterly 36(3):286-300.
738 Sociological Methods & Research 50(2)

Lowe, Sarah R., Jean E. Rhodes, and Mary C. Waters. 2015. “Understanding Resi-
lience and Other Trajectories of Psychological Distress: A Mixed-methods Study
of Low-income Mothers Who Survived Hurricane Katrina.” Current Psychology
34(3):537-50.
Luker, Kristin. 2008. Salsa Dancing into the Social Sciences: Research in an Age of
Info-glut. Cambridge, MA: Harvard University Press.
Maxwell, Joseph. 1992. “Understanding Validity in Qualitative Research.” Harvard
Educational Review 62(3):279-301.
Miles, Matthew B. and Michael Huberman. 1994. Qualitative Data Analysis.
Thousand Oaks, CA: Sage Press.
Morris, Katherine Ann and Nicole M. Deterding. 2016. “The Emotional Cost of
Distance: Geographic Social Network Dispersion and Post-traumatic Stress in
Survivors of Hurricane Katrina.” Social Science and Medicine 165(1):56-65.
Price, Heather and Christian Smith. 2017, August. “Process and Reliability for Cultural
Model Analysis Using Semi-structured Interviews.” Paper presented at the Annual
Meeting of the American Sociological Association, Montreal, Québec, Canada.
Ragin, Charles, Joane Nagel, and Patricia White. 2004. “Workshop on Scientific
Foundations of Qualitative Research.” NSF Report 04-219, National Science
Foundation, Arlington, VA.
Small, Mario L. 2011. “How to Conduct a Mixed-methods Study: Recent Trends in a
Rapidly Growing Literature.” Annual Review of Sociology 37(1):57-86.
Silbey, Susan. 2009. “In Search of Social Science.” Pp. 76-82 in Workshop on
Interdisciplinary Standards for Systematic Qualitative Research, edited by M.
Lamont and P. White. Arlington, VA: National Science Foundation.
Strauss, Anselm L. 1987. Qualitative Analysis for Social Scientists. New York:
Cambridge University Press.
Strauss, Anselm L. and Juliet M. Corbin. 1990. Basics of Qualitative Research:
Grounded Theory Procedures and Techniques. London, England: Sage.
Timmermans, Stefan and Iddo Tavory. 2012. “Theory Construction in Qualitative
Research: From Grounded Theory to Abductive Analysis.” Sociological Theory
30(3):167-86.
Tollette, Jessica. 2013. “Nueva New Orleans: Race Relations and (Im)migration in
the Post-Katrina South.” Poster presented at the Annual Meeting of the Population
Association of America, New Orleans, LA.
Wacquant, Loı̈c. 2002. “Scrutinizing the Street: Poverty, Morality, and the Pitfalls of
Urban Ethnography.” American Journal of Sociology 107(6):1468-532.
Waters, Mary C. 1990. Ethnic Options: Choosing Identities in America. Berkeley:
University of California Press.
Waters, Mary C. 1999. Black Identities: West Indian Immigrant Dreams and
American Realities. Cambridge, MA: Harvard University Press.
Deterding and Waters 739

Waters, Mary C. 2016. “Life after Hurricane Katrina: The Resilience in Survivors of
Katrina (RISK) Project.” Sociological Forum 31:750-69.
Waters, Mary C., Patrick J. Carr, Maria J. Kefalas, and Jennifer Ann Holdaway. 2011.
Coming of Age in America: The Transition to Adulthood in the Twenty-first
Century. Berkeley: University of California Press.
Weiss, Robert S. 1994. Learning from Strangers: The Art and Method of Qualitative
Interview Studies. New York: Free Press.
Weitzman, Eben A. and Matthew B. Miles. 1995. Computer Programs for Qualitative
Data Analysis: A Software Sourcebook. Thousand Oaks, CA: Sage.
White, Michael J., Maya D. Judd, and Simone Poliandri. 2012. “Illumination with a
Dim Bulb? What Do Social Scientists Learn by Employing Qualitative Data
Analysis Software (QDAS) in the Service of Multi-method Designs?” Sociologi-
cal Methodology 42(1):43-76.
Znaniecki, Florian. 1934. The Method of Sociology. New York: Farrar and Strauss.

Author Biographies
Nicole M. Deterding produced this work while a National Poverty Center Postdoc-
toral Fellow with the Institute for Research on Poverty, University of Wisconsin–
Madison. Her research uses mixed methods to examine the role of postsecondary
education and training in the lives of economically vulnerable adults. Empirical
articles on which this article is based have recently appeared in Sociology of Educa-
tion and Social Science and Medicine.

Mary C. Waters is the John L. Loeb Professor of Sociology at Harvard University.

Her large-scale interview studies have examined the integration of immigrants and
their children; the transition to adulthood; the measurement and meaning of racial and
ethnic identity; and the social, demographic, and psychological impact of natural
disasters. She is the coauthor, most recently, of The Art and Science of Social
Research (2017, W. W. Norton).