CONTENT ANALYSIS

"Content analysis" has evolved into an umbrella label that includes various procedures for making reliable, valid inferences from qualitative data, including text, speech, and images. These procedures have improved and expanded due to numerous developments in recent years since this encyclopedia's first edition.

Traditionally, "content analysis" has referred to systematic procedures for assigning prespecified codes to text, such as interviews, newspaper editorials, open-ended survey answers, or focus-group transcripts, and then analyzing patterns in the codings. Some projects will count each specific occurrence within a text, while others will have coders tally the number of column inches assigned a code. Either way, the procedure usually employs a "top down" strategy, beginning with a theory and hypotheses to be tested, developing reliable coding categories, applying these to coding-specified bodies of text, and finally testing the hypotheses by statistically comparing code indexes across documents.

With the increasing popularity of qualitative sociology, content analysis has also come to refer to "grounded" inductive procedures for identifying patterns in various kinds of qualitative data including text, illustrations, and videos. For example, the data might include observers' detailed notes of children's behaviors under different forms of supervision, possibly supplemented with videotapes of those same behaviors. While traditional content analysis usually enlisted statistical analyses to test hypotheses, many of these researchers do not start with hypotheses, but carefully search for patterns in their data.

However, rather than just produce statistical analyses or search for patterns, investigators should also situate the results of a content analysis in terms of the contexts in which the documents were produced. A content-analysis comparison of letters to stockholders, for example, should take into consideration the particular business sectors covered and the prevailing economic climates in which they were written. An analysis of American presidential nomination acceptance speeches should consider that they changed dramatically in form once they started to be broadcast live on national radio. A content analysis may have reliable coding, but the inferences drawn from that coding may have little validity unless the researcher factors in such shaping forces.

Like any expanding domain, there has been a tendency for content analysis to segment into specialized topics. For example, Roberts (1997) focuses on drawing statistical inferences from text, including Carley's networking strategies and Gottschalk's clinical diagnostic tools. There also has been a stream of instructional books, including several series published by Sage, that focus on particular kinds of qualitative data such as focus-group transcripts. A technical literature also has developed addressing specialized computer software and video-analysis techniques. Nevertheless the common agenda is analyzing the content of qualitative data. Inasmuch as our lives are shaped by different forms of media, and inasmuch as different analytic procedures can complement one another in uncovering important insights, it makes sense to strive toward an integration, rather than fragmentation.

Many advances in content-analysis procedures have been made possible by the convenience and power of desktop and laptop computers. In addition, an overwhelming proportion of text documents are now generated on computers, making their text files computer accessible for content analysis. And a revolution in hand-held analogue and digital video cameras, together with computer-based technology for editing and analyzing videotapes, makes new research procedures feasible.

Consider, for example, new possibilities for analyzing responses to open-ended questions in survey research. For years, survey researchers have been well aware that closed-ended questions require respondents to frame how they think about an issue in terms of a question's multiple choices, even when the choice options had little to do with how a respondent views an issue. But the costs and time involved in analyzing open-ended responses resulted in such questions rarely being used. Even when they were included in a survey, the interviewers usually just recorded capsule summaries of the responses that omitted most nuances of what was said.

Contrast this then with survey research using today's audio information-capturing technologies. Telephone survey interviewers are guided by instructions appearing on a computer screen. Whenever an open-ended question appears, the interviewer no longer needs to type short summaries of the responses. Instead, a computer digitally captures an audio recording of each open-ended response, labels it, and files it as a computer record. Any audio response can later be easily fetched and replayed, allowing a researcher, for example, to identify a "leaky voice," that is, one indicative of the respondent's underlying emotion or attitude. And the full audio responses are then available to be transcribed to text, including, if desired, notations indicating hesitations and voice inflections. Until computer voice recognition is completely reliable, transcribing usually remains a manual task. But with spreadsheet software (such as Excel) no longer restrictively limiting the amount of text in any one cell, the text of each individual's entire response to a question can be placed in a spreadsheet cell, thus capturing both the closed-ended and open-ended data for a survey into a convenient, single spreadsheet for researchers to analyze.

With the survey data in this convenient form, researchers can then code open-ended responses manually, putting their assigned codes in additional spreadsheet columns. As a teaching exercise, it is instructive to assign students a task such as identifying gender differences among a thousand responses to a broad open-ended question, such as a question asking respondents' views about peace or family values. Students first might sort the spreadsheet by gender in order to read separately samples of male responses and female responses and obtain a sense of what possible gender differences exist. They then develop coding instructions that capture these differences and apply the codes to the entire set of responses. This coding, of course, is better done without knowledge of the respondents' genders, with responses in a random order, and on different respondents than those used to develop the codes. After coding several hundred responses, however, students usually begin to glaze over and soon the most ardent humanist student is asking whether the computer could possibly be of help in assigning codes.

For some kinds of coding, computer help is indeed available in the form of computer programs that assign codes. Such codings can be treated as advisory and then manually confirmed, augmented perhaps by also assigning a weight. Or they may be used as is after being spot-checked for accuracy. Not only can a computer complete huge amounts of tedious coding in minutes, possibly assigning many different types of codings to each text, but these codings may uncover statistically significant frequency differences that human coders would not uncover, if only because computer analysis is so even-handed and untiring. The static created by occasional miscodings may be more than offset by gains from a reliable consistency in making many codings.

Computer coding assignments are usually based on the occurrence of words, particular senses of words, or multiword idioms appearing in the text. For example, the word "father" in a text might be coded as "male," "family member," etc. as well as possibly "authority-role." Computer content-analysis software may search the contexts of words in the text to ferret out and correctly code common word senses. For example, for a national study of people's perceptions of African-American young males on several open-ended questions, it was particularly important for the computer to identify correctly each respondent's usages of such multi-meaning words as "race," "white," "black," and "color."

In addition to developing their own coding categories, researchers may enlist existing computer-scored categories that are relevant to the task at hand. For example, it might be hypothesized that one group being studied is more optimistic and its responses will reflect more "positive thinking" while another is more negative or pessimistic. To code "positive thinking" a researcher may want the computer to apply an existing content-analysis category that includes over 1200 words, word roots, word senses, phrasal verbs, and idioms, thus essentially covering most expressions of "positive-thinking" that occur as infrequently as three times per million words of ordinary English text. A similar category exists for negative-thinking, allowing the investigator to check whether the groups being studied differ in their coded positive thinking, negative thinking, or both. And by enlisting such standard categories, the results obtained in one study can be readily compared with results found in other studies.

Once data has been captured in a convenient format for computer use, they can be repeatedly analyzed. For example, should our now glazed-over students have any energy left after analyzing the responses by gender, they could be given an additional assignment of identifying and coding rural-urban differences in these same open-ended responses. Given so many analyses that can be made, it makes sense to let the computer do what it can, saving manual labor for those types of codings that would be hard to have a computer assign. Even multimedia qualitative-analysis software such as HyperResearch includes some rudimentary tools for automatic assignments.

Moreover, desktop computer software has also become available that identifies patterns in text without having to develop coding categories. For example, SPSS's TextSmart uses an algorithm that groups respondents into clusters based on word-use co-occurrences within responses and then maps these clusters into a two-dimensional grid that uses colors to represent each cluster group. If such an automated inductive procedure can produce additional valid insights that other techniques are likely to overlook, then why not use it too?

Pioneering work in inductive automatic categorizing, such as Iker's (1969), usually enlisted procedures based upon correlation matrixes, such as factor analysis. These procedures tended not to be particularly suited to analyzing text both because of the shape of word-usage frequency distributions as well as the limited number of words that a correlation matrix could feasibly handle. TextSmart, based upon a word-distance measure, provides much more suitable solutions. Further automatic categorizing procedures may be expected from artificial intelligence, as well as from categorizing techniques being developed for Internet search engines.

Content-analysis research strategies can thus now easily be multipronged, spanning from completely automatic inductive procedures to manual coding. But even manual coding these days is likely to utilize computer software to help coders manage information. Consider these changes in costs and convenience: Unlike mainframe computing of the 1960s and 1970s, when the cost of an hour of computer time was about the same as a coder's wage for several weeks, the marginal cost of using a desktop computer is essentially the electricity it uses. Today's desktop computer is likely to be more than fivefold faster at content-analysis coding than those mainframe computers ever were. They also can access much larger dictionaries and other information in their RAM than was ever feasible on a user partition of a mainframe computer, thus making their coding more accurate and comprehensive. Moreover, a single CD-ROM full of text to be analyzed is easily popped into a desktop computer, whereas in the days of mainframe computing, a comparable amount of text would have to be keypunched on over 3,000 boxes of IBM cards.

Given today's convenience and low cost of computer-based procedures, there is no reason to limit an analysis to one approach, especially if insights gained from one approach will differ and often complement those gained from another. Instead of being limited by technology, the limits now may lie in the skills, proclivities, and comfort zones of the researchers. Research teams, rather than individual researchers, may prove the best solution, for only in a team made up of people with complementary strengths is one likely to find the full range of statistical, conceptual, intuitive, experiential, and perhaps clinical strengths needed to carry out penetrating, comprehensive content-analysis projects. Moreover, some researchers will prefer to learn from the main trends while others will learn more from studying outlying cases. Some will learn from bottom-line numbers while others will learn more from innovative graphics that highlight information patterns. Some will focus on current data while others will contextualize data historically by comparing them with data in archives. Data that has been gathered and assembled at considerable cost, especially data-gathering that imposed on many respondents, merit as thorough and comprehensive analyses as these various procedures collectively offer.

Unfortunately, however, an "either-or" assumption about how to do content analysis has continued to be supported both by books and computer software. Authors who do an excellent job of describing one approach to content analysis, such as Boyatzis (1998), give an impression that an either-or decision has to be made about which approach to use. Some software—especially that ported from mainframe computers or developed for early desktop computers—still may steer or even limit researchers who use it to just one approach. For example, some software packages create specialized data formats such as "classification trees" that then in effect constrain the user to analyses that can be readily derived from that format. Software reviews such as Lewis's (1998) excellent comparison of ATLAS/ti and NUD-IST software have been explicit about what assumptions a researcher buys into when utilizing each package.

Additional leverage in analyzing qualitative information has stemmed from computer-based tools, such as newer versions of HyperResearch and ATLAS/ti, that integrate the handling of multiple media (text, illustrations, and video). Especially as more software comes from countries where there are expert programming skills and programming labor is relatively inexpensive, we can expect ambitious content-analysis software, of which TextAnalyst from Russia (www.megaputer.com) may be a forerunner.

Given continuing content-analysis software developments, those who would like to learn what is currently available are advised search Internet sites rather than rely upon even recently published materials. One recommended starting point is the Georgia State University content-analysis site (www.gsu.edu/~wwwcom/content.html), which gives links to software web sites (including software mentioned in this article), indexes recent content-analysis publications, and has a mailing list of more than 700 members. Technical reviews of relevant language-analysis tools, such as Berleant's (1995), occasionally appear in computational linguistics journals and web sites. For training, the University of Essex Summer School in Social Science Data Analysis and Collection (www.essex.ac.uk), as part of a program of a European consortium, has been offering a content-analysis module for years as part of its program.

Given the developments described here, some of the contributions that content analysis should be able to make to sociological research include:

1. A major shift from reliance upon closedended questions to an appropriate use of open-ended questions that lets people be heard in the ways they frame issues, as well as the way they think and feel about them, as discussed in detail by Stone (1997)
2. A better understanding of both print and television media and its impact on public opinion, both in setting agendas and in influencing opinion intensity, as laid out in Neuman (1989). This will involve research that compares the content of media with the content of opinions. Not only will survey research data be archived and accessible from Internet servers, but full-display media will be accessible from Lexis-Nexis and on-line editions supplied by media providers, as well as television news archives such as those at Vanderbilt University
3. Better use of historical qualitative data, including both text and graphic materials, to address such issues as how economic cycles impact ideology, as examined by Namenwirth and Weber (1987), or to uncover cycles of creativity, as demonstrated by Martindale (1990)
4. Investigations, several of which are already underway, of both intranet communication patterns within organizations as well as Internet communications, including analyses of the content of communications over those networks.

There is also, however, good reason for caution. Never before in history has so much qualitative information been available electronically. High-volume image scanning will also further increase the amount of information that can be electronically accessed and content-analyzed. Quite understandably, those agencies responsible for limiting terrorist activities may look on content-analysis procedures as possibly providing early warnings that could save lives. But these procedures can also become tools for a "big-brother" monitoring society. Sociologists have an important role in anticipating these problems and helping resolve them.

references

Berleant, Daniel 1995 "Engineering 'Word Experts' for Word Disambiguation." Natural Language Engineering 1 (4):339–362.

Boyatzis, Richard E. 1998 Transforming Qualitative Information: Thematic Analysis and Code Development. Thousand Oaks, Calif.: Sage.

Iker, Howard, and Norman Harway 1969 "A Computer Systems Approach Toward the Recognition and Analysis of Content." In George Gerbner, Ole Hosti, Klaus Krippendorff, William Paisley, and Philip Stone, eds., The Analysis of Communication Content. New York: Wiley.

Lewis, R. Berry 1998 "ATLAS/ti and NUD-IST: A Comparative Review of Two Leading Qualitative Data Analysis Packages." Cultural Anthropology Methods 10 (3):41–47.

Martindale, Colin 1990 The Clockword Muse: The Predictability of Artistic Change. New York: Basic Books.

Namenwirth, J. Zvi, and Robert Philip Weber 1987 Dynamics of Culture. Winchester, Mass.: Allen and Unwin.

Neuman, Richard 1989 "Parallel Content Analysis: Old Paradigms and New Proposals." Public Communication and Behavior 2:205–289.

Roberts, Carl W. (ed.) 1997 Text Analysis for the SocialSciences: Methods for Drawing Statistical Inferences fromTexts and Transcripts. Mahwah, N.J.: Lawrence Erlbaum.

Stone, Philip J. 1997 "Thematic Text Analysis: New Agendas for Analyzing Text Content." In Carl W. Roberts, ed., Text Analysis for the Social Sciences: Methods for Drawing Statistical Inferences from Texts andTranscripts. Mahwah, N.J.: Lawrence Erlbaum.

PHILIP STONE