Data Analysis is the process of systematically applying statistical and/or logical techniques to describe and illustrate, condense and recap, and evaluate data. According to Shamoo and Resnik (2003) various analytic procedures “provide a way of drawing inductive inferences from data and distinguishing the signal (the phenomenon of interest) from the noise (statistical fluctuations) present in the data”..
While data analysis in qualitative research can include statistical procedures, many times analysis becomes an ongoing iterative process where data is continuously collected and analyzed almost simultaneously. Indeed, researchers generally analyze for patterns in observations through the entire data collection phase (Savenye, Robinson, 2004). The form of the analysis is determined by the specific qualitative approach taken (field study, ethnography content analysis, oral history, biography, unobtrusive research) and the form of the data (field notes, documents, audiotape, videotape).
An essential component of ensuring data integrity is the accurate and appropriate analysis of research findings. Improper statistical analyses distort scientific findings, mislead casual readers (Shepard, 2002), and may negatively influence the public perception of research. Integrity issues are just as relevant to analysis of non-statistical data as well.
Considerations/issues in data analysis
There are a number of issues that researchers should be cognizant of with respect to data analysis. These include:
- Having the necessary skills to analyze
- Concurrently selecting data collection methods and appropriate analysis
- Drawing unbiased inference
- Inappropriate subgroup analysis
- Following acceptable norms for disciplines
- Determining statistical significance
- Lack of clearly defined and objective outcome measurements
- Providing honest and accurate analysis
- Manner of presenting data
- Environmental/contextual issues
- Data recording method
- Partitioning ‘text’ when analyzing qualitative data
- Training of staff conducting analyses
- Reliability and Validity
- Extent of analysis
Having necessary skills to analyze
A tacit assumption of investigators is that they have received training sufficient to demonstrate a high standard of research practice. Unintentional ‘scientific misconduct' is likely the result of poor instruction and follow-up. A number of studies suggest this may be the case more often than believed (Nowak, 1994; Silverman, Manson, 2003). For example, Sica found that adequate training of physicians in medical schools in the proper design, implementation and evaluation of clinical trials is “abysmally small” (Sica, cited in Nowak, 1994). Indeed, a single course in biostatistics is the most that is usually offered (Christopher Williams, cited in Nowak, 1994).
A common practice of investigators is to defer the selection of analytic procedure to a research team ‘statistician’. Ideally, investigators should have substantially more than a basic understanding of the rationale for selecting one method of analysis over another. This can allow investigators to better supervise staff who conduct the data analyses process and make informed decisions
Concurrently selecting data collection methods and appropriate analysis
While methods of analysis may differ by scientific discipline, the optimal stage for determining appropriate analytic procedures occurs early in the research process and should not be an afterthought. According to Smeeton and Goda (2003), “Statistical advice should be obtained at the stage of initial planning of an investigation so that, for example, the method of sampling and design of questionnaire are appropriate”.
Drawing unbiased inference
The chief aim of analysis is to distinguish between an event occurring as either reflecting a true effect versus a false one. Any bias occurring in the collection of the data, or selection of method of analysis, will increase the likelihood of drawing a biased inference. Bias can occur when recruitment of study participants falls below minimum number required to demonstrate statistical power or failure to maintain a sufficient follow-up period needed to demonstrate an effect (Altman, 2001).
Inappropriate subgroup analysis
When failing to demonstrate statistically different levels between treatment groups, investigators may resort to breaking down the analysis to smaller and smaller subgroups in order to find a difference. Although this practice may not inherently be unethical, these analyses should be proposed before beginning the study even if the intent is exploratory in nature. If it the study is exploratory in nature, the investigator should make this explicit so that readers understand that the research is more of a hunting expedition rather than being primarily theory driven. Although a researcher may not have a theory-based hypothesis for testing relationships between previously untested variables, a theory will have to be developed to explain an unanticipated finding. Indeed, in exploratory science, there are no a priori hypotheses therefore there are no hypothetical tests. Although theories can often drive the processes used in the investigation of qualitative studies, many times patterns of behavior or occurrences derived from analyzed data can result in developing new theoretical frameworks rather than determined a priori (Savenye, Robinson, 2004).
It is conceivable that multiple statistical tests could yield a significant finding by chance alone rather than reflecting a true effect. Integrity is compromised if the investigator only reports tests with significant findings, and neglects to mention a large number of tests failing to reach significance. While access to computer-based statistical packages can facilitate application of increasingly complex analytic procedures, inappropriate uses of these packages can result in abuses as well.
Following acceptable norms for disciplines
Every field of study has developed its accepted practices for data analysis. Resnik (2000) states that it is prudent for investigators to follow these accepted norms. Resnik further states that the norms are ‘…based on two factors:
(1) the nature of the variables used (i.e., quantitative, comparative, or qualitative),
(2) assumptions about the population from which the data are drawn (i.e., random distribution, independence, sample size, etc.). If one uses unconventional norms, it is crucial to clearly state this is being done, and to show how this new and possibly unaccepted method of analysis is being used, as well as how it differs from other more traditional methods. For example, Schroder, Carey, and Vanable (2003) juxtapose their identification of new and powerful data analytic solutions developed to count data in the area of HIV contraction risk with a discussion of the limitations of commonly applied methods.
If one uses unconventional norms, it is crucial to clearly state this is being done, and to show how this new and possibly unaccepted method of analysis is being used, as well as how it differs from other more traditional methods. For example, Schroder, Carey, and Vanable (2003) juxtapose their identification of new and powerful data analytic solutions developed to count data in the area of HIV contraction risk with a discussion of the limitations of commonly applied methods.
While the conventional practice is to establish a standard of acceptability for statistical significance, with certain disciplines, it may also be appropriate to discuss whether attaining statistical significance has a true practical meaning, i.e., ‘clinical significance’. Jeans (1992) defines ‘clinical significance’ as “the potential for research findings to make a real and important difference to clients or clinical practice, to health status or to any other problem identified as a relevant priority for the discipline”.
Kendall and Grove (1988) define clinical significance in terms of what happens when “… troubled and disordered clients are now, after treatment, not distinguishable from a meaningful and representative non-disturbed reference group”. Thompson and Noferi (2002) suggest that readers of counseling literature should expect authors to report either practical or clinical significance indices, or both, within their research reports. Shepard (2003) questions why some authors fail to point out that the magnitude of observed changes may too small to have any clinical or practical significance, “sometimes, a supposed change may be described in some detail, but the investigator fails to disclose that the trend is not statistically significant ”.
Lack of clearly defined and objective outcome measurements
No amount of statistical analysis, regardless of the level of the sophistication, will correct poorly defined objective outcome measurements. Whether done unintentionally or by design, this practice increases the likelihood of clouding the interpretation of findings, thus potentially misleading readers.
Provide honest and accurate analysis
The basis for this issue is the urgency of reducing the likelihood of statistical error. Common challenges include the exclusion of outliers, filling in missing data, altering or otherwise changing data, data mining, and developing graphical representations of the data (Shamoo, Resnik, 2003).
Manner of presenting data
At times investigators may enhance the impression of a significant finding by determining how to present derived data (as opposed to data in its raw form), which portion of the data is shown, why, how and to whom (Shamoo, Resnik, 2003). Nowak (1994) notes that even experts do not agree in distinguishing between analyzing and massaging data. Shamoo (1989) recommends that investigators maintain a sufficient and accurate paper trail of how data was manipulated for future review.
The integrity of data analysis can be compromised by the environment or context in which data was collected i.e., face-to face interviews vs. focused group. The interaction occurring within a dyadic relationship (interviewer-interviewee) differs from the group dynamic occurring within a focus group because of the number of participants, and how they react to each other’s responses. Since the data collection process could be influenced by the environment/context, researchers should take this into account when conducting data analysis.
Data recording method
Analyses could also be influenced by the method in which data was recorded. For example, research events could be documented by:
a. recording audio and/or video and transcribing later
b. either a researcher or self-administered survey
c. either closed ended survey or open ended survey
d. preparing ethnographic field notes from a participant/observer
e. requesting that participants themselves take notes, compile and submit them to researchers.
While each methodology employed has rationale and advantages, issues of objectivity and subjectivity may be raised when data is analyzed.
Partitioning the text
During content analysis, staff researchers or ‘raters’ may use inconsistent strategies in analyzing text material. Some ‘raters’ may analyze comments as a whole while others may prefer to dissect text material by separating words, phrases, clauses, sentences or groups of sentences. Every effort should be made to reduce or eliminate inconsistencies between “raters” so that data integrity is not compromised.
Training of Staff conducting analyses
A major challenge to data integrity could occur with the unmonitored supervision of inductive techniques. Content analysis requires raters to assign topics to text material (comments). The threat to integrity may arise when raters have received inconsistent training, or may have received previous training experience(s). Previous experience may affect how raters perceive the material or even perceive the nature of the analyses to be conducted. Thus one rater could assign topics or codes to material that is significantly different from another rater. Strategies to address this would include clearly stating a list of analyses procedures in the protocol manual, consistent training, and routine monitoring of raters.
Reliability and Validity
Researchers performing analysis on either quantitative or qualitative analyses should be aware of challenges to reliability and validity. For example, in the area of content analysis, Gottschalk (1995) identifies three factors that can affect the reliability of analyzed data:
- stability , or the tendency for coders to consistently re-code the same data in the same way over a period of time
- reproducibility , or the tendency for a group of coders to classify categories membership in the same way
- accuracy , or the extent to which the classification of a text corresponds to a standard or norm statistically
The potential for compromising data integrity arises when researchers cannot consistently demonstrate stability, reproducibility, or accuracy of data analysis
According Gottschalk, (1995), the validity of a content analysis study refers to the correspondence of the categories (the classification that raters’ assigned to text content) to the conclusions, and the generalizability of results to a theory (did the categories support the study’s conclusion, and was the finding adequately robust to support or be applied to a selected theoretical rationale?).
Extent of analysis
Upon coding text material for content analysis, raters must classify each code into an appropriate category of a cross-reference matrix. Relying on computer software to determine a frequency or word count can lead to inaccuracies. “One may obtain an accurate count of that word's occurrence and frequency, but not have an accurate accounting of the meaning inherent in each particular usage” (Gottschalk, 1995). Further analyses might be appropriate to discover the dimensionality of the data set or identity new meaningful underlying variables.
Whether statistical or non-statistical methods of analyses are used, researchers should be aware of the potential for compromising data integrity. While statistical analysis is typically performed on quantitative data, there are numerous analytic procedures specifically designed for qualitative material including content, thematic, and ethnographic analysis. Regardless of whether one studies quantitative or qualitative phenomena, researchers use a variety of tools to analyze data in order to test hypotheses, discern patterns of behavior, and ultimately answer research questions. Failure to understand or acknowledge data analysis issues presented can compromise data integrity.
Gottschalk, L. A. (1995). Content analysis of verbal behavior: New findings and clinical applications. Hillside, NJ: Lawrence Erlbaum Associates, Inc
Jeans, M. E. (1992). Clinical significance of research: A growing concern. Canadian Journal of Nursing Research, 24, 1-4.
Lefort, S. (1993). The statistical versus clinical significance debate. Image, 25, 57-62.
Kendall, P. C., & Grove, W. (1988). Normative comparisons in therapy outcome. Behavioral Assessment, 10, 147-158.
Nowak, R. (1994). Problems in clinical trials go far beyond misconduct. Science. 264(5165): 1538-41.
Resnik, D. (2000). Statistics, ethics, and research: an agenda for educations and reform. Accountability in Research. 8: 163-88
Schroder, K.E., Carey, M.P., Venable, P.A. (2003). Methodological challenges in research on sexual risk behavior: I. Item content, scaling, and data analytic options. Ann Behav Med, 26(2): 76-103.
Shamoo, A.E., Resnik, B.R. (2003). Responsible Conduct of Research. Oxford University Press.
Shamoo, A.E. (1989). Principles of Research Data Audit. Gordon and Breach, New York.
Shepard, R.J. (2002). Ethics in exercise science research. Sports Med, 32 (3): 169-183.
Silverman, S., Manson, M. (2003). Research on teaching in physical education doctoral dissertations: a detailed investigation of focus, method, and analysis. Journal of Teaching in Physical Education, 22(3): 280-297.
Smeeton, N., Goda, D. (2003). Conducting and presenting social work research: some basic statistical considerations. Br J Soc Work, 33: 567-573.
Thompson, B., Noferi, G. 2002. Statistical, practical, clinical: How many types of significance should be considered in counseling research? Journal of Counseling & Development, 80(4):64-71.