Data Extraction Prepared for The Agency for Healthcare

Data Extraction Prepared for: The Agency for Healthcare Research and Quality (AHRQ) Training Modules for Systematic Reviews Methods Guide www. ahrq. gov

Systematic Review Process Overview

Learning Objectives To describe why data extraction is important To identify challenges in data extraction To describe the general layout of a data extraction form To suggest methods for collecting data accurately and efficiently To discuss the pros and cons for querying original authors

Why Is Data Extraction Important? To summarize studies in a common format to facilitate synthesis and coherent presentation of data To identify numerical data for meta-analyses To obtain information to assess more objectively the risk of bias in and applicability of studies To identify systematically missing or incorrectly assessed data, outcomes that are never studied, and underrepresented populations

On Data Extraction (I) Extracted data should: Accurately reflect information reported in the publication Remain in a form close to the original reporting, so that disputes can be easily resolved Provide sufficient information to understand the studies and to perform analyses Extract only the data needed, because the extraction process: Is labor intensive Can be costly and error prone Different research questions may have different data needs

On Data Extraction (II) Data extraction involves more than copying words and numbers from the publication to a form. Clinical domain, methodological, and statistical knowledge is needed to ensure the right information is captured. Interpretation of published data is often needed. What is reported is sometimes not what was carried out. Data extraction and evaluation of risk of bias and of applicability typically occur at the same time.

Data Extraction: A Boring Task? “It is an eye-opening experience to attempt to extract information from a paper that you have read carefully and thoroughly understood only to be confronted with ambiguities, obscurities, and gaps in the data that only an attempt to quantify the results reveals. ” — Gurevitch and Hedges (1993) Gurevitch J, Hedges LV. In: Design and analysis of ecological experiments; 1993.

Comparative Effectiveness Reviews: Clarifying Research Terminology In the Evidence-based Practice Center Program, we often refer to two types of tables: Evidence Tables Essentially are data extraction forms Typically are study specific, with data from each study extracted into a set of such tables Are detailed and typically not included in main reports Summary Tables Are used in main reports facilitate the presentation of the synthesis of the studies Typically contain context-relevant pieces of the information included in study-specific evidence tables Address particular research questions

What Data To Collect? Use key questions and eligibility criteria as a guide Anticipate what data summary tables should include: To describe studies To assess outcomes, risk of bias, and applicability To conduct meta-analyses Use the PICOTS framework to choose data elements: Population Intervention (or exposure) Comparator (when applicable) Outcome (remember numerical data) Timing Study design (study setting)

Data Elements: Population, Intervention, and Comparator Population-generic elements may include patient characteristics, such as age, gender distribution, and disease stage. More specific items may be needed, depending upon the topic. Intervention or exposure and comparator items depend upon the extracted study. Study types include randomized trial, observational study, diagnostic test study, prognostic factor study, family-based or population-based genetic study, et cetera.

Data Elements: Outcome (I) Outcomes should be determined a priori with the Technical Expert Panel. Criteria often are unclear about which outcomes to include and which to discard. Example: mean change in ejection fraction versus the proportion of subjects with an increase in ejection fraction by ≥ 5 percent Record different definitions of “outcome” and consult with content experts before making a decision about which definition to use.

Data Elements: Outcome (II) Apart from outcome definitions, quantitative data are needed for meta-analysis: Dichotomous variables (e. g. , deaths, patients with at least one stroke) Count data (e. g. , number of strokes, counting multiple ones) Continuous variables (e. g. , mm Hg, pain score) Survival data Sensitivity, specificity, receiver operating characteristic Correlations Slopes

Data Elements: Timing and Study Design The data elements to be extracted vary by type of study. Consider collecting this information when recording study characteristics for randomized trials: Number of centers (multicenter studies) Method of randomization (adequacy of allocation concealment) Blinding Funding source Whether or not an intention-to-treat analysis was used

Always Provide Instructions Provide “operational definitions” (instructions) indicating exactly what should be extracted in each field of the form. Make sure that all data extractors understand the operational definitions the same way. Pilot-test the forms on several published papers. Encourage communication to clarify even apparently mundane questions.

Single Versus Double Extraction Independent extraction of data by at least two experienced reviewers is ideal but is also resource intensive. There is a tradeoff between cost and the quality of data extraction. Data extraction often takes longer than 2 hours per paper. A reduction in the scope of the work may be necessary if independent data extraction is desired. Careful single extraction by experienced reviewers, with or without crosschecking of selected items by a second reviewer, is a good compromise.

Developing Data Extraction Forms (Evidence Tables) To address all needs, a generic data extraction form will have to be very comprehensive. Although there are common generic elements, forms need to be adapted to each topic or study design to be most efficient. Organization of information in the PICOTS (population, intervention, comparator, outcome, timing, and setting) format is highly desirable. Balance the structure of the form with the flexibility of its use. Anticipate the need to capture unanticipated data. Use an iterative process and have several individuals test the form on multiple studies.

Common Problems Encountered When Creating Data Extraction Forms (Evidence Tables) (I) Forms have to be constructed before any serious data extraction is underway. Original fields may turn out to be inefficient or unusable when coding begins. Reviewers must: be as thorough as possible in the initial set-up, reconfigure the tables as needed, and use a dual review process to fill in gaps.

Evidence Tables: Example (I) First Draft Second Draft

Evidence Tables: Example (II) Final Draft

Common Problems Encountered When Creating Data Extraction Forms (Evidence Tables) (II) Lack of uniformity among outside reviewers: No matter how clear and detailed are the instructions, data will not be entered identically by one reviewer to the next. Solutions: Develop an evidence table guidance document—instructions on how to input data. Limit the number of core members handling the evidence tables to avoid discrepancies in presentation.

Sample Fields From a Table Guidance Document: Vanderbilt University Evidence-based Practice Center In the “country, setting” field, data extractors should list possible settings that could be encountered in the literature: Academic medical center(s), community, database, tertiary care hospital(s), specialty care treatment center(s), substance abuse center(s), level I trauma center(s), et cetera. In the “study design” field, data extractors should list one of the following: Randomized controlled trial, cross-sectional study, longitudinal study, case-control study, et cetera.

Example: Two Reviewers Extract Different Data Reviewer A Reviewer B

Samples of Final Data Extraction Forms (Evidence Tables) For evidence reports or technology assessments that have many key questions, data extraction forms may be several pages long. The next few slides are examples of data extraction forms. Remember, there is more than one way to structure a data extraction form. Trikalinos TA, et al. AHRQ Technology Assessment. Available at: http: //www. cms. gov/determinationprocess/downloads/id 48 TA. pdf.

Examples: Differential Data Extraction by Two Reviewers Trikalinos TA, et al. AHRQ Technology Assessment. Available at: http: //www. cms. gov/determinationprocess/downloads/id 48 TA. pdf.

Characteristics of the Index Test and Reference Standard Trikalinos TA, et al. AHRQ Technology Assessment. Available at: http: //www. cms. gov/determinationprocess/downloads/id 48 TA. pdf.

Results (Concordance/Accuracy) Trikalinos TA, et al. AHRQ Technology Assessment. Available at: http: //www. cms. gov/determinationprocess/downloads/id 48 TA. pdf.

Results (Nonquantitative) Trikalinos TA, et al. AHRQ Technology Assessment. Available at: http: //www. cms. gov/determinationprocess/downloads/id 48 TA. pdf.

Tools Available for Data Extraction and Collection Pencil and paper Word processing software (e. g. , Microsoft Word) Spreadsheet (e. g. , Microsoft Excel) Database software (e. g. , Microsoft Access, Epi Info™) Dedicated off-the-shelf commercial software Homegrown software

Extracting the Data Who should extract the data? Domain experts versus methodologists What extraction method should be used? Single or double independent extraction followed by reconciliation versus single extraction and independent verification Should data extraction be blinded (to authors, journal, results)? Berlin J, for the University of Pennsylvania Meta-analysis Blinding Study Group. Lancet 1997; 350: 185 -6.

Challenges in Data Extraction Problems in data reporting Inconsistencies in published papers Data reported in graphs

Examples of Data Reporting Problems (I) “Data for the 40 patients who were given all 4 doses of medications were considered evaluable for efficacy and safety. The overall study population consisted of 10 (44%) men and 24 (56%) women, with a racial composition of 38 (88%) whites and 5 (12%) blacks. ”

Examples of Data Reporting Problems (II)

Examples of Data Reporting Problems (III)

Inconsistencies in Published Papers Let us extract the number of deaths in two study arms, at 5 years of followup. . .

Results Text Overall Mortality […] 24 deaths occurred in the PCI group, […] and 25 in the MT group […] Dead PCI MED (205) 24 (203) 25 MED and MT = medical treatment; PCI = percutaneous coronary intervention

Overall Mortality Figure Dead PCI MT (205) (203) 24 28 25 35 MT = medical treatment PCI = percutaneous coronary intervention

Clinical Events Table Dead PCI MT (205) (203) 24 25 28 35 32 33 CABG = coronary artery bypass graft MT = medical treatment PCI = percutaneous coronary intervention

Why Do Such Problems Exist? Because so few research reports give effect size, standard normal deviates, or exact p-values, the quantitative reviewer must calculate almost all indices of study outcomes. Little of this calculation is automatic, because results are presented in a bewildering variety of forms and are often obscure. Green BF, Hall JA. Annu Rev Psychol 1984; 35: 37 -53.

Using Digitizing Software Engauge Digitizer, an open-source software: Each data point is marked with an “X, ” and the coordinates are given in a spreadsheet. Source Forge Web site. Engauge Digitizer. Available at: http: //sourceforge. net/projects/digitizer/files/Engauge%20 Digitizer/.

Additional Common Issues Missing information in published papers Publications with at least partially overlapping patient subgroups Potentially fraudulent data

Conclusions Data extraction is laborious and tedious. To err is human: data extractors will identify errors and will err themselves. Interpretation and subjectivity are unavoidable. Data are often not reported in a uniform manner (e. g. , quality, location in paper, metrics, outcomes, numerical value vs. graphs).

Key Messages Key questions will guide reviewers in choosing which information to extract. There is no single correct way to record extracted data. Extracting data requires familiarity with the content and knowledge of epidemiological principles and statistical concepts. Be persistent: Often, one can extract more information than the paper initially appears to contain (e. g. , by digitizing graphs). Be comprehensive: Try to verify the same piece of information from different places in the same article. Sometimes there are surprising inconsistencies. Inconsistencies indicate suboptimal reporting quality at least.

References (I) Berlin J, for the University of Pennsylvania Meta-analysis Blinding Study Group. Does blinding of readers affect the results of meta-analysis? Lancet 1997; 350: 185 -6. Green BF, Hall JA. Quantitative methods for literature reviews. Annu Rev Psychol 1984; 35: 37 -53. Gurevitch J, Hedges LV. Meta-analysis: combining the results of independent experiments. In: Scheiner AM and Gurevich J, eds. Design and analysis of ecological experiments. New York: Chapman & Hall; 1993. p. 347 -70. Source Forge Web site. Engauge Digitizer. Available at: http: //sourceforge. net/projects/digitizer/files/Engauge %20 Digitizer/.

References (II) Trikalinos TA, Ip S, Raman G, et al. Home Diagnosis of Obstructive Sleep Apnea-Hypopnea Syndrome. Technology Assessment (Prepared by Tufts–New England Medical Center Evidencebased Practice Center). Rockville, MD: Agency for Healthcare Research and Quality; August 2007. Available at: http: //www. cms. gov/determinationprocess/dow nloads/ id 48 TA. pdf.

Authors This presentation was prepared by Joseph Lau, M. D. , and Thomas Trikalinos, M. D. , Ph. D. , members of the Tufts–New England Medical Center Evidence-based Practice Center, and Melissa L. Mc. Pheeters, Ph. D. , M. P. H. , and Jeff Seroogy, B. S. , members of the Vanderbilt University Evidence-based Practice Center. The information in this module is currently not included in Version 1. 0 of the Methods Guide for Comparative Effectiveness Reviews (available at: http: //www. effective healthcare. ahrq. gov/rep. Files/2007_10 Draft. Methods. G uide. pdf).