Arch Fam Med
Institution: STANFORD Univ Med Center  | My Account | E-mail Alerts | Access Rights | Sign In
  Vol. 9 No. 10, November 2000 TABLE OF CONTENTS
  •  Online Features
  Original Contribution
 This Article
 •Send to a friend
 • Save in My Folder
 •Save to citation manager
 Citing Articles
 •Citation map
 •Citing articles on HighWire
 •Citing articles on Web of Science (42)
 •Contact me when this article is cited
 Related Content
 •Similar articles in this journal
 Topic Collections
 •Medical Practice
 •Medical Ethics
 •Statistics and Research Methods
 •Alert me on articles by topic

Selection Bias From Requiring Patients to Give Consent to Examine Data for Health Services Research

Steven H. Woolf, MD, MPH; Stephen F. Rothemich, MD; Robert E. Johnson, PhD; David W. Marsland, MD

Arch Fam Med. 2000;9:1111-1118.


Background  New rulings nationwide require health services researchers to obtain patient consent before examining personally identifiable data. A selection bias may result if consenting patients differ from those who do not give consent.

Objective  To compare patients who consent, refuse, and do not answer.

Design  Patients completing an in-office survey were asked for permission to be surveyed at home and for their records to be reviewed. Survey responses and practice billing data were used to compare patients by consent status.

Setting  Urban family practice center.

Patients  Of 2046 eligible patients, 1106 were randomly selected for the survey, were approached by staff, and agreed to participate. Approximately 87% of the nonparticipants were eliminated through a randomization process.

Main Outcome Measure  Consent status.

Results  A total of 33% of patients did not give consent: 25% actively refused, and 8% did not answer. Consenting patients were older, included fewer women and African Americans, and reported poorer physical function than those who did not give consent (P<.05). Patients who did not answer the question were older, included more women and African Americans, and were less educated than those who answered (P<.02). Visits for certain reasons (eg, pelvic infections) were associated with lower consent rates. On multivariate analysis, older age, male sex, and lower functional status were significant predictors of consent.

Conclusions  Patients who release personal information for health services research differ in important characteristics from those who do not. In this study, older patients and those in poorer health were more likely to grant consent. Quality and health services research restricted to patients who give consent may misrepresent outcomes for the general population.

 Jump to Section
 •Author information

CONCERNS ABOUT patient privacy have prompted governments and health care organizations to impose new rulings that require researchers to obtain explicit consent from patients before they can access personally identifiable data and medical records. Although in principle the duty to safeguard privacy is unassailable,1 having to obtain written authorization from patients introduces problems for primary care, health services, and public health research.

The first problem is logistical: locating and contacting patients can be daunting, if not impossible, and may render certain vital classes of studies infeasible.2 Second, restricting data to patients who provide authorization may compromise external validity: patients who are available and willing to provide consent may differ in important respects from other patients. In such cases, inferences based on this subset of patients may mischaracterize the quality and effectiveness of health care for the general population.

For example, a 1997 Minnesota statute requires researchers to obtain patient consent before examining medical records.3 A health plan in that state, participating in a Food and Drug Administration surveillance study, could not examine the records of patients using a new drug (to investigate reports of the occurrence of seizures) without obtaining patient permission. Despite 2 rounds of letters (and telephone calls to nonrespondents), only 52% of patients responded, and only 19% gave consent.4 Unreported by the authors is whether this subgroup of patients differed in important ways from other patients taking the drug.

Yawn et al,5 implementing the same law to conduct outcomes research at a health care system, confirmed that refusal was nonuniform, occurring more often among women, young persons, and patients visiting for certain reasons. They noted, for example, that the refusal rate among female smokers whose fetus or infant died following preterm birth or from sudden infant death syndrome was 4 times greater than among other patients (B. P. Yawn, e-mail communication, 1999). They warned that this "authorization bias" could distort outcomes data. For example, if data from women withholding consent were excluded, an epidemiologic study of maternal smoking and child mortality might not expose its association with prematurity and sudden infant death syndrome. Yawn et al noted that the restriction itself—barred access to data on nonconsenting patients—made the bias "undefined and undefinable."5 We know of no other studies that formally examined this phenomenon.

We acquired an opportunity to explore this matter further when our institutional review board (IRB) instructed us to ask patients for consent before contacting them at home for surveys or examining their medical records. The IRB permission to access limited demographic and health status data, even for patients who do not consent to further participation, enabled us to compare the characteristics of consenting and nonconsenting patients.

 Jump to Section
 •Author information


The Department of Family Practice of Virginia Commonwealth University (VCU) has established a practice-based research network that collects prospective data on primary care patients and practices. Eventually, data on patient risk factors, provider characteristics, performance indicators, and outcomes will derive from a variety of sources (eg, patients, medical records, billing data). Currently, the primary data come from a survey that patients complete while awaiting their appointments with physicians. The Health Assessment Survey (HAS) asks about demographics (age, sex, race, education, income), functional status, cigarette use, and (for women) recent Papanicolaou testing.6


This study was conducted at the first site in the network, a university-affiliated family practice clinic in downtown Richmond, Va, averaging 10 700 annual visits. The practice includes large proportions of low-income and minority patients; approximately 34% are either Medicaid beneficiaries or uninsured and 71% are African American.


The institutional review board required no explicit consent for the HAS. A statement at the beginning of the HAS reviewed how the results would be used, and patients (or the parent or guardian of minors) decided whether to proceed. Although completed forms left the practice when they were forwarded to VCU, the only personal identifier they contained was the patient's medical record number, which meant little outside the practice and, therefore, functioned as a key code. (Under federal statute, this permitted the IRB to alter consent procedures7 and waive documentation.8)

We linked HAS responses with administrative visit data (birth date, visit date, Current Procedural Terminology 4 and International Classification of Diseases, Ninth Revision, Clinical Modification9 [ICD-9-CM] codes). The practice exported this information to VCU researchers in an electronic file from which personally identifiable data were removed. We used the medical record number key code to link this information with HAS responses. The inability to link this information to the personal identity of the patient permitted the IRB to waive consent.7

Future plans call for augmenting these data with mail or telephone surveys and with audits of medical records. Because these studies will require access to personal information (eg, mailing address), our IRB instructed us to obtain explicit patient authorization. We therefore added a question soliciting this consent to the end of the HAS survey:

"Our practice is working with Family Practice researchers at Virginia Commonwealth University to ensure that you receive the best possible care. We are asking you to give them permission to contact you by phone or mail and to review your medical record to inquire further about your health and the medical care you are receiving (see the blue information sheet for details). All obtained information will be kept strictly confidential by the VCU researchers and us. (Mark one oval.)

O I give permission to be contacted and for my records to be reviewed to obtain information about the quality of care I have received.

O I do not give permission to be contacted or for my records to be reviewed."

The blue sheet (Figure 1), attached to the HAS, provided study details required by the IRB for informed consent.

View larger version (118K):
[in this window]
[in a new window]
Figure 1. One-page fact sheet accompanying survey.


We monitored consent rates for 29 weeks (May through November 1999). During that time, 2046 patients visiting the practice met HAS eligibility criteria: visiting a physician, not making their first visit, age of 14 years or older (minors had to be accompanied by a parent or guardian), and HAS not taken for 1 year. Of these 2046 patients, approximately 60% were selected randomly to take the survey; we applied this random filter to limit the burden on office staff. Approximately 6% of eligible patients did not participate because of inability (eg, too ill), staff workload or oversights, or patient refusal (at this clinic we previously documented an 89% participation rate among approached patients6). The 1106 patients (54%) who were approached by office staff and agreed to take the HAS constitute the sample for this study. These patients were younger and included more women than the 940 nonparticipants (P<.01).


There were 3 possible responses to the consent question: active consent, active denial, or no answer (passive denial). The last 2 groups formed a combined category of patients who, actively or passively, did not give consent.


Age and sex were available in practice billing data, and race, educational status, and income were provided by the HAS. The latter, adapted from the US Census questionnaire and the National Health Interview Survey, asked patients to choose 1 of 7 levels of education and 1 of 8 ranges of annual combined family income.

The ICD-9-CM diagnoses recorded at the index visit were classified with Clinical Classification Software (CCS).10 Severity of illness, based on ICD-9-CM codes for the past year, was evaluated with the Johns Hopkins Adjusted Clinical Groups (ACG) Case-Mix System.11 After classifying diagnoses into ambulatory diagnostic groups, the ACG branching algorithm assigns patients to 1 of 93 mutually exclusive ACG categories.12 Functional status was measured with the 36-Item Short-Form Health Survey,13 included in the HAS. We used the 36-Item Short-Form Health Survey Physical Component Summary (PCS) and Mental Component Summary (MCS) to aggregate its 8 physical and mental functional status domains.14 Higher PCS and MCS scores reflect higher functional status.


Results were analyzed with SAS software.15 The statistical significance of differences between means was assessed by analysis of variance (t test). To determine if differences in summary measures across groups deviated significantly from what would be expected if groups were assigned randomly, we used an asymptotic approximation to the permutation t test: the 2-sample t test corrected for sampling from a finite population (the sampling frame).16 Differences between percentages were assessed with the {chi}2 test. Associations between the consent responses (consent given, consent denied) and demographic and health variables were assessed with multivariate logistic regression, thereby producing variable-specific adjusted odds ratios. The model included indicators for missing data.

We imputed consent responses for patients not answering the question by likening their response to their nearest neighbor (as measured by the Mahalanobis distance17) in the responding group based on age, sex, the number of missing survey items, and the number of the last completed item before the consent question. This nonparametric discrimination method correctly classified 75% of the responding group (Goodman-Kruskal {gamma}, 0.84).

 Jump to Section
 •Author information


The mean age of the 1106 patients who participated in the HAS was 41 years; 78% were female, 55% had a high school or higher education, and 69% were above the poverty level in Richmond, Va, for a family of 4 (combined annual family income, >$15 000). Twenty-six percent of respondents were current cigarette smokers.


Of the 1106 HAS participants, 743 (67%) answered yes to the consent question (ie, giving researchers permission to contact them for further surveys and review their records). A total of 278 (25%) actively denied consent, and 85 (8%) did not answer. Approximately 80% of the patients in the latter group also did not answer preceding questions (compared with 41% for patients who did answer the consent question), suggesting that many gave no answer because they did not reach the end of the form.


Patients who gave consent (n = 743) were significantly older (P<.01) and included fewer women (P<.001) and African Americans (P<.01) than patients who did not give consent (n = 363) (Table 1). Within the latter group of nonconsenters, we found significant demographic differences between patients who actively denied consent (n = 278) and those who passively denied by not answering (n = 85). Those who did not answer were older (P<.001), had less education (P<.01) and income (P = .05), and included more African Americans (P<.01) than those who actively denied consent.

View this table:
[in this window]
[in a new window]
Table 1. Characteristics of Respondents by Consent Status*

Because these data indicated that the active and passive denial groups were demographically distinct, we excluded patients who did not answer and contrasted the responses of patients who actively denied consent (n = 278) with those of patients who gave consent (n = 743). Patients who actively denied consent were younger (P<.001), included more women (P<.01), and were more educated (P = .05) than those who gave consent (Table 1). Refusal rates decreased with age (Figure 2), with the 18- to 24-year-old age group having the highest refusal rate (34%).

View larger version (28K):
[in this window]
[in a new window]
Figure 2. Responses to consent question by age. An association between age and consent is apparent for adults aged 18 to 64 years. Age groups that do not seem to follow this pattern (ages 14-17, 65-74, and >=75 years) had small sample sizes, with wide confidence intervals limiting the precision of the mean.

The 85 patients who did not answer (neither actively giving nor denying consent) differed significantly from those who answered (n = 1021); they were older (P = .01), included more women (P<.01), were less educated (P = .02), and included more African Americans (P<.001). Although statistical significance was not reached, they appeared to be less affluent than patients who answered the question (P = .07). Current smoking rates were similar among all groups.


Functional Status

Patients who gave consent had significantly lower mean PCS scores than patients who did not consent (P<.001) (Table 1). When patients who did not answer (passive denials) were excluded from the latter group, the difference in PCS scores was larger, and the difference in mean MCS scores approached significance (P = .07). Patients with the highest physical function scores were significantly more likely to refuse (Figure 3). The PCS scores were significantly lower for patients who did not answer the question than for those who actively denied consent (P = .02).

View larger version (16K):
[in this window]
[in a new window]
Figure 3. Consent responses and physical functional status. Denial rates are highest in patients with mean 36-Item Short-Form Health Survey Physical Component Summary (PCS) scores above the 75th percentile. Corresponding PCS scores by percentile interval were as follows: 0 to 29.8 (0%-12.5%), greater than 29.8 to 36.0 (>12.5%-25%), greater than 36.0 to 41.8 (>25%-37.5%), greater than 41.8 to 46.2 (>37.5%-50.0%), greater than 46.2 to 50.0 (>50.0%-62.5%), greater than 50.0 to 53.0 (>62.5%-75.0%), greater than 53.0 to 55.9 (>75.0%-87.5%), and greater than 55.9 to 100 (>87.5%).

Diagnoses Recorded at Index Visit

Denial rates were highest when patients visited for contraception (CCS 11.1, 46.7%), urinary disorders (CCS 10.1, 39.4%), uncomplicated diabetes (CCS 3.2, 37.5%), headache (CCS 6.5, 37.0%), or female genital disorders (CCS 10.3, 35.4%). The CCS categories were distributed similarly for those who gave and denied consent, although patients who denied consent were more likely to visit for certain reasons: miscellaneous conditions of the nervous system (CCS 6.9), parasitic and other infections (CCS 1.4), and contraception (CCS 11.1) (P = .02, .06, and .09, respectively). Within the broad category of female genital disorders (CCS 10.3), there were no significant differences, but denials were more likely among women with pelvic infections (CCS 10.3.2) (P = .02). A total of 20% of patients who did not answer the question visited for diabetes (CCS 3.2-3.3), a much higher proportion than among those who answered (5.7%) (P<.01).

Severity of Illness

The ACG distributions of patients who gave and denied consent did not differ significantly. Some differences were noted, however, in 2 of the 32 ambulatory diagnostic groupings on which they are based. Patients who gave consent were more likely than those who denied consent to have visited for "stable chronic medical conditions" (eg, type 2 diabetes mellitus) (59.2% vs 50.2%, P = .01), and the rate was even higher (67.5%) for those who did not answer. The difference for "unstable chronic medical conditions" (16.1% vs 11.5% for those who gave and denied consent, respectively) was smaller and did not reach statistical significance (P = .07).

Accounting for Patients Who Did Not Answer

We do not know how many patients did not answer the last question because they did not reach the end of the form, nor can we predict their answers. Although this group had unique demographic and health characteristics, we doubt that their answers would have changed our overall findings. Using nearest neighbor estimates, we predict that only 53 (62%) of the 85 patients who did not answer would have given consent. When we combined these 53 patients with the 743 who actively gave consent and the remaining 32 nonresponders with the 278 who actively denied consent, we found no difference in the previously described demographic and health profiles of the 2 populations.

Multivariate Logistic Regression Analysis

Multivariate logistic regression analysis revealed a significant and independent association between giving consent and increased age (P = .01), male sex (P = .01), and lower mean PCS scores (P = .03) (Table 2). Patients were also less likely to give consent if they did not answer questions on income (P<.01), smoking (P<.01), or functional status (PCS and MCS, P<.01). We repeated the multivariate logistic regression after reassigning patients in the passive denial category to either the active consent or active denial category, using imputed values from the nearest neighbor analysis. It revealed the same set of significant associations, except that missing data on smoking status lost significance.

View this table:
[in this window]
[in a new window]
Table 2. Odds Ratios for Consent (vs No Consent) From Multivariate Logistic Regression Analysis

 Jump to Section
 •Author information

This study found that patients who gave consent differed in important demographic and health characteristics from those who did not consent. The demographic variables that differed between these groups (eg, age, sex, race) have important associations with health outcomes. Furthermore, patients who denied consent differed in functional status and in the health conditions for which they sought care. Patients in poorer health were more likely to consent. In our model, for example, the probability of obtaining consent would be 89% for a 65-year-old man with a PCS of 30 but only 70% for a 35-year-old woman with a PCS of 60. Differences in some indices followed this pattern but did not reach statistical significance. Whether this reflects a lack of difference or inadequate power is unclear. For example, the CCS category with the most patients (CCS 10, genitourinary disorders) had only 165 patients. With this sample size, the power to detect a true consent rate that was 10% higher or lower than the observed value (67.5%) was only 49%. Power would be even smaller for less populated categories.

To the extent that these data are generalizable, they indicate that research restricted to patients who give explicit written consent may mischaracterize the health status of the larger population. Whether statistical adjustments for these discordances could compensate for this selection bias is unclear. It is also uncertain how this selection bias interrelates, both in magnitude and synergism, to other biases, such as those introduced by nonresponse and nonparticipation. For example, patients who took the HAS were younger and included more women than those who did not participate.6 In effect, the selection bias we report represents the last of a series of biases that filter out patients from the original population.

Both the internal and external validity of this study are contestable. The internal validity—the extent to which our findings are valid for our clinic population—may be limited by the wording and location of the consent question on the HAS form and by the setting in which the question was asked. These factors might explain our high refusal rate. When Yawn et al5 requested patients' consent to comply with Minnesota law, only 3.6% refused and 4.5% were undecided. Similarly, the Mayo Clinic reported a consent rate of 96% among patients who returned forms.2 Both programs sought access only to medical records, however. Our question combined 2 requests: permission to examine medical records and permission to be contacted for surveys. Our high rejection rate is not unlike the large proportion of patients who do not return mail surveys (ie, our question may simply have preselected these nonresponders).

The topics on the HAS that precede the consent question and their wording may have influenced consent. For example, more patients may have given consent if preceding questions addressed a condition that they had. The reading level and cultural orientation of the text may have dissuaded many of our patients with limited education or minority backgrounds. Potentially sensitive questions on the HAS may have discouraged consent. Patients who did not answer the question about income were, on multivariate analysis, significantly less likely to give consent.

Locating the consent question at the end of the form may have contributed to nonresponse. Patients who did not give consent were, on multivariate analysis, less likely to have answered the preceding questions. Patients who did not answer tended to be older, more ill, and less educated and would be expected to proceed more slowly through a survey. Moreover, asking the consent question while patients waited for appointments might yield different answers than inquiring at a less stressful time. Finally, consent may have an independent association with visit frequency, an association we hope to analyze in future studies.

The external validity of the study—the extent to which our findings can be extrapolated elsewhere—may be limited. Patients attending a family practice center in downtown Richmond, Va, a predominantly African American and low-income population, differ from patients in other settings. Moreover, much of the information that patients read in our fact sheet concerned our study and research network and would not apply when other researchers solicit consent. We would expect patients' willingness to offer consent and the characteristics of those who refuse to differ by setting.

Yawn et al,5 for example, described a different profile for patients who denied consent. In their study, refusal was most common among women, persons aged 41 to 64 years, and patients visiting for mental illness, eye care, trauma, and gynecologic problems. Their patient population differed from ours. They also solicited consent differently. In addition to outpatient clinics, authorization was obtained at the emergency department and hospital. A receptionist or hospital registration clerk requested consent as part of normal registration. The form and accompanying material addressed only the use of medical records (with no mention of surveys) and gave patients greater reassurance that their anonymity would be protected.

Despite these differences, both our study and the work by Yawn et al reached the same conclusion: patients who give consent differ in important characteristics from those who do not. We know of no other studies in the medical setting that have examined this bias. Studies in school settings have also shown that having to obtain parental consent to survey children leads to underrepresentation of at-risk youth.18 To the extent that this occurs in medicine, studies of the effectiveness and quality of care that require explicit consent could produce skewed results, potentially misleading clinicians, policymakers, and consumers.

The net harms to society of requiring consent should be weighed against the benefits. If obtaining consent selectively omits data that mask a problem in women receiving timely mammograms, for example, then data privacy could come at the cost of higher breast cancer mortality. When this selection bias is extended across all studies for which written authorization might be required—public health surveillance, outcomes research, quality assurance—the potential grows for misjudgments of epidemiologic trends, risk rates, and the effectiveness and safety of tests and treatments. Some public health groups have warned about the harms to society from such data errors.19

Nonetheless, strong ethical arguments support the need for consent, at least for the release of personally identifiable data.20 Patients deserve to know when unauthorized individuals examine their medical records or have access to their address or other personal details. Electronic incursions into personal information21 and violations of privacy by unauthorized entities (eg, employers,22 marketing firms23) have heightened public sensitivity to this issue. Access to personal genetic information has raised further concerns about the need for consent, as illustrated by recent experience in Iceland.24-25 According to a Louis Harris poll, 27% of Americans believe that their medical records have been improperly disclosed, and 64% object to researchers gaining access without explicit consent.26

Strong public sentiment prompted federal action by Congress27 and the Clinton administration (privacy regulations on electronically transferred medical data,28 bioethics commission on human subjects research29). As of 1999, 5 state legislatures had passed or were considering comprehensive health information statutes.30 Federal officials have recently terminated research at major academic centers, often because of privacy violations.31

The dilemma posed by the selection bias we observed reflects a tension between competing ethical duties: the rights of individuals to privacy and, conversely, of society to accurate monitoring of the effectiveness and safety of health care.2 Which duty should prevail is an ethical30 and legal issue32 worthy of thoughtful discourse by providers, ethicists, researchers, policymakers, and patients. The first step is to confirm the existence of the selection bias. Further work by investigators in other research settings and with larger sample sizes will help elucidate this problem, its magnitude, and the extent to which it can be attenuated by better methods for obtaining consent.

Since this article was accepted for publication, the Institute of Medicine published a report on protecting data privacy in health services research. The report warned that "the requirement of consent would likely lead to bias and invalid findings, because those who opt out might differ systematically from those giving consent."33 Our findings validate this concern.

 Jump to Section
 •Author information

Accepted for publication August 9, 2000.

This research was supported in part by a Robert Wood Johnson Generalist Physician Faculty Scholars Award to Dr Rothemich.

Presented in part at the 27th Annual Meeting of the North American Primary Care Research Group, San Diego, Calif, November 10, 1999, and the 2000 Annual Meeting of the Association for Health Services Research, Los Angeles, Calif, June 26, 2000.

We thank Barbara P. Yawn, MD, MSc, and Finn-Aage Esbensen PhD, MA, for their helpful contributions to the manuscript.

Reprints: Steven H. Woolf, MD, MPH, Department of Family Practice, Medical College of Virginia, Virginia Commonwealth University, 3712 Charles Stewart Dr, Fairfax, VA 22033.

From the Department of Family Practice, Virginia Commonwealth University (Drs Woolf, Rothemich, and Marsland), and the Department of Mathematical Sciences (Dr Johnson), Virginia Commonwealth University, Richmond. Dr Johnson is now with the Department of Biostatistics.

 Jump to Section
 •Author information

1. Donaldson MS, ed, Lohr KN, ed. Health Data in the Information Age: Use, Disclosure, and Privacy. Washington, DC: National Academy Press; 1994.
2. Melton III LJ. The threat to medical-records research. N Engl J Med. 1997;337:1466-1470. FREE FULL TEXT
3. Access to Health Records, Subdivision 3-Patient Consent to Release of Records. Minnesota Law 144.335. Effective 1/1/97.
4. McCarthy DB, Shatin D, Drinkard CR, Kleinman JH, Gardner JS. Medical records and privacy: empirical effects of legislation. Health Serv Res. 1999;34(1, pt 2):417-425.
5. Yawn BP, Yawn RA, Geier GR, Xia Z, Jacobsen SJ. The impact of requiring patient authorization for use of data in medical records research. J Fam Pract. 1998;47:361-365. ISI | PUBMED
6. Rothemich SF, Woolf SH, Johnson RE, Marsland DW. Evaluating the administration of an office-based health survey in a primary care practice. J Clin Epidemiol. 2000;53(10):1002-1012.
7. Federal Policy for the Protection of Human Subjects (1994) (codified at 45 CFR §46.116(d)). 56. Federal Register 28003. (June 18, 1991).
8. Federal Policy for the Protection of Human Subjects (1994) (codified at 45 CFR §46.117(c)2). 56 Federal Register 28002, (June 18, 1991).
9. International Classification of Diseases, Ninth Revision, Clinical Modification. Washington, DC: Public Health Service, US Dept of Health and Human Services; 1988.
10. Clinical Classification Software (CCS) Summary and Download: Summary and Downloading Information. Rockville, Md: Agency for Health Care Policy and Research. Available at: http://www.ahrq.gov/data/hcup/ccs.htm. Accessed February 5, 2000.
11. Starfield B, Weiner J, Mumford L, Steinwachs D. Ambulatory care groups: a categorization of diagnoses for research and management. Health Serv Res. 1991;26:53-74. ISI | PUBMED
12. The Johns Hopkins University ACG Case Mix Adjustment System: Documentation for PC-DOS and UNIX. Baltimore, Md: Johns Hopkins University School of Hygiene and Public Health, Health Services Research and Development Center; 1998. Further details available at: http://www.acg.jhsph.edu/. Accessed February 5, 2000.
13. Ware JE Jr, Kosinski M, Keller SD. SF-36 Physical & Mental Health Summary Scales: A User's Manual. Boston, Mass: Health Institute, New England Medical Center; 1994.
14. Ware JE Jr, Kosinski M, Bayliss MS, McHorney CA, Rogers WH, Raczek A. Comparison of methods for the scoring and statistical analysis of SF-36 health profile and summary measures: summary of results from the Medical Outcomes Study. Med Care. 1995;33(4 suppl):AS264-AS279.
15. SAS Statistical Software [computer program]. Cary, NC: SAS Institute Inc; 1996.
16. Cox DR, Hinkley DV. Theoretical Statistics. New York, NY: Chapman & Hall Ltd; 1974.
17. Morrison DF. Multivariate Statistical Methods. New York, NY: McGraw-Hill Inc; 1976.
18. Esbensen FA, Miller MH, Taylor TJ, He N, Freng A. Differential attrition rates and active parental consent. Eval Rev. 1999;23:316-335. FREE FULL TEXT
19. American College of Epidemiology. Statement on Health Data Control, Access, and Confidentiality. Available at: http://www.acepidemiology.org/data.html. Accessed February 14, 2000.
20. Woodward B. Challenges to human subject protections in US medical research. JAMA. 1999;282:1947-1952. FREE FULL TEXT
21. Hagel III J, Rayport JF. The coming battle for customer information. Harv Bus Rev. 1997;75:53-55, 58, 60-61. ISI | PUBMED
22. Studdert DM RAND Institute for Civil Justice. Direct contracts, data sharing and employee risk selection: new stakes for patient privacy in tomorrow's health insurance markets. Am J Law Med. 1999;25:233-265. ISI | PUBMED
23. Lo B, Alpers A. Uses and abuses of prescription drug information in pharmacy benefits management programs. JAMA. 2000;283:801-806. FREE FULL TEXT
24. Gulcher JR, Stefansson K. The Icelandic Healthcare Database and informed consent. N Engl J Med. 2000;342:1827-1830. FREE FULL TEXT
25. Annas GJ. Rules for research on human genetic variation: lessons from Iceland. N Engl J Med. 2000;342:1830-1833. FREE FULL TEXT
26. Harris/Equifax. Health Care Information Privacy: A Survey of the Public and Leaders. Conducted for Equifax. New York, NY: Louis Harris & Associates; 1993.
27. Health Insurance Portability and Accountability Act of 1996. §264[00a9](2), Pub L No. 104-191, 110 Stat 1936, 2033 (1996).
28. Standards of Privacy of Individually Identifiable Health Information: Proposed Rule, 64. Federal Register 55918-60065 (1999) (codified at 45 CFR §160164).
29. Executive Order 12975 of October 3, 1995. Available at: http://www.bioethics.gov/about/eo12975.htm. Accessed February 15, 2000.
30. Hodge JG Jr, Gostin LO, Jacobson PD. Legal issues concerning electronic health information: privacy, quality, and liability. JAMA. 1999;282:1466-14671. FREE FULL TEXT
31. Phillips DF. IRBs search for answers and support during a time of institutional change. JAMA. 2000;283:729-730. FREE FULL TEXT
32. Gostin L. Health care information and the protection of personal privacy: ethical and legal considerations. Ann Intern Med. 1997;127:683-690. FREE FULL TEXT
33. Institute of Medicine. Protecting Data Privacy in Health Services Research: Committee on the Role of Institutional Review Boards in Health Services Research Data Privacy Protection. Washington, DC: National Academy Press; 2000.


Up to a quarter of the Australian population may have suboptimal health literacy depending upon the measurement tool: results from a population-based survey
Barber et al.
HEALTH PROMOT INT 2009;24:252-261.

Feasibility and cost of obtaining informed consent for essential review of medical records in large-scale health services research
Noble et al.
J Health Serv Res Policy 2009;14:77-81.

Evaluating Predictors of Geographic Area Population Size Cut-offs to Manage Re-identification Risk
El Emam et al.
J. Am. Med. Inform. Assoc. 2009;16:256-266.

Holding personal information in a disease-specific register: the perspectives of people with multiple sclerosis and professionals on consent and access
Baird et al.
J. Med. Ethics 2009;35:92-96.

Regulatory Challenges for the Resuscitation Outcomes Consortium
Tisherman et al.
Circulation 2008;118:1585-1592.

The effect of privacy legislation on observational research
Gershon and Tu
CMAJ 2008;178:871-873.

Extracting information from hospital records: what patients think about consent
Campbell et al.
Qual Saf Health Care 2007;16:404-408.

Impact of privacy legislation on the number and characteristics of people who are recruited for research: a randomised controlled trial.
Trevena et al.
J. Med. Ethics 2006;32:473-477.

Mother's consent to linkage of survey data with her child's birth records in a multi-ethnic national cohort study
Tate et al.
Int J Epidemiol 2006;35:294-298.

Recruiting patients to medical research: double blind randomised trial of "opt-in" versus "opt-out" strategies
Junghans et al.
BMJ 2005;331:940.

Cluster Randomized Trials of Professional and Organizational Behavior Change Interventions in Health Care Settings
Grimshaw et al.
The ANNALS of the American Academy of Political and Social Science 2005;599:71-93.

Practice-Based Research Network Studies in the Age of HIPAA
Pace et al.
Ann Fam Med 2005;3:S38-S45.

Patterns of Consent in Epidemiologic Research: Evidence from Over 25,000 Responders
Dunn et al.
Am J Epidemiol 2004;159:1087-1094.

Impracticability of Informed Consent in the Registry of the Canadian Stroke Network
Tu et al.
NEJM 2004;350:1414-1421.

The privacy paradox: laying Orwell's ghost to rest
Upshur et al.
CMAJ 2001;165:307-309.

Protecting the Privacy of Family Members in Research
Levinson et al.
JAMA 2001;285:1960-1963.

© 2000 American Medical Association. All Rights Reserved.