Search

Objectives To develop a patient risk adjustment model for experience of care (EOC) quality measures for long-term care hospitals (LTCHs) that includes mode of survey administration. To assess presence of nonresponse bias in the adjusted facility-level scores. Design Tested three modes of collecting the EOC data: mail-only, mixed (i.e., mail with telephone follow-up), and in-facility. Used sequential modeling and impact analysis, specified a risk and mode adjustment model and evaluated presence of non-response after adjustment. Setting Long-term care hospitals Participants 1,364 patients and 69 LTCHs. Intervention Not applicable Main Outcome Measures Risk and mode adjusted responses to twenty-eight survey questions and six facility-level scores derived from survey responses. Results Mode of data collection, and patient risk variables of age, sex, overall health, overall mental health, marital status, education, race, and whether a proxy responded were significant were included in the model. Clinical variables were not significant. The in-facility mode was associated with significantly higher performance scores than the other modes. When the recommended risk and mode adjustment model was applied, nonresponse bias was not observed in any mode. Conclusions LTCH EOC data should be adjusted for patient risk variables including mode of data collection.

Background Survey research is common practice in podiatry literature and many other health-related fields. An important component of the reporting of survey results is the provision of sufficient information to permit readers to understand the validity and representativeness of the results presented. However, the quality of survey reporting measures in the body of podiatry literature has not been systematically reviewed. Objective To examine the reporting of response rates and nonresponse bias within survey research articles published in the podiatric literature in order to provide a foundation with regard to the development of appropriate research reporting standards within the profession. Methods This study reports on a secondary analysis of survey research published in the Journal of the American Podiatric Medical Association, the Foot, and the Journal of Foot and Ankle Research. 98 surveys published from 2000 to 2018 were reviewed and data abstracted regarding the report of response rates and non-response bias. Results 67 surveys (68.4%) report a response rate while only 36 articles (36.7%) mention non-response bias in any capacity. Conclusions The findings suggest that there is room for improvement in the quality of reporting response rates and nonresponse in the body of podiatric literature involving survey research. Both nonresponse and response rate should be reported to assess survey quality. This is particularly problematic for studies that contribute to best practices.

We evaluate how nonresponse affects conclusions drawn from survey data and consider how researchers can reliably test and correct for nonresponse bias. To do so, we examine a survey on labor market conditions during the COVID-19 pandemic that used randomly assigned financial incentives to encourage participation. We link the survey data to administrative data sources, allowing us to observe a ground truth for participants and nonparticipants. We find evidence of large nonresponse bias, even after correcting for observable differences between participants and nonparticipants. We apply a range of existing methods that account for nonresponse bias due to unobserved differences, including worst-case bounds, bounds that incorporate monotonicity assumptions, and approaches based on parametric and nonparametric selection models. These methods produce bounds (or point estimates) that are either too wide to be useful or far from the ground truth. We show how these shortcomings can be addressed by modeling how nonparticipation can be both active (declining to participate) and passive (not seeing the survey invitation). The model makes use of variation from the randomly assigned financial incentives, as well as the timing of reminder emails. Applying the model to our data produces bounds (or point estimates) that are narrower and closer to the ground truth than the other methods.

From February to April 2020, as COVID-19 hit the U.S. economy, the official unemployment rate (UR) climbed from 3.5 percent—the lowest in more than 50 years—to 14.7—the highest since current measurement began in January 1948. This unprecedented, speedy quadrupling of UR coincided with major disruptions in survey-data-collection procedures and a dramatic, differential drop in response rates. To what extent did measurement issues contribute to this quadrupling? We revisit two recently studied potential biases in the Current Population Survey: rotation group bias (Krueger, Mas and Niu, 2017) and difficulty-of-reaching bias (Heffetz and Reeves, 2019). We extend the original analyses to the years prior to the crisis and focus on the six months of peak UR, from April to September 2020. Our ballpark estimates suggest that the peak official UR figure could be biased by up to ∼1.5 percentage points in either direction.

Traditional urban park research has used self-reported surveys and activity logs to examine relationships between health benefits, park use, and park features. An alternative approach uses participating mapping methods. This study sought to validate and expand on previous participatory mapping research methods and findings and address spatial scaling by applying these methods to a large urban park system. Key challenges for spatial scaling included ambiguity in park classification and achieving representative sampling for larger and spatially-disbursed urban residents. We designed an internet-based public participation GIS (PPGIS) survey and used household and volunteer sampling to identify the type and locations of urban park benefits. Study participants (n = 816) identified locations of physical activities and other urban park benefits (psychological, social, and environmental) which were analyzed by park type. Consistent with previous suburb-scale research, we found significant associations between urban park type and different urban park benefits. Linear parks were significantly associated with higher intensity physical activities; natural parks were associated with environmental benefits; and community parks were associated with benefits from social interaction. Neighborhood parks emerged as significantly associated with psychological benefits. The diversity of park activities and benefits were positively correlated with park size. Distance analysis confirmed that physical benefits of parks were closest to participant domicile, while social and environmental benefits were more distant. These results validate previous suburb-scale findings despite greater variability in park types and sample populations. Future urban park research using participatory mapping would benefit from greater effort to obtain participation from under-represented populations that can induce nonresponse bias, and analyses to determine whether system-wide results can be disaggregated by suburb or neighborhood to address social inequities in urban park benefits Refereed/Peer-reviewed

Background A wealth of research has examined the prevalence and impact of adverse childhood experiences (ACEs) via various research methodologies. Some of these studies have also examined the presence of nonresponse bias, showing minimal nonresponse bias effects. More recently, many states and the District of Columbia have used the Behavioral Risk Factor Surveillance System (BRFSS) to examine ACEs, however, limited research exists on the impact of nonresponse bias in ACE studies using the BRFSS. Methods This study used data from the 2014–2015 South Carolina BRFSS to examine nonresponse bias to the ACE module. Results Significant differences between responders and non-responders were found for sex, age, race/ethnicity, education, income, and rurality. Findings indicate that marginalized populations were more likely to be under-represented in ACE survey data because of nonresponse, potentially limiting targeted prevention and intervention efforts. Conclusion Future research should examine differences in health and social outcomes between responders and non-responders to the ACE module in the BRFSS and ways to increase responses from marginalized groups.

Introduction The purpose of this systematic review was to examine literature on workplace factors associated with neck pain or symptoms in computer users performing clerical functions. Methods A systematic search of the Cochrane, Medline, CINAHL, and EMBASE databases was conducted for observational and experimental studies published since 2000. This review applied the case definition of The Bone and Joint Decade 2000-2010 Task Force on Neck Pain and Its Associated Disorders. Results Seven hundred twenty-nine studies were identified. Seven hundred and two studies were excluded. Twenty-seven studies fulfilled inclusion criteria and were assessed for risk of bias. Cross-sectional studies were commonly at risk from nonresponse bias and lack of adequate case definitions. Experimental studies were mostly at risk of bias due to confounding and participant recruitment methods. Conclusions Neck pain was not significantly associated with high job demands, low skill discretion, low decision authority, or low peer support. However, when these variables were combined with increased duration of computing tasks, or ergonomic demands, they reached significance. Supervisor support was found to be the only significant buffer capable of preventing these variables reaching significance in female office workers.

Background Survey-based research is essential for evaluating the outcomes of health care in an era of patient-centered care. However, many such studies are hampered by poor response rates in completion of study questionnaires, thus limiting the generalizability of any findings. The objectives of this analysis were to identify independent variables associated with nonresponse to surveys following breast reconstruction to improve future patient-reported outcomes research. Materials and methods The Mastectomy Reconstruction Outcomes Consortium is a prospective cohort study involving 11 leading medical centers from the United States and Canada. Nonresponse rates for surveys assessing satisfaction with breast, satisfaction with care (BREAST-Q), depression (Patient Health Questionnaire-9), and anxiety (Generalized Anxiety Disorder-7) were measured at 1 y and 2 y postoperatively. Clinical complication rates were compared between responders and nonresponders, and multivariable models were used to assess predictors of nonresponse. Results Among 2856 women in the analytic cohort, 1882 (65.9%) underwent implant-based, 817 (28.6%) received autologous, and 157 (5.5%) underwent latissimus dorsi myocutaneous flap breast reconstructions. Nonresponse rates to surveys at 1 y and 2 y were 27.8% and 34.4%, respectively. Race, ethnicity, and annual household income were associated with nonresponse to surveys. Women who underwent implant-based procedures were less likely to complete long-term surveys. Conclusions As survey-based research plays an increasingly prominent role in evaluating the outcomes of breast reconstruction, we found socioeconomic and procedure-related differences in survey response rates. Investigators must consider systematic differences in response rates among particular groups of women on the generalizability and validity of findings and perform rigorous nonresponse bias analyses.

To develop models of transport mode choice, mobility cross-sectional survey or panel data can be used. However, the extent to which data from these sources yield accurate parameter values and probabilities is influenced by nonresponse possibly leading to a nonresponse bias. The main research objective of the current study is to assess whether the inclusion of initial nonresponse in a nested logit mode choice model leads to changes in parameter values and more adequate estimated probabilities. The results show that not taking account of nonresponse may lead to a negligibly small overestimation of the choice for car as passenger and bicycle along with an underestimation of the choice for car as driver and e-bike of the same magnitude. Based on the models in this paper, it is not possible to conclude that including the willingness to participate in a mode choice model leads to substantial improvements, but more research is needed to fully assess the value of including willingness.

For many scientific disciplines that rely on surveys and voluntary participation (e.g., organizational behavior, psychology), nonresponse bias (NRB) has been shown to bias estimates, create range restriction, and lead to both Type I and Type II errors. The present research endeavors to fill a methodological gap in the entrepreneurship literature by presenting the following findings: (a) response rates in entrepreneurship research (39%) are higher than firm-level research in management, but lower than individual-level management research, (b) there is virtually no evidence that response rate has any meaningful or consistent influence on relationships in entrepreneurship, and (c) there is little evidence of selective reporting when response rates are low. These promising findings should give entrepreneurship researchers cautious optimism that at least in the aggregate, NRB has limited impact.

Online reviews suffer from self-selection biases. One of these is under-reporting bias: those who have an extreme experience, either positive or negative, are more likely to review a product than those who have a moderate experience, and consequently are overrepresented in the sample of reviews. In this paper we study whether under-reporting bias decreases the effectiveness of the mean star-rating as a criterion for choosing between competing products of different but unobservable qualities. We formulate a model of consumer choice and the decision to review, and simulate this model in a variety of scenarios. We find that if under-reporting bias decreases the effectiveness of the mean star-rating, this is usually only for products that are of either extremely poor or extremely good quality. Even in these cases, the effectiveness is reduced only when the extent of bias is extreme and the variance in quality is low. Otherwise, the presence of under-reporting bias generally does not decrease, and in fact, often enhances the effectiveness of the mean star-rating as a measure of relative quality.

Context Achieving adequate response rates from family members of critically ill patients can be challenging, especially when assessing psychological symptoms. Objectives To identify factors associated with completion of surveys about psychological symptoms among family members of critically ill patients. Methods Using data from a randomized trial of an intervention to improve communication between clinicians and families of critically ill patients, we examined patient-level and family-level predictors of the return of usable surveys at baseline, three months, and six months ( n = 181, 171, and 155, respectively). Family-level predictors included baseline symptoms of psychological distress, decisional independence preference, and attachment style. We hypothesized that family with fewer symptoms of psychological distress, a preference for less decisional independence, and secure attachment style would be more likely to return questionnaires. Results We identified several predictors of the return of usable questionnaires. Better self-assessed family member health status was associated with a higher likelihood and stronger agreement with a support-seeking attachment style with a lower likelihood, of obtaining usable baseline surveys. At three months, family-level predictors of return of usable surveys included having usable baseline surveys, status as the patient's legal next of kin, and stronger agreement with a secure attachment style. The only predictor of receipt of surveys at six months was the presence of usable surveys at three months. Conclusion We identified several predictors of the receipt of surveys assessing psychological symptoms in family of critically ill patients, including family member health status and attachment style. Using these characteristics to inform follow-up mailings and reminders may enhance response rates.

Measurement accuracy of respondents' opinions, perceptions, attitudes and behaviors is vital for better managerial suggestions, theoretical conclusions and advancement of science. There is a growing concern, voiced frequently in other disciplines that a high rate of response bias exists and this casts serious doubts on generalizability of conclusions. Based on limited research on the topic, this paper examines the extent to which response bias is acknowledged as a limitation in tourism, travel and hospitality research. Additional analyses were carried out to explore i) researchers' attentiveness to cautions when citing a biased study and ii) editorial policies of journals against bias. The results revealed that i) acknowledgement of response bias was low, ii) that researchers’ negligence to reported response bias limitations was high, and iii) that editorial policies of the examined journals were far from encouraging authors to discuss and include response bias related limitations. Implications were discussed.

Survey researchers have long hypothesized that social isolation negatively affects the probability of survey participation and biases survey estimates. Previous research, however, has relied on proxy measures of isolation, such as being a marginalized group member within a population. We re-examine the relationship between social isolation and survey participation using direct measures of social isolation derived from social network data; specifically, instrumental research and expressive friendship connections among faculty within academic departments. Using a reconceptualization of social isolation, we find that social network isolation is negatively associated with unit response. Among women (a numerical minority group within the organization), we further find that social group isolation (i.e., lacking instrumental network connections to men, the majority group in the organization) is negatively associated with survey participation. Finally, we show that some survey estimates are systematically biased due to nonparticipation from socially isolated people.

Background Limited study has been done on proxy responses for non-respondents with subjective cognitive decline (SCD) on the Behavioral Risk Factor Surveillance System (BRFSS). Objective To directly compare results for survey respondents with SCD with those for proxies provided for non-respondents with SCD. Methods Publicly available 2011 BRFSS data from 120,485 households in 21 states were analyzed using Stata. Respondents ages 40 and older with SCD (n = 10,831) were compared with proxy responses for non-respondents ages 40 and older with SCD (n = 4296) living in households where the respondent did not have SCD. Outcome measures included functional difficulties associated with their SCD, needing help, receiving informal care, talking with a healthcare provider about their SCD, getting treatment, and having a dementia diagnosis. Logistic regression for each outcome controlled for age, household income, state of residence, and number of household adults. Results Non-respondents were significantly more likely than respondents by Pearson chi square tests with alpha = 0.05 to report all 6 outcomes. Adjusted odds ratios comparing non-respondents with respondents ranged from 2.61 (95% confidence interval: 2.22–3.07) for needing help, to 8.99 (6.60–12.24) for a dementia diagnosis and confirmed unadjusted results. Conclusion Respondent results only represent adults capable of answering a telephone survey. To represent all household adults and avoid nonresponse bias that may under-represent the true population parameters by as much as 70%, results must include both respondents and non-respondents. Other measures may be similarly affected if they inhibit one's ability to respond to a telephone survey (e.g. disability, stroke).

Papers submitted to the journals of academic librarianship sometimes include errors in research design or statistical methods that limit the extent to which the results can be generalized beyond the sample of study participants. This guide discusses how to avoid common problems associated with survey design, sampling, and significance testing (hypothesis testing). The first section, on surveys, covers generalization, random error, investigator bias, response bias, survey design, and alternative methods of gathering data. The second section, on sampling, covers the need for sampling, the importance of defining the population and the cases, sampling strategies, sampling bias (selection bias, volunteer bias, survivorship bias, and nonresponse bias), the need to compare sample and population characteristics, and sample size. The third section, on significance testing, covers the need for significance tests, their inability to account for bias, levels of measurement, types of tests (independent-samples, paired-samples, and one-sample tests), t-tests and nonparametric alternatives, p values and alpha levels, hypothesis testing, and the importance of magnitude or effect size. The paper concludes with guidance on non-statistical methods of generalizing from limited data.

Purpose Patients receiving radiation therapy are subject to high anxiety levels. Radiation Therapists (RTs) are uniquely positioned to support patients as they progress through treatment, but little is known about therapists' ability to perform this important task. This research was designed to describe RTs' ability to communicate with cancer patients during emotional interactions and to determine what intrinsic and extrinsic factors influence that ability. Methods Single-centre focus groups were used to develop themes for a questionnaire. That questionnaire was then distributed throughout 13 Ontario cancer centres. Themes evaluated included intrinsic factors (demographics, value, and so forth) and extrinsic factors (work environment, resources, and so forth). A validated clinical vignette was used to evaluate actual respondent communication skill and a validated scale reported various aspects of emotional intelligence. Results One hundred ninety-nine responses were received from 13 cancer centres distributed across Ontario. Response rate was low (31%), but wave analysis indicated a lack of nonresponse bias. Based on the clinical vignette, RTs showed a high degree of skill when communicating with emotional patients, reporting minimal use of blocking or controlling behaviours. Factors shown to significantly influence the ability to communicate with emotional patients included personal sociability and years of experience, organizational provision of time, private space, and support ( P Conclusions RT respondents were particularly skilled at communicating with emotional patients. That skill level was significantly influenced by both personal and organizational factors. The most important of these were experience and time. These findings are consistent with the literature for other professions and are likely generalizable throughout Canada and similar international practice environments.

Information on the cost of fishing is integral in measuring fisheries performance and correctly evaluating the economic impacts of fisheries regulations and management rules. However, limited availability of cost data often renders such analyses either incomplete or implausible. This study discusses a survey effort that was undertaken in the Northeast US to collect fisheries annual fixed cost data directly from commercial fishermen. Stratified random samples of fishing vessel owners were surveyed in 2012 and 2013. An overview of the survey procedures, sample selection and administration is given here. The data cleaning process, non-response bias and data summaries are also presented. A modeling framework is discussed for estimating fisheries cost data. Robust cost models in combination with regular data collection efforts will allow for more complete economic and financial analyses relevant to fisheries management and conservation.

Background Block scheduling optimizes utilization of valuable operating room (OR) resources. While most studies investigate block scheduling and surgical workloads from a systems perspective, surgeons’ perspectives have yet to be assessed. Such information could provide systems-level solutions to help mitigate prevalent issues such as surgeon burnout. This study surveyed attending surgeons on satisfaction with OR block scheduling and to assess the impact of surgeon characteristics on OR block allocation/satisfaction. Methods This was a cross-sectional investigation of attending surgeons at three academic/tertiary-care hospitals. Sensitivity analyses were completed to evaluate for nonresponse bias. Surgeons’ allocated block times versus their workloads, as well as surgeons’ satisfaction with their block times, were investigated. Associations between survey responses and surgeon characteristics were evaluated using univariable and multivariable analyses. Results Surveys were emailed to 264 surgeons, of whom 159 (60.2%) participated. Nonresponse bias was not found to impact study results. Senior faculty (odds ratio [OR]:2.4; 99.5% confidence interval [CI]:1.1-5.2; p=0.002) and fellowship-trained surgeons (OR:9.5; 99.5% CI:3.7-24.4; p Conclusions Surgical block time allocation and satisfaction at three academic centers varied by surgeons’ gender, position, and training. Furthermore, our results demonstrated that surgical block scheduling can substantially impact surgeons’ personal/career goals and burnout levels. Greater transparency and incorporation of surgeon input into block allocation may improve surgeons’ professional development and patient care.

Objective To evaluate the use of school gardens accompanying nutrition education for 3rd through 5th graders provided by the UF/IFAS Extension Family Nutrition Program (FNP) through process evaluation. Use of Theory or Research Prior research establishes the benefits of school gardens, including improved preference for healthy foods and increased physical activity but are difficult to maintain because of limited time, funding, and engagement. In addition, there is variability in how gardens are utilized. Target Audience Florida 3rd to 5th grade teachers (n = 209) of classrooms that received FNP nutrition education and garden support during the 2018-2019 school year. Program Description FNP provides and helps to maintain over 200 gardens annually through financial and staff support. Gardens are only supported at schools where FNP provides nutrition education. Evaluation Methods An online survey tool was developed through an expert panel consisting of FNP programmatic and evaluation staff and aimed to measure the use and support of gardens. Sixteen questions queried teachers on their classroom's use of the garden, outside support, integration in educational curriculum, and production. Results Teachers (n = 35) representing over 500 students reached by FNP nutrition education completed the survey. Students spent an average of 2,639 minutes in the garden throughout the school year. This ranged from 480 to 5,400 minutes. A majority of teachers report using the garden as a teaching tool outside of FNP-provided nutrition education (88%). All respondents reported that students were involved in at least planting, weeding or harvesting. Once harvested, food is most commonly used for taste tests (91%). Twenty-three percent of respondents reported that schools partially fund gardens. Conclusions There is variability in how SNAP-Ed funded school gardens are utilized and supported. Possible implications include using highly utilized gardens as examples when working with schools and working through barriers such as time and resources. The results also demonstrate that gardens should not be treated as dichotomous variables when used in evaluation. Limitations include a low response rate, possibly caused by time constraints faced by teachers. This could cause a nonresponse bias. In-person surveys delivered by local staff may increase response rates.

There is a disciplinary assumption in our field that surveys with low response rates produce biased estimates, which leads to the use of simple rules for judging the quality of survey data. Surveys with “low” response rates fail this “response rate test” and become difficult to publish. Most of our research methods texts list these rules: e.g., “A response rate below 60% is a disaster, and even a 70% response rate is not much more than minimally acceptable”. Editors embrace this view, and often reject out of hand any study failing to reach this conventional standard. We argue that our field’s use of response rate rules in evaluating scholarship is based more on disciplinary custom than on survey science. In this paper, we describe the long-term downward trend in response rates and address confusion about nonresponse bias and its relation to response rates. Using Groves and Peytcheva’s (2008) meta-analytic data, we present evidence about the magnitude of the estimate- and study-level relationships between response rates and two different measures of nonresponse bias in univariate estimates. We then discuss several consequences of using the “response rate test” to judge data quality.

While elite polarization in US politics has been well established, there has been substantial debate about whether US mass opinion has polarized in recent decades. Cavari and Freedman (2018) suggest that the apparent polarisation of the US public may in part be an artefact of declining survey response rates. Cavari and Freedman argue that declining response rates may have led to only the most politicised — and therefore polarized — respondents being willing to give their opinions. We argue that while the mechanism posited is highly plausible, the results of Cavari and Freedman’s analysis are driven by spurious time trend correlations. After accounting for these there is no evidence for a response rate effect on polarization.

Earnings nonresponse in household surveys is widespread, yet there is limited knowledge of how nonresponse biases earnings measures. We examine the consequences of nonresponse on earnings gaps and inequality using Current Population Survey individual records linked to administrative earnings data. The common assumption that earnings are missing at random is rejected. Nonresponse across the earnings distribution is U-shaped, highest in the left and right tails. Inequality measures differ between household and administrative data due in part to nonresponse. Nonresponse biases earnings differentials by race, gender, and education, particularly in the tails. Flexible copula-based models can account for nonrandom nonresponse.

Pediatricians report many barriers to caring for children with mental health (MH) problems. The American Academy of Pediatrics (AAP) has focused attention on MH problems, but the impact on perceived barriers is unknown. We examined whether perceived barriers and their correlates changed from 2004 to 2013.In 2004, 832 (52%) of 1600 and in 2013, 594 (36.7%) of 1617 of randomly selected AAP members surveyed responded to periodic surveys, answering questions about sociodemographics, practice characteristics, and 7 barriers to identifying, treating/managing, and referring child/adolescent MH problems. To reduce nonresponse bias, weighted descriptive and logistic regression analyses were conducted.Lack of training in treatment of child MH problems (∼66%) and lack of confidence treating children with counseling (∼60%) did not differ across surveys. Five barriers (lack of training in identifying MH problems, lack of confidence diagnosing, lack of confidence treating with medications, inadequate reimbursement, and lack of time) were less frequently endorsed in 2013 (all P .01), although lack of time was still endorsed by 70% in 2013. In 2004, 34% of pediatricians endorsed 6 or 7 barriers compared to 26% in 2013 (P .005). Practicing general pediatrics exclusively was associated with endorsing 6 or 7 barriers in both years (P .001).Although fewer barriers were endorsed in 2013, most pediatricians believe that they have inadequate training in treating child MH problems, a lack of confidence to counsel children, and limited time for these problems. These findings suggest significant barriers still exist, highlighting the need for improved developmental and behavioral pediatrics training.

In two important studies, Tourangeau (2017) and Brick and Tourangeau (2017) reanalyzed Groves and Peytcheva’s (2008) meta-analytic data on nonresponse bias. They found that: 1) 23% of the variation in nonresponse bias (absolute relbias) is between studies, and 2) there is a sizable and robust study-level correlation between bias and response rates, which is much larger than the estimate-level correlation (more than twice as large with analytic weights applied). Initially, however, these authors examined two different measures of nonresponse bias, absolute differences and absolute relbias, because of concerns the absolute relbias measure was misleading. Unfortunately, only the absolute relbias results were included in the two published articles. In the current study, I replicate and extend their published results, fixing an error in their analysis. I also provide the omitted results for absolute differences. The published and omitted results lead to dissimilar conclusions. In the omitted findings for absolute differences, much more of the variation (40-46%) in nonresponse bias is at the study level, but the study-level correlation between bias and response rates tends to be much weaker, weaker even than the estimate-level correlation, is sometimes positive, is rarely statistically significant, and is highly sensitive to model specification. The dissimilarities in the results for absolute differences versus absolute relbias raise important new questions for future research.

This study evaluates the magnitude of biases in the measurement of income inequality due to the tendency of households at the top ends of national income distributions not to complete surveys. This unit nonresponse bias is evaluated and corrected for through a recently proposed probabilistic model utilizing information on regional nonresponse rates and income distributions. To this end, regional nonresponse rates were collected from individual national statistical agencies and are reported in this study. The study also uses 72 microdatasets from 34 high and middle-income countries across the world harmonized and made available by Luxembourg Income Study. The large number of surveys is used to infer consequences of unit nonresponse for the measurement of inequality in the industrialized and middle-income world and in major world regions. Analysis performed on survey data disaggregated geographically into regions for which nonresponse rates are available yields sizeable, systematic corrections for nonresponse. Nonresponse bias is estimated to affect national Gini coefficients typically by 1–8 percentage points, but sometimes by 15 percentage points or more, in line with previous estimates. Winsorizing performed by Luxembourg Income Study in producing key figures affects negatively not only Gini estimates but also the estimated corrections for nonresponse. This suggests that specific measurement issues such as the nonresponse bias evaluated here are unlikely to be identified and corrected fully in winsorized income distributions. The degree of geographic disaggregation appears to affects the nonresponse correction negatively, corroborating existing evidence. This phenomenon has not been explained satisfactorily so far and calls for further inquiry.34 surveys, encompassing 576,700 household records in 418 regions, are used to derive back-of-the-envelope estimates of the Ginis for world regions and worldwide. Gini coefficient for North America – of 44.72 after correction of unit nonresponse – appears quite robust to nonresponse, being merely 0.34 percentage points above the uncorrected statistic. Ginis for Europe and Latin America (excluding low income countries), however, are quite sensitive to the correction, rising from 32.32 to 37.72 (5.40 pc.pt.) in the European Single Market, from 35.38 to 42.92 (7.55 pc.pt.) in Europe including former Soviet-Union and Balkan nations, and from 49.71 to 61.12 (11.41 pc.pt.) in Latin America. Correction for unit nonresponse also matters for estimating the worldwide Gini. The corrected estimates, 57.6–59.3, are 3–4.7 percentage points above the uncorrected value, 54.6. Overall, these numbers from across countries, world regions and worldwide appear quite consistent with each other and suggest that unit nonresponse biases Gini coefficients downward by 1–8 percentage points.

Background Intensive longitudinal methods are used to study the daily-level, within-person associations between substance use and its predictors and consequences. By definition, these designs require greater respondent effort than more traditional longitudinal designs, the result of which may be increased unit nonresponse and, more importantly, nonresponse bias. The present study contrasts the nonresponse properties of a measurement burst design with those of a single 1-year follow-up (“control”) design, with a particular emphasis on the retention of young adults who did not plan to graduate from college. Methods High school seniors (N = 318) from three Midwestern schools completed an in-school baseline survey in spring 2012. Respondents were then randomized into a measurement burst or control group. Four, eight, and twelve months after baseline, young adults in the measurement burst group received a 30-min follow-up web survey, followed by 14 days of web-based daily surveys. Young adults in the control group received only a 30-min follow-up web survey 12 months after baseline. Response rates, predictors of nonresponse, and relative nonresponse biases for measures of sociodemographics, college plans, and substance use were compared between the two measurement groups. Results Compared to the control design, the measurement burst design had greater relative nonresponse bias for statistics measuring substance use, but lower relative nonresponse bias for the statistic measuring college plans. Conclusion Intensive longitudinal methods have the potential to retain non-college attending young adults during the transition to adulthood. Nonresponse adjustment weights should be used to correct for any detected bias.

For as long as colleges and universities have been conducting student evaluations of teaching (SET), faculty have questioned the validity of the information collected. Substantial effort has been expended to improve the SET process, and researchers have communicated a consistent understanding of why students evaluate teachers as they do. Most of these conclusions have been based on an analysis of SET data gathered at the end of the semester by a sample of students who may not represent all students enrolled in the class. This clearly creates the potential for a sample selection bias that puts into question much of what we have learned about why students evaluate their instructors as they do. In this paper, we gather a unique data set from individual students who provide either online SET forms or in-class SET forms. We use these responses to address the concerns regarding the validity of SET data which results from the low response rates often associated with online SET evaluations. The Heckman (1979) sample selection bias correction estimator is used to quantify the bias resulting from the low online response rate and finds this bias to be negligibly small.

Objectives Understanding attitudes to mental health issues can inform public health interventions. However, low response rates may contribute to nonresponse bias. In a randomized controlled trial we examined the effect of sending a prenotification postcard before the questionnaire and the placement of a short message on the survey envelope (teaser) on response rates to a mailed questionnaire about bulimia nervosa “mental health literacy”. Study Design and Setting Questionnaires were mailed to 3,010 adults (50.6% female and 49.4% male) aged 18–65 years. In a 2 (pre-notification–present; absent) by 2 (teaser–present; absent) design, questionnaire recipients were randomly allocated to the experimental strategies. Outcomes considered were response rate, response time, and cost. Results The overall response rate was 22.0%. Significant main effects showed higher response rates for the use of prenotification (present = 23.6%; absent = 20.3%), among female participants, and older participants. A significant interaction of teaser by gender indicated lower response rates for men who received the teaser but not for women. Older participants returned the questionnaire more promptly than younger participants. Females—but not males—who received the teaser were slower to return the questionnaire. Higher response rates for participants receiving the postcard compensated for increased costs, particularly for males and older participants. Conclusion Response rates to a mental health postal survey can be increased through the use of prenotification.

Nonresponse is a persistent problem in surveys because results from respondents only are subject to nonresponse bias. Many methods have been developed to deal with ignorable (missing at random) nonresponse data. In this paper, we provide a method to assess and adjust nonignorable (not missing at random) nonresponse bias in a small area estimation problem. We propose a bivariate Bayesian hierarchical linear mixed model to estimate both satisfaction rate and response rate. This model uses spatial dependencies among subdomains and auxiliary information from sample units to assess and adjust nonresponse bias. In addition, it explicitly includes a parameter that indicates whether the nonresponse is ignorable or not. The method is used to analyze the 2001 Missouri Deer Hunter Attitude Survey (MDHAS). The result shows that the nonresponse in MDHAS is nonignorable. Hunter age and the number of deer harvested have strong effects on satisfaction and response rates, and spatial dependencies are strong amongst counties of hunters’ residences. The estimated satisfaction rates are lower after adjusting for nonresponse bias.

Background and objectives Collecting Patient-Reported Outcomes (PROs) is an important way to get first-hand information by patients on the outcome of treatments and surgical procedure they have undergone, and hence about the quality of the care provided. However, the quality of PRO data cannot be given for granted and cannot be traced back to the dimensions of timeliness and completeness only. While the reliability of these data can be guaranteed by adopting standard and validated questionnaires that are used across different health care facilities all over the world, these facilities must take responsibility to assess, monitor and ensure the validity of PROs that are collected from their patients. Validity is affected by biases that are hidden in the data collected. This contribution is then aimed at measuring bias in PRO data, for the impact that these data can have on clinical research and post-marketing surveillance. Methods We considered the main biases that can affect PRO validity: Response bias, in terms of Acquiescence bias and Fatigue bias; and Non-Response bias. To assess Acquiescence bias, phone interviews and online surveys were compared, adjusted by age. To assess Fatigue bias, we proposed a specific item about session length and compared PROs scores stratifying according to the responses to this item. We also calculated the intra-patient agreement by conceiving an intra-interview test-retest. To assess Non-Response bias, we considered patients who participated after the saturation of the response-rate curve as proxy of potential non respondents and compared the outcomes in these two strata. All methods encompassed common statistical techniques and are cost-effective at any facility collecting PRO data. Results Acquiescence bias resulted in significantly different scores between patients reached by either phone or email. In regard to Fatigue bias, stratification by perceived fatigue resulted in contrasting results. A relevant difference was found in intra-patient agreement and an increasing difference in average scores as a function of interview length (or completion time). In regard to Non-Response bias, we found non-significant differences both in scores and variance. Conclusions In this paper, we present a set of cost-effective techniques to assess the validity of retrospective PROs data and share some lessons learnt from their application at a large teaching hospital specialized in musculoskeletal disorders that collects PRO data in the follow-up phase of surgery performed therein. The main finding suggests that response bias can affect the PRO validity. Further research on the effectiveness of simple and cost-effective solutions is necessary to mitigate these biases and improve the validity of PRO data.

Objectives: To evaluate the presence of productivity goals among licensed rehabilitation clinicians and their relationship with observed unethical behavior. Design: Exploratory, cross-sectional survey. Setting: Online. Participants: Licensed physical therapy clinicians (N=3446). Intervention: Not applicable. Main Outcome Measure: Participants completed an electronic survey regarding use of clinical productivity goals. They rated the frequency in which they observed 6 unethical behaviors on a 7-point Likert scale in their practice setting from 1=never to 7=always. An overall observed unethical behavior score was calculated by summing these scales. Results: The response rate was 12.8% (N=3446), with analyses showing low risk of nonresponse bias. Many respondents (73.9%) had a formal productivity goal. Most (89.4%) reported observing some form of unethical behavior, but many (68.6%) reported it occurred “rarely” or “never.” Those in skilled nursing facility (SNF) settings reported higher frequencies of observance and were 4.1 times more likely to report more unethical behavior than the median compared with all other settings. A positive correlation existed between expected productivity rate and rate of unethical behaviors observed (ρ=0.225; P

Background Surveys are a common method of data collection within health service research. An essential aspect of reporting survey research is ensuring that sufficient information is provided to enable readers to determine the validity and representativeness of research findings. Method This study reports a secondary analysis of survey research published in Radiography and Clinical Radiology between 2001 and 2010. The purpose of the study was to evaluate trends in response rates and establish how non-response bias was being addressed. Results Analysis of non-response bias was undertaken in 9.4% (n = 9/96) of studies. Where analysis was performed, strong reliance on demographic characteristics to determine sample representativeness was noted (n = 8/9; 88.9%). Conclusion The findings of this study suggest that non-response bias is not being adequately addressed within published imaging related survey research and more needs to be done to encourage a rigorous approach to the analysis and reporting of survey results.

This paper examines antecedent Latin American citizens' attitudes toward globalization, taking into account the effect of the extent of globalization at the national level. The model incorporates the effect of nonresponse and accounts for the hierarchical nature of individual opinions on globalization. Multilevel mixed-effects model estimations using both macro and micro level data show that there is a high potential for nonresponse bias in globalization studies, specifically in countries where digital access is still limited. The degree of acceptance of globalization among Latin American citizens is heterogeneous with important determinants at the country level being the degree of cultural globalization and its interaction with variables at the individual level. Young male individuals, with high levels of education, income and access to internet and international TV channels are likely to favor globalization in Latin America.

Figures from postelection surveys often grossly overestimate election turnout. Two distinct phenomena are responsible for this gap: overrepresentation of actual voters and vote misreporting by actual nonvoters among survey respondents. Previous accounts of turnout bias are inconclusive in that they either focus on a single component, or fail to separate between the two. In this paper, we formally decompose turnout bias in election surveys into its constituent parts, assess their empirical prevalence and heterogeneity using an extensive collection of 49 vote validation studies from six countries, and employ Bayesian meta regression techniques to account for cross-study differences. Our results indicate that both election and survey characteristics such as actual voter turnout and survey response rates differentially affect the components of turnout bias. We conclude with a discussion of the threats and potentials of our findings for survey-based comparative electoral research.

How high is unemployment? How low is labor force participation? Is obesity more prevalent among men? How large are household expenditures? We study the sources of the relevant official statistics --- the Current Population Survey (CPS), the Behavioral Risk Factor Surveillance System (BRFSS), and the Consumer Expenditure Survey (CEX) --- and find that the answers depend on whether we look at easy- or at difficult-to-reach respondents, measured by the number of call and visit attempts made by interviewers. A challenge to the (conditionally-) random-nonresponse assumption, these findings empirically substantiate the theoretical warning against making population-wide estimates from surveys with low response rates.

This article reports on two different methods applied in the same survey (N = 1881) to measure the impact of the carsharing system car2go on other transportation modes in Ulm, Germany. The first method calculated how the mobility behavior of respondents would hypothetically be at the present time if car2go was not available. The second method determined the respondents’ past mobility behavior before using car2go. Confounding circumstances were corrected in both approaches through different mechanisms. Comparable methods calculating carsharing impacts have only been applied individually in past studies. This is the first study applying two measurement methods within the same survey, which enables a triangulation. As other influencing parameters were equal (e.g. sampling frame, nonresponse bias, mode of asking, point in time of the survey), the deviating results are assumed to have resulted from the different measurement techniques. The findings indicate a primacy effect (disproportionally high selection of first answer options) having influenced the first measurement and an overestimation of the impact on total kilometers travelled in the second measurement. The comparative findings of this dual-measurement could contribute to research designs of greater precision in future work on carsharing impacts.

Objectives To display and discuss the reasons and consequences of length and time-dependent bias. They might occur in presence of a time-dependent study entry or a time-dependent exposure which might change from unexposed to exposed. Study Design and Setting Recalling the popular study of Oscar nominees and using a real-data example from hospital epidemiology, we give innovative and easy-to-understand graphical presentations of how these biases corrupt the analyses via distorted time-at-risk. Cumulative hazard plots and Cox proportional hazards models were used. We are building bridges to medical disciplines such as critical care medicine, hepatology, pharmaco-epidemiology, transplantation medicine, neurology, gynecology and cardiology. Results In presence of time-dependent bias, the hazard ratio (comparing exposed with unexposed) is artificially underestimated. The length bias leads to an artificial underestimation of the overall hazard. When both biases coexist it can lead to different directions of biased hazard ratios. Conclusion Since length and time-dependent bias might occur in several medical disciplines, we conclude that understanding and awareness are the best prevention of survival bias.

Background The continuing decline in response rates to household surveys is a concern for the health and social sciences as it increases the risk of selective non-response biasing the estimates of interest. Aim To assess non-response bias in a postal survey measuring drinking behaviour, experience of harm and opinion on local government alcohol policies among residents in six New Zealand communities. Methods The Continuum of Resistance model, which suggests that late respondents to a survey are most similar to non-respondents on the measures of interest, was used to guide our investigation. Results Men, younger people, those of Māori descent and those living in more deprived areas were less likely to respond to our survey than women, older people, those not of Māori descent and those living in comparatively affluent areas. Late respondents more closely resembled non-respondents demographically than early respondents. The prevalence of binge drinking and experience of assault was higher, and support for restrictive local government alcohol policies lower, among late respondents. Assuming the drinking behaviour and alcohol-related experiences of non-respondents were the same as those of late respondents, prevalence was under-estimated by 3.4% (relative difference: 13%) and 2.1% (relative difference: 21%) for monthly binge drinking and assault respectively. There was little evidence of bias in estimates of policy support. Contribution to the Field The findings add to a growing body of evidence suggesting that surveys under-estimate risk behaviour because of selective non-response. Notably, public opinion may not be subject to such misestimation. This is an abstract of a presentation at Safety 2012, the 11th World Conference on Injury Prevention and Safety Promotion, 1-4 October 2012, Michael Fowler Center, Wellington, New Zealand. Full text does not seem to be available for this abstract.

This study evaluates the effectiveness of Motivational Enhancement Therapy/Cognitive Behavioral Therapy-5 (MET/CBT-5) when delivered in community practice settings relative to standard community-based adolescent treatment. A quasi-experimental strategy was used to adjust for pre-treatment differences between the MET/CBT-5 sample (n = 2293) and those who received standard care (n = 458). Results suggest that youth who received MET/CBT-5 fared better than comparable youth in the control group on five out of six 12-month outcomes. A low follow-up rate (54%) in the MET/CBT-5 sample raised concerns about nonresponse bias in the treatment effect estimates. Sensitivity analyses suggest that while modest differences in outcomes between the MET/CBT-5 nonrespondents and respondents would yield no significant differences between the two groups on two of the six outcomes, very large differences in outcomes between responders and nonresponders would be required for youth receiving MET/CBT-5 to have fared better had they received standard outpatient care.

Statistical accident data plays an important role for traffic safety development involving the road system, vehicle design, and driver education. Vehicle manufacturers use data from accident mail surveys as an integral part of the product development process. Low response rates has, however, lead to concerns on whether estimates from a mail survey can be trusted as a source for making strategic decisions. The main objective of this paper was to investigate nonresponse bias in a mail survey addressing driver behaviour in accident situations. Insurance data, available for both respondents and nonrespondents were used to analyze, as well as adjust for nonresponse. Response propensity was investigated by using descriptive statistics and logistic regression analyses. The survey data was then weighted by using inverse propensity weights. Two specific examples of survey estimates are addressed, namely driver vigilance and driver's distraction just before the accident. The results from this paper reveal that driver age and accident type were the most influential variables for nonresponse weighting. Driver gender and size of town where the driver resides also had some influence, but not for all survey variables investigated. The main conclusion of this paper is that nonresponse weighting can increase confidence in accident data collected by a mail survey, especially when response rates are low. Weighting has a moderate influence on this survey, but a larger influence may be expected if applied on a more diverse driver population. The development of auxiliary data collection can further improve accident mail survey methodology in future.

Low participation is a potential source of bias in population-based studies. This article presents use of inverse probability weighting (IPW) in adjusting for non-participation in estimation of brain volumes among subjects with schizophrenia. Altogether 101 schizophrenia subjects and 187 non-psychotic comparison subjects belonging to the Northern Finland 1966 Birth Cohort were invited to participate in a field study during 1999–2001. Volumes of grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) were compared between the 54 participating schizophrenia subjects and 100 comparison subjects. IPW by illness-related auxiliary variables did not affect the estimated GM and WM mean volumes, but increased the estimated CSF mean volume in schizophrenia subjects. When adjusted for intracranial volume and family history of psychosis, IPW led to smaller estimated GM and WM mean volumes. Especially IPW by a disability pension and a higher amount of hospitalisation due to psychosis had effect on estimated mean brain volumes. The IPW method can be used to improve estimates affected by non-participation by reflecting the true differences in the target population.

Objectives To examine nonparticipation in a survey by linking it with register information and identify potential nonresponse bias of inequalities in health status among aging people. Study Design and Setting Cross-sectional questionnaire survey with clinical checkups carried out in 2002 among persons born in 1926–1930, 1936–1940, and 1946–1950 in Southern Finland. The sample was linked with register information from Statistics Finland and analyzed in terms of participation and health status as measured by medicine reimbursements. Results Participation in the survey was more frequent among those who were older, female, married or cohabiting, higher educated and nonurban residents, and those with higher income and moderate health. Among nonrespondents, women were less healthy than men, whereas among respondents, the results were reversed. Among nonrespondents, better income was associated with unfavorable health. Poor health was generally more common among nonrespondents than respondents in several subgroups. Conclusion Differences in response rates were found in sociodemographic factors, health, and socioeconomic position. Favorable health was generally more frequent among respondents than nonrespondents. In particular, health inequalities by gender and income differed between respondents and nonrespondents. Thus, nonresponse may lead to bias in analyses of health inequalities among aging people.

International research shows that both work–family conflict and time pressure are increasing in Western societies. With these increased pressures, precisely those people who are combining work and family obligations may be missing from survey estimates. This study investigates whether the measurement of work–family conflict (WFC) and time pressure is influenced by nonresponse bias. Using the Flemish “Work, Family and Time use in Flanders”-survey and the Basic Question survey for nonrespondents of this survey, we investigate whether nonrespondents score higher on subjective and objective indicators of WFC and time pressure than respondents. Results show that nonrespondents are indeed those sample units who experience significantly more WFC and time pressure, demonstrating that WFC and time pressure related nonresponse biases exist. Implications and suggestions to reduce bias in data collections are discussed.

Objective There has been little investigation of non-response bias in web-based health surveys. We hypothesised that non-respondents have a higher prevalence of risk behaviours than respondents. Method In 2005, random samples of students aged 17–25 years from 12 New Zealand tertiary institutions (n = 7130) were invited to complete a web-based health behaviour survey, with three e-mail reminders. Early respondents (before 2nd reminder) were compared with late respondents (after 2nd reminder). Late respondents served as a proxy for non-respondents. Results 2607 students (37%) responded early, 676 (9%) responded late, and 3847 (54%) did not respond. There were differences between early and late respondents in high school binge drinking (38% vs 47%, p = 0.002) and non-compliance with physical activity guidelines (12% vs 18%, p = 0.004). Differences in overweight/obesity (26% vs 31%, p = 0.058), smoking (18% vs 22%, p = 0.091) and non-compliance with dietary guidelines (76% vs 77%, p = 0.651) were non-significant but point estimates were in the expected direction. Estimated bias in prevalence of risk behaviours was an absolute difference of 1–4% and a relative difference of 0–21%. Conclusion Respondents whose participation was hardest to elicit reported more risk behaviour. Assuming non-respondents' behaviour is similar or more extreme than that of late respondents, prevalence will have been substantially underestimated.

Rammstedt, Goldberg, and Borg (2010) demonstrated that the Big Five factor structure is sensitive to educational bias and that this bias seems to be due to a more pronounced tendency for acquiescent responding among less well-educated respondents. The present follow-up study investigated whether using an interview assessment mode in which the item presentation is standardized and thus differences in reading ability are controlled for, reduces the bias in the factorial structures. Results based on a large and heterogeneous sample clearly contradict this assumption: The findings unambiguously replicate those reported by Rammstedt et al. in that lower educated respondents have a higher and more variable tendency for acquiescent responding which directly affects the factor structure of the Big Five measure.

Objective Using an appropriate method to handle cases with missing data when performing secondary analyses of survey data is important to reduce bias and to reach valid conclusions for the target population. Many published secondary analyses using child health data sets do not discuss the technique employed to treat missing data or simply delete cases with missing data. Missing data may threaten statistical power by reducing sample size or, in more extreme situations, estimates derived by deleting cases with missing values may be biased, particularly if the cases with missing values are systematically different from those with complete data. The aim of this study was to determine which of 4 techniques for handling missing data most closely estimates the true model coefficient when varying proportions of cases are missing data. Methods We performed a simulation study to compare model coefficients when all cases had complete data and when 4 techniques for handling missing data were employed with 10%, 20%, 30%, or 40% of the cases missing data. Results When >10% of the cases had missing data, the reweight and multiple imputation techniques were superior to dropping cases with missing scores or hot deck imputation. Conclusions These findings suggest that child health researchers should use caution when analyzing survey data if a large percentage of cases have missing values. In most situations, the technique of dropping cases with missing data should be discouraged. Investigators should consider reweighting or multiple imputation if a large percentage of cases are missing data.

Exit polls are best known for their use in election forecasting. In recent years, however, some prominent mistaken predictions have been made, undermining public confidence in the accuracy of both exit polls and survey methods. Nonresponse bias has been claimed as being one of the main reasons for inaccurate projections. Traditionally, the issue has been handled through an age–race–sex adjustment at the national and state levels. An alternative solution is suggested and detailed in this paper. A two-step strategy is proposed to reduce nonresponse bias and improve predictions. First, “vote-remembering” (vote recall) is used to correct party proportion estimates at polling locations; second, this is used to estimate party proportions at precinct level through a regression estimator. The method is gauged by forecasting the 2003 and 2007 Corts Valencianes elections using raw data from the exit polls conducted by SigmaDos for Generalitat Valenciana. In light of the results, this procedure considerably improves raw data projections and shows a substantial improvement over industry (SigmaDos) forecasts. It therefore represents an interesting alternative that could easily be adopted for exit polling in any country where precinct-level voting data exist.

Background: After-death surveys are an important source of information about the quality of end-of-life care, but response rates generally are low. Our goal was to understand the potential for nonresponse bias in survey studies of family members after a patient’s death in the hospital ICU by identifying differences in patient demographics and delivery of palliative care between patients whose families respond to a survey about end-of-life care and those whose families do not. Methods: We performed a cohort study of patients who died in the ICU at 14 hospitals. Surveys were mailed to family members 1 to 2 months after the patient’s death. Chart abstraction was completed on all patients, assessing demographic characteristics and previously validated indicators of palliative care. Results: Of the 2,016 surveys sent to families, 760 were returned, for a response rate of 38%. Patients whose family members returned the surveys were more likely to be white (88% vs 74%, respectively; p < 0.001); to be older (71 years vs 69 years, respectively; p 0.015); and to have received more indicators of palliative care, including medical record documentation of family present at death, involvement of spiritual care, and dying after a decision to limit life-sustaining therapies (p < 0.05). Conclusions: Patients whose family members responded to a survey about end-of-life care were more likely to be white, older, and have indicators of palliative care documented in the medical record. Because these patients likely received higher quality palliative care, these findings suggest that the response bias in end-of-life care research is toward overestimating the quality of palliative care. Trial registration: ClinicalTrials.gov Identifier: NCT00685893 (CHEST 2009; 136:1496–1502) Abbreviation: DNR do not resuscitate

We link information on graduates from many cohorts to their high-school and college records and demographics to infer the impact of college major on earnings. We develop an estimator to handle potential non-response bias and identify non-response using an affinity measure—the potential respondent’s link to the survey organization. This technique is generally applicable for adjusting for unit non-response. In the earnings model estimated using the identified (for non-response bias) selectivity adjustments, adjusted earnings differentials across college majors are below half as large as unadjusted differentials and ten percent smaller than those that do not account for selective non-response.