Reweighting UK Biobank corrects for pervasive selection bias due to volunteering

BACKGROUND

Biobanks typically rely on volunteer-based sampling. This results in large samples (power) at the cost of representativeness (bias). The problem of volunteer bias is debated. Here, we (i) show that volunteering biases associations in UK Biobank (UKB) and (ii) estimate inverse probability (IP) weights that correct for volunteer bias in UKB.

METHODS

Drawing on UK Census data, we constructed a subsample representative of UKB's target population, which consists of all individuals invited to participate. Based on demographic variables shared between the UK Census and UKB, we estimated IP weights (IPWs) for each UKB participant. We compared 21 weighted and unweighted bivariate associations between these demographic variables to assess volunteer bias.

RESULTS

Volunteer bias in all associations, as naively estimated in UKB, was substantial-in some cases so severe that unweighted estimates had the opposite sign of the association in the target population. For example, older individuals in UKB reported being in better health, in contrast to evidence from the UK Census. Using IPWs in weighted regressions reduced 87% of volunteer bias on average. Volunteer-based sampling reduced the effective sample size of UKB substantially, to 32% of its original size.

CONCLUSIONS

Estimates from large-scale biobanks may be misleading due to volunteer bias. We recommend IP weighting to correct for such bias. To aid in the construction of the next generation of biobanks, we provide suggestions on how to best ensure representativeness in a volunteer-based design. For UKB, IPWs have been made available.

Large Questionnaire Survey on Sleep Duration and Insomnia Using the TV Hybridcast System by Japan Broadcasting Corporation (NHK)

Background: Japanese people are known to have the shortest sleep duration in the world. To date, no study has assessed a large Japanese population for insomnia and sleep duration. Methods: We performed an Ιnternet-based survey in association with a national television (TV) program. Questionnaire data were collected not only through personal computers, tablets, and smartphones, but also through the Hybridcast system, which combines broadcasts over airwaves with broadband data provided via the Internet using the TV remote controller. The Athens Insomnia Scale (AIS) was used to assess insomnia. Results: A total of 301,241 subjects participated in the survey. Participants slept for an average of 5.96 ± 1.13 h; the average AIS score was 6.82 ± 3.69. A total of 26.1% of male and 27.1% of female participants had both insomnia (AIS ≥ 6) and short sleep duration (<6 h). Responses were recorded through the Hybridcast system for 76.4% of the elderly (age ≥ 65 years) subjects and through personal computers, tablets, or smartphones for 59.9–82.7% of the younger subjects (age ≤ 65 years). Conclusions: Almost a quarter of the Japanese participants presented short sleep duration and insomnia. Furthermore, the Hybridcast system may be useful for performing large internet-based surveys, especially for elderly individuals.

Multilevel Regression and Poststratification Versus Survey Sample Weighting for Estimating Population Quantities in Large Population Health Studies

Multilevel regression and poststratification (MRP) is a model-based approach for estimating a population parameter of interest, generally from large-scale surveys. It has been shown to be effective in highly selected samples, which is particularly relevant to investigators of large-scale population health and epidemiologic surveys facing increasing difficulties in recruiting representative samples of participants. We aimed to further examine the accuracy and precision of MRP in a context where census data provided reasonable proxies for true population quantities of interest. We considered 2 outcomes from the baseline wave of the Ten to Men study (Australia, 2013-2014) and obtained relevant population data from the 2011 Australian Census. MRP was found to achieve generally superior performance relative to conventional survey weighting methods for the population as a whole and for population subsets of varying sizes. MRP resulted in less variability among estimates across population subsets relative to sample weighting, and there was some evidence of small gains in precision when using MRP, particularly for smaller population subsets. These findings offer further support for MRP as a promising analytical approach for addressing participation bias in the estimation of population descriptive quantities from large-scale health surveys and cohort studies.

Drug utilization in the Maastricht Study: A comparison with nationwide data

Within the southern region of the Netherlands, the Maastricht Study is an on-going observational prospective population-based cohort study that focuses on the etiology of Type 2 diabetes mellitus (T2DM). Representativeness of the participating population is a crucial but often an unknown factor in population-based cohort studies such as the Maastricht Study. We therefore aimed to assess the representativeness of the study population by comparing drug utilization of the participants of the Maastricht Study with the general population of the Netherlands.Since T2DM patients were oversampled in this study, a sampling method was applied in order to ensure a similar distribution of T2DM over the study population. Drug use in the study population was compared with drug use in the population of the Netherlands, using a Z-test to compare 2 independent proportions.In general, drug use in the study was similar compared with national data. However, in the age group 65 to 74 years total drug use was lower in the study population (833/1000 persons) versus nationwide data (882/1000 persons). The use of pulmonary medications was lower (104/1000 persons vs 141/1000 persons) and the use of hypnotics/anxiolytics was higher (90/1000 persons vs 36/1000 persons) in the Maastricht Study as compared with national data.Drug use in the Maastricht Study population is largely comparable to that in the total Dutch population aged 45 to 74. Therefore, data on drug use by participants in the Maastricht Study can be used to perform studies assessing outcomes associated with drug use.

Assessing the Role of Selection Bias in the Protective Relationship Between Caregiving and Mortality

Caregivers have lower mortality rates than noncaregivers in population-based studies, which contradicts the caregiver-stress model and raises speculation about selection bias influencing these findings. We examined possible selection bias due to 1) sampling decisions and 2) selective participation among women (baseline mean age = 79 years) in the Caregiver-Study of Osteoporotic Fractures (Caregiver-SOF) (1999-2009), an ancillary study to the Study of Osteoporotic Fractures (SOF). Caregiver-SOF includes 1,069 SOF participants (35% caregivers) from 4 US geographical areas (Baltimore, Maryland; Minneapolis, Minnesota; the Monongahela Valley, Pennsylvania; and Portland, Oregon). Participants were identified by screening all SOF participants for caregiver status (1997-1999; n = 4,036; 23% caregivers) and rescreening a subset of caregivers and noncaregivers matched on sociodemographic factors 1-2 years later. Adjusted hazard ratios related caregiving to 10-year mortality in all women initially screened, subsamples representing key points in constructing Caregiver-SOF, and Caregiver-SOF. Caregivers had better functioning than noncaregivers at each screening. The association between caregiving and mortality among women invited to participate in Caregiver-SOF (41% died; adjusted hazard ratio (aHR) = 0.73, 95% confidence interval (CI): 0.61, 0.88) was slightly more protective than that in all initially screened women (37% died; aHR = 0.83, 95% CI: 0.73, 0.95), indicating little evidence of selection bias due to sampling decisions, and was similar to that in Caregiver-SOF (39% died; aHR = 0.71, 95% CI: 0.57, 0.89), indicating no participation bias. These results add to a body of evidence that informal caregiving may impart health benefits.

Multilevel regression and poststratification as a modeling approach for estimating population quantities in large population health studies: A simulation study

There are now a growing number of applications of multilevel regression and poststratification (MRP) in population health and epidemiological studies. MRP uses multilevel regression to model individual survey responses as a function of demographic and geographic covariates. Estimated mean outcome values for each demographic-geographic respondent subtype are then weighted by the proportions of each subtype in the population to produce an overall population-level estimate. We recently reported an extensive case study of a large nationwide survey and found that MRP performed favorably compared to conventional survey sampling weights for the estimation of population descriptive quantities in a highly selected sample. In this study, we aimed to evaluate, by way of a simulation experiment, both the accuracy and precision of MRP versus survey sampling weights in the context of large population health studies. While much of the research into MRP has been focused on U.S. political and social science, we considered an alternative population structure of smaller size and with notably fewer geographic subsets. We explored the impact on MRP performance of sample size, model misspecification, interactions, and the addition of a geographic-level covariate. MRP was found to achieve generally superior performance in both accuracy and precision at both the national and state levels. Results were generally robust to model misspecification, and MRP performance was further improved by the inclusion of a geographic-level covariate. These findings offer further evidence that MRP provides a promising analytic approach for addressing participation bias in the estimation of population descriptive quantities from large-scale health surveys and cohort studies.

Impact of different privacy conditions and incentives on survey response rate, participant representativeness, and disclosure of sensitive information: a randomized controlled trial

BACKGROUND

Anonymous survey methods appear to promote greater disclosure of sensitive or stigmatizing information compared to non-anonymous methods. Higher disclosure rates have traditionally been interpreted as being more accurate than lower rates. We examined the impact of 3 increasingly private mailed survey conditions-ranging from potentially identifiable to completely anonymous-on survey response and on respondents' representativeness of the underlying sampling frame, completeness in answering sensitive survey items, and disclosure of sensitive information. We also examined the impact of 2 incentives ($10 versus $20) on these outcomes.

METHODS

A 3X2 factorial, randomized controlled trial of 324 representatively selected, male Gulf War I era veterans who had applied for United States Department of Veterans Affairs (VA) disability benefits. Men were asked about past sexual assault experiences, childhood abuse, combat, other traumas, mental health symptoms, and sexual orientation. We used a novel technique, the pre-merged questionnaire, to link anonymous responses to administrative data.

RESULTS

Response rates ranged from 56.0% to 63.3% across privacy conditions (p = 0.49) and from 52.8% to 68.1% across incentives (p = 0.007). Respondents' characteristics differed by privacy and by incentive assignments, with completely anonymous respondents and $20 respondents appearing least different from their non-respondent counterparts. Survey completeness did not differ by privacy or by incentive. No clear pattern of disclosing sensitive information by privacy condition or by incentive emerged. For example, although all respondents came from the same sampling frame, estimates of sexual abuse ranged from 13.6% to 33.3% across privacy conditions, with the highest estimate coming from the intermediate privacy condition (p = 0.007).

CONCLUSION

Greater privacy and larger incentives do not necessarily result in higher disclosure rates of sensitive information than lesser privacy and lower incentives. Furthermore, disclosure of sensitive or stigmatizing information under differing privacy conditions may have less to do with promoting or impeding participants' "honesty" or "accuracy" than with selectively recruiting or attracting subpopulations that are higher or lower in such experiences. Pre-merged questionnaires bypassed many historical limitations of anonymous surveys and hold promise for exploring non-response issues in future research.

Age biases in a large HIV and sexual behaviour-related internet survey among MSM

BACKGROUND

Behavioural data from MSM are usually collected in non-representative convenience samples, increasingly on the internet. Epidemiological data from such samples might be useful for comparisons between countries, but are subject to unknown participation biases.

METHODS

Self-reported HIV diagnoses from participants of the European MSM Internet Survey (EMIS) living in the Czech Republic, Germany, The Netherlands, Portugal, Sweden and the United Kingdom were compared with surveillance data, for both the overall diagnosed prevalence and for new diagnoses made in 2009. Country level prevalence and new diagnoses rates per 100 MSM were calculated based on an assumed MSM population size of 3% of the adult male population. Survey-surveillance discrepancies (SSD) for survey participation, diagnosed HIV prevalence and new HIV diagnoses were determined as ratios of proportions. Results are calculated and presented by 5-year age groups for MSM aged 15-64.

RESULTS

Surveillance derived estimates of diagnosed HIV prevalence among MSM aged 15-64 ranged from 0.63% in the Czech Republic to 4.93% in The Netherlands. New HIV diagnoses rates ranged between 0.10 per 100 MSM in the Czech Republic and 0.48 per 100 in The Netherlands. Self-reported rates from EMIS were consistently higher, with prevalence ranging from 2.68% in the Czech Republic to 12.72% in The Netherlands, and new HIV diagnoses rates from 0.36 per 100 in Sweden to 1.44 per 100 in The Netherlands. Across age groups, the survey surveillance discrepancies (SSD) for new HIV diagnoses were between 1.93 in UK and 5.95 in the Czech Republic, and for diagnosed prevalence between 1.80 in Germany and 4.26 in the Czech Republic.Internet samples of MSM were skewed towards younger age groups when compared to an age distribution of the general adult male population. Survey-surveillance discrepancies (SSD) for EMIS participation were inverse u-shaped across the age range. The two HIV-related SSD were u- or j-shaped with higher values for the very young and for older MSM. The highest discrepancies between survey and surveillance data regarding HIV-prevalence were observed in the oldest age group in Sweden and the youngest age group in Portugal.

CONCLUSION

Internet samples are biased towards a lower median age because younger men are over-represented on MSM dating websites and therefore may be more likely to be recruited into surveys. Men diagnosed with HIV were over-represented in the internet survey, and increasingly so in the older age groups. A similar effect was observed in the age groups younger than 25 years. Self-reported peak prevalence and peak HIV diagnoses rates are often shifted to higher age groups in internet samples compared to surveillance data. Adjustment for age-effects on online accessibility should be considered when linking data from internet surveys with surveillance data.

Secular trends in cardiovascular risk factors with a 36-year perspective: observations from 38- and 50-year-olds in the Population Study of Women in Gothenburg

OBJECTIVES

To study secular trends in cardiovascular risk factors in four different cohorts of women examined in 1968-1969, 1980-1981, 1992-1993 and 2004-2005.

DESIGN

Comparison of four representative cohorts of 38- and 50-year-old women over a period of 36 years.

SETTING

Gothenburg, Sweden with approximately 450,000 inhabitants.

SUBJECTS

Four representative samples of 38- and 50-year-old women were invited to free health examinations (participation rate 59-90%, n =1901).

MAIN OUTCOME MEASURES

Body mass index (BMI), systolic and diastolic blood pressure (SBP, DBP), leisure time exercise, use of antihypertensive medication, smoking, levels of haemoglobin, b-glucose, s-cholesterol, s-triglycerides and HDL-cholesterol.

RESULTS

There was no significant difference in mean BMI from 1968-1969 versus 2004-2005. Mean leisure time exercise was significantly higher in later born cohorts; in 1968, around 15% were physically active compared with 40% in 2004. SBP and DBP, mean s-cholesterol and s-triglyceride levels were significantly lower in both 38- and 50-year-old cohorts in 2004-2005 versus 1968-1969. HDL-cholesterol (not measured until 1992-1993), showed a significantly higher mean level in 2004-2005. Reduction of risk factors was apparent in women with a high as well as low level of physical activity. Smoking declined most in women with high levels of physical activity.

CONCLUSIONS

Several cardiovascular risk factors related to lifestyle have improved in middle-aged women from the 1960s until today. Most of the positive trends are observed in women with both low and high physical activity.