ORIGINAL PAPER

æ

Participation bias in longitudinal studies: experience from the Population Study of Women in Gothenburg, Sweden Lauren Lissner1,2, Ingmar Skoog3, Kate Andersson1, Nils Beckman3, Valter Sundh1, Magda Waern3, Dimitri Edin Zylberstein1, Calle Bengtsson and Cecilia Bjo¨rkelund1 Department of Primary Health Care, Sahlgrenska Academy at Go¨teborg University, 2Nordic School of Public Health, Go¨teborg, 3Neuropsychiatric Unit, Institute of Clinical Neurosciences, Sahlgrenska Academy at Go¨teborg University, Sweden. Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

1

Scand J Prim Health Care 2003;21:242
/247. ISSN 0281-3432




Objective / To describe a cohort study of women receiving a series of comprehensive health examinations over 32 years. Design / Longitudinal population study based on a randomised sample of the female population from defined age cohorts. Setting / City of Go¨teborg, Sweden. Subjects / Subjects were 38, 46, 50, 54 or 60 years old at the start of the study in 1968. Re-examinations were performed in 1974, 1982 and 1992. Non-participants in the most recent examination, initiated in 2000, were offered home visits. Main outcome measures / Participation, anthropometric and blood pressure changes. Results / At the end of the 32-year follow-up, 64% of the original participants were alive, and low participation among survivors was a problem. An acceptable participation rate (71% of those alive) was obtained after home visits were offered. Surviving non-participants already had elevated cardiovascular risk factors at onset of the study in 1968, along with lower educational level and lower socioeconomic status. Home visited subjects were similar to non-participants with










‘‘The Population Study of Women in Gothenburg’’ (Go¨teborg in Swedish) started in 1968
/69, with health examinations of representative samples of 1462 women from five age cohorts (1). Follow-up examinations were conducted in 1974
/75, 1980
/81 and 1992
/93 (2
/4). Among the strengths of the study are the unusually high participation rate at baseline (90%), the comprehensiveness of examinations and access to national registers ensuring virtually complete followup even among subsequent drop-outs. Participation after 1968
/69 remained relatively high among survivors (2
/4). However, certain differences between participants and non-participants started to emerge in the 1980s (3). For instance, in 1968
/69 systolic blood pressure was significantly higher in subsequent drop-outs than in participants on both occasions (3). At follow-up in 1992
/93, mortality in the original cohort was similar to that of the population as a whole, i.e. probably not affected by the health examinations themselves (4). We now report on a follow-up examination started in 2000. Baseline characteristics 1968
/69 are comScand J Prim Health Care 2003; 21

regard to anthropometry and blood pressure, but did not differ from participants with regard to social indicators. Thirty-two-year longitudinal data demonstrate clear ageing effects for several important variables, which should, however, be considered in the context of documented differences with non-participants at the baseline examination. Conclusions / Longitudinal studies in elderly populations provide important data on changes during the ageing process. However, participation rates decline for a number of reasons and generalisations should be made with care. Moreover, including home visits in the protocol can both increase participation and reduce participation bias in elderly cohorts.




Key words: epidemiology, longitudinal data, participation, prospective studies, women. Lauren Lissner, Department of Primary Health Care, Sahlgrenska Academy at Go¨teborg University, Box 454, SE-405 30 Go¨teborg, Sweden. E-mail: [email protected]

pared in participants versus non-participants still alive in 2001. We also present longitudinal data on selected anthropometric and cardiovascular risk factors in women who participated in all five examinations. It was hypothesised that as attrition rates increase,

Non-participation is potentially a source of bias in longitudinal population research. . The Population Study of Women in Gothenburg achieved satisfactory participation rates (71%) after 32 years by offering optional home visits. . 32-year follow-up curves demonstrate changes in BMI, abdominal obesity, height and blood pressure during the ageing process. . Differences between participants and dropouts were observed which emphasise the importance of optimising participation rates in population studies. DOI 10.1080/02813430310003309

Participation bias in longitudinal studies

Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

differences between participants and surviving nonparticipants may become more apparent.

METHODS Population-based cohort in 1968
/69 In 1968
/69, a representative sample of 1622 women living in the city of Go¨teborg were identified by the Revenue Office Register and received written invitations to free health examinations, followed up by telephone calls to schedule visits. The sampling method (1) was based on date of birth (women born 1908, 1914, 1918, 1922 and 1930 on days ending with 6, 12, 18, 24 or 30), in order to sample a representative cross-section of women in the community in the age groups studied. A total of 1462, of women, aged 38, 46, 50, 54 and 60, accepted the invitation, resulting in a participation rate of 90.1%. The survey was performed during a 12-month period, examining women born at the beginning of the year first, thus reducing the influence of age differences within each cohort. Similar procedures were applied in all the subsequent examinations, making the intervals between examinations as even as possible. A personal history describing for example social background, education, exercise, smoking and alcohol consumption was obtained during a standardised interview. Anthropometric measurements, i.e. height, weight, waist and hip circumferences, were taken according to protocols described previously (1
/4) from which body mass index (BMI) and waist-hip ratio (WHR) were calculated. Blood samples were drawn in the fasting state for determination of serum triglyceride and total cholesterol concentrations. Serum samples were stored for future (then undetermined) analyses, e.g. homocysteine. Systolic and diastolic blood pressures were taken after 5 min of rest in a sitting position. For the purpose of the present analysis, women were classified into selected risk factor categories as follows.

243

‘‘Current smokers’’ were identified as women who currently smoked or had stopped smoking during the previous year. Subjects were classified as ‘‘inactive’’ if they reported being physically active less than 4 h a week during leisure time. Women completing only ‘‘basic’’ education (approx. 6 years, depending on cohort) were classified as less educated. Alcohol consumption was categorised according to whether subjects reported drinking any alcoholic beverage (beer, wine or liquor) more than once a week. Finally, women reported their own occupations and, if married, their husbands’ occupations, according to Carlsson’s standard occupations grouping system (5). The analysis reported here compares Group I (high socioeconomic status) with all others, using husband’s occupational group if available, and woman’s own occupation otherwise. In 1974
/75, 1980
/81 and 1992
/93, all women examined in 1968
/69 were offered new examinations (2
/4). Details of the recruitment, age-specific participation rates, and re-examination protocols have also been presented elsewhere (2
/4). Main examination in 2000
/2001 From September 2000 until June 2001, all women who had participated in 1968
/69 and were alive according to the national person registry were invited to new examinations, conducted in the same premises and with partly the same staff as in 1992
/93 (4). Initial contact was by mail, followed by telephone call. An updated interview describing medical history, drug consumption and lifestyle characteristics was administered by physicians and research nurses. All women underwent psychiatric, dental and physical examinations. Subgroups of women also underwent audiologic examinations (40%), dietary interviews (60%), Activities of Daily Living (ADL) function assessment (60%), movement examination (60%) and a psychological interview (25%). Anthropometric measurements were identical to the earlier examinations and also included four skinfold

Table I. Number of participants in the original examination and status after 32 years. Birth cohort

1908

1914

1918

1922

1930

Total

Age if alive in 2000
/2001 Participants 1968
/1969 Available for the study (alive1 15 June 2001) Clinical examination Home visit Non-participant Participation (%) including home visits

92 y 81 12 1 7 4 66

86 y 180 69 20 24 25 65

82 y 398 236 113 63 60 75

78 y 431 300 155 47 98 67

70 y 372 312 205 26 81 74

1462 929 494 167 268 71

1

With the exception of two subjects who were ‘‘available’’ for the study (one clinical examination, one home visit) but died before 15 June 2001 and are included in participation (%) group. Scand J Prim Health Care 2003; 21

244

L. Lissner et al.

Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

measures. Blood pressures were taken in the lying and sitting positions and after standing for 60 sec. Peak expiratory flow (PEF) was measured with a PEF meter, and vital capacity (VC) and forced expiratory volume with a spirometer. Blood samples were drawn in the fasting state. Home visits during 2000
/2002 During the examinations September 2000
/June 2001, many women, especially from the oldest cohorts, declined participation because of physical impairment or frailty, despite being offered taxi transportation. Others had moved from the Go¨teborg region. We therefore offered all non-participants home visits with a reduced protocol. The home visits were mainly performed September 2001
/April 2002 with a twostep procedure. First, a trained psychiatric research nurse performed a full neuropsychiatric examination and took a modified medical history. The second nurse visit included blood and urine sampling, ECG, anthropometric measurements, PEF, blood pressure and pulse rate, and questions regarding drug consumption and ADL. Two-thirds of the home examinations took place in the Go¨teborg area (within the local telephone zone), and the remaining third in other parts of Sweden by travelling teams. Most home visits (86%) took place after the main examination cycle (2001
/2002) was completed. The initial home visits (14%) were conducted simultaneously with the clinical examinations (2000
/2001).

Statistical analyses Subjects from the original cohort were stratified into four participation categories: clinical visit, home visit, non-participant or deceased. Baseline (1968
/1969) characteristics of participants subsequently attending the main (clinical) examination were calculated. Using a least squares linear regression model, age-adjusted contrasts were estimated for the clinical group versus the surviving non-participants. As age cannot be assumed to be linearly related to all other factors in the model, age group has been entered in the model as dichotomous ‘‘dummy’’ variables. Similar comparisons were made between clinical participants and those receiving home visits.

RESULTS Participation Table I describes the participation status of the five original age cohorts at the time of the 32-year followup study. The overall participation rate, including 494 visits to the clinical setting and 167 home-visits, was 71%, after excluding all subjects who died before the clinical examination period ended on 15 June 2001. Characteristics of drop-outs Table II compares baseline characteristics (1968
/ 1969) of participants versus non-participants. The first data column contains statistics describing baseline characteristics among women who later partici-

Table II. Baseline characteristics (data collected in 1968
/1969) of participants in 2000
/2001 clinical follow-up, and age-adjusted differences with surviving non-participants and with subjects undergoing home visits in 2000
/2002. Differences reflect values in nonparticipants (A) or home-visits (B) minus values in participants. Mean values in non-participants and home visits, respectively, may be estimated by adding columns (A)/(B) and (A)/(C). Variables measured/ reported in 1968
/1969 BMI (kg/m2) WHR (cm/cm) Height (cm) Systolic blood pressure (mmHg) Diastolic blood pressure (mmHg) Serum triglycerides (mmol/l) Serum cholesterol (mmol/l) Serum homocysteine (mmol/l) Current smoker (%) Socioeconomic group middle/low (%) Education basic or less (%) Physically inactive in leisure time (%) Alcohol more than once/ week (%)

(A) Participants in 2000
/2001 (B) Difference between surviving (C) Difference between home visits 2000
/2002 and clinical examination (n/494, non-participants 2000
/2001 and mean/SD or%) participants participants 23.3 (3.2) 0.72 (0.04) 164.3 (5.9) 126.4 (16.2) 80.1 (8.7) 1.11 (0.58) 6.60 (1.38) 11.3 (4.1)

0.77** 0.009* /0.38 5.44**** 2.69**** 0.03 0.12 /0.09

34.4 88.2

7.4* 5.7*

5.3 /0.09

63.6 13.6

11.0** 5.4(*)

5.8 2.9

34.8

/4.7

P-levels for age adjusted differences: (*) B/0.1; *B/0.05; **B/0.01; *** B/0.001; **** B/0.0001. Scand J Prim Health Care 2003; 21

0.82** 0.013** /0.54 3.56* 2.52** 0.03 0.03 0.24

7.5

Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

Participation bias in longitudinal studies

pated in the 2000
/2001 clinical examinations, in all age groups combined. The second column displays age-adjusted baseline differences between participants in the clinical examinations and surviving non-participants. This adjustment corrects for confounding from known differences in age distributions in the groups under comparison, with significance levels corresponding to age-adjusted differences. The third column displays corresponding differences between clinical participants and the home-visit group. For each variable, age-adjusted means in surviving non-participants may be estimated by adding columns (A) and (B). Likewise, mean values in the home-visited group are estimated as the sum of (A) and (C). Two patterns of non-participation were detected. For several biological parameters (i.e. BMI, WHR, blood pressure), both non-participants and subjects who received home visits had higher values in 1968
/ 1969 than those who participated in the clinical follow-up. In contrast, serum cholesterol and triglyceride levels at the baseline examination did not vary significantly between clinical participants and either of the comparison groups. Several lifestyle related characteristics, such as smoking, physical inactivity, less education and lower social group were more common in non-participants than in clinical participants. How-

245

ever, subjects agreeing to home visits did not differ from those attending the regular clinical examination with respect to any of these ‘‘lifestyle’’ variables. Longitudinal measures on participants Fig. 1 illustrates serial changes in height, BMI, WHR and systolic blood pressure over 32 years in those women from birth cohorts 1930, 1922 and 1918 who participated in all five examinations. These three birth cohorts, each shown with a separate curve, are of particular methodological value as they have health examinations at ages 50 or 70 in common. These figures provide an example of the longitudinal data that have been collected over three decades, at the same time underscoring that participants in all five examinations represent a highly selected group, both in terms of having survived and choosing to participate. Mean values at baseline are displayed with open symbols for subsequent non-participants who were still alive at the end of the 32-year follow-up period, to underscore the presence of self-selection bias, independent of attrition of study group due to mortality. The curves describing changes in WHR may be of special interest because there have been few longitudinal studies describing this important risk factor for extended periods of time.

Fig. 1. Longitudinal plots showing BMI, height, WHR and systolic blood pressure in the three youngest age cohorts examined between 1968
/1969 and 2000
/2001. The solid diamonds and solid line refer to women born in 1930 (n /185). The solid squares and dotted line refer to women born 1918 (n/144). The solid triangles and dashed line refer to women born 1918 (n /104). The open symbols refer to baseline data on subsequent nonparticipants in corresponding age cohorts. Scand J Prim Health Care 2003; 21

Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

246

L. Lissner et al.

DISCUSSION Over the past 50 years, a large number of prospective population studies have been initiated in different parts of the world, the Framingham Study being an early example (6). As far as we know there has been no other study of women with a combination of representative sampling, long follow-up, comprehensive examination protocols and high participation rates at baseline comparable to Population Study of Women in Gothenburg. Results from this study have provided much population-based research and reference material (e.g. 7
/13). In this article, we present new longitudinal curves describing changes in risk factors during the ageing process, in a now elderly sample of women who have survived to advanced ages and chosen to participate in five comprehensive health examinations. However, limitations must be considered, the most important being that the cohort becomes increasingly selected over time. The most dramatic participation difference observed in this study involves participation in clinical visits by the oldest and youngest cohorts, respectively, where 9% versus 66% of the surviving original participants returned after 32 years. These figures must be considered together with total survival rates of only 15% in the oldest cohort compared to 84% in the youngest. After including the home visits, the participation rate ranged from 65% to 75% across age groups. It should be noted that the overall participation of 71% is an approximation of what 32-year follow-up rates might have been had it been possible to offer all the home visits during the clinical examination cycle. Problems of attrition are well known in longitudinal research, particularly as study populations reach advanced ages. The present article documents that subjects who continue to participate are highly selected. An interesting pattern of non-participation emerged: independent of age, surviving non-participants were likely to have had higher BMI, WHR and blood pressure 32 years earlier than those who reattended the clinical examination, in addition to being less educated and smoking more. Importantly, when combining the home and clinical visits into a total participation category, most of the differences with non-participants were attenuated (data not shown), as would be expected since the home-visited subjects were similar to non-participants in a number of ways. Note that differences in blood pressure between clinical participants and non-participants were observed as early as 1980
/1981, but an increasing number and variety of differences were observed as drop-out increased. The 32-year participation experience of this study highlights the need to offer home visits to elderly Scand J Prim Health Care 2003; 21

subjects, even if this changes the protocol, in order to obtain an acceptable participation rate together with less selection bias. Many elderly may still decline participation, but it should be noted that total nonparticipation rates in this study are similar to current rates typically experienced in much younger cohort studies in Go¨teborg (14). In conclusion, participation bias is an important issue to consider, both in designing and conducting prospective studies as well as in clinical interpretation of data and reference material from longitudinal studies. ACKNOWLEDGEMENTS This study was funded by grants from the Swedish Research Council (6652, 04578, 11267), the Swedish Council for Working Life and Social Research (2835, 2646), The Alzheimer’s Association Stephanie B. Overstreet Scholars (IIRG-00-2159) and the Bank of Sweden Tercentary Foundation. REFERENCES 1. Bengtsson C, Hallberg L, Ha¨llstro¨m T, Hultborn A, Isaksson B, Lennartsson J, et al. The population study of women in Go¨teborg 1974
/75: The second phase of a longitudinal study. General design, purpose and sampling results. Scand J Soc Med 1978;6:49
/54. 2. Bengtsson C, Blohme´ G, Hallberg L, Ha¨llstro¨m T, Isaksson B, Korsan-Bengtsen K, et al. The study of women in Gothenburg 1968
/1969: A population study. General design, purpose and sampling results. Acta Med Scand 1973;193:311
/8. 3. Bengtsson C, Gredmark T, Hallberg L, Ha¨llstro¨m T, Isaksson B, Lapidus L, et al. The population study of women in Gothenburg 1980
/81: The third phase of a longitudinal study. Comparison between participants and non-participants. Scand J Soc Med 1989;17:141
/5. 4. Bengtsson C, Ahlqvist M, Andersson K, Bjo¨rkelund C, Lissner L, So¨derstro¨m M. The prospective population study of women in Gothenburg, Sweden, 1968
/69 to 1992
/93. A 24-year follow-up study with special reference to participation, representativeness, and mortality. Scand J Prim Health Care 1997;15:214
/9. 5. Carlsson G. Social mobility and class structure. Lund, Sweden: GWK Gleerup, 1958. 6. Dawber TR. The Framingham Study: The epidemiology of atherosclerotic disease. Cambridge, MA: Harvard University Press, 1980. 7. Bengtsson C. Ischaemic heart disease in women. A study based on a randomized population sample of women and women with myocardial infarction in Go¨teborg, Sweden. Acta Med Scand 1973;Suppl 549. 8. Ha¨llstro¨m T. Mental disorder and sexuality in the climacteric. A study of psychiatric epidemiology. Go¨teborg: Scandinavian University Books, 1973. 9. Bjo¨rkelund C, Hulte´n B, Larsson B, Lissner L, Rothenberg E, Bengtsson C, et al. Nya vikt-la¨ngdtabeller fo¨r medela˚lders och a¨ldre. Vikten o¨kar mer a¨n la¨ngden (Weight is increasing faster than height: new weight
/height tables for the middle-aged and elderly). Lakartidningen 1997;94:332
/5.

Participation bias in longitudinal studies

prospective cohort study in Gothenburg, Sweden. BMJ 1999;319:890
/3. ¨ , Wedel H, Bjo¨rkelund C, Bengtsson 13. Cabrera C, Helgesson O C, Lissner L. Socioeconomic status and mortality in Swedish women: Opposing trends for cardiovascular disease and cancer. Epidemiology 2001;12:532
/6. 14. Wilhelmsen L, Johansson S, Rosengren A, Wallin I, Dotevall A, Lappas G. Risk factors for cardiovascular disease during the period 1985
/1995 in Go¨teborg, Sweden. The GOT-MONICA Project. J Intern Med 1997;242:199
/ 211.

Scand J Prim Health Care Downloaded from informahealthcare.com by 61.147.103.151 on 05/20/14 For personal use only.

10. Bengtsson C, Bjo¨rkelund C, Lapidus L, Lissner L. Associations of serum lipid concentrations and obesity with mortality in women: 20-year follow-up of participants in prospective population study in Gothenburg, Sweden. BMJ 1993;307:1385
/8. 11. Ahlqwist M, Bengtsson C, Lapidus L, Lindstedt G, Lissner L. Concentrations of blood, serum and urine components in relation to number of amalgam tooth fillings in Swedish women. Community Dent Oral Epidemiol 1995;23:217
/21. 12. Ro¨dstro¨m K, Bengtsson C, Lissner L, Bjo¨rkelund C. Preexisting risk factor profiles differ in subsequent users versus non-users of hormone replacement therapy: A

247

Scand J Prim Health Care 2003; 21