The importance of a lifecourse perspective in understanding the onset and progression of chronic disease is now well accepted.1 This approach requires information on exposures from early in life, ideally from birth, through childhood and into adulthood. Studies that follow up participants over decades are, however, inherently susceptible to attrition. Unavoidable losses through death or emigration (for geographically defined samples) accumulate over time and individuals may refuse to participate further or become untraceable. Such losses reduce the statistical power of those studies and may introduce bias.

The 1958 birth cohort (National Child Development Study) is the second oldest of the British birth cohorts and is a valuable resource for lifecourse health research in several areas including, for example, obesity, asthma and mental health.2 At age 45 years, a biomedical survey was completed with detailed and objective measures of health, function and disease risk.3 4 It is anticipated that this survey will become a baseline from which the onset and progression of chronic diseases and the ageing processes of later life can be followed. Earlier surveys of the cohort have been examined for sample bias. The 16-year sample underrepresented socially, educationally and behaviourally disadvantaged children, but physical attributes varied little.5 These biases continued into adulthood, with the additional loss of non-white ethnic groups;6^–8 however, in most respects the magnitude of bias has remained modest.

Evaluation of bias in the 45-year sample has yet to be demonstrated, and given that these biomedical data will be used extensively in immediate and long-term lifecourse studies it is important to assess the nature and extent of any sample biases. Accordingly, we describe response in the 1958 birth cohort to date. Our aim was to examine whether the 45-year sample is representative of the total cohort and to establish the extent to which any bias was present before the 45-year biomedical survey.

METHODS

Study sample

The cohort comprises 17 638 individuals born in one week in 1958 in England, Scotland and Wales and recruited into the Perinatal Mortality Study,9 and a further 920 immigrants with the same birth week.2

Contact and data collection

For the Perinatal Mortality Study, mothers were interviewed by health workers with access to medical records. Cohort members still resident in Great Britain were recontacted through their schools at ages seven, 11 and 16 years; it was during these surveys that immigrants were recruited. Information was collected from parents and teachers; cohort members underwent medical examination and cognitive testing. In adulthood (ages 23, 33 and 42 years) information was obtained by research interviewers conducting home visits. The 45-year biomedical survey comprised a home interview by a research nurse, two short self-completed questionnaires (one sent in advance, the second completed during the nurse visit), blood samples, and a saliva sample obtained after the interview. Ethical approval for the survey was obtained from South East Multi-centre Research Ethics Committee (ref. 01/1/44).

Contact in adulthood was not attempted for those who had not participated in any childhood follow-up survey. In addition, those who had not participated since age 16 years, lacked a valid address, had previously displayed threatening behaviour, were incapable of giving informed consent, or were in the armed forces were not contacted at the 45-year survey.10

Classification of response

Deaths were ascertained through receipt of death certification (the cohort is flagged in the National Health Service Central Registry (NHSCR)) or notification to the study team. Those not resident in Great Britain at the time of the survey were classified as emigrants, residents were defined as the eligible sample (ie deaths were excluded). Eligible cohort members who could not be contacted or refused to participate were classified as non-responders. Those providing data were participants; those for whom contact was not attempted were classified separately.

The biomedical survey consisted of four survey components (physical measurements, mental health interview, blood sampling and saliva sampling) plus two self-completed questionnaires. Separate consent was sought for each component or, in the case of the physical component, for each measurement (vision, blood pressure, hearing threshold, height, weight, waist and hip circumference, and lung function). Blood samples were analysed for fibrinogen, C-reactive protein, tissue plasminogen activator, von Willebrand factor, glycosylated haemoglobin, total cholesterol, high-density lipoprotein cholesterol, triglycerides, total immunoglobulin E, and insulin-like growth factor 1. Further consents were sought for the extraction of DNA for genetic analysis, and for immortalised cell line culture. Two saliva samples, collected early and mid-morning on a day after the interview by the cohort member, were mailed for cortisol analysis. Questionnaires were collected by the nurse at the end of the interview or returned by post if not completed.

Key characteristics

Ethnic group was self-reported at ages 42 or 33 years or, if not available, observer-rated at age 16 or 11 years.

Childhood characteristics (measured at birth or 7 years) were selected from social, behavioural, cognitive, physical and maternal behaviour domains. Social characteristics comprised mother’s husband’s social class (Registrar General’s classification), recorded at birth and categorised into non-manual, manual, or no male head of household. Duration of mother’s education was classified as up to, or over, the statutory school leaving age (age 14 years for those born before April 1933, age 15 years for those born subsequently). Housing tenure, hospitalisation for any reason, and episodes in social care were recorded at age 7 years. Teachers rated cohort members’ behaviour at age 7 years using the Bristol Social Adjustment Guide;11 externalising and internalising behaviour scores underwent square root transformation before categorisation into normal (lowest 50% of scores), problem (top 13% of scores) and intermediate (remainder) groups.12 13 Cognitive ability was assessed using the Southgate Reading Test14 and Problem Arithmetic Test;15 poor reading or mathematical ability approximates the lowest 10% of scores. Physical characteristics comprised birthweight, height and body mass index (BMI) at age 7 years. Babies weighing less than 2500 g were classified as low birthweight. Children below the fifth centile were identified as short stature (⩽1.15 m for boys and ⩽1.12 m for girls); overweight was defined as a BMI of 17.92 kg/m² or greater for boys and 17.75 kg/m² or greater for girls, according to International Obesity Task Force recommendations.16 Smoking after the fourth month of pregnancy and breastfeeding for at least one month represented maternal behaviour.

Adult characteristics, collected at age 42 years, included occupational status, social class (manual or non-manual, based on current or most recent occupation) and marital status. BMI was calculated from self-reported height and weight, those with a BMI of 30 kg/m² or greater were classified as obese. Cigarette smoking was defined as current, ever or never, (reports of never smoking were verified using data at ages 23 and 33 years); usual frequency of alcohol consumption was self-reported.

Analysis

The number of cohort members who were dead, migrant, not contacted, non-responders or participants was ascertained for each survey from birth to 45 years; response was expressed as a percentage of the eligible sample, and of the sample for whom contact was attempted. Consent (n, %) to each biomedical survey component, including individual physical measures, was calculated. The number of participants with valid values (that is, after invalid and improbable values were excluded) for each survey component, usable DNA specimens, and the number completing each questionnaire was also calculated. Because there are multiple physical measures and blood analytes, highest and lowest valid values are presented. For saliva, results are presented for the baseline (early morning) cortisol measure only.

We examined the distribution of childhood and adult characteristics in the 45-year sample, relative to the total cohort at 45 years. Representativeness of the 45-year sample was expressed as percentage bias ((45-year sample% − total surviving cohort%)/total surviving cohort%).5 Deaths were excluded from the calculations because of differential mortality;17 emigrants were included, as they are eligible to rejoin the cohort on return to Great Britain. To establish the extent of bias before the 45-year survey, we calculated percentage bias between the 42-year sample and the total cohort surviving up to 42 years. To establish whether the selected childhood and adult characteristics represented different constructs we used multivariate logistic regression models predicting 45-year drop-out (odds ratios, 95% confidence intervals). Birth, seven-year, and 42-year variables were modelled separately because of differences in the level of missing data between surveys. Analyses were conducted using Stata 9.2 (Stata Corp., College Station, Texas, USA).

RESULTS

Survey response

The total number of cohort members increased from 17 638 at birth to 18 558 by age 16 years after supplementation with immigrants (table 1). The eligible sample has declined over time as a result of death and emigration, although early loss was offset by the recruitment of immigrants. By age 45 years, 6.7% of the cohort had died, with over half the deaths occurring before the age of seven years, mostly in the first months of life.18 Emigration increased throughout childhood to age 33 years, but subsequently decreased; returning emigrants being eligible to participate. Despite losses, 86.3% of the total cohort remained eligible for contact at the 45-year survey.

Non-response increased with age to 42 years; the most pronounced increase occurring at 23 years. The 45-year survey had the lowest proportion of non-responders in adulthood (22.7% of the eligible sample), but participation did not increase because contact was not attempted for 18.8% of the eligible sample. Accordingly, with this latter group omitted, response at 45 years was comparable to other adult surveys (72.1%). After excluding permanent refusals (n  =  1038), and those in whom contact was not attempted for other reasons (n  =  3004), contact details were issued for fieldwork for 12 069 individuals, of whom 98 were subsequently found to have died or emigrated, thus contact was attempted for 11 971; 9377 participated in the 45-year survey.

Participation within the 45-year survey

Consent was lowest for the blood and saliva components (93.7% and 97.7% of participants, respectively); virtually all participants consented to the mental health and physical components, with little variation in consent for individual physical measurements (table 2). Of 9377 participants, 8585 (91.6%) consented to all four interview components. Of 8790 consenting participants, 8302 successfully provided a blood sample, the majority also provided consent for DNA analysis and immortalised cell line cultures (table 2).

All those who consented to the mental health interview completed it; completion was high for the pre-interview (92.8%) and within-interview questionnaires (98.3%). The number with valid values for physical measurements and blood analytes varied. For physical measurements, height and weight had the greatest number with valid values, whereas lung function and hearing had the fewest; of the analytes, fibrinogen and tissue plasminogen activator had the fewest values. Nevertheless, 84.6% of participants had valid values for all seven physical measures and 78.1% for all 10 analytes. Fewer participants had a valid salivary cortisol measure (69%).

45-Year sample representativeness

Table 3 compares the 45-year sample with the total surviving cohort at 45 years on gender, ethnicity and childhood characteristics. In general, bias is modest; however, for all of the characteristics studied it is the disadvantaged group that is underrepresented in the sample. Only 3.2% of the total cohort were non-white, by age 42 years the proportion had decreased, a trend that continued to 45 years; there was little difference in gender. Social characteristic biases were modest, with underrepresentation of those from manual class backgrounds, whose mother did not remain at school, or living in rented housing, although biases were larger for extreme groups, notably those born into households with no male head or with periods in social care. The 45-year sample underrepresented those with externalising (percentage bias −16.9%) or internalising (−14.3%) behavioural problems, poorer reading (−20.7%) or maths (−13.6%) scores, and shorter stature (−14.3%), at age 7 years. Smaller biases were found for overweight at age 7 years and those not breastfed; there was little difference between the 45-year sample and the total cohort with respect to low birthweight and maternal smoking. In terms of adult characteristics, 45-year participants were less likely to be in a manual social class, obese or current smokers. The 45-year sample differed little from the total cohort in marital status and drinking frequency (table 4).

Table 3 shows that in most instances bias in the 45-year sample appears to have arisen less through refusal to participate than through other types of non-participation; emigration also contributed to ethnic bias (4.4% of emigrants were non-white; data not shown). Comparison of the 42-year sample with the total cohort shows biases in childhood characteristics in the same direction as those for the 45-year sample, although less pronounced in magnitude. In logistic regression models of non-response to the 45-year survey, the odds ratios for birth characteristics before and after simultaneous adjustment for other birth factors, show persistent, although attenuated elevated risk for most factors (table 5). Similarly, for 7-year characteristics, the elevated odds ratio for non-response was reduced, but remained after adjustment for most factors.

DISCUSSION

Avoidable, and to a lesser extent, unavoidable losses have contributed to sample attrition in the 1958 birth cohort. Nevertheless, the 45-year sample is broadly representative of the total surviving cohort with respect to childhood social class, housing tenure, physical and maternal factors, and key adult characteristics. There was greater underrepresentation of more extreme groups, such as those with no male head of household at birth, experience of social care in early childhood, or of non-white ethnicity, and of those with childhood cognitive and behavioural problems. These biases were present in the study sample before the 45-year survey, albeit to a lesser extent. Within the 45-year survey consent for all four interview components was high. Despite sample attrition and some specific biases reported here the study remains of demonstrable value in lifecourse epidemiology. 3 4 13 19^–22

Methodological issues

Ascertainment of deaths since the 1980s, when the cohort was first flagged at NHSCR, is considered virtually complete, but before this date there is some scope for underestimation, as the study was reliant on notification by relatives or others. NHSCR records of emigration are dependent on the (voluntary) return of the emigrant’s NHS card or notification of cessation of benefits for emigrating claimants;23 however, emigration was also self-reported to the study.

The characteristics chosen to assess sample representativeness were not exhaustive, but were selected based on previous analyses of sample attrition in this cohort, and to represent the major areas of information covered by the study.5 7 8 The selection of childhood characteristics was restricted to the first two surveys when attrition was minimal, nevertheless some losses had occurred by age 7 years and therefore bias may be underestimated for 7-year characteristics, and even more so for adult characteristics (age 42 years).

Percentage bias was used to assess the representativeness of the 45-year sample rather than statistical testing because the large sample size would result in minor differences achieving significance.5^–7 Small differences in extreme groups can, however, be exaggerated when expressed as percentage bias and results should be interpreted accordingly. The effects of premature mortality on the representativeness of the 45-year sample are not considered here, as deaths were excluded from analyses.

Comparison with other studies

Only 2% of the 45-year sample classified themselves as non-white, compared with 7.5% of 45–49 year olds in the 2001 England and Wales census population.24 It is recognised that the 1958 cohort is not ethnically representative of the current British population. Approximately half of non-white cohort members were recruited to the study as immigrants at ages 7, 11 and 16 years; discontinuation of this supplementation in subsequent surveys and differential attrition has limited the ethnic representativeness of the cohort. In other respects the cohort differs only slightly from contemporary British 45–49 year olds, with broadly comparable proportions who are married (70% versus 68.5%), in paid employment (82.2% versus 79.3%), and home-owners (78.3% versus 79.5%) in the 45-year sample and census populations, respectively.24

Apart from modest variations in perinatal mortality and emigration rates, the overall pattern of attrition in the 1958 cohort is similar to that in other British birth cohorts.25 26 At 45 years, response in the 1958 cohort is only slightly lower than that in the 1946 cohort at age 43 years (58.6% versus 65.3% of the surviving sample living in Great Britain),26 despite the exclusion of illegitimate births from the 1946 cohort, who showed lower participation in our sample (at least as indicated by those born into households with no male head).

In the 1958 cohort, the 45-year sample was found here to be largely representative of the surviving cohort with respect to key childhood characteristics. Bias was modest, except for some extreme groups, and was already evident in the 42-year sample, to a lesser degree. Previous examinations of bias in the cohort have also found those with childhood cognitive and behavioural problems to be underrepresented.5^–7 In addition, bias was observed for adult social class, obesity and smoking behaviour. Other studies following large groups of individuals over time have also found socially disadvantaged, academically less able, and problem behaviour children,26^–30 or the unemployed,31 smokers,32 33 and immigrants,29 to show greater vulnerability to attrition.

Within the 45-year biomedical survey, consent to all four survey components was high. Similar consent rates for biomedical measurements were reported among participants of the 2003 Health Survey for England nurse visit, suggesting that once individuals consent to biomedical interview they are highly likely to consent to individual procedures.34 Lifelong association with the study may have encouraged cohort members to provide biological specimens. The number of valid values obtained varied between survey components. The use of technically complicated instruments for some physical measurements, problems taking and assaying blood, and the reliance on cohort members to collect and return saliva samples are likely to have affected whether a valid value was obtained.

In conclusion, the 1958 cohort shows attrition over time; however, the 45-year sample remains broadly representative of the total surviving cohort in terms of key childhood and adult characteristics. In general, 45-year survey participants were willing to participate in all aspects of the survey, and valid health data were obtained for the majority. Those biases found in the 45-year sample were largely present in previous surveys of the cohort and are common to other longitudinal studies. Because most lifecourse epidemiology tends to rely on several measures from different surveys sample bias may be exacerbated. It will be important to examine and possibly take account of the biases presented here in the interpretation and analysis of some future studies based on this cohort, for example, through multiple imputation,35^–37 or weighting strategies. The 1958 cohort study nevertheless remains an invaluable resource for lifecourse epidemiology. Although bias was modest, there was a consistent underrepresentation of disadvantaged groups who are most likely to experience ill health. Focus of resources on those most likely to respond in ongoing studies such as the 1958 cohort is of doubtful value in maintaining response over time. Renewed efforts to contact all eligible cohort members could improve the representativeness and thereby the value of future survey samples.