The Boyd Orr study is an historical cohort based on the Carnegie (Boyd Orr) Survey of Diet and Health in Pre-War Britain, 1937–1939 ^[17]. In the original survey, 4,999 boys and girls aged 0–19 years in 16 centers across the UK underwent physical measurements, the family completed a week-long dietary inventory and detailed assessments of socioeconomic environment were made. Ethical approval for the Boyd Orr study was obtained from the Multicentre Research Ethics Committee for Scotland. In 1997, a follow-up study re-established contact with the subjects using the National Health Service Central Register and its equivalent in Edinburgh. The follow-up study identified 3,182 individuals from the original sample of families who were traced, alive and residing in Britain. These subjects were sent a detailed health and lifestyle questionnaire and 1,648 responses were received. In 2002, all 732 surviving study members aged 63–83 years who lived near clinics in Bristol, London, Wisbech, Aberdeen and Dundee, and had previously consented to clinical follow-up, were contacted; of these 405 (55%) - 182 men and 223 women - agreed to take part in a detailed clinical examination and questionnaire, where fasting blood samples were taken and the get-up and go walking speed ^[18] and flamingo balance ^[19] physical performance tests were carried out.

The Caerphilly Prospective Study (CaPS) recruited 2,512 men aged 45–59 years between 1979 and 1983 from the town of Caerphilly, South Wales and the adjacent villages (http://www.epi.bris.ac.uk/caerphilly/caerphillyprospectivestudy.htm) ^[20]. Ethical approval for the different phases of the study was given by the Ethics Committee of the Division of Medicine of the former South Glamorgan Area Health Authority and Gwent Research Ethics Committee. For the second examination (phase II 1984–1988), the original cohort was supplemented with 447 men of a similar age who had moved into the defined area; however, 561 men were lost from the cohort giving a total of 2,398 men who participated in phase II. Since then, the men have been examined on three further occasions: phases III (1989–1993), IV (1993–1996) and V (2002–2004). At each phase, the men completed lifestyle questionnaires and fasting blood samples were taken. A total of 1,195 men aged 66–86 years attended the phase V clinic (or had a home visit) where the physical performance tests were undertaken. All participants in both cohorts gave informed written consent.

In Boyd Orr, IGF-I and IGF-II, the major binding proteins (IGFBP-2 and IGFBP-3) and insulin were measured from blood samples taken from the 405 participants in the 2002 clinic. Samples were taken after an overnight fast. The samples were spun and frozen to −20°C within one hour and transferred within three weeks to a −80°C freezer ^[21]. In CaPS, fasting insulin was assayed for previous studies and was measured using a radioimmunoassay based on a double antibody for insulin, available for 1,005 men with physical performance data. Insulin was used from phase I if available, or if the men did not attend phase I, data from phase II was used (n = 186). Serum IGF-I, IGF-II and IGFBP-3 were measured on 739 samples stored at −80°C and selected from phase II, approximately 19 years prior to the physical performance tests. Serum IGF-I, IGF-II, and IGFBP-3 levels were measured using in-house double-antibody radio immunoassays in both cohorts ^[22]. The radioimmunoassay measures total IGF-I and IGF-II levels, including all forms that have undergone minor fragmentation. In Boyd Orr, total levels of IGFBP-2 were measured by one-step sandwich enzyme-linked immunosorbent assays (DSL-10-7100; Diagnostic Systems Laboratories, Webster, TX); insulin was measured with an insulin-specific assay that does not cross-react with proinsulin and split forms of proinsulin. In Boyd Orr, the average coefficients of variation for intraassay variability for IGF-I, IGF-II, IGFBP-3, and IGFBP-2 were 6.7, 10.0, 3.9, and 5.0% and for interassay variation were 9.7, 14.0, 8.1, and 7.1%. Interassay coefficients of variation were 5.3%, 4.2% and 4.4% for low, medium and high values of insulin. In CaPS, the average coefficients of variation for intraassay variability for IGF-I, IGF-II, IGFBP-3 and phase II insulin were 7.0, 7.9, 5.6 and 7.5% and for interassay variation were 8.8, 9.7, 8.8 and 8.9%.

Information on potential confounders was collected from questionnaires administered at research clinics; in Boyd Orr this was from 1997 and 2002, and in CaPS this was from phases I (1979–1983), II (1984–1988) and III (1989–1993) and has been described in detail previously ^[23]. In Boyd Orr, adult social class was recorded in 1997 and corresponded to the job participants spent most of their working lives. For men, adult social class was classified into categories based on the 1966 Classification of Occupations. For women, the social class of their spouse or partner was used, if available, otherwise their own occupation was used. Adult social class was classified into three categories in Boyd Orr (I/II, professional and managerial; III, skilled; IV/V, partly skilled, unskilled, other or unclassifiable). In CaPS, social class in adulthood was based on participants' present or last job at phase II (1984–1988) and divided into four categories (I/II, professional and managerial; III, skilled non-manual; III, skilled manual; IV/V, partly skilled, unskilled, other or unclassifiable). Health behaviors were recorded at the 2002 Boyd Orr clinic and CaPS phase II. Questions on smoking behavior were used to assign people into three groups (current; past; and never smokers). Alcohol consumption was coded into three groups (frequent [at least weekly]; occasional; and never). Exercise scores were derived from questions on the frequency physical activities were carried out. The overall scores were split into three approximately equal sized groups ranging from least to most exercise. Adult diet was estimated from food frequency questionnaires and averaged over multiple time periods to reduce measurement error (Boyd Orr 1997 and 2002; CaPS phases I to III). To calculate nutrient intake, frequency of consumption of a food was multiplied by the nutrient content of a standard portion of that food using food tables ^[24]. In CaPS, the average portion sizes used were calculated from seven-day weighed dietary records completed by a one-third sample of the Caerphilly cohort at phase I ^[25]. Milk consumption, calcium, protein, fat and energy intake have been linked to IGFs in many cohorts ^[15], ^[26]–^[28], so these were considered as potential dietary confounding variables. Intake of milk was split into four groups ranging from lowest to highest intake; calcium, protein, fat and energy intake were considered as continuous variables.

Body mass index (BMI) was derived from height and weight measurements (weight [kg]/height [m]²), which were recorded at the research clinics when the blood samples were taken. We control for BMI because a previous study in the Boyd Orr cohort found a negative association between BMI and IGFBP-2, an inverted U-shape relationship between BMI and IGF-I, and a positive relationship of BMI with IGF-II and insulin ^[12]. Increased BMI was associated with slower walking times and poor balance in both cohorts ^[23]. Information on major co-morbidities (stroke, cancer, diabetes and angina) was collected from self-reported questionnaires. Derived variables were created to indicate a diagnosis of each disease by the time of the physical performance examination. In Boyd Orr: a diagnosis of diabetes was derived from a combination of self-report and fasting glucose measurement; cancer prevalence at the time of the measurement of physical performance was based on either self-reported cancer or a notification of cancer registration received from the National Health Service Central Register. Angina was based on the Rose chest pain questionnaire ^[29].

All analyses were carried out using Stata version 11. IGF values and levels of insulin were converted into z-scores, which accounted for any baseline and laboratory related variation in IGF levels: age at blood sample; sex (Boyd Orr); duration of sample storage time; the laboratory technician who performed the assays (CaPS); and the week of the assay (CaPS). The z-scores have a mean of zero and a standard deviation (SD) of one, enabling the comparison for an equivalent one SD change in each IGF measurement. Due to skewed distributions, IGFBP-2 and insulin and were natural log transformed before the z-scores were created. The IGF-I∶BP-3 molar ratio may be a marker of IGF-I bioavailability. Based on the molecular weight of IGF-I (7500) and IGFBP-3 (40,000, mean of glycosylated variants), the molar ratio was calculated by multiplying IGF-I∶BP-3 by 5.33 (40,000∶7500) ^[21]. Blood samples for IGF measurements (CaPS phase II) were only available for 739 men with the physical performance outcomes (CaPS phase V). Any differences for those with and without stored blood samples were examined with chi-squared tests. Correlations between continuous variables were calculated using Pearson's correlation coefficients.

Walking times from the get-up and go test were natural log-transformed, due to a skewed distribution. Linear regression models were used to investigate associations of IGFs with the log-transformed walking time and regression coefficients were multiplied by 100 to represent percentage change in walk time ^[30]. Since a longer time to complete the get-up and go test implies person was slower, a positive coefficient of percentage change represents a slower walking time and a negative estimated coefficient of percentage change represents a faster walking time. Over a third of participants achieved the maximum balance of 30 seconds; therefore, the flamingo test was dichotomized at the lowest 20% of performers, using a cut-point of five seconds. The flamingo test was modeled using logistic regression, with the outcome of interest being the inability to balance for five seconds (coded with a value of one in models and termed ‘poor balance’ in the text). We tested for interactions between the cohorts to assess whether it was suitable to pool the data. Minimally adjusted models control for age at the time of the physical performance test and sex (Boyd Orr); further adjusted models additionally control for research centre, adult social class, diet, exercise, smoking status and alcohol consumption; final models also controlled for BMI and co-morbidities (stroke, cancer diabetes and angina) at the time of the IGF measurement. Non-linear associations for the IGF variables were investigated by comparing models with and without quadratic terms.

All participants in CaPS were male and there was no evidence of a sex-interaction for IGFs with physical performance in Boyd Orr (p-values for the interaction terms ranged from 0.2 for IGF-II to 0.9 for IGF-I); therefore, data were pooled for men and women. In CaPS, a one SD increase in IGF-I was associated with a 2.0% faster get-up and go time after 19 years' follow-up in minimally-adjusted models (95% CI 0.4%, 3.7%; p = 0.017) (Table 3). In the cross-sectional Boyd Orr analysis there was a 1.1% faster get-up and go time per SD increase in IGF-I. However, the confidence interval was consistent with chance (95% CI 1.3% slower, 3.6% faster; p = 0.4). There was no evidence of an interaction between cohorts (Boyd Orr and CaPS) and associations of IGF-I with the get-up and go test, so the data were pooled (p interaction = 0.5). The combined estimate for a SD change in IGF-I was a 1.7% faster get-up and go time (95% CI 0.3%, 3.1% faster; p = 0.017). The coefficient was attenuated to 1.2% faster get-up and go time (0.2% slower, 2.3% faster; p = 0.09) when controlling for confounding variables and to 1.0% faster get-up and go time (95% CI 0.3% slower, 2.4% faster; p = 0.14) after controlling for BMI and co-morbidities (stroke, cancer, diabetes and angina). There was no evidence of a non-linear association between IGF-I and the get-up and go test (p for including a quadratic term = 0.4 in Boyd Orr; p = 0.7 CaPS). There was no evidence of an association between IGF-II and the get-up and go test in either cohort. In Boyd Orr, a one SD increase in IGFBP-2 was associated with 2.6% slower get-up and go times (95% CI 0.4%, 4.8% slower; p = 0.021), but this was only seen when controlling for BMI and co-morbidities. In Boyd Orr, a one SD increase IGFBP-3 in old age was associated with a 3.0% faster get-up and go time (95% CI 0.9%, 5.2% faster; p = 0.007) in fully adjusted models. There was no evidence of any association between IGFBP-3 and get-up and go time in CaPS and there was evidence of a difference between the Boyd Orr and CaPS studies (p interaction = 0.012). A one SD increase in insulin was associated with a 1.3% slower get-up and go time in the confounder-adjusted model (pooled data; 95% CI 0.0%, 2.7%; p = 0.07), but this was attenuated after adjustment for BMI and co-morbidities (coefficient 0.3% faster; 95% CI 1.1% slower, 1.8% faster; p = 0.6). Mutually controlling for IGF-I and insulin did not alter findings (results not shown).

Associations of the insulin-IGF system with the flamingo test are shown in Table 4. The estimated odds suggest a lower risk of poor balance for a one SD increase in IGF-I, but the confidence intervals are consistent with chance (Boyd Orr OR = 0.88 in age- and sex-adjusted model; 95% CI 0.65, 1.20; p = 0.4; CaPS OR = 0.87; 0.73, 1.04; p = 0.13). There was no evidence of an interaction for the associations of IGF-I and cohort for balance ability so the data were pooled (p interaction = 0.9). The combined estimate gave an odds ratio of poor balance of 0.87 (95% CI 0.75, 1.01; p = 0.08). The estimated odds ratio was only slightly attenuated when socioeconomic position, diet, health behaviors, BMI and co-morbidities were controlled for. There was no evidence of associations of IGF-II, IGFBP-2, IGFBP-3 or the IGF-I∶BP-3 molar ratio with balance ability in either cohort, or when data were pooled across cohorts (Table 4). In CaPS, a one SD increase in log transformed insulin was associated with a higher odds of poor balance (OR 1.16; 95% CI 0.99, 1.35; p = 0.059). A similar association for insulin was not observed in Boyd Orr (p interaction = 0.6; giving a pooled estimate of 1.13; 95% CI 0.98, 1.29; p = 0.08 in confounder-adjusted models), which was attenuated when BMI and co-morbidities were controlled for. Mutually controlling for IGF-I and insulin did not alter findings (results not shown).