Participants wore heart rate (HR) monitors (Polar Electro Ltd, Kemple, Finland) continuously during the waking hours over 4 days following their visit to the study centre. Resting energy expenditure (REE) was calculated by the Weir formula 31 using VO₂ and VCO₂ measurements obtained in the fasting state, after approximately 10 minutes of supine rest using the same indirect calorimetry system as previously described 26. Physical activity energy expenditure (PAEE) was measured using the flex HR method, which has been described in detail elsewhere 3233. Below the flex HR point, energy expenditure (EE) was assumed to be equivalent to REE. EE above the flex point was predicted from the individual HR-EE regression line. PAEE was calculated by subtracting REE from the estimated average daily energy expenditure, and thereafter averaged over the 4-day period.

Participants attended the study centre after an overnight fast and a sample of venous blood was taken. Total cholesterol was measured using standard enzymatic methods, as previously described 26. Glycosylated haemoglobin (HbA_1c) was measured on fresh EDTA blood samples using high-performance liquid chromatography. Weight was measured on standard calibrated scales and height was measured using a rigid stadiometer. BMI was calculated as weight (kg) divided by height squared (m²). Systolic and diastolic blood pressure were measured in a seated position using an automated Accutorr sphygmomanometer (Accutorr, Cambridge, UK) three times at one-minute intervals; the mean of these measurements was used in the analysis 26. Aerobic fitness (VO_2max) was predicted as oxygen uptake at estimated maximal HR (220 minus age) by extrapolation of the regression line established during the individual calibration for the relationship between oxygen consumption and HR during a submaximal graded treadmill exercise test.

Self-report measures of well-being and quality of life were collected, including the SF-36 health survey 34, the short form of Spielberger's State Anxiety Inventory (STAI) 35 and a measure of current health status (EuroQol). In brief, the SF-36 is a health survey of 36 items which can be used to generate psychometrically-based physical and mental health summary measures. The STAI consists of six, 4-point Likert scales, and has been shown to correlate well (r > 0.9) with the results of the longer questionnaire from which it was derived 35, which is in turn associated with indicators of anxiety 36. The STAI used in this study measures state (current state of anxiety or mood) rather than trait (recurring or individual characteristics of anxiety or mood) anxiety 35. The EuroQol is a health "thermometer" which asks responders to mark on a scale of 0 to 100 their current health status, from "Worst imaginable health state (0)" to "Best imaginable health state (100)". We also collected data on cognitive predictors of physical activity, based on the Theory of Planned Behaviour, that were hypothesized to be directly associated with behaviour change 29. Items were developed according to Theory of Planned Behaviour guidelines 37 and measured on a Likert-type scale ranging from 1 (strongly disagree) to 5 (strongly agree) 38, and averaged to calculate overall scores. The questionnaire included two items to assess perceived behavioural control ('it would be difficult for me to be more physically active in the next 12 months even if I wanted to' (score reversed), 'I am confident that I could be more physically active in the next 12 months, if I wanted to') and intention ('I intend to be more physically active in the next 12 months', 'it is likely that I will be more physically active in the next 12 months'). Cronbach's alpha for internal consistency was 0.54 for perceived behavioural control, and 0.77 for behavioural intention at baseline.

Past activity behaviour was measured using the EPIC-Norfolk physical activity questionnaire (EPAQ-2), which has previously been validated using objective heart rate monitoring 39. Participants answered questions about smoking status and weekly alcohol consumption in a general health questionnaire.

Descriptive summary statistics were calculated separately for men and women using means and SDs at baseline and follow-up. T-tests were used to examine whether there were any differences in baseline characteristics between those with and without follow-up data. We used linear regression to model baseline demographic, clinical, behavioural and psychosocial variables separately against PAEE at follow-up. These models were adjusted for baseline PAEE to describe change in this variable over time. We also ran a multivariable regression model to examine PAEE change, mutually adjusting for all significant predictor variables (p < 0.05), to establish which variables were independently associated with the outcome. The residuals of all linear regression models were checked to ensure they were normally distributed. Models were also run separately by trial arm and by sex. As results were similar in the three trial arms and by sex (data not shown), the data were pooled and results from linear regression models conducted in the whole cohort are presented. Type I error was set at 0.05 for all tests. All data were analyzed in continuous form using Stata Version 10.0. (STATA Corp., College Station, Texas, USA).

In general, our results do not support the limited literature showing associations between certain socio-demographic, psychosocial and behavioural variables, and change in physical activity behaviour 1442. Younger age is usually associated with increased physical activity, both cross-sectionally and over time 20. We did not show a significant relationship with age, though the range was quite restricted (30-50 years). Furthermore, most previous results come from observational studies rather than intervention studies, such as the ProActive trial, in which all participants were encouraged to increase their activity levels. Prospective studies that have examined the relationship between baseline predictors and level of physical activity at follow-up have demonstrated associations with self-efficacy 1441434445; perceived behavioural control 46; past behaviour 43; perceived benefits or views of exercise 4547; and the participant's belief that physical activity was important for their health 13. However, the questions derived from the Theory of Planned Behaviour to characterize perceived behavioural control and intention about becoming more physically active over the next year did not predict change in physical activity in this cohort. These findings suggest that such measures may not be useful in identifying subgroups of participants who are more or less likely to make long-term changes in their activity levels. The positive association between fitness and change in PAEE is a novel finding. As fitness is partially genetically determined 48, for two people starting at the same physical activity level but with different fitness levels, it may be easier for the individual with a higher fitness level to increase their physical activity.

One of the challenges in comparing our results to those found in other studies is the difference in methods used to measure physical activity. Previous studies usually measured current or future self-reported behaviour, and frequently focused on sporting activities 15, while we examined change in everyday objective physical activity behaviour. When physical activity behaviour and physical activity determinants are both measured using the same method e.g. by self-report questionnaire, associations may be due, at least in part, to commonality in response patterns to these measures ("common method variance") 49. Consequently, when physical activity is measured by a different method (e.g. heart rate monitoring), part of the correlation explained by common method variance disappears, producing lower, non-significant or non-existent associations. Indeed, studies using objectively measured behaviour are less likely to observe significant associations with predictors, such as those based on the theory of planned behaviour, than those using self-reported behaviour 22. However, when we examined the relationship between the same predictors and change in self-reported physical activity, we found only a few significant associations (for male sex, and higher fitness and BMI at baseline), indicating that common method variance cannot fully explain the lack of association with objectively measured physical activity. A number of other reasons for the lack of association between baseline variables and physical activity change might also be considered.

First, previous work has focused on selected populations with a short follow-up 15. We examined a sample of middle-aged, slightly overweight individuals, identified through primary care registers via a first degree family history of diabetes. This group represents a sub-group of the population who would benefit from increased overall physical activity. The longer than usual follow-up period in the ProActive trial (one year) may account for some of the differences found between this and other studies. Second, although our sample size was larger than many other physical activity promotion trials, there may have been a lack of heterogeneity in the cohort. This would have reduced our power to detect any associations with physical activity. However, the large standard deviations and ranges for change in PAEE indicate that this is unlikely to be the case.

Third, we did not measure some of the variables that have been shown to be associated with physical activity behaviour, including environmental and social factors, and may therefore have focused on the weaker determinants of objectively measured physical activity behaviour. King et al, for example, found that neighborhood environment is a potential predictor of physical activity maintenance over two years, though a self-report measure of physical activity was used 14. Wider collective determinants of physical activity may act independently or interact with psychosocial correlates to influence physical activity behaviour, though these relationships have not been clearly elucidated 5051. Furthermore, some of the associations may have been weaker in our study because of the low reliability of the measures, e.g. Cronbach's alpha was only 0.54 for perceived behavioural control.

Fourth, since all participants were at risk due to their family history of disease, we cannot examine whether this factor was an important predictor of change. Similarly, all participants were subject to trial recruitment procedures and objective physical activity measurement (albeit without feedback), and the effects of this may have diluted associations reported in other studies. However, individuals in observational studies have also been recruited and measured, and these processes have been linked to behaviour change 52. Objective measurement of physical activity in the ProActive cohort may have increased the saliency of this behaviour and prompted self-monitoring, an effective behaviour change technique in its own right.

Fifth, the precision in the methods we used to assess PAEE may also have contributed to the lack of observed associations. PAEE from individually calibrated heart rate monitoring is an integrated measure of energy expenditure above rest calculated from free-living heart rate data. In sedentary populations, much of the daytime is spent in the region around the flex heart rate, which is used to discriminate between rest and physical activity. The association between heart rate and energy expenditure is less precise in this region, which may reduce the accuracy of predicted PAEE on an individual level 53. Finally, we conducted multiple significance tests (> 20) between baseline predictors and change in PAEE which may have led to an increased risk of Type 1 errors. It is unclear whether the few significant observations represent real or chance associations. The exploratory and post-hoc nature of our analyses mean that results should be interpreted with caution.

However, the ProActive trial allowed objective measurement of physical activity over a period of 12 months in a well-defined and accessible group of individuals who were at increased risk of diabetes and who would benefit from increases in physical activity. There was a high follow-up rate (88%). Objective measurement of physical activity reduces the error and bias commonly associated with self-report measures, and PAEE has been extensively validated in the laboratory and during free-living conditions 5354. Furthermore, we measured a wide range of potential predictors from a variety of domains, including clinical, behavioural, demographic and psychosocial variables.

Systematic application of methods aimed at increasing our understanding of how predictors from different domains influence physical activity change over time remains scarce 42. Yet, the potential benefit of examining these variables for physical activity, diet and weight control is well recognized in the behavioural and social sciences literature 55. Identifying predictors of change in physical activity, regardless of whether the predictors are causally related to change, could help the targeting of appropriate sub-groups for intervention and improve efficacy. We could target those who were likely to change with brief interventions, such as action planning or goal setting in combination with self-monitoring, or target those who were unlikely to change with more intensive interventions 56. Results from this analysis suggest that trying to predict who will change their physical activity behaviour is not straightforward. Ideally we would target people on the basis of both their risk of future health problems and the likelihood that they would change in response to intervention, but further research is needed to improve our understanding of determinants of change in physical activity behaviour. In addition, the lack of associations between traditional determinants of self-reported physical activity and behaviour change in this cohort suggests these factors may not be the most appropriate to inform intervention development.