The primary consideration of physical activity assessment in cancer epidemiological studies is to ensure appropriate and consistent categorisation of participants according to their total physical activity level. Our findings suggest fair agreement between the EPIC questionnaire and accelerometer measurements in the ranking of physical activity level, and satisfactory long-term repeatability of the EPIC questionnaire over a 10-month interval.
Estimates of total physical activity, encompassing occupational, recreational and household activity from the EPIC questionnaire, are based on a four-level 'total physical activity index'. This index was positively correlated (ρ = 0.29) with increasing accelerometer measurements, suggesting that the index is suitable for ranking overall level of total physical activity. However, in our study, the index appeared better at distinguishing the most active participants than those in the lower activity categories. The more active participants may have clearly defined patterns of physical activity that are more easily recalled, which could partly explain the greater agreement. There were similar positive associations for both the Cambridge index (ρ = 0.32) and the occupational level classification (ρ = 0.37). For each of these three indices there was a clear distinction between the most active and the least active group. Previously, studies have used heart rate monitoring , diaries  or indirect tests of validity based on predicted energy requirements [9, 17], to assess the validity of the EPIC questionnaire. Heart rate monitoring was found to be associated with the Cambridge index but not with total activity that included lower-intensity activities . Using accelerometry as an objective measure, we showed that the EPIC questionnaire does distinguish levels of total physical activity.
However, although the EPIC instrument can suitably rank participants according to physical activity level, there remains a considerable level of measurement error when assessing an individual's physical activity. Physical activity questionnaires with similar measurement characteristics to the EPIC questionnaire have been shown to lead to substantial attenuation of relative risk estimates for associations between physical activity and an outcome of interest, assuming the measurement errors are non-differential . For example, an attenuation factor of 0.13 was estimated for the past-year total physical activity questionnaire developed by Friedenreich and colleagues, which translates to observing a relative risk of 1.10 instead of a true relative risk of 2.00 . The correlation of 0.26 between the physical activity questionnaire and accelerometer measurement in their study  was similar to our results for the EPIC questionnaire, suggesting that a similar degree of attenuation may also be present when using the EPIC physical activity questionnaire to examine associations with disease outcomes.
The correlation between non-occupational physical activity from the EPIC questionnaire and accelerometer measurement was 0.21 overall, suggesting weak to fair agreement. It may, however, underestimate the true level of agreement because the accelerometer measurements include occupational physical activity, which is not necessarily well correlated with non-occupational activity since people who are physically sedentary at work might compensate by doing more recreational activity or vice-versa. In this population, for example, the correlation between these two components from the LTPAQ questionnaire was -0.12 (95% CI -0.27, 0.02). Thus, non-occupational physical activity is likely to be a better measure of total activity among those who are not employed, which is supported by our stratified results (ρ = 0.17 for full-time workers, ρ = 0.30 for non-workers/casual-workers). Despite this probable underestimation, the correlation of EPIC with accelerometer measurements is within the range of validity coefficients that have been shown with other self-report measures of adults' habitual or global physical activity: generally 0.14 to 0.36 [24, 30, 31]. Although we had limited power to evaluate the correlations between the accelerometer and EPIC measures among different subgroups, our results suggest that the EPIC questionnaire may be more accurate at ranking non-occupational physical activity levels among participants who were male, had a lower BMI, were younger, or were not full-time workers, which is consistent with other recent research [29, 30].
The weighted kappa coefficients for the repeatability of the total physical activity index (0.62) and the Cambridge index (0.66) indicate good agreement  in classification of physical activity over a 10-month period. Wareham et al  reported slightly lower repeatability for the Cambridge index (0.60) over 18–21 months. The overall test re-test correlation of 0.65 for the continuous measure of non-occupational activity also indicates satisfactory long-term repeatability. Other questionnaires assessing past-year activity have shown similar estimates of repeatability [30, 31]. Using a longer instrument interspersed with the EPIC physical activity questions, Pols et al  observed test re-test correlations over 5–11 months ranging from 0.47 to 0.89. The repeatability of the EPIC questionnaire may be underestimated in our study because of the long (10-month) interval between the first and second administration. Differences in self-reported activity between the repeat measures may thus reflect true changes in physical activity levels during the year in addition to recall error.
Measures of total and vigorous physical activity generally have higher repeatability coefficients than light-moderate intensity activities because they are usually more easily recalled [31, 33–35]. However, we observed a higher repeatability correlation for household activity (0.73) than for recreational activity (0.58). Some previous studies have shown that light-moderate intensity household activities that are well-defined and routinely-performed, such as laundry, cooking, washing dishes or gardening, are easier to recall and have better measurement characteristics compared to more variable activities of similar intensity, such as walking [7, 14, 33, 34, 36, 37].
Long-term exposures are thought to be more important than recent exposures in the aetiology of most cancers. Our data comparing the EPIC and Friedenreich LTPAQ questionnaires suggest that recent physical activity partially reflects lifetime activity, as recalled by the participants. The correlations were significantly higher for household activity than for recreational activity (0.46 vs. 0.21 respectively, Pdiff = 0.008), which may reflect the more variable nature of recreational activities throughout life compared to household activities that are regularly performed . The slightly higher correlation (0.34 vs. 0.26) that was observed when we used the average of the two EPIC administrations (baseline and follow-up) suggests that repeat administration of the EPIC questionnaire would reduce intra-individual variation in physical activity . Lack of strong agreement between the EPIC and LTPAQ questionnaires may reflect true differences between past-year and lifetime activity, in addition to different modes of administration. The LTPAQ has been shown to have high repeatability  but it is also a self-reported measure and may have similar measurement errors as those of the EPIC questionnaire.
Participants in our study are comparable to the EPIC cohort with regards to age, employment status and BMI [6, 38, 39], but the overall level of self-reported non-occupational activity was slightly higher in our study population . Other factors, such as environmental and cultural differences between Europe and Australia, may affect the generalisability of our results to the EPIC cohort.
Our study has several strengths, including a large sample size, a fairly representative population, a high retention rate during follow-up, and the use of an objective validation measure that overcomes many of the inherent limitations in self-report methods . Accelerometry is a valid and widely-used measure of total physical activity in adults [20, 40], and unlike heart-rate monitoring, is able to detect low-moderate intensity activities. We used the Actigraph accelerometer, which has been shown to have little variability across individual units, and high overall reliability . Three 7-day accelerometer monitoring periods were used during the 10-month study period, to ensure an accurate assessment of usual physical activity during the reference period, and to capture seasonal variations in physical activity. When comparing total MET-hours/week obtained from the first, second and third weeks of accelerometer measurement, the correlations were all in the range of 0.72–0.74, suggesting relatively little intra-individual variability in physical activity levels during the study period. Previous research has demonstrated that three to five days of accelerometer monitoring is sufficient to estimate habitual physical activity in adults reliably .
However, accelerometers are not a perfect gold standard measure of physical activity. Accelerometers alone cannot provide contextual information about the type or purpose of specific activities (e.g. work versus recreational activity), and they are limited in their ability to monitor upper body movements, water activities and movements with a weak vertical component such as cycling [40, 42, 43]. They may also underestimate some household-based activities involving upper body movements . The choice of prediction equation and cut-points to categorise accelerometer time in different intensity categories may also influence results, although there is no optimal equation . We chose the Swartz method  because it was derived using a broad age-group and range of field activities that best reflected our study population. We also focused on 'total' activity rather than intensity-specific activity.
For future use in epidemiological studies, some minor changes could be incorporated into the EPIC questionnaire that may improve its measurement characteristics and distinguish better between people who are sedentary or moderately inactive. Suggested improvements include i) capturing the frequency and duration of occupational activity, ii) changing the question on vigorous activity to mention 'breathing much harder than normal' rather than focus on 'sweating' (which can be weather dependent), and iii) splitting the housework activities into two or more categories (e.g. active childcare, cooking, cleaning), to assist with recall and to allow more precise estimation of intensity levels.