Physical Activity Throughout Adolescence and Peak Hip Strength in Young Adults

Key Points Question Is the amount of time spent in moderate to vigorous–intensity and light-intensity physical activity throughout adolescence associated with a clinical marker of hip strength in young adult men and women? Findings In this cohort study of 2569 young people who received repeated accelerometer assessments beginning at age 12 years, more time spent in moderate to vigorous–intensity physical activity in adolescence was associated with greater hip bone mineral density at age 25 years, whereas more time spent in light-intensity physical activity was not associated with bone mineral density at age 25 years. Meaning The findings indicate that higher-intensity physical activity in early life may be important for maximizing peak adult hip strength and protecting against osteoporosis in later life.


Latent trajectory modeling
Latent trajectory models were used to derive trajectory subgroups (latent classes) for time spent in MVPA and LPA from age 12 to 25 years (assessed at mean ages 12, 14, 16 and 25 years). These models aim to classify individuals into distinct subgroups that share similar trajectories such that individuals within a group are more similar than individuals between groups (1)(2)(3)(4)(5)(6)(7)(8). Models were applied separately to males and females and to MVPA and LPA to derive trajectories for time spent in each intensity (eFigure 1). Random intercepts were used to allow variation in baseline activity (at age 12 years). Growth (i.e. change with age in time spent at MVPA and LPA) was captured by random linear and quadratic slopes. Factor loadings were fixed at the mean age at physical activity assessment.
In initial model parameterisations, the individual growth trajectories within classes were assumed to be homogenous by fixing the variance and covariance estimates for the growth factors within each class to zero. We used this model structure to identify the optimal number of classes by testing models in a stepwise fashion starting with a 1-class model (i.e. a model which assumes there are no subgroups and that all individuals follow the same trajectory over time) up to models with 6 latent trajectory classes. To aid model convergence, all the residual variances and variance-covariance matrix were fixed across latent classes. The class-specific mean trajectories from each of these models are shown in eFigure 2 and eFigure 3 for MVPA and LPA respectively in males and in eFigure 4 and eFigure 5 for MVPA and LPA respectively in females.
The models with the favoured number of subgroups (latent classes) were chosen based on a combination of theory, interpretability, meaningfulness and fit indices. Interpretability and meaningfulness were informed by the smallest class size whereby we disfavoured classes with too few participants. Fit indices included the Bayesian Information Criterion (BIC), sample-size adjusted BIC, Vuong-Lo-Mendell-Rubin test (VLMR) and entropy. Where model indices were in conflict with each other, we selected the model with the lower number of classes that was still theoretically meaningful. The characteristics of the initial models with varying number of classes are presented in eTable 1. Following this process, we selected 3class MVPA models (i.e. trajectory subgroups) and 3-class LPA models for both males and females.
After identifying the optimal number of classes, we considered alternative specifications of within-class heterogeneity. We tested models that allowed for within class heterogeneity in the intercept, followed by models allowing variation in individual linear slopes and finally models that allow within class heterogeneity in both the intercept and linear slope. The classspecific mean trajectories from each of these 3-class models with varying model structures are shown in eFigure 6 and eFigure 7 for MVPA and LPA respectively in males and in eFigure 8 and eFigure 9 for MVPA and LPA respectively in females. The characteristics of each of these 3-class models with varying model structures are presented in eTable 2.
The final models were selected based on the lowest BIC or the highest entropy where two models had similar BIC. Model selection was based on a trade-off between efficiency and validation, with the aim of summarising distinct trajectories of the data as parsimonious as possible and not just the maximisation of model fits. For MVPA, we favoured models that allowed for within class variance in the intercept and linear slope in males and models without within class heterogeneity in females (eTable 2). For LPA in both males and females, we favoured models that allowed for within class variance in the linear slope (eTable 2).
The results of the final favoured MVPA and LPA models in terms of how they relate to measured variables are shown in eTable 3 for males and eTable 4 for females. eFigure 10 shows the observed MVPA and LPA individual trajectories colour-coded by most likely class from the final favoured MVPA and LPA models. This figure helps visualise the extent to which individual variability is explained by the latent group trajectory as well as the extent of overlap between observations of individuals from different groups. The figures show that there was good separation between trajectory classes for both MVPA and LPA. To check the validity of the MVPA trajectory subgroups in females, and in particular the Low Adolescent-High Adult MVPA group, we compared how adult i.e. age 25 BMI and fat mass as well as change in these between age 16 and age 25 vary between subgroups. The results showed that, as expected, the 8% of females in the Low Adolescent-High Adult MVPA group had lower BMI and fat mass at age 25 as well as slower gains in BMI and fat mass between ages 16 and 25 than the other 2 groups. This provides further support for the validity of these subgroupings.
Latent trajectory models were estimated using the expectation-maximisation (EM)-algorithm in Mplus version 8 (Muthen & Muthen). The Mplus code used to fit the models is available at https://github.com/aelhak/PA-Hip-BMD. To investigate and avoid multiple local solutions, we specified the number of initial stage random sets of starting values to generate, and the number of final stage optimizations to use. Specifically, 3500 random sets of starting values for the initial stage and 350 final stage optimizations, along with 25 initial stage iterations were used. We fitted all the models to the analytic sample i.e. those with at least 1 MVPA and LPA measure from any one age in addition to complete data on all confounders and adult hip outcomes. To check stability of the identified groups we repeated models in the maximal sample. eFigure 11 shows that the favoured models resulted in very similar trajectories when they were re-estimated in the maximal sample sizes.
Latent trajectory models handle missing physical activity data using full information maximum likelihood estimation, which allows inclusion of all participants with at least 1 MVPA and LPA measure under the missing at random assumption. eFigure 12 shows the MVPA and LPA missing data patterns and proportions. eTable 5 shows that among those included in the latent trajectory models, the probability of missing accelerometer data was related to the confounders in the subsequent regression models, indicating the data may be consistent with missing at random. eTable 6 shows that those missing all four accelerometer assessments, who were as a result excluded from the study analysis, were socioeconomically different from the analytical sample, which might limit the generalisability of the associations of MVPA and LPA with adult hip outcomes. Participants with missing data on covariates (17.8% of those potentially eligible) were excluded, which might introduce a bias if they had systematically different adult hip bone density and geometrical measurements.

Sensitivity analysis for uncontrolled confounding
Because the observational associations between adolescent physical activity and adult hip strength may be biased by uncontrolled confounding, we used a negative outcome control study to explore the possibility that our results were biased by uncontrolled confounding (eFigure 13). A negative control study is one that includes the same participants as the 'real' study but examines the association of the 'real' exposure with a negative control outcome (or vice versa) (9,10). Under the assumptions that the negative control outcome (or exposure) has been correctly selected such that confounders (measured or unmeasured) relate similarly to it as the 'real' outcome (or exposure), and that there is no plausible causal link between the 'real' exposure and negative control outcome, a similar (non-null) association in the negative control outcome would suggest that the main (hip strength) result may be explained by uncontrolled confounding (9 10). Negative control studies have increasingly been used to explore bias due to uncontrolled confounding in observational studies (9)(10)(11)(12).
We used adult leg length from age 25 years as a negative control outcome. Leg length was measured in centimetres and was calculated by subtracting seated height (measured using a Harpenden sitting height table and anthropometer board) from standing height (measured using a Harpenden wall-mounted stadiometer). Leg length was selected as negative control outcome because it is known to be sensitive to early life environment (13)(14)(15). This is supported by evidence that legs grow rapidly between infancy and puberty with most of the growth being completed by puberty (16,17), therefore factors that influence leg growth are largely occurring in infancy and early childhood. This means that adult leg length therefore likely shares similar measured and unmeasured (including genetic) determinants as adult hip strength ( Figure 1). Because we considered it implausible that leg length would be influenced by physical activity through adolescence (intensity or impact) therefore, any association with between adolescent physical activity and adult leg length would be likely due to confounding and suggest the same may be true for our main (real i.e. hip strength) analyses (9,10).
For the study, the main (i.e. adult hip strength) analysis and the negative control analyses (i.e. adult leg length) were performed on the same sample (i.e. n=981 males and n=1588 females for the MVPA and LPA analyses, and n=478 for the vertical impacts analyses) Table 7