The effect of long-term adherence to physical activity recommendations in midlife on plasma proteins associated with frailty in the Atherosclerosis Risk in Communities (ARIC) study

The target trial emulation framework entails specifying a high-level protocol (e.g., eligibility, intervention strategies, follow-up, outcome assessment, and statistical analysis) of a hypothetical randomized trial, i.e., a target trial, that answers the research question of interest before describing how the observed data will be used to mimic the target trial [32]. This framework guides the estimation of the causal effect using the gold standard for estimating the effect of an intervention strategy and allows for transparent evaluation of causal estimates.

The target trial protocol for this paper is summarized in Table 2. Briefly, to assess the effect of long-term adherence to the recommended ≥ 150 min/week of MVPA on frailty-associated plasma proteins, the target trial randomizes participants who are capable of engaging in MVPA to either: (i) receive the intervention strategy that ensures the achievement and maintenance of the recommended MVPA level throughout the follow-up, or (ii) a “natural course” strategy that allows participants to engage in a habitual MVPA level, which may or may not be below the recommended level. Participants assigned to either strategy are free to engage in different levels of MVPA as long as they meet the strategy’s MVPA threshold (≥ 150 min/week for the intervention strategy and ≥ 0 min/week for “the natural course” strategy). Such strategies are an example of a representative intervention [33, 34].

The effect of long-term adherence to the recommendation is a per-protocol effect [35] as it compares outcomes between individuals adhering to two strategies, intervention versus “natural course”, throughout follow-up. Under the “natural course” strategy, regardless of what level of MVPA these participants engage in during follow-up, they are always adherent to this strategy because no constraints are placed on their MVPA level. Under the intervention strategy, participants are no longer adherent to the assigned strategy when they engage in < 150 min/week of MVPA at any time during follow-up. Because post-randomization factors can influence whether participants adhere to the assigned strategy or remain in the trial, appropriate statistical methods such as the parametric g-formula and IPW are needed to account for these factors even in a trial with non-adherence [35, 36].

We emulated the target trial described above using 14,898 participants at Visit 1 (1987–1989) of the ARIC study who were followed until Visit 3 (1993–1995, Fig. 1A). The ARIC study is an ongoing community-based cohort study that enrolled middle-aged participants (aged 45–64 years) from 4 communities across the United States: Washington County, MD; Forsyth County, NC; the northwestern suburbs of Minneapolis, MN; and Jackson, MS [37].

Bild vergrößern

The emulated trial closely follows the protocol of the target trial with three exceptions (Table 2). The deviations (assumption that no participants had limitations in engaging in PA, exclusion of races with a small number of participants in the sample, and approximation of long-term adherence to PA by measurements at two study visits only) and the rationales are described in Supplemental Methods. To emulate the randomization at baseline, based on the assumed causal structure of the observed data depicted in Fig. 2, we assumed that whether a participant met the PA guidelines at baseline was as if randomized after conditioning on time-fixed confounders (i.e., age, sex, race, and education) and the baseline values of time-varying confounders (i.e., smoking status, number of chronic conditions, body mass index [BMI], glomerular filtration rate [eGFR], and total cholesterol). At follow-up, as we would in the target trial, we censored participants who did not return to study visits, had missing confounders and protein measurements, or had missing physical activity measurements at Visit 3. We assumed that adherence to the assigned strategy and censoring were both as if randomized, conditional on the baseline time-fixed and time-varying confounders, the post-baseline values of the time-varying confounders (at the current and additionally, for continuous confounders, the previous two visits), baseline MVPA, and protein levels at the last visit (if available; proteins were not measured at Visit 1). We used the ICE approach [38] of the parametric g-formula method and the IPW method [34] (described below) to appropriately emulate the randomization at baseline and properly handle the confounders associated with non-adherence to their assigned strategy or censoring. These two methods also emulate the representative intervention. Under the specified intervention strategy, both methods implicitly assign participants an MVPA level as a random draw from the observed MVPA distribution among all participants who have observed MVPA ≥ 150 min/week and the same confounder history [33, 34, 38]. This implicit assignment was performed at each visit.

Bild vergrößern

PA was measured in ARIC using the Modified Baecke Physical Activity Questionnaire, which has been validated against the doubly labeled water, the gold standard for measuring physical activity energy expenditure [39]. Participants were asked to report up to 4 sports or exercises in which they most frequently engaged at Visit 1 and Visit 3. For each reported sport or exercise, the participants reported the duration (hours/week) and the frequency (number of months/year) [40]. The sport and exercise questions have shown strong correlations with vigorous physical activity recorded by a 48-h PA diary [41]. To assess whether a participant met the PA guidelines, the intensity of each reported sport or exercise was assigned a metabolic equivalent of task (MET) based on the 2011 Compendium of Physical Activities [42]. Sports or exercises with MET > 3 but ≤ 6 were considered moderate, and those with MET > 6 were considered vigorous. For each sport or exercise that met moderate or vigorous intensity levels, the weekly level (minutes/week) was estimated by multiplying the frequency and duration by 4.35 weeks/month and dividing by 52.18 weeks/year. A participant’s MVPA in minutes/week was then calculated as the sum of the weekly levels of all moderate-intensity sports or exercises and two times the weekly levels of all vigorous-intensity sports or exercises [40]. Participants not reporting any sport or exercise meeting the MVPA intensity were assigned zero for MVPA minutes/week. Participants with MVPA minutes/week ≥ 150 were considered to meet the PA guidelines.

The 45 proteins used in this analysis were selected based on evidence from our previous study on the midlife proteomics of frailty in later life [28]. Frailty was defined by the 5 criteria of physical frailty phenotype: (a) weight loss defined by percent weight decrease and current BMI < 18.5 kg/m; (b) weakness by grip strength below sex and BMI-specific cut-points; (c) slowness defined by usual gait speed below sex- and height-specific cut-points; (d) exhaustion defined by self-report; and (e) low PA was ascertained as ranking in the lowest quintile of the Modified Baecke Physical Activity Questionnaire sports and exercise index (see [27] and [28] for more details).

The proteins were measured using the SomaScan platform (Version 4.0, SomaLogic, Inc., Boulder, Colorado). The SomaScan platform quantifies the relative abundances of the plasma proteins and protein complexes using single strands of DNA with chemically modified nucleotides, called modified aptamers or “SOMAmers”, which act as protein-binding reagents with defined three-dimensional structures and unique nucleotide sequences. The abundances of the SOMAmers are quantified using dynamic DNA detection technology and represent the levels of the proteins in plasma [43]. The 45 proteins were significantly differentially expressed in midlife (Visit 3, 1993–1995) among participants who were frail (meeting at least 3 of the 5 criteria) in late-life (Visits 5–7, 2011–2019) compared to participants who were robust (meeting none of the 5 criteria) in late-life after Bonferroni correction and adjustment for age, sex, race-center, education, family income, drinking and smoking status, dietary protein intake, total cholesterol, eGFR, history of hypertension, diabetes, coronary heart disease, heart failure, cancer, and chronic lung disease, and functional limitation (Fig. 3) [28].

Bild vergrößern

The ICE and IPW methods can estimate the per-protocol effect of a representative intervention under the assumptions of exchangeability (or no unmeasured confounding), positivity, consistency, and no model misspecification. (see Supplemental Methods for discussions on these assumptions) [33, 34, 38]. These two methods condition, at each time point, on the time-varying confounders up to that time point, properly handling the “exposure-confounder feedback”, i.e., a feedback loop of MVPA → confounder → MVPA. An overview of these methods has been described [44].

We selected confounders that, based on our knowledge, have associations with both moderate-to-vigorous physical activity and plasma protein levels independent of other confounders. We used the following functional forms for the confounders (see Supplemental Methods for the measurements of confounders) in the ICE and IPW models: age (continuous with a linear term), sex (men/women), race (Black/White participant), education (nominal categorical, less than completed high school, high school/GED/vocational school, or any college), smoking status (ever or never smoker), number of chronic conditions (nominal categorical, 0, 1, or ≥ 2, including namely, hypertension, diabetes, coronary heart disease, heart failure, stroke, cancer, and chronic lung disease), BMI (continuous with a restricted cubic spline of 3 internal knots at quartiles), eGFR (continuous with a restricted cubic spline of 2 internal knots at tertiles), and total cholesterol (continuous with a linear term). MVPA minutes/week were modeled linearly in both methods. Protein level at the last visit was modeled using a restricted cubic spline with 3 internal knots at quartiles. The functional forms were selected based on residual plots and likelihood ratio tests (data not shown). Individual chronic conditions were not included as confounders due to positivity violations. Time-fixed confounders (i.e., age, sex, race, and education) were measured at Visit 1. All other confounders were treated as time-varying, and measures at all three visits were used in the analysis.

We estimated the 95% confidence intervals using 1000 samples from the non-parametric cluster bootstrap [45], where individuals were sampled with replacement and all of their follow-up observations were retained. For the IPW method, to ensure satisfactory control of confounding by the confounders including MVPA minutes/week and protein level from the last visit, we checked the standardized mean differences (SMD) of these variables (1) between participants who achieved ≥ 150 min/week of MVPA and those who did not, respectively at Visit 1 and Visit 3 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring) and (2) between participants who were censored and those who remained under study, respectively at Visit 1 and Visit 2 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring). The SMDs were calculated before and after applying the final weights at respective visits [46, 47]. We used SMD within ± 0.1 as good confounder balance and within ± 0.25 as acceptable confounder balance [48, 49]. We also checked the distribution of the final weights at Visit 3 as recommended in previous literature [50]. All protein levels used in this analysis were log 2-transformed and standardized using their standard deviation (SD) at Visit 3. The detailed procedures of the two methods, including the g-formula, the identifying assumptions, the model specifications, and the confounder measurements, are described in Supplemental Methods.

To check that residual confounding and selection bias by individual chronic conditions, drinking status, family income, and histories of binary and categorical confounders (all excluded from the confounder list due to positivity violation) were reasonably controlled by the included confounders, we checked the SMDs across levels of adherence to the recommended MVPA and levels of censoring as described earlier.

Next, we performed the same ICE and IPW analyses after excluding participants who had major chronic conditions (i.e., coronary heart disease, heart failure, stroke, cancer, chronic lung disease, eGFR < 30 ml/min/1.73 m²) or underweight (BMI < 18.5 kg/m²) at baseline (n = 2951, 19.8%, Figure S1). By restricting the study sample to a healthier subset of individuals, the variability of unmeasured confounders may be reduced and therefore better controlled.

Finally, we performed a tipping analysis [51] (Supplemental Methods) at Visit 1 to assess how strongly an unmeasured confounder at or before Visit 1 must be associated with MVPA at Visit 1 (i.e., the exposure) and the protein levels at Visit 3 (i.e., the outcomes) to bring the effects of exposure on the outcomes to the null (i.e., the tipping point). This tipping analysis was not a direct assessment of the robustness to unmeasured confounding of the effects of interest in this paper. However, robustness to unmeasured confounding suggested by the tipping analysis may indicate the robustness of our effects of interest. To provide a context of plausible unmeasured confounders that would bring the effects to the null (e.g., what they could be and how strongly they are associated with the exposure and the outcome), we used the confounding of Visit 2 BMI, eGFR, and protein level on the effects of MVPA at Visit 3 on protein levels at Visit 3 to compare with the tipping points for the effects of MVPA at Visit 1 on protein levels at Visit 3. We discuss the rationale and the detailed procedures for the tipping analysis in the Supplemental Methods.

To address the unobserved outcome due to death, we performed another sensitivity analysis using the same ICE and IPW procedures after excluding all observations from participants who died before Visit 3 (n = 794, 5.3%, Figure S2). Death before Visit 3 was defined by not attending Visit 3 and having a date of death on or prior to February 5th, 1996, the last observed Visit 3 date. This approach can result in selection bias [52], but similar estimates to the main analysis can suggest that our findings were robust to death-related selection bias.

The 95% confidence intervals in all sensitivity analyses were obtained using the same non-parametric cluster bootstrap as in the main analysis. All analyses were performed in R (version 4.3.1) [53].

The baseline (Visit 1) characteristics of the 14,898 participants are summarized in Table 3 by whether they met PA guidelines (≥ 150 min/week of MVPA) at baseline. Participants who met the guidelines had similar baseline age, cholesterol, and eGFR, but lower BMI, compared to those who did not meet the guidelines. Women, participants self-identified as Black, and participants who had less than a completed high school education were less likely to meet PA guidelines. The prevalence of coronary heart disease, heart failure, stroke, cancer, and lung disease was low (< 10%). Participants with more chronic conditions were less likely to meet PA guidelines.

The transitions between meeting PA guidelines and censoring are summarized in Fig. 1B. About one-third of the participants met PA guidelines at Visit 1. Of the participants who met PA guidelines at Visit 1, 42.1% maintained ≥ 150 min/week of MVPA at Visit 3, compared to 23.8% who decreased to < 150 min/week at Visit 3. The remaining 34.1% of the participants were censored. Of the participants who did not meet the guidelines at Visit 1, 43.8% maintained < 150 min/week of MVPA at Visit 1, compared to 16.7% who increased to ≥ 150 min/week at Visit 3. The remaining 39.5% of the participants were censored.

Because the weight estimation was based on a different model for each outcome protein that included only the history of the outcome protein, not other proteins, this resulted in different weights for different outcome proteins. Therefore, the confounder balance was examined separately for each outcome protein. Figures 4 and 5 depict the balance before and after weighting for CACNA2D3 and HTRA1 as the outcomes. The balance for all other proteins can be found in Figs. S3 and S4. For all the 45 protein outcomes, after applying the IPW final weights, good balance (SMD within ± 0.1 SD) was achieved for all included confounders, MVPA minutes/week, and protein levels at all available visits (i) between participants who achieved ≥ 150 min/week of MVPA and those who did not, both at Visit 1 and Visit 3 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring) and (ii) between participants who were censored and those who remained under study, both at Visit 1 and at Visit 2 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring). Moreover, the final weights at Visit 3 had a mean close to 1 for all protein outcomes (Table S1).

Bild vergrößern

Bild vergrößern

For both the ICE and IPW methods, we found small differences in the population-level average for the 45 frailty-associated proteins comparing all participants achieving and maintaining ≥ 150 min/week of MVPA at Visits 1 and 3 to the “natural course” strategy (ranged from 0.04 to 0.11 SD, Fig. 6 and Table S2). The largest increase in the population-level average was observed with voltage-dependent calcium channel subunit alpha-2/delta-3 (CACNA2D3; difference_ICE = 0.10 SD, 95% CI: 0.05, 0.14; difference_IPW = 0.11 SD, 95% CI: 0.05, 0.15). The largest decrease in the population-level average was observed with high-temperature requirement serine protease A1 (HTRA1; difference_ICE =  − 0.09 SD, 95% CI: − 0.13, − 0.05; difference_IPW =  − 0.08, 95% CI: − 0.12, − 0.04). Other proteins with larger increased population-level averages included neurogenic locus notch homolog protein 1 (NOTCH1), neural cell adhesion molecule 1 (NCAM1), and contactin-1 (CNTN1). Other proteins with larger decreased population-level averages included C-reactive protein (CRP), transmembrane protein 132D (TMEM132D), and vesicle-fusing ATPase (NSF). All these proteins had 95% CIs not overlapping with zero.

Bild vergrößern

The effect estimates from the ICE and IPW methods mostly agree with wider CIs from the IPW method. A few proteins with smaller effect sizes had larger discrepancies between the methods but the CIs were all consistent with null effects (e.g., leptin [LEP, difference_IPW =  − 0.001, 95% CI: − 0.05, 0.05; difference_ICE =  − 0.02, 95% CI: − 0.06, 0.01]; insulin [INS; difference_IPW = 0.02, 95% CI: − 0.05, 0.12; difference_ICE =  − 0.02, 95% CI: − 0.06, 0.04); and cystatin-SA [CST2; difference_IPW =  − 0.02, 95% CI: − 0.14, 0.06; difference_ICE = 0.02, 95% CI: − 0.03, 0.06]).

The directions of the effects of PA on the frailty-associated proteins were mostly consistent with the known directions of associations between the proteins and frailty and the known benefits of PA on frailty. For example, a higher level of CACNA2D3 was associated with a lower risk of frailty in the previous study (i.e., protective, Fig. 6), and meeting the PA guidelines at both Visits 1 and 3 increased the population level of CACNA2D3. The two exceptions were prostate-associated microseminoprotein (MSMP) and interleukin-1 receptor antagonist protein (IL1RN). However, both had 95% CI overlapping with zero, consistent with null effects.

The balance of individual chronic conditions, drinking status, family income, and lagged values of the binary and categorical confounders was improved after weighting, even though these variables were not included in the IPW models (Figs. 7 and 8, S5, and S6). For all protein outcomes, after applying the IPW final weights, SMDs of most confounders were within ± 0.1 SD (i) comparing participants who achieved ≥ 150 min/week of MVPA and those who did not, both at Visit 1 and at Visit 3 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring) and (ii) comparing participants who were censored to those who remained under study, both at Visit 1 and at Visit 2 (conditional on MVPA ≥ 150 min/week at Visit 1 and no prior censoring). The exceptions were drinking status and family income for (i) and family income, stroke, and lung disease for (ii), but all had SMS within ± 0.2 SD.

Bild vergrößern

Bild vergrößern

Excluding participants with major chronic conditions at Visit 1 produced mostly consistent results with the main results using either method (Fig. 9). The effects for some proteins were slightly attenuated, e.g., CACNA2D3 and HTRA1. However, the effects for other proteins were strengthened using the IPW method, e.g., insulin-like growth factor-binding protein 1 (IGFBP1), amyloid-like protein 1 (APLP1), transmembrane protein 132B (TMEM132B), and delta and Notch-like epidermal growth factor-related receptor (DNER). CST2 and INS now had more similar effect estimates and the same effect directions between the ICE and IPW methods. For the IPW method, good balance (SMDs within ± 0.1 SD) was achieved for all the included confounders for all protein outcomes (Figs. S7 and S8 for CACNA2D3 and HTRA1; plots for other proteins not shown). Most excluded confounders achieved good balance (SMDs within ± 0.1 SD) in the sensitivity analysis sample. Other excluded confounders (drinking status, family income, lung disease, stroke, coronary heart disease, and cancer) achieved satisfactory balance (SMDs within ± 0.25 SD, Figs. S9 and S10 for CACNA2D3 and HTRA1; plots for other proteins not shown).

Bild vergrößern

The tipping points for the effect of MVPA minutes/week at Visit 1 on the levels of CNCNA2D3 and HTRA1 at Visit 3 are presented in Fig. 10. An unmeasured confounder, with similar confounding strength as Visit 2 measures of BMI or eGFR on Visit 3 MVPA and Visit 3 CNCNA2D3/HTRA1, would not bring the effect of Visit 1 MVPA on these two proteins at Visit 3 to the null. However, an unmeasured confounder, with similar confounding strength as a Visit 2 measure of CNCNA2D3 on Visit 3 MVPA and Visit 3 CNCNA2D3 level, would bring the effect of Visit 1 MVPA on Visit 3 CNCNA2D3 to the null. An unmeasured confounder, with similar confounding strength as a Visit 2 measure of HTRA1 on Visit 3 MVPA and Visit 3 HTRA1 level, would bring the effect of MVPA minutes/week at Visit 1 on HTRA1 at Visit 3 to the null. Similar patterns were observed for all other proteins (Fig. S11).

Bild vergrößern

The effect estimates after excluding participants who died before Visit 3 were close to the main estimates with small attenuations for some proteins, e.g., CACNA3D3 and CRP (Fig. 11). For the IPW method, good balance (SMDs within ± 0.1 SD) was achieved for the included confounders for all protein outcomes. Good balance (SMDs within ± 0.1 SD) was also achieved for all excluded individual chronic conditions except those for coronary heart disease. Coronary heart disease, drinking status, and family income had SMDs within ± 0.2 SD (Figs. S12–S15 for CACNA2D3 and HTRA1; plots for other proteins not shown).

Bild vergrößern