Association between usual protein intake and muscle function in older U.S. adults: a target-trial emulation using NHANES 2011–2018

Abstract

Background

The link between long-term protein intake and muscle performance in older adults has been hard to define, partly because most studies rely on short dietary windows and are vulnerable to confounding and measurement noise. In this work, we attempted to estimate the usual protein intake and functional limitation among U.S. adults aged ≥ 60 years using a target-trial emulation framework with overlap weighting and semiparametric estimators.

Methods

Data were drawn from four NHANES survey cycles (2011–2018), including 5,736 adults aged ≥ 60 years with complete exposure, outcome, and covariate data. Usual protein intake (g/kg/day) was derived from available 24-hour recalls to approximate habitual intake. The primary outcome was PFQ-defined mobility limitation across cycles; grip strength (2011–2014) was analyzed separately as a secondary outcome. Causal contrasts across predefined intake categories (<0.8, 0.8– < 1.0, 1.0– < 1.2, ≥1.2 g/kg/day) were evaluated using covariate-balancing propensity score overlap weighting (ATO estimand) followed by marginal structural models. Doubly robust sensitivity analyses were conducted using augmented inverse probability weighting and targeted maximum likelihood estimation with generalized linear models. Simulation extrapolation (SIMEX) was applied to assess potential bias from dietary measurement error. Exploratory analyses evaluated hs-CRP as a potential mediator and tested effect modification by vitamin D status and physical activity.

Results

Mean usual protein intake was 0.93 g/kg/day, and approximately 42% of participants consumed at least 0.8 g/kg/day, the current Recommended Dietary Allowance (RDA) for the general adult population. In the prespecified overlap-weighted marginal structural model (ATO estimand), higher intake was associated with lower odds of mobility limitation, although the primary contrast comparing ≥ 1.2 versus < 0.8 g/kg/day was modest and not statistically significant (OR 0.89, 95% CI 0.54–1.47). A doubly robust binary contrast yielded a −6.6 percentage-point difference in predicted limitation (95% CI −25.8 to 12.7), consistent in direction but imprecise. In cycle-specific analyses, the inverse association was more pronounced in 2015–2018 (OR 0.80, 95% CI 0.65–0.98). Spline models suggested a steeper decline in predicted limitation below approximately 1.0–1.1 g/kg/day, with a flatter trajectory at higher intakes. Exploratory mediation models indicated a potential indirect component through hs-CRP, though these estimates were not overlap-weighted and should be interpreted cautiously.

Conclusions

Higher usual protein intake was directionally associated with lower odds of mobility limitation among older U.S. adults within a target trial emulation framework, although the primary overlap-weighted estimates were modest and imprecise. Evidence of nonlinearity suggests that intakes near 1.0–1.1 g/kg/day may mark a range where predicted limitation declines more steeply, but uncertainty increases at higher intake levels. Given the cross-sectional design and residual potential for confounding, these findings should be interpreted cautiously. Prospective studies are needed to determine whether sustained protein intake in this range meaningfully preserves functional capacity over time.

1. Introduction

Loss of muscle strength and mobility are central features of sarcopenia, a progressive age-related condition characterised by reduced muscle mass, diminished strength, and impaired physical performance. Sarcopenia is now recognised as a major contributor to frailty, disability, hospitalisation, and mortality in older adults [1,2]. Adequate dietary protein is essential for maintaining skeletal-muscle homoeostasis through stimulation of muscle protein synthesis (MPS) and repair [3,4]. Although 0.8 g/kg/day remains the RDA for adults in general, several expert groups have advocated higher targets for older individuals, often in the range of 1.0–1.2 g/kg/day, to support muscle maintenance. These recommendations reflect physiologic changes with aging, including reduced anabolic sensitivity to protein and greater vulnerability to muscle loss. Sex-specific differences may also be relevant. Men generally have higher absolute muscle mass, whereas women experience more rapid relative declines after midlife; hormonal differences, body composition, and dietary patterns may influence protein requirements and responsiveness. Prior observational analyses have reported potential sex-specific associations between protein intake and grip strength, although findings have been inconsistent [5,6]. Yet, despite decades of research, the population-level relationship between habitual protein intake and muscle function remains uncertain.

Previous cross-sectional analyses of NHANES 2011–2018 and other datasets have reported weak or null associations, once age, adiposity, and total energy are controlled [5,7]. These discrepancies may reflect exposure misclassification and residual confounding rather than a genuine lack of effect [8,9].

Two methodological challenges complicate inference. First, single-day or short-term dietary recalls may not adequately capture long term intake, and within-person variation biases associations toward the null [10]. Second, conventional regression models may not fully address confounding behavioural and clinical factors that jointly affect diet and muscle function such as inflammation, comorbid disease, and physical activity.

Recent advances in causal inference offer tools to improve analytic transparency in observational research. Target-trial emulation combined with marginal structural models (MSMs) and targeted maximum likelihood estimation (TMLE) allows estimation of defined causal contrasts under explicit assumptions, even in complex observational data [11,12]. In parallel, usual intake modelling, approaches attempt to better approximate habitual intake by accounting for within-person variability. Accordingly, we drew on four NHANES cycles (2011–2018) to emulate a target trial design to evaluate the association of usual protein intake on muscle function among U.S. adults aged 60 years and older. Usual protein intake (g/kg/day) was derived from available 24-hour recalls approximating habitual intake. We applied overlap-weighted MSMs and complementary semiparametric estimators to improve robustness. We also examined systemic inflammation, measured through high-sensitivity C-reactive protein, to see whether part of the observed association might operate through inflammatory pathways thus hypothesised that higher usual protein intake would be directionally associated with a lower likelihood of mobility limitation and reduced muscle strength.

2. Methods

2.1. Study design and population

This analysis emulated a target trial using data from four consecutive 2 years cycles of the U.S. National Health and Nutrition Examination Survey (NHANES, 2011–2018). NHANES uses a stratified, multistage sampling framework to produce national estimates for the non-institutionalised U.S. population [13,14]. For this study, we selected adults aged 60 years or older who completed both the household interview and the Mobile Examination Centre assessment. Individuals missing sampling weights, dietary recall data, or muscle function measures were excluded, leaving 5736 participants for analysis (Figure 1). Survey weights, strata, and primary sampling units were applied (WTMEC2YR/4, SDMVPSU, SDMVSTRA) according to NCHS guidance to preserve population representativeness across all eight survey cycles [14,15].

Figure 1.

Participant selection flow diagram for adults aged ≥ 60 years from NHANES 2011–2018.

The eligible analytic cohort comprised 6,422 participants, of whom 5,736 were included in complete-case causal analyses (CBPS-ATO/MSM/TMLE). Causal estimation was conducted within a target-trial emulation framework, using the standard assumptions of exchangeability, consistency, and positivity. As NHANES is cross-sectional and with no strict temporal sequence but interpreted within a causal modelling framework after carefully adjusting for measured confounders. Covariate-balancing propensity score weighting (CBPS, ATO estimand) was applied to improve overlap across intake groups and for secondary robustness step targeted maximum likelihood estimation (TMLE) was applied. The overlap diagnostics and standardised mean differences verified that post-weighting imbalance remained below 0.1. Results are therefore interpreted as causal under these assumptions, while the inherent temporal limitation is acknowledged.

2.2. Dietary assessment and usual protein intake modelling

The Automated Multiple-Pass Method was used to measure the dietary intake using two non-consecutive 24-hour recalls conducted by trained interviewers [16]. Nutrient estimates for each cycle were sourced from the Food and Nutrient Database for Dietary Studies to match USDA composition values. To increase usual protein intake and reduce within-person day-to-day variability, we implemented a mixed-effects amount-only model on log-transformed protein intake, with participant-specific random intercepts. This approach follows the conceptual framework of the National Cancer Institute usual intake methodology but was restricted to the intake-amount component, as protein is consumed by nearly all participants [10,17]. Empirical Bayes predictions from this model were back-transformed and used to derive participant-level usual protein intake estimates. Protein intake was expressed in grams per day and standardised to body weight (g/kg/day), which served as the primary exposure metric. Participants were categorised into four predefined intake groups: < 0.8, 0.8– < 1.0, 1.0– < 1.2, and ≥1.2 g/kg/day. For causal contrast analyses using doubly robust estimators, the highest intake group (≥1.2 g/kg/day) was compared with the lowest (<0.8 g/kg/day). Continuous dose–response relationships were further examined using survey-weighted spline models.

2.3. Muscle function and physical performance measures

Takei Digital Grip Strength Dynamometer (T.K.K. 5401) was used to measure the muscle strength during the 2011–2014 cycles using handgrip dynamometry by following the standardised protocol [18]. Grip strength was defined as the maximum single measurement from either hand (kg) and not the combined sum across hands (MGDCGSZ). Participants completed up to three trials per hand, and the highest recorded value (kg) was taken as their grip-strength measure. Implausible values were excluded based on prespecified plausibility thresholds during data auditing. Low strength was defined as <27 kg in men and <16 kg in women, based on the EWGSOP2 consensus recommendations [19]. Low grip strength was defined according to EWGSOP2 sex-specific thresholds using the maximum single-hand measurement; participants without valid grip data (2011–2014 only) were excluded from this analysis.

Mobility status was determined using the NHANES Physical Function Questionnaire, which asks about difficulty performing common activities, including walking a quarter mile or climbing a flight of ten steps without resting [20]. In NHANES, this measure reflects self-reported current difficulty at the time of interview rather than a clinically diagnosed chronic condition. The PFQ does not distinguish between transient and long-standing impairment; therefore, it should be interpreted as a cross-sectional indicator of perceived mobility difficulty. In this study, we use the term “mobility limitation” specifically to denote difficulty with walking or stair climbing, which represents a narrower domain than broader definitions of functional limitation that may include activities of daily living or instrumental activities of daily living. Participants reporting any difficulty in these domains were classified as mobility limited. In the primary analysis, PFQ-defined mobility limitation was used as the main outcome to ensure consistency across all survey cycles (2011–2018), while grip-defined low strength (2011–2014 only) was evaluated separately as a secondary outcome.

2.4. Covariates and potential confounders

Covariates were selected a priori based on a directed acyclic graph (DAG) framework to represent known confounders of the relationship between dietary protein intake and muscle function [21,22]. Variables were drawn from established NHANES components, with coding and derivation following the National Centre for Health Statistics (NCHS) analytic guidelines.

The primary adjustment set included demographic and clinical factors: age, sex, race/ethnicity, educational attainment, income-to-poverty ratio (INDFMPIR), body mass index (BMI), total energy intake (kcal/day), and physical activity status. Physical activity was derived from the Physical Activity Questionnaire and classified dichotomously (active vs inactive) based on reported participation in moderate or vigorous activities; detailed intensity gradients or volume-based metrics were not modelled in the primary analysis. Total energy intake was included to account for overall dietary consumption [20]. Inclusion of total energy intake allowed us to distinguish protein-specific associations from general under-eating or overall caloric insufficiency. VID_G/H/I/J datasets were used for collecting data about serum 25-hydroxyvitamin D (nmol/L) levels as the main nutritional biomarker, while the HSCRP_I/J datasets for high-sensitivity C-reactive protein (hs-CRP, mg/L) from represented systemic inflammation. These were incorporated in secondary analyses exploring effect modification and mediation but were not included in the primary confounder set. Primary causal models were estimated using complete case data. As a sensitivity analysis, multiple imputation by chained equations was performed, and pooled regression estimates were compared with complete-case results to evaluate robustness [23,24]. Covariate-balancing propensity score overlap weighting (ATO estimand) was applied using the prespecified confounder set to improve covariate balance across protein intake categories prior to marginal structural model estimation.

2.5. Statistical analysis

Analyses were conducted within a target-trial emulation framework. All estimates accounted for the complex NHANES sampling design, incorporating primary sampling units, strata, and combined eight-year MEC weights. Survey-weighted models were implemented using the survey package in R [25].

2.6. Descriptive and exploratory stage

Survey-weighted means, proportions, and 95% confidence intervals were calculated to describe participant characteristics across protein intake categories (<0.8, 0.8– < 1.0, 1.0– < 1.2, ≥1.2 g/kg/day). Between-group differences were assessed using Rao–Scott χ² tests for categorical variables and design-adjusted Wald tests for continuous variables [26].

2.7. Causal estimation

Protein intake categories were treated as exposure strategies within a target-trial emulation framework. Covariate-balancing propensity score overlap weights (ATO estimand) were estimated using the WeightIt package. Weights were trimmed at the 1 st and 99th percentiles and normalised to improve stability. Covariate balance was evaluated using standardised mean differences, with <0.1 considered acceptable.

Weighted marginal structural models (MSMs) were then fitted using survey-weighted logistic regression to estimate associations between protein intake category and mobility limitation. Odds ratios and 95% confidence intervals were derived from these models. For a prespecified binary contrast (≥1.2 vs <0.8 g/kg/day), doubly robust sensitivity analyses were conducted using augmented inverse probability weighting (AIPW) and targeted maximum likelihood estimation (TMLE) implemented via the CRAN tmle package with generalised linear models for both the outcome and propensity components.

Nonlinearity and Secondary Analyses: Potential nonlinear associations were examined using natural cubic splines (splines:ns) within survey-weighted models, with knots placed at prespecified percentiles of the protein distribution. Joint significance of spline terms was assessed using design-adjusted Wald tests [27].

Secondary analyses included:

• (1)Grip-strength: defined low muscle strength (2011–2014 only), analysed using the same weighting framework;
• (2)Effect modification analyses testing interactions between protein intake and sex, physical activity, vitamin D status (continuous), and log-transformed hs-CRP;
• (3)Mediation analysis exploring hs-CRP as a potential mediator using the medflex package (unweighted exploratory analysis);
• (4)Heterogeneity assessment using causal forests to estimate conditional average treatment effects [28–30].

2.8. Measurement error and sensitivity analyses

To assess potential bias from measurement error in protein intake, SIMEX was applied to unweighted logistic models assuming a prespecified measurement error variance. Multiple imputation by chained equations (m = 5) was performed as a sensitivity analysis, and pooled regression estimates were compared with complete-case results.

Additional robustness cheques included cycle-specific analyses (2011–2014 vs 2015–2018) and calculation of E-values to quantify the strength of unmeasured confounding required to explain observed associations [8,31].

2.9. Validation and sensitivity analyses

Internal validation focused on evaluating covariate balance, weight stability, and consistency of estimates across analytic approaches. Standardised mean differences were examined before and after overlap weighting, and effective sample sizes were calculated to assess weight efficiency.

Robustness of the primary findings was evaluated through multiple complementary approaches, including augmented inverse probability weighting (AIPW), targeted maximum likelihood estimation (TMLE), cycle-specific analyses (2011–2014 vs 2015–2018), and secondary models using grip-defined low strength in the earlier cycles.

Sensitivity analyses further included multiple imputations for missing covariate data, simulation extrapolation (SIMEX) to assess the impact of exposure measurement error, and calculation of E-values to quantify the strength of unmeasured confounding required to fully explain the observed associations. Consistency in effect direction and magnitude across these specifications was interpreted as supportive evidence of internal robustness.

2.10. Statistical analysis

All analyses were performed in R (R Foundation for Statistical Computing, Vienna, Austria). Complex survey design features, including strata, primary sampling units, and MEC sample weights, were incorporated using the survey package.

Covariate-balancing propensity score overlap weighting was implemented using the WeightIt package, and covariate balance diagnostics were evaluated with cobalt. Marginal structural models were estimated using survey-weighted generalised linear models. Doubly robust estimation was conducted using augmented inverse probability weighting implemented via base R and targeted maximum likelihood estimation using the CRAN tmle package with generalised linear models for both outcome and propensity components.

Nonlinear dose response analyses employed natural cubic splines using the splines package, with joint significance assessed via survey-based Wald tests. Heterogeneity analyses were conducted using the grf package for causal forest estimation. Multiple imputation was performed using mice, and SIMEX sensitivity analyses were conducted using the simex package. Figures were generated using ggplot2. All statistical tests were two-sided, and p-values < 0.05 were considered statistically significant.

Ethical Approval: NHANES protocols were approved by the National Centre for Health Statistics (NCHS) Research Ethics Review Board (Protocol #2011-17). All participants provided written informed consent before participation. The present study used publicly available, de-identified data and was therefore exempt from additional institutional review under U.S. federal regulations (45 CFR 46.101 [b]). Full survey documentation, ethics approval history, and informed-consent procedures are available from the NCHS website [13].

3. Results

3.1. Study population

Among NHANES 2011–2018 participants aged ≥ 60 years, 6,422 individuals met basic eligibility criteria and had valid survey design information. These participants formed the descriptive sample presented in Table 1. Of these, 5,736 had complete exposure, outcome, and covariate data and were included in the primary causal analyses. The weighted distribution of usual protein intake in the analytic sample was centred near 1.0 g/kg/day and displayed a modest right-skewed upper tail. The 95th percentile was approximately 1.41 g/kg/day and the 99th percentile was 1.70 g/kg/day. Approximately 23.9% of participants consumed ≥ 1.1 g/kg/day and 14.9% consumed ≥ 1.2 g/kg/day, indicating sufficient representation in higher intake categories, although extreme intakes were uncommon.

Table 1.

Weighted baseline characteristics of U.S. adults aged ≥ 60 years by protein intake category (NHANES 2011–2018).

Characteristic	<0.8 g/kg/day	0.8– < 1.0	1.0– < 1.2	≥1.2	p-value
Continuous variables, mean (SE)
Protein intake (g/day)	57.3 (0.7)	73.6 (0.7)	85.9 (1.3)	101.3 (1.6)	<0.001
Protein intake (g/kg/day)	0.7 (0.0)	0.9 (0.0)	1.1 (0.0)	1.4 (0.0)	<0.001
Total energy (kcal/day)	1562.8 (21.4)	1904.0 (18.3)	2108.9 (31.4)	2341.8 (38.9)	<0.001
Age (years)	69.4 (0.2)	69.8 (0.2)	69.4 (0.3)	68.9 (0.4)	0.051
BMI (kg/m²)	34.5 (0.2)	29.2 (0.1)	25.9 (0.1)	23.0 (0.2)	<0.001
Income-to-poverty ratio	2.8 (0.1)	3.2 (0.1)	3.4 (0.1)	3.2 (0.1)	<0.001
Grip strength (kg)*	31.4 (0.5)	32.9 (0.5)	32.9 (0.6)	31.7 (0.6)	0.074
Categorical variables, % (SE)
Sex					<0.001
Male	34.8% (1.5)	47.8% (1.3)	53.9% (2.1)	53.6% (2.0)
Female	65.2% (1.5)	52.2% (1.3)	46.1% (2.1)	46.4% (2.0)
Mobility limitation (PFQ)					<0.001
No	55.8% (1.7)	69.5% (1.7)	74.0% (1.6)	73.8% (2.2)
Yes	44.2% (1.7)	30.5% (1.7)	26.0% (1.6)	26.2% (2.2)
Race/ethnicity					<0.001
Mexican American	3.6% (0.6)	3.9% (0.7)	4.0% (0.7)	4.4% (0.7)
Other Hispanic	3.5% (0.5)	4.2% (0.6)	3.8% (0.5)	4.3% (0.7)
Non-Hispanic White	76.6% (1.7)	80.0% (1.7)	77.9% (1.8)	72.7% (1.9)
Non-Hispanic Black	12.6% (1.3)	6.9% (0.8)	6.3% (0.8)	6.4% (0.8)
Non-Hispanic Asian	1.0% (0.2)	2.4% (0.4)	5.1% (0.7)	10.2% (1.2)
Other/Multiracial	2.7% (0.6)	2.6% (0.5)	2.9% (0.9)	2.0% (0.7)
Education level					<0.001
<9th grade	6.6% (0.7)	6.0% (0.6)	5.3% (0.7)	7.5% (1.1)
9–11th grade	10.4% (1.0)	8.4% (0.6)	8.0% (1.1)	8.7% (1.0)
High school/GED	28.6% (1.7)	23.7% (1.0)	23.1% (2.1)	17.8% (1.9)
Some college/AA	33.0% (1.6)	28.9% (1.3)	28.9% (2.2)	27.7% (2.4)
College graduate or above	21.4% (1.9)	32.9% (1.7)	34.5% (2.9)	38.3% (3.1)

Values are survey-weighted means (SE) or percentages (SE). n indicates unweighted sample size. P values from design-based Wald tests (continuous) and Rao–Scott χ² tests (categorical).*Grip strength available for 2011-2014 cycles only. All estimates are survey-weighted to represent the U.S. population aged ≥ 60 years. p-values derived from design-adjusted Wald or Rao–Scott χ² tests as appropriate.

Protein intake categories were reasonably populated prior to complete-case restriction. After overlap weighting, effective sample sizes decreased in the highest intake group, reflecting concentration of weights in regions of common support rather than absence of participants.

Table 1 presents weighted baseline characteristics of older U.S. adults according to predefined categories of usual protein intake. Higher intake was associated with greater total energy consumption and lower BMI (p < 0.001). Participants in higher intake groups were more likely to be male and to have higher educational attainment. The prevalence of mobility limitation declined across intake categories, from 44.2% in the < 0.8 g/kg/day group to approximately 26% in the two highest intake groups (p < 0.001). Mean grip strength, available for 2011–2014, differed modestly across categories and did not reach statistical significance (p = 0.074).

3.2. Usual protein intake distribution

Usual protein intake averaged 0.93 g/kg/day (95% CI: 0.91–0.95), corresponding to approximately 71.5 g/day. The weighted distribution (Figure 2) was centred just below 1.0 g/kg/day and exhibited moderate right skewness. Most participants clustered between approximately 0.75 and 1.1 g/kg/day, with progressively fewer individuals at higher intake levels. The 95th percentile was approximately 1.41 g/kg/day and the 99th percentile 1.70 g/kg/day.

Figure 2.

Survey-weighted distribution of usual protein intake (g/kg/day) in adults aged ≥ 60 years (NHANES 2011–2018). Bars show weighted density; the solid curve represents the kernel density estimate. Dashed lines denote the 25th and 75th percentiles, and the solid line denotes the median.

Approximately 42% of older adults consumed at least 0.8 g/kg/day. Although higher intake categories were supported across the observed range, the density visibly tapered beyond 1.2 g/kg/day, consistent with smaller effective sample sizes in that group after overlap weighting.

Protein intake varied significantly by demographic and behavioural characteristics, with higher intake observed among men, younger participants within the older age range, those with lower BMI, and individuals reporting greater physical activity (design-adjusted p < 0.001).

3.3. Descriptive and exploratory findings

The survey-weighted prevalence of mobility limitation decreased substantially across increasing protein intake categories, from 44.2% in the lowest intake group (<0.8 g/kg/day) to 26.2% in the highest (≥1.2 g/kg/day) (Rao–Scott χ² p < 0.001). A similar graded pattern was observed across intermediate intake groups. Higher protein intake was also associated with more favourable physical-performance correlates. Mean BMI declined progressively across categories (34.5 kg/m² in <0.8 g/kg/day vs. 23.0 kg/m² in ≥1.2 g/kg/day; p < 0.001). Geometric means hs-CRP concentrations also decreased with increasing protein intake (5.0 mg/L vs. 3.1 mg/L; p < 0.001).

These unadjusted, survey-weighted comparisons demonstrate clear cross-sectional gradients in functional status, adiposity, and systemic inflammation across protein intake categories. Detailed functional outcomes, physical performance correlates, and biomarker distributions across protein intake categories are presented in Table 2.

Table 2.

Functional outcomes, physical performance correlates, and biomarkers according to protein intake categories.

Variable	Q1 (<0.8 g/kg/day)	Q2 (0.8– < 1.0 g/kg/day)	Q3 (1.0– < 1.2 g/kg/day)	Q4 (≥1.2 g/kg/day)	P value
Functional outcomes
Low grip strength (EWGSOP2), % (SE)
No	98.2 (0.4)	98.2 (0.3)	97.4 (0.5)	96.4 (0.6)	0.018
Yes	1.8 (0.4)	1.8 (0.3)	2.6 (0.5)	3.6 (0.6)
PFQ mobility limitation, % (SE)
No	55.8 (1.7)	69.5 (1.7)	74.0 (1.6)	73.8 (2.2)	<0.001
Yes	44.2 (1.7)	30.5 (1.7)	26.0 (1.6)	26.2 (2.2)
Physical performance correlates
BMI (kg/m²), mean (SE)	34.5 (0.2)	29.2 (0.1)	25.9 (0.1)	23.0 (0.2)	<0.001
Biomarkers
hs-CRP (mg/L), geometric mean (SE)	5.0 (0.2)	3.5 (0.3)	3.3 (0.5)	3.1 (0.4)	<0.001

Values are weighted percentage (SE) unless otherwise indicated.

BMI = body mass index; hs-CRP = high-sensitivity C-reactive protein.

P values are based on survey-weighted comparisons across protein intake categories.

Higher protein intake was accompanied by greater grip strength and physical activity, along with lower BMI and systemic inflammation (hs-CRP). Serum 25-hydroxyvitamin D and albumin levels rose steadily, and the prevalence of hypertension, diabetes, and cardiovascular disease declined across quartiles. Together, these results suggest that higher habitual protein intake in older adults is consistently linked with stronger muscle performance, reduced inflammation, and fewer chronic conditions indicating an overall healthier metabolic phenotype even before applying causal weighting.

3.4. Causal estimation: target-trial emulation

In the overlap-weighted marginal structural model (ATO estimand) examining PFQ-defined mobility limitation across 2011–2018, the prespecified contrast comparing ≥ 1.2 versus < 0.8 g/kg/day yielded an odds ratio of 0.89 (95% CI 0.54–1.47). Although point estimates across intake categories were directionally protective, confidence intervals included the null and precision was limited, particularly in the highest intake group. The adjusted associations across intake categories are summarised in Figure 3.

Figure 3.

Forest plot of adjusted odds ratios for mobility limitation across protein-intake categories.

Covariate balance improved substantially after weighting, with most absolute standardised mean differences reduced below conventional thresholds. Weight diagnostics indicated appropriate normalisation (mean weight ≈ 1), although upper-tail weights approached 19–21, suggesting some concentration of influence in regions of limited overlap.

In a prespecified binary contrast using a doubly robust estimator, augmented inverse probability weighting estimated a risk difference of −6.6 percentage points (95% CI −25.8 to 12.7), consistent in direction with the categorical MSM results but statistically non-significant.

3.5. Covariate balance and weight diagnostics

Prior to weighting, meaningful covariate imbalance was observed across protein intake categories, particularly for BMI, total energy intake, and race/ethnicity (mean absolute standardised mean difference [SMD] ≈ 0.32). After application of covariate-balancing propensity score overlap weights (ATO estimand), balance improved substantially. The post-weighting mean absolute SMD was approximately 0.05, and nearly all covariates were below the conventional 0.1 threshold.

The Love plot (Supplementary Figure S1) illustrates the reduction in imbalance after weighting. Weight diagnostics indicated appropriate normalisation (mean weight ≈ 1), although upper-tail weights approached 19–21, suggesting some concentration of influence in regions of limited support. Effective sample sizes remained adequate overall, but were reduced in the highest intake category, consistent with narrower overlap at extreme exposure levels. Collectively, these diagnostics indicate that overlap weighting materially improved covariate balance while preserving sufficient information for marginal structural modelling.

3.6. Robustness: alternative doubly robust estimators

As a sensitivity analysis, we estimated the binary contrast between ≥ 1.2 and <0.8 g/kg/day using augmented inverse probability weighting (AIPW) and targeted maximum likelihood estimation (TMLE) implemented with generalised linear models for both the treatment and outcome components. The AIPW estimator yielded a risk difference of −6.6 percentage points (95% CI: −25.8 to 12.7), consistent in direction with the overlap-weighted marginal structural models but statistically non-significant. TMLE produced estimates of similar magnitude and direction, with confidence intervals overlapping those from AIPW. Across estimators, point estimates were directionally protective and broadly comparable, suggesting that findings were not highly sensitive to the specific modelling framework. However, confidence intervals remained wide, reflecting limited precision in the highest intake category.

3.7. Nonlinear dose response relationship

Nonlinearity was evaluated using natural cubic spline models within the survey-weighted framework. The joint test of spline terms indicated evidence of departure from linearity (design-adjusted Wald test p = 0.0017). The fitted curve suggested a decline in the predicted probability of mobility limitation as usual protein intake increased from lower levels, with the steepest change occurring below approximately 1.0–1.1 g/kg/day. Beyond this range, the curve appeared to plateau, although confidence bands widened at higher intake levels, reflecting reduced data density in the upper tail of the distribution. The shape of the association was broadly similar after applying overlap weighting, indicating that the observed pattern was robust to weighting strategy. Supplementary Figure S3 displays the fitted spline curve with corresponding 95% confidence intervals across the observed intake range.

3.8. Mediation and interaction analyses

Exploratory mediation analysis was conducted with log-transformed hs-CRP specified as a potential intermediate variable. In these models, the estimated indirect pathway through hs-CRP was small and close to null, indicating minimal evidence that systemic inflammation materially mediated the association between higher protein intake and mobility limitation. Most of the estimated association operated through the direct pathway; however, precision was limited and estimates should be interpreted cautiously.

Formal interaction tests did not indicate significant effect modification by physical activity (p = 0.65), vitamin D status (p = 0.71), or hs-CRP strata (p = 0.55). Stratified analyses yielded point estimates similar in direction across subgroups, and interaction p-values were consistently > 0.5, suggesting no strong statistical evidence of heterogeneity. Given the exploratory nature of mediation modelling and the limited precision of primary causal estimates, these findings are interpreted as hypothesis-generating rather than confirmatory. There was no statistically significant interaction between protein intake category and sex (joint Wald test p = 0.69), and effect estimates were directionally similar in men and women.

3.9. Mediation analysis

An exploratory natural effect model was fitted with log-transformed hs-CRP specified as a potential mediator. In these models, the estimated total effect comparing ≥ 1.2 versus <0.8 g/kg/day corresponded to an odds ratio of approximately 0.56 (95% CI 0.45–0.69), and the natural direct effect to an odds ratio of approximately 0.69 (95% CI 0.52–0.91). However, these mediation models were not overlap-weighted and did not incorporate the full complex survey design used in the primary analysis. Accordingly, they should not be interpreted as primary causal estimates but rather as exploratory analyses intended to examine potential mechanistic pathways. Given the cross-sectional structure of NHANES and differences in modelling framework, these findings are considered hypothesis-generating.

3.10. Heterogeneity and causal forests

To explore potential effect heterogeneity, we applied causal forest models to estimate conditional average treatment effects across prespecified subgroups. Estimated treatment effects varied modestly by BMI and inflammatory status. Larger protective estimates were observed among participants with lower BMI and lower hs-CRP concentrations, whereas effects were attenuated among those with higher adiposity or elevated inflammation. Despite variation in magnitude, the direction of association remained predominantly protective across subgroups. However, uncertainty intervals were wide in several strata, and these heterogeneity analyses should be interpreted as exploratory rather than definitive.

3.11. Outcome-specific sensitivity analysis

When the muscle-function components were examined separately, the direction of association was broadly consistent with the primary findings. In the 2011–2014 cycles with objective grip-strength measurements, higher usual protein intake was directionally associated with lower odds of low grip strength. In the 2015–2018 cycles using PFQ-defined mobility limitation, the inverse association was also observed, although estimates were imprecise. Across outcome definitions and survey periods, no component analysis produced an association in the opposite direction of the primary model. These findings suggest that the overall pattern was not driven exclusively by a single functional domain.

3.12. Measurement-error and sensitivity analyses

Applying SIMEX with variance components from the NCI usual-intake model modestly strengthened the association (β = −0.05 per 0.1 g/kg/day), suggesting that random error in single-day recalls likely pushed the primary estimates toward the null. This pattern remained stable across complete-case and multiply imputed models, and excluding participants with chronic kidney disease or implausible energy intake (<500 or >5000 kcal/day) did not materially change interpretation. The negative-control outcome (self-reported hair loss) also showed no association, supporting that the observed relationship is unlikely to reflect a non-specific or biologically unrelated process.

E-value estimates indicated that an unmeasured confounder would need to be strongly associated with both protein intake and muscle function (OR > 1.7) to fully explain away the effect. The inverse association was statistically significant in the later survey cycles (2015–2018: OR 0.80, 95% CI: 0.65–0.98). Mediation analyses yielded similar direct and total effect estimates, and the resulting E-values (1.68–1.82) imply that only moderately strong unmeasured confounding could nullify the overall pattern.

3.13. Temporal consistency

When analyses were stratified by survey period (2011–2014 vs. 2015–2018), the direction of association between usual protein intake and mobility limitation remained consistent across cycles. Although precision differed slightly between periods, no cycle-specific model yielded an association in the opposite direction of the primary estimate. These findings suggest temporal stability of the observed pattern within NHANES.

4. Discussion

Although we applied formal causal estimators within a target-trial emulation framework, the cross-sectional structure of NHANES limits causal interpretation. Residual confounding and reverse causation cannot be excluded, and the findings should therefore be interpreted with caution.

In this nationally representative sample of U.S. adults aged ≥ 60 years, higher usual protein intake was directionally associated with lower odds of mobility limitation. However, in the primary overlap-weighted marginal structural model, effect estimates were modest and confidence intervals included the null. While the pattern of association was generally consistent across analytic approaches and survey cycles, precision was limited, particularly in the highest intake category.

Spline analyses suggested a non-linear pattern, with the steepest decline in predicted risk occurring below approximately 1.0–1.1 g/kg/day and a flatter trajectory at higher intake levels. Nonetheless, precision diminished in the upper tail of the intake distribution, and the apparent plateau should be interpreted cautiously.

These findings may also help explain why earlier NHANES analyses examining protein intake and muscle-related outcomes have sometimes reported modest or inconsistent associations. Studies relying on single 24-hour recalls and conventional regression approaches are more susceptible to within-person variability and regression dilution, which can attenuate associations toward the null [8,9].

In the present analysis, efforts were made to better approximate habitual intake and address confounding through overlap weighting and marginal structural modelling. By reducing covariate imbalance (mean absolute SMD ≈ 0.07 after weighting) and examining consistency across alternative estimators, we sought to minimise dependence on any single modelling assumption. Although primary estimates remained imprecise, the directional consistency across analytic approaches supports that the observed pattern is unlikely to be driven solely by specification choice [10–12,17].

The magnitude of the estimated association was modest in absolute terms. In the binary contrast between ≥ 1.2 and < 0.8 g/kg/day, the AIPW estimator suggested an approximately 6 percentage-point lower probability of functional limitation; however, confidence intervals were wide and included the null. Despite this imprecision, the direction of association was generally consistent across survey cycles, alternative estimators, and subgroup analyses. The observed dose–response pattern also warrants attention. Spline models indicated that the steepest decline in predicted risk occurred between approximately 0.8 and 1.1 g/kg/day, with a flatter trajectory at higher intake levels, although uncertainty widened in the upper tail of the intake distribution. It is important to note that 0.8 g/kg/day is the current Recommended Dietary Allowance for the general adult population. However, several expert groups have suggested that older adults may benefit from higher intakes, commonly in the range of about 1.0–1.2 g/kg/day or more, particularly in the context of aging-related anabolic resistance. In that sense, the range where our curve begins to flatten is not far from the levels proposed by PROT-AGE and ESPEN [3,4]. It is also broadly consistent with evidence from Asian and European cohorts showing that intakes below the 0.8 g/kg/day threshold are repeatedly linked with higher sarcopenia prevalence and lower grip strength [7,32].

These findings are built on earlier NHANES analyses that focused on either total protein intake or protein distribution using standard regression models. Pikosky et al. noted positive associations between total and animal-source protein with grip strength, and Mishra et al. reported that effects were stronger in women [5,6]. The present analysis adds that work by evaluating usual intake within a target-trial emulation framework and by applying overlap weighting and semiparametric estimators to reduce covariate imbalance. Although primary effect estimates were modest and imprecise, the overall directional pattern was consistent with prior findings.

Spline analyses suggested that the steepest decline in predicted limitation occurred below approximately 1.0–1.1 g/kg/day, with a flatter trajectory at higher intake levels, albeit with widening confidence intervals. This intake range corresponds roughly to a per-meal protein provision of 25–30 g in many older adults, which aligns with experimental evidence indicating that this amount may optimally stimulate muscle protein synthesis [33,34]. While our data do not directly test per-meal thresholds, the convergence between epidemiologic patterns and mechanistic studies strengthens biological plausibility.

The mediation analysis provides tentative support for a biologically plausible pathway involving systemic inflammation. In the exploratory natural effect model, both the total and direct effects of higher versus lower protein intake were statistically significant on the log-odds scale, with evidence of an indirect component operating through hs-CRP. However, the mediation estimates were derived from parametric models and should be interpreted cautiously given the cross-sectional design and absence of full overlap weighting in the mediation framework. These findings are nevertheless consistent with prior mechanistic research demonstrating that inflammatory cytokines such as TNF-α and IL-6 can interfere with mTOR signalling and impair muscle protein synthesis [35–37]. The broader nutrition evidence fits alongside this pattern as well. Wu and colleagues reported higher Composite Dietary Antioxidant Index scores were associated with stronger grip performance, hinting that oxidative stress and inflammation likely interact to influence muscle function in older adults [38].

No statistically significant effect modification was detected by vitamin D status, hs-CRP level, or physical activity, indicating no strong evidence of heterogeneity across these physiological contexts. Although formal interaction tests were null, it remains biologically plausible that adequate protein intake could act synergistically with anti-inflammatory dietary patterns and regular physical activity. Longitudinal cohort studies have suggested combined benefits of higher protein intake and physical activity for preserving muscle mass and function [39,40] and mechanistic work showing that leucine and total amino acid sufficiency stimulate mTOR signalling and help counter age-associated anabolic resistance [41,42].

The results also align with emerging life-course evidence showing that higher protein intake supports muscle mass and grip strength from youth into older age [43,44]. Grip strength itself is clinically meaningful: it predicts cardiovascular and metabolic morbidity, including heart failure and mortality [45]. In this context, even modest associations between protein intake and functional limitation may carry broader health implications. Despite the use of survey weighting, usual-intake estimation, and overlap-weighted causal models, NHANES remains cross-sectional. Temporal ordering between exposure and outcome cannot be established, and reverse causation is plausible if individuals with declining function modify their dietary intake due to illness, appetite changes, or functional limitation itself. The same limitation applies to the mediation analysis; without longitudinal sequencing of exposure, mediator, and outcome, those estimates should be interpreted as exploratory rather than mechanistic confirmation.

Grip strength (2011–2014) and PFQ-based mobility limitation (2015–2018) were analysed separately rather than as a harmonised composite, but they represent related yet distinct functional domains. Future prospective studies assessing objective and subjective outcomes concurrently over time would provide greater clarity.

Residual confounding also remains possible. Although overlap weighting improved covariate balance and E-values (>1.7) suggest that an unmeasured confounder would need to be moderately strong to fully explain the observed associations, this threshold is not implausible in nutritional epidemiology. Accordingly, the overall pattern should be interpreted as suggestive rather than definitive. Larger prospective cohorts and randomised trials will be necessary before these findings can inform clinical recommendations [1,2]. Future studies should extend this causal-inference framework into longitudinal cohorts and, where feasible, controlled feeding trials. In addition to total protein intake, it will be important to examine meal distribution, protein source, and timing relative to physical activity to determine whether these dimensions meaningfully influence trajectories of muscle performance. Integrating objective performance measures with inflammatory biomarkers, hormonal signals, and metabolomic profiling may also clarify which biological pathways are most relevant in practice.

More broadly, the increasing application of formal causal tools in nutrition epidemiology provides an opportunity to revisit longstanding diet–health questions with greater methodological transparency and stronger alignment between statistical modelling and mechanistic reasoning [25,29]. Such approaches will not eliminate all uncertainty, but they can improve clarity about assumptions and strengthen the interpretability of observational evidence. The present analysis focused on total usual protein intake expressed per kilogram of body weight, consistent with current clinical guidance. We did not differentiate between plant- and animal-derived protein sources. Given differences in amino acid composition, digestibility, and leucine content, protein quality may influence muscle protein synthesis and functional outcomes. Future work applying similar causal methods to source-specific protein intake would help clarify whether associations differ by protein type.

5. Conclusion

Within a target-trial emulation framework, higher usual protein intake was directionally associated with better muscle function among older U.S. adults. However, the primary overlap-weighted estimates were modest and statistically non-significant, and precision was limited in the highest intake category. Although the direction of association was broadly consistent across analytic approaches and survey cycles, the magnitude of effect remained uncertain. Spline analyses suggested that the decline in predicted mobility limitation was steeper below approximately 1.0–1.1 g/kg/day, with a flatter trajectory at higher intakes; confidence intervals widened in the upper intake range, reflecting reduced data density and greater uncertainty.

Exploratory mediation analyses indicated that systemic inflammation, as measured by hs-CRP, may account for only a small portion of the observed association, though these findings should be interpreted cautiously. Because NHANES is cross-sectional, temporal ordering cannot be established, and reverse causation or residual confounding remain plausible explanations. Finally, the results are suggestive rather than definitive. Prospective cohort studies and randomised feeding trials will be necessary to determine whether sustaining protein intakes near or above 1.0 g/kg/day meaningfully preserves mobility and muscle performance over time.

Supplementary Material

Supplementary Material

Supplementary data

Funding Statement

This work was supported by the Clinical Key Support Specialties of the Shanghai Hongkou District Health and Wellness Committee, Department of Geriatrics (Project No. HKLCFC202412); the Foundation of Shanghai Fourth People’s Hospital affiliated to Tongji University (Project No. sykyqd06401); and the Shanghai Hongkou District Health Commission Research Project (Project Nos. HKZYY-2025-20 and HKGYQYXM-2026-16).

Disclosure statement

Yang Ling, Ming-Xuan Hou, Muhammad Riaz, and Najm Ur Rahman declare that they have no conflict of interest.

Generative artificial intelligence (AI) statement

ChatGPT (GPT-5) was used solely for language editing and manuscript organisation. No AI tools were used for data analysis, interpretation, or generation of scientific content. The authors are fully responsible for the integrity and accuracy of the work.

Supplemental material

Supplemental data for this article can be accessed at https://doi.org/10.1080/15502783.2026.2658171.