Abstract
Background The Canada’s Food Guide 2019 (CFG) provides universal recommendations to individuals aged 2 years or older. The extent to which these recommendations are appropriate for older adults is unknown. Although ideal, conducting a large randomized controlled trial is unrealistic in the short term. An alternative is the target trial emulation framework for causal inference using observational data.
Objective Our aim is to describe the protocol of a target trial emulation in older adults with emphasis on key aspects of a hypothetical sustained diet and physical activity intervention.
Methods To emulate the target trial, non-experimental data from the NuAge prospective study (n=1753, adults aged 67 years or older) will be used. NuAge includes 4 yearly measurements of dietary intakes, covariates and outcomes. The per protocol causal contrast will be the primary causal contrast of interest to account for non-adherence. The sustained intervention strategy will be modelled using the parametric g-formula. In the hypothetical trial, participants would be instructed to meet sex-specific minimal intakes for vegetables and fruits, whole grains, animal- and plant-based protein foods, milk & plant-based beverages and unsaturated fats. Eligibility criteria, follow-up, intervention, outcomes, and causal contrast will be similar in the emulation to the target trial except for minor modifications. We will attempt to emulate randomization of treatment by adjusting for baseline covariates and pre-baseline dietary habits.
Conclusion Emulating a target trial will provide the first evidence of the adequacy of CFG 2019 recommendations for older adults in relation to health outcomes.
Introduction
The latest edition of Canada’s Food Guide (CFG) was published in 2019 (Health Canada 2019a). Compared with the previous edition, key changes include the removal of the pre-specified number of servings to consume each day, a shift towards qualitative (“eat plenty of…”) instead of quantitative recommendations and the provision of universal recommendations instead of age and sex-specific recommendations. In addition, CFG recommendations primarily aim at reducing chronic disease risk. Indeed, the evidence behind CFG’s recommendations concern mitigation of cardiovascular disease, cancer and type 2 diabetes risk (Health Canada 2019b). However, evidence from a nationally representative survey of adults aged 65 years or older from Canada suggested that greater adherence to recommendations was insufficient to meet calcium, vitamin D and folate requirements (Brassard and Chevalier 2023). In Canada, one third of community-dwelling older adults are at high nutrition risk (Ramage-Morin and Garriguet 2013; Mills et al. 2023), highlighting the importance of maintaining adequate nutritional status in this stratum of the population. Similarly, in absence of specific recommendations on the amount of protein foods to eat, older adults may be eating less protein than required to maintain muscle mass (Bauer et al. 2013; Deutz et al. 2014; Paddon-Jones et al. 2015). Canada’s Food Guide also provide brief physical activity recommendation, but without explicit acknowledgement of the importance of these recommendations for older adults (Health Canada 2019a). Indeed, performing a minimal amount of physical activity is paramount to maintaining muscle mass (Bauer et al. 2013; Deutz et al. 2014). Thus, the universal recommendations in CFG may not be appropriate for older adults since they face unique challenges to consuming a healthy diet (Remond et al. 2015), and may require specific nutritional strategies (Shlisky et al. 2017).
Ideally, a randomized controlled trial (RCT) would be conducted to investigate the adequacy of CFG recommendations in older adults. However, such RCT is unlikely to be conducted in the short term. An alternative is the target trial emulation framework for causal effect estimation using observational data (Hernan and Robins 2016; Kutcher et al. 2021; Tobias and Lajous 2021; Hernan et al. 2022). Informally, the target trial emulation framework is intended to highlight and address design issues of observational data analysis that aims to estimate a causal effect by explicitly emulating a hypothetical trial (Hernan and Robins 2016).
To the best of our knowledge, the target trial framework has not been used to assess the effect of adhering to CFG’s recommendations. Accordingly, the objectives of the present study are, first, to describe the protocol of the emulation of a target trial using observational data from the NuAge study (Gaudreau et al. 2007), a cohort of 1753 adults aged 67 to 84 years at baseline, and, second, to address key aspects of the target trial emulation in the context of a sustained lifestyle intervention strategy involving diet and physical activity. Key aspects are the description of the sustained lifestyle intervention strategy, the attempt to emulate randomization, as well as assumptions and limitations specific to diet intervention. Of note, more general introductory texts to the target trial framework are available elsewhere (Hernan and Robins 2016; Kutcher et al. 2021; Hernan et al. 2022).
Methods
Research Question and Hypothesis
Explicitly acknowledging the causal nature of a research question is prerequisite to causal effect estimation using observational data (Hernan 2018; Goetghebeur et al. 2020; Haber et al. 2022). The general objective of this study will be to examine the adequacy of CFG’s universal dietary recommendations for older adults. Expressed as a counterfactual statement, our objective is to answer the following question:
What would be the difference in a given health outcome at the end of the follow-up, if all eligible participants had increased their adherence to CFG recommendations on healthy food choices compared with, instead, if they had maintained their habitual diet?
Specifically, among adults aged 67 to 84 years followed over 3 years, and compared with maintenance of habits, we aim to:
1. Estimate the causal effect of adhering to CFG’s dietary recommendations on markers of muscle health (e.g., physical function, muscle strength), general health (e.g., waist circumference, blood pressure, glucose), and cognitive health (i.e., Modified Mini-Mental State Exam).
Hypothesis: adhering to recommendations positively influence general and cognitive health but has no influence on muscle health.
2. Estimate the causal effect of adhering to a reformulation of CFG’s dietary recommendations, including more protein foods and a minimal physical activity recommendation, to amplify the positive health effects.
Hypothesis: increasing the consumption of protein-rich foods positively influence muscle health, and meeting minimal physical activity recommendation (30 minutes or more per day) further amplifies positive health effects.
Study Design and Sample
Data from the NuAge prospective cohort study will be used to emulate the target trial (Gaudreau et al. 2007). The NuAge cohort comprised 1753 generally healthy community-dwelling adults aged 67 to 84 years at baseline and followed over 3 years. The baseline and each annual follow-up included comprehensive assessment on sociodemographic data, diet, physical activity, functional status as well as physical and mental health (Gaudreau et al. 2007).
The NuAge Database and Biobank (Database and Biobank 2024) has received approval by the Research Ethics Board of the Centre intégré universitaire de santé et de services sociaux de l’Estrie—Centre hospitalier universitaire de Sherbrooke. Finally, the NuAge study sample is relevant to our research question. The target population of CFG recommendations comprise all individuals aged 2 years or older, which is compatible with the target sample of the NuAge study of generally healthy older adults from the greater Montreal, Sherbrooke and Laval areas, the province of Quebec, Canada.
Target trial
The first step of a target trial emulation is the description of the target trial, i.e., the protocol of the hypothetical randomized controlled trial we would like to conduct (Hernan and Robins 2016; Hernan et al. 2022). Table 1 presents the target trial and its emulation using observational data from the NuAge study. Key differences between the target trial and its emulation are that participants will be required to provide complete dietary assessment and covariate data at baseline (eligibility component); and that we will attempt to emulate the randomized assignment by adjusting for dietary intakes before baseline as well as baseline covariates (assignment component).
Each part of the target trial and its emulation are described below.
Eligibility Criteria
The inclusion criteria for the target trial are the same as for the NuAge study (Gaudreau et al. 2007). In the emulation, participants will be required to have at least one 24-hour dietary recall completed at baseline with 500 calories or more, as well as complete covariate data at baseline, as identified below.
Hypothetical Interventions
The hypothetical intervention strategies evaluated will be:
no change in dietary habits or physical activity (“control” intervention);
adherence to CFG’s recommendations;
adherence to CFG’s recommendations including reformulation (i.e., higher intake of protein foods);
adherence to CFG’s recommendations including reformulation (i.e., higher intake of protein foods) and performing at least 30 minutes of aerobic physical activity.
Physical activity recommendations are not traditionally at the forefront of CFG’s recommendations. However, CFG does mention that “at least 150 minutes of moderate-to vigorous-intensity aerobic physical activity per week […] is recommended to achieve health benefits” (Health Canada 2019a). Thus, recognizing the key role of exercise to maintaining health and muscle for older adults, the 4th hypothetical intervention includes a formal physical activity recommendation. In the target trial, the physical activity corresponds to performing aerobic exercise of light to vigorous-intensity aerobic for at least 30 minutes per day (Bauer et al. 2013).
The challenge of a well-defined nutritional intervention
Emulating a well-defined dietary intervention for CFG’s recommendations is challenging. First, recommendations in the latest edition of CFG are qualitative and flexible (e.g., “Eat plenty of vegetables and fruits”; Health Canada (2019a); Brassard et al. (2022b)). Thus, multiple suitable, but nonetheless very different, dietary patterns can achieve adherence to recommendations. Second, CFG’s recommendations target both food intakes (e.g., vegetables and fruits) and nutrients (e.g., saturated fats). The nutrient-based recommendations can be achieved by intervening on the consumption of multiple different foods. For example, to achieve the hypothetical intervention “decreasing consumption of calories from saturated fats” one could decrease saturated fats from dairy, nuts and low nutritive value foods altogether. Arguably, the relationship beetween these food categories and a given health outcome may vary greatly.
To estimate a causal effect using observational data, the hypothetical interventions must be sufficiently well-defined or that until “no meaningful variation” of that intervention remains (Hernan 2016; Chiu 2022). In other words, the hypothetical diet interventions should be elaborated until no additional dietary characteristics are deemed impactful in regards of the outcome of interest. Another consideration is that the modelling of hypothetical interventions should ideally be conducted with dietary intakes expressed using the same units. For example, mixing food intakes expressed in servings and grams in a statistical model may cause poor estimation of causal effects (Tomova et al. 2022a). Finally, the statistical approach used to account for ‘total energy’ or total food intake also affects the causal effect of interest and should be consistent with the research question (Kipnis et al. 1993; Chiu 2022; Tomova et al. 2022b, 2022a).
Diet Simulations
For the present study, adherence to CFG’s recommendations was defined based on simulated diets generated by Health Canada (Health Canada 2022) and summarized in Supplemental Table 1. The simulated diets were designed to meet both CFG’s recommendations on healthy food choices and nutrient requirements (Dietary Reference Intake).
These diets achieve near perfect Healthy Eating Food Index (HEFI)-2019 scores (>78/80) through relatively high intakes of recommended foods (i.e., vegetables and fruits, whole-grain foods, protein foods, and unsweetened milk and plant-based beverages with protein) and null intakes of foods not recommended (i.e., non-whole grain foods, other low nutritive value foods, juice, sugary drinks and alcohol, and fatty foods rich in saturated fats). The HEFI-2019 score indicates the extent to which dietary intakes are consistent with CFG’s recommendations on healthy food choices (Brassard et al. 2022b, 2022a).
Note that the HEFI-2019 could have been used as main exposure to measure adherence to CFG. However, the use of a composite score metric would not fully satisfy the criterion of a well-defined intervention to estimate a causal effect. First, high HEFI-2019 scores, and high adherence to CFG’s recommendations, can be achieved through many different strategies or dietary patterns. In the context of observational data, the specific strategies through which individuals achieve high HEFI-2019 score would be based on dietary habits and patterns self-selected by the participants. This approach is similar to asking hypothetical trial participants to modify their intakes without clearly indicating how, which would obscure the causal effect estimated. Second, the HEFI-2019 score includes recommendations both on foods and on nutrients. As described above, mixing servings and grams in a statistical models may cause poor estimation of causal effects (Tomova et al. 2022a).
We stress that the diets simulated by Health Canada were not actually consumed by older adults. For this reason, the simulated values for vegetables and fruits, whole-grain foods and plant-based protein foods exceed the 99th percentile of the distribution of usual intakes of these food categories as estimated in adults aged 65 years or more from the Canadian Community Health Survey 2015 - Nutrition (Brassard and Chevalier 2023). In Table 2, the target intakes vegetables and fruits, whole grains, and plant-based protein foods for the adhering to Canada’s Food Guide 2019 recommendations intervention were revised to correspond, at most, to the 90th percentile of the distribution of usual intakes among Canadians aged 65 years or more in 2015 (Brassard and Chevalier 2023).
Implementation
In the target trial, the sustained intervention strategy could be implemented as follows:
the participant’s usual dietary intakes and physical activity would be assessed by research dietitians at each study visit;
if reported food intakes and duration of physical activity were equal to or above the prespecified thresholds (Table 2), no change would be suggested to the participants’ diet or physical activity If food intakes or duration of physical activity were below the prespecified thresholds, participants would be instructed to increase food consumption to exactly the prespecified portions or increase physical activity duration to 30 minutes per day (when applicable);
if changes are required, participants would be instructed to decrease consumption of foods not recommended to the extent of the increase in (2). For example, if a 2-serving increase in vegetables and fruits is required to meet the pre-specified intervention thresholds, participants would be instructed to substitute 2 servings of vegetables and fruits for non-whole grain foods, other low nutritive value foods, juice, sugary drinks and alcohol, and fatty foods rich in saturated fats.
In the emulation, the substitution will be implemented by including “total intakes” as a covariate and leaving out foods not recommended from the model. In this approach to account for “total energy”, all model coefficients reflect the action of increasing intakes of recommended foods and a concomitant decrease in any of the foods not recommended (Kipnis et al. 1993). On the one hand, this approach is potentially confusing (Tomova et al. 2022b, 2022a), since the default interpretation of model coefficients is the action of increasing the intake of each food included in the model, while decreasing the intake of foods not in the model (Kipnis et al. 1993). On the other hand, the standard model is generally consistent with the implementation of dietary intervention in feeding trials (Willett 2013; Tobias 2022; Willett et al. 2022). The standard model also reduces the number of variables to be considered as intervention variables. Otherwise, 4 additional dietary components would have to be modelled for foods not recommended (i.e., non-whole grain foods, other low nutritive value foods, juice, sugary drinks and alcohol, and fatty foods rich in saturated fats). Finally, the explicit description of the intervention strategies in the target trial protocol clarifies the estimand of interest, as done previously (Chiu et al. 2021; Chiu 2022). Of note, nutrient-based recommendations in CFG (i.e, saturated fats, free sugars and sodium) are not explicitly modelled to avoid the problems associated with mixed-unit models (Tomova et al. 2022a). In the target trial, we assume that nutrient-based targets would be met by reducing consumption of foods not recommended (i.e., non-whole grain foods, other low nutritive value foods, juice, sugary drinks and alcohol, and fatty foods rich in saturated fats). In that regard, food-level substitution analyses in Canadians support this assumption for saturated fats (Harrison et al. 2021, 2022).
The “additional protein” intervention consists of increasing the intake of both animal-based and plant-based protein foods by 1.5 and 0.5 RA per day, respectively, as well as milk and plant-based beverages with sufficient protein by 0.5 RA per day. In terms of amount of food, this corresponds to adding 150 grams of cooked unprocessed red meat, fish or poultry or 3 small eggs, 25 grams of nuts and seeds and 125 ml of milk or plant-based beverages with sufficient protein, while proportionally decreasing the intake of foods not recommended. In a previous RCT (Daly et al. 2014), older women aged 60 to 90 years were able to consume an additional 160 grams of cooked lean red meat without substitution, hereby supporting the feasibility of the protein intervention in the present hypothetical study.
Assignment
Random allocation (randomization) will be emulated by adjusting for dietary intakes in the year prior to the intervention, as well as adjusting for covariates at the start of the study. Covariates were identified using the causal diagrams depicted in Figure 1 based on background knowledge of the relationship between the hypothetical lifestyle intervention and outcomes.
Dietary components that are the foundation for healthy eating in CFG include intakes of vegetables and fruits, whole grains, protein foods (plant- and animal-based protein foods, milk and plant-based beverages with protein), unsaturated oils and fats (Health Canada 2019a).
Covariates include age at baseline, biological sex, region, education, living alone, smoking and drinking (alcohol) habits, major chronic diseases (i.e., hypertension, diabetes, cancer, heart disease), number of medications, supplement use (e.g., vitamins and minerals), living alone, and height and weight.
Text Box 1
Z, baseline covariates: age, sex, region, education, history of smoking, height, former cancer.
P, prior exposure (i.e., exposure of time-varying intervention prior to baseline): dietary habits before baseline
L, (time-varying) covariates: weight, number of medications, supplement use, living alone, major chronic diseases (hypertension, diabetes, cancer, heart disease), smoking and alcohol habits
X, (time-varying) treatment: diet and physical activity habits
Y, end of follow-up outcome: muscle health, general health, cognitive health.
Contrary to dietary habits, data on physical activity habits before baseline were not collected in the NuAge study. In this case, the potential effect of prior physical activity habits will not be accounted for in the models that aim at emulating the sustained physical activity intervention strategy. For the models emulating the sustained diet intervention strategy only, physical activity habits during the study will be used as covariate, hence mitigating the confounding of prior physical activity, at least to some extent.
A successful emulation of randomization requires that there is no unmeasured confounding. However, this is never guaranteed with observational data. Thus, we emphasize our assumptions that 1) the causal graph accurately depicts the relationship under study; and 2) the covariates included are a sufficient set of covariates to address confounding.
Outcomes
The primary outcomes will be the mean end of follow-up values for muscle strength (i.e., handgrip using vigorimeter, elbow flexor, knee extensor) and physical function (i.e., normal and fast walking, “timed up-and-go”). Outcome values were measured according to standardized protocol in NuAge (Gaudreau et al. 2007).
For secondary outcomes, mean end of follow-up values for a set of relevant variables will be considered by domains:
General health: waist circumference, blood pressure (systolic, diastolic), blood glucose, estimated glomerular filtration rate;
Cognition: the modified Mini-Mental State Examination (3MS) score.
Time Zero and Follow-up
In the target trial of a sustained lifestyle intervention, participants would be met at baseline and then regularly to ensure that diet and physical activity habits are consistent with the intervention assigned by the random allocation. The hypothetical diet and physical activity intervention would be assigned and initiated at baseline. In the emulation, annual follow-ups with comprehensive diet, physical activity and covariate data collection are available to emulate the hypothetical intervention. Hence, participants would be followed from study baseline (time 0; i.e., the time at which the intervention strategy would also be assigned and would begin), at each year (time 1 and 2) and until the end of the study (time 3). We also assume that the diet and physical activity habits measured at each follow-up time adequately reflect the habits during the full year.
The end of follow-up outcome measurements will be used to estimate the effect of the sustained lifestyle intervention strategy. Measures of dietary intakes and physical activity throughout the study (i.e., time 0 to time 3) will be used to emulate the sustained lifestyle intervention. Dietary intakes in the year prior to the intervention will be estimated using the frequency questionnaire completed at study baseline (time 0). xMissing covariate data at a given follow-up will be carried forward once, after which participants will be considered as having incomplete follow-up.
Causal Contrast
The estimand of interest in this study, the target causal effect of a sustained lifestyle intervention strategy, is
that is, the expected value of a given health outcome Y at the end of follow-up if all participants had increased their adherence to CFG recommendations on healthy food choices and physical activity, when applicable, at all three time points (Xk = 1, “always intervene”) vs., instead, if all participants had maintained their habitual diet and physical activity (Xk = 0, “never intervene”). The estimand also indicates that all participants completed the intervention (C = 0), i.e., in absence of incomplete follow-up.
The causal contrasts of interest are the observational analogues of “intention-to-treat” and “per protocol” (Hernan and Robins 2016). Given the observational design, participants are not expected to have followed a treatment strategy unknown to them at the time of data collection. Therefore, the primary analysis will be the per protocol contrast of a sustained lifestyle intervention strategy. In the per protocol analysis, non-adherence to the hypothetical interventions can be accounted for. In the target trial, participants with a condition after baseline that would have prevented or limited participation in a hypothetical lifestyle intervention would be allowed to discontinue the intervention (e.g., lengthy hospitalization, prolonged bed rest, incident cancer). In the emulation, if such conditions occur in a sufficiently large number of participants, these participants will be “excused” from following the hypothetical intervention (Dickerman et al. 2019). In other words, participants that would have been unable to pursue the study due to major events will not be considered as having incomplete follow-up if they attended the annual assessment. Allowing participants to discontinue adhering to the (hypothetical) intervention strategy mitigates confounding by the disease burden (Dickerman et al. 2019).
The intention-to-treat analysis will be a secondary analysis of a hypothetical point intervention, e.g., dietary counselling at baseline only. Note that it will not be possible to conduct an intention-to-treat analysis identical to that of a controlled study where the interest is to estimate the effect of being assigned to an intervention (Hernan and Robins 2016). However, it is possible to conduct an “observational analogue” of the intention-to-treat analysis. In the observational analogue, the intention-to-treat analysis aims to estimate the impact of a hypothetical intervention which adherence is measured at baseline only.
Statistical Analysis
For the per protocol analysis, non-adherence to the hypothetical interventions and incomplete follow-up will be considered using g methods (Hernan and Robins 2016). G-methods and the parametric g-formula in particular allow for adequate consideration of the time-varying confounding in the presence of feedback between the intervention and the confounding variables. In other words, these methods allow estimation of a total causal effect when the time-varying confounding variables (e.g., body weight or medication) are affected by a previous intervention (i.e., the hypothetical sustained lifestyle intervention). Time-varying covariates will be: major health problems, number of medications, dietary supplement use, physical activity, smoking and weight.
In the context of an observational study, the (parametric) g-formula is used with parametric models to predict the joint-history of prior diet and physical activity habits (i.e., the hypothetical sustained intervention strategy) and confounding variables. For example, linear regression models are used to predict continuous covariates (e.g., body weight), while logistic regressions are used to predict binary or categorical variables (e.g., indicator variable for dietary supplement use). The per protocol causal contrast of hypothetical intervention presented in Table 2 is then emulated based on Monte Carlo simulated data generated using the g formula algorithm (McGrath et al. 2020). More precisely, “threshold interventions” that depend on the reported dietary intakes or amount of physical activity (Taubman et al. 2009; Young et al. 2014) and the parametric g-formula (Robins 1986; McGrath et al. 2020) will be used to emulate the intervention of “consuming at least servings of food” and “doing at least minutes of light to vigorous physical activity”.
Figure 2 presents the causal DAG of the hypothesized relationship between a sustained lifestyle intervention strategy ( ) and an end of follow-up outcome, for one follow-up after baseline (year 1). The model is limited to year 1 for clarity but the hypothesized causal structure extends to additional follow-ups. The exposure of interest is the joint and cumulative effect of a sustained diet and physical activity intervention strategy measured at baseline and at follow-ups.
In the context of this target trial emulation (Figure 2), includes dietary habits prior to the baseline assessment. can have an effect on baseline dietary habits (e.g., prior healthy habits increase likelihood of baseline healthy habits) and dietary habits throughout the target trial emulation (e.g., prior healthy habits increase likelihood of adhering to healthy habits). also influences baseline and time-varying confounding. Finally, given the long-term effect of chronic exposure, potentially also affect directly.
The intention-to-treat analysis is similar to the per protocol analysis. However, only baseline diet and physical activity habits and covariates are considered as well as pre-baseline diet and physical activity. In both the per protocol and intention-to-treat analyses, loss to follow-up (e.g., non-response or missing follow-up, health outcomes not measured) will be accounted for using g-methods such as inverse probability weighting.
Dietary Assessment
In the NuAge study, diet during the year before baseline was assessed using one semi-quantitative food-frequency questionnaire (Shatenstein et al. 2005). Dietary intakes at baseline (time 0) and each annual follow-ups (time 1, 2 and 3) were assessed using 3 repeated face-to-face interviewer-administered 24-hour dietary recall.
Dietary intakes measured using 24-h dietary recall are more accurate (i.e., less systematic error or bias) than food-frequency questionnaire (Freedman et al. 2014, 2015; Thompson et al. 2015) but are particularly affected by random measurement error (i.e., within-individual random error) (Thompson et al. 2015). When several variables measured with errors are considered simultaneously in a regression model, the regression coefficients may be biased in any direction (Keogh et al. 2020). To account for random measurement error, the National Cancer Institute Markov Chain Monte Carlo (MCMC) multivariate method could be applied (Zhang et al. 2011). However, the combination of the parametric g-formula and multivariate measurement error correction using the NCI MCMC method is not feasible. The NCI MCMC method estimates time-invariant measurement error-corrected intakes, while the g formula algorithm is designed for time-varying exposures.
Recognizing the importance of accounting for measurement error, the correction for measurement error will be reserved for the secondary intention-to-treat analysis. For the intention-to-treat contrast, the time-varying values of exposure and the time-varying values of confounding are not considered. Thus, the NCI MCMC method will be used to obtain measurement error-corrected estimates of the relationship between dietary intakes measured at baseline and outcome at the end of follow-up. Of note, the three 24-h dietary recalls collected at each time point contribute to reducing random errors, at least to some extent, even in absence of measurement error correction.
Sensitivity analysis to assess the impact of measurement error
For the secondary intention-to-treat analysis, results based on the measurement error-corrected and uncorrected dietary intakes will be compared. The difference between the estimated relationships will allow to extrapolate the impact of not accounting for random errors in the primary per protocol analysis.
Physical activity assessment
Physical activity throughout the study was assessed using the Physical Activity Scale for the Elderly (PASE) questionnaire (Washburn et al. 1993, 1999). Rather than the total PASE score, specific questions estimating the total time of physical activities will be used to be consistent with the intervention strategy.
Covariates and subgroups
To the extent permitted by the number of observations for each outcome, continuous covariates will be modelled using restricted cubic splines with 3 to 5 knots (at percentiles 10-50-90, 5-35-65-95 or 5-27.5-50-77.5-95) (Kyle et al. 2019). Categorical covariates will be modelled to ensure a sufficient sample size in each level.
The effect of dietary changes will be estimated for the full sample. The sample will also be stratified by (biological) sex to reflect both potential biological differences and, to some extent, gender differences (although not reported).
Variance estimation
Variance will be estimated using a minimum of 200 bootstrap sample replicates to consider uncertainty at each step of the estimation (Bland and Altman 2015).
Software and code
The main statistical analyses will be conducted using R (version 4.3.1 or greater) and the gfoRmula package (McGrath et al. 2020; Lin et al. 2023). The manuscript results will be generated using Quarto markdown. Codes for main analyses and generation of manuscript results will be shared in a publicly available code repository.
Discussion
In this study protocol, we have described a target trial to assess the effect of adhering to CFG’s recommendations on healthy food choices. The emulation will be performed using data from the NuAge Database and Biobank (Database and Biobank 2024). Benefiting from the flexibility of observational data, we also aim to compare adherence to multiple reformulations of CFG’s recommendations, including the effect of increasing the intake of protein-rich foods and the amount of aerobic physical activity on selected health outcomes. We have also described the rationale for using simulated diets to emulate the adherence to CFG’s recommendations, the process of selecting covariates - to attempt - to emulate randomization with causal diagrams, and the challenges of addressing random measurement error.
We emphasize that the purpose of emulating a target trial using observational data is to improve the quality of observational analysis (Hernan and Robins 2016; Hernan et al. 2022). In other words, the target trial framework aims to support the coherence between the (causal) research question and the observational data analysis (Hernan 2018). However, estimating causal effects with non-experimental observational data depends on strong assumptions. The key assumptions are that there are no unmeasured confounder, no measurement error and no model misspecification (e.g., functional form of covariates, model outcome distribution) (Chiu et al. 2021). We first recognize that the absence of residual or unmeasured confounding cannot be guaranteed. The extent to which this assumption is sufficiently satisfied depends on the appraisal of covariates considered. In that regard, we have used graphical tools, DAG, to explicitly describe our analytical assumptions and to identify confounding variables (Greenland et al. 1999; Tennant et al. 2021; Lipsky and Greenland 2022). Second, the absence of measurement error assumption will not be satisfied considering the use of dietary intake data measured with 24-h dietary recalls. On one hand, 24-h dietary recalls have the least systematic error (bias) compared with other common instruments such as food-frequency questionnaire (Freedman et al. 2014, 2015). On the other hand, 24-h dietary recalls are largely affected by within-individual random errors (Thompson et al. 2015), which can cause bias in any direction in multivariable models (Keogh et al. 2020), as in the present study. This issue is mitigated, at least to some extent, by using average data from three repeated 24-h dietary recalls at each follow-up. Sensitivity analyses comparing estimates based on measurement-error corrected and uncorrected baseline dietary intakes will be used to assess the impact of random measurement errors. Third, the absence of model misspecification will be assessed by examining differences between the observed value of time-varying covariates and the predicted value of time-varying covariates as modeled with the g-formula.
Strengths of this study and protocol include the explicit emulation of a hypothetical trial, the thorough description of the emulation of the sustained dietary intervention, and the use of background knowledge and DAG to derive a sufficient set of confounders. Limitations must be addressed. First, the sample size of the NuAge study is relatively limited (n=1753), although comprehensive nutrition and covariate data was collected. Second, the target food intakes based on diet simulations from Health Canada exceed the 99th percentile of the usual intake distribution of Canadians 65 years or older from Canada in 2015 (Brassard and Chevalier 2023). The a posteriori revision of the dietary intervention targets will be needed if observed dietary intakes in NuAge are too far from targets (Table 2), as was done in a previous nutrition target trial emulation (Ibsen et al. 2023). Third, the presence of random measurement error associated with 24-h dietary recalls may bias estimates. Finally, the target trial emulation cannot replace an actual RCT. Evidence from RCT will be required to confirm the value of either CFG recommendations or the enhanced CFG recommendations.
In conclusion, the target trial framework is relevant to estimate the causal effect of adhering to CFG’s recommendations using non-experimental data when a RCT is impractical (Hernan and Robins 2016; Chiu et al. 2021). Coupled with key assumptions, including the absence of unmeasured confounding, the absence of measurement error and no model misspecification, we believe the emulation will provide timely evidence regarding the effect of adhering to CFG’s recommendations in older adults and inform on the added value of a reformulation.
Data availability and code
The data and code used for this manuscript will be made available at https://github.com/didierbrassard/NuAge_protocol
Acknowledgements
Conflict of Interest and Funding Disclosure
DB was a casual employee of Health Canada (2019-2020), held a doctoral training award from the Fonds de recherche du Québec – Santé (2019-2021) and holds a Canadian Institutes of Health Research (CIHR) Fellowship award (MFE-181852) to conduct the present study. SC receives research funding from the CIHR, Fonds de recherche du Québec, Canadian Foundation for Dietetics Research, Canadian Foundation for Innovation and Canadian Cancer Society. None of these agencies has funded nor was involved in this work. NP is the NuAge Database Administrator; NP and SC serve as NuAge Steering Committee Members; DB has no conflicts of interest.
The NuAge Study was supported by a research grant from the Canadian Institutes of Health Research (CIHR; MOP-62842). The NuAge Database and Biobank are supported by the Fonds de recherche du Québec (FRQ; 2020-VICO-279753), the Quebec Network for Research on Aging, a thematic network funded by the Fonds de Recherche du Québec - Santé (FRQS) and by the Merck-Frosst Chair funded by La Fondation de l’Université de Sherbrooke.
References
