Dietary Carbohydrate Quality and Physical Functioning

A Longitudinal Analysis in the New England Centenarian Study

Tools: SAS 9.4 · R · GEE · Multi-source data harmonization
Data: New England Centenarian Study (NECS) · N = 438 · Longitudinal
Skills demonstrated: Novel exposure engineering · ETL pipeline · GEE macro development · multi-dataset harmonization · stratified longitudinal modeling · R visualization

---

Overview

This project examines the relationship between dietary carbohydrate quality and physical functioning in older adults using longitudinal data from the New England Centenarian Study (NECS) — one of the world's largest studies of centenarians and their family members.

Six dietary carbohydrate quality exposures were derived from food frequency questionnaire (FFQ) data and linked to Instrumental Activities of Daily Living (IADL) scores across multiple study visits. Associations were estimated using Generalized Estimating Equations (GEE) across three progressively adjusted models and stratified by sex, diabetes status, and chronic disease burden.

This was conducted as independent graduate research at the Jean Mayer USDA Human Nutrition Research Center on Aging, under the supervision of Dr. Andres Ardisson Korat (May 2026).

---

Research questions

1. Are dietary carbohydrate quality measures associated with longitudinal IADL functioning in older adults?
2. Do associations differ by sex, diabetes status, or chronic disease burden?
3. Is carbohydrate quality (fiber density, GI, GL, carb:fiber ratio) more predictive than carbohydrate quantity alone?

---

Data

• Source: New England Centenarian Study (NECS), Boston University
• Access: Restricted — available via formal data use agreement (see /data/data_source.md)
• Analytic sample: N = 438 after exclusion criteria
• Structure: Longitudinal, multiple visits per participant

Datasets harmonized: FFQ · Nutrient database · Cereal standardization file · Demographics · IADL outcome data

---

Exposure development

A novel carbohydrate quality index was engineered from raw FFQ data through a multi-stage pipeline:

1. Cereal standardization — Free-text cereal entries were systematically reviewed and assigned to 11 standardized groups based on composition and processing characteristics. GI and GL values were assigned using the University of Sydney Glycemic Index Research Database.

2. Food-level calculations — For each FFQ item: daily carbohydrate intake computed, GI assigned, GL calculated as GL = (GI × carbohydrates) / 100

3. Participant-level exposure derivation:

   • Dietary glycemic index (carbohydrate-weighted average GI)
   • Dietary glycemic load (sum of food-level GL)
   • Total carbohydrate intake (g/day)
   • Total dietary fiber (g/day)
   • Fiber density (g fiber per 1,000 kcal)
   • Carbohydrate-to-fiber ratio (lower = higher quality)

4. Energy adjustment — All exposures except fiber density were energy-adjusted using the residual method

See /code/03_exposure_construction.sas for full derivation code.

---

Statistical methods

Model: Generalized Estimating Equations (GEE)

• Normal distribution, identity link
• Exchangeable working correlation structure
• Cluster variable: participant ID
• Time variable: year of visit − year of enrollment

Model progression:

• Model 1: time + exposure + age + sex + education
• Model 2: Model 1 + CVD + stroke + diabetes + cancer
• Model 3: Model 2 + smoking + alcohol use

Each exposure modeled separately. Stratified analyses by sex, diabetes status, and chronic disease burden using Model 3 specification. A reusable SAS macro was developed to run all three model specifications efficiently across all six exposures (see /code/04_gee_models_macro.sas).

---

Key findings

• Fiber density was significantly associated with better IADL functioning in the fully adjusted model (β = 0.70, p < 0.001)
• Carbohydrate-to-fiber ratio (higher = lower quality) was significantly associated with lower IADL scores (β = −0.19, p < 0.001), consistent across model adjustments
• Glycemic index showed no meaningful association across models
• Carbohydrate quantity alone was not a significant predictor after full adjustment — quality matters more than quantity
• Sex-stratified analyses showed associations were significant among men but not women
• Associations were strongest among participants with existing chronic disease, potentially due to ceiling effects in the disease-free group (high-functioning cohort)

---

Analytic pipeline

Raw NECS datasets (FFQ, nutrients, cereal, demographics, IADL)
        ↓
01_data_cleaning.sas        — variable recoding, exclusions, QC
        ↓
02_cereal_processing.sas    - cleaning cereal FFQ responses, creating participant-level cereal GI/GL dataset
        ↓
03_gi_gl_construction.sas   — constructing dietary exposures, computing total GL/weighted GI
        ↓
04_dataset_merging_cleaning.sas — merge datasets to create analytic dataset
        ↓
05_exposure_construction.sas    - residual energy adjustment, quartiles, diagnostics
        ↓
06_gee_visualizations.R         — gee models and figures using ggplot2

---

Repository structure

├── code/
│   ├── 01_data_cleaning.sas
│   ├── 02_cereal_processing.sas
│   ├── 03_gi_gl_construction.sas
│   ├── 04_dataset_merging_cleaning.sas
│   ├── 05_exposure_construction.sas
│   └── 06_gee_visualizations.R
├── outputs/
│   ├── figure1_gee_maineffects_model3.png
│   ├── figure2_stratified_by_sex.png
│   ├── figure3_correlation_heatmap.png
│   ├── figure4_distribution_carbtofiber.png
│   └── figure5_predictedIADL_carbtofiber.png
├── docs/
│   ├── Wallen_NECS_CarbQuality_Report.pdf
│   └── data_dictionary.md
└── data/
    └── data_source.md

---

Visualizations

Figure 1 — GEE model results: carbohydrate quality exposures vs. IADL (fully adjusted)

This forest plot displays beta coefficients and 95% confidence intervals from fully adjusted generalized estimating equation (GEE) models examining associations between carbohydrate quality measures and longitudinal Instrumental Activities of Daily Living (IADL scores.

The carbohydrate to fiber ratio demonstrated the strongest inverse association with IADL function, whereas fiber density showed a modest poitive association. GI and GL estimates were closer to the null with wider confidence intervals.

Figure 2 — Sex-Stratified GEE Associations with IADL

This stratified forest plot compares fully adjusted GEE model estimates seperately among male and female participants. The visualization highlights potential sex-specific differences in the relationship between dietary carbohydrate quality and functional aging outcomes.

Across several exposures, male participants showed larger magnitude effect estimates and widen confidence intervals, suggesting greater variability is associations within stratified models.

Figure 3 — Correlations Among Dietary Variables

This correlation heatmap displays Pearson correlation coefficients among major dietary carbohydrate quality variables included in the longitudinal modeling workflow. This visualization helps to assess potential collinearity among exposure variables prior to regression modeling.

The strongest inverse correlation was observed between fiber density and carbohydrate:fiber ratio (r = -0.66), reflecting their conceptual overlap as indicators of carbohydrate quality. Glycemic load and total carbohydrate intake were highly positively correlated (r = 0.91), while glycemic index demonstrated relatively weaker correlations with the remaining variables.

Figure 4 — Distribution of Carbohydrate to Fiber Ratio

This density plot shows how the carb:fiber ratio is distributed across the study sample. Quartiles were created to model predicted IADL score and functional decline across the follow-up. The three dashed red vertical lines are the quartile cut-points — they divide the sample into four equally sized groups, from Q1 on the left (best carbohydrate quality, lowest ratio) to Q4 on the right (worst quality, highest ratio).

The distribution is right-skewed, with most of the participants clustering in the moderate quality range, but there's a meaningful tail of individuals with very high carb:fiber ratios. These are the Q4 participants who consume a lot of carbohydrates relative to fiber.

Figure 5 — Functional Decline by Carbohydrate Quality

This line plot shows predicted IADL trajectories extracted from a linear mixed-effects model, adjusted for age, sex, and education. The lines here are smoothed and represent the average expected trajectory for each group, holding all other variables constant.

All four quartile groups begin at the same IADL level confirming there were no meaningful baseline differences. Over time, the lines fan out and by year 10, Q4 — the group with the worst carbohydrate quality — has declined nearly a full additional IADL point compared to Q1.

---

Author

Olivia Wallen
MS, Nutrition Epidemiology and Data Science · Tufts University (2026)
Graduate Researcher, Jean Mayer USDA HNRCA
LinkedIn · Email