Analyzing Synthetic Patient Data: What Drives HbA1c?

§ · The question, set first

Before touching the data we fixed one outcome to test: of age, body weight, and household income, which actually move a person's HbA1c — and is there any relationship at all? HbA1c is the standard three-month blood-sugar marker; clinically it climbs with age and adiposity, and the diagnostic bands are well defined (normal below 5.7%, prediabetes 5.7–6.4%, diabetes 6.5% or higher). A calibrated synthetic dataset should reproduce those relationships. If it does, you can prototype an analysis on it and expect the same directions you would see in real patient data.

§ · The data, and how each field varies

Five thousand synthetic US adults, generated at seed 42 and loaded into pandas:

import pandas as pd
df = pd.read_csv("persons.csv")            # 5,000 rows, 77 columns
df[["age", "weight_kg", "bmi", "a1c_value"]].describe()

The mean BMI of 29.5 matches the US adult average from NHANES almost exactly, and HbA1c is right-skewed (median 5.5%, a long tail to 12.9%) — the shape of a population that is mostly non-diabetic with a diabetic minority.

Income is captured as a bracket, not a raw salary, so we look at its distribution rather than a mean:

§ · Does anything move HbA1c?

Correlate each candidate against HbA1c. Age, weight, and BMI are continuous, so we use Pearson; income is an ordinal bracket, so we rank it 1–5 and use Spearman:

rank = {"under25k":1, "r25to50k":2, "r50to100k":3, "r100to150k":4, "over150k":5}
df["income_rank"] = df["household_income_bracket"].map(rank)

df[["a1c_value", "age", "bmi", "weight_kg"]].corr()["a1c_value"]           # Pearson
df[["a1c_value", "income_rank"]].corr(method="spearman").iloc[0, 1]        # Spearman

Weight tracks HbA1c a little less tightly than BMI, exactly as expected — BMI normalizes weight for height, so it is the cleaner signal. Income is flat: Spearman +0.004.

§ · Which one drives it, holding the others fixed?

Simple correlations can be confounded, so we fit a standardized regression — HbA1c on age, BMI, and income together, all z-scored, so the coefficients are directly comparable:

from sklearn.linear_model import LinearRegression
from sklearn.preprocessing import StandardScaler
X = StandardScaler().fit_transform(df[["age", "bmi", "income_rank"]])
y = StandardScaler().fit_transform(df[["a1c_value"]]).ravel()
LinearRegression().fit(X, y).coef_        # -> [0.281, 0.219, -0.036]

Model R² = 0.13 — age and BMI explain a real slice of HbA1c; the rest is clinical status and individual variation. The partial effects barely differ from the raw correlations because age and BMI are nearly independent here (r ≈ 0.01), so neither steals the other's credit. Their pushes on HbA1c simply add up.

§ · The gradients, in plain numbers

Correlations hide the shape. Group means show it — and confirm the direction is monotonic, not noise:

df.groupby(pd.cut(df.age, [17, 34, 49, 64, 120]))["a1c_value"].mean()
df.groupby(pd.cut(df.bmi, [0, 25, 30, 100]))["a1c_value"].mean()
df.groupby("diabetes_status")["a1c_value"].mean()

Both climb cleanly, step by step. By income bracket, HbA1c is flat — 5.80, 5.76, 5.77, 5.73, 5.76 from lowest to highest. No gradient.

And grouped by clinical status, the means fall inside the correct diagnostic bands — the sharpest fidelity check of all:

The diabetic group averages 7.44% against 5.32% for non-diabetics — a 2.1-point gap, and every mean sits in its clinically correct range.

§ · The outcome

HbA1c is driven by age, body mass, and clinical status — not by income. Age (β +0.28) and BMI (β +0.22) each raise it independently and additively; income moves it essentially not at all (β −0.04, flat group means). The diabetic subgroup runs 2.1 points higher, and every group mean lands in its correct diagnostic band.

That is exactly what real US epidemiology says: HbA1c rises with age and adiposity, the diabetes and prediabetes prevalences here (14.8% and 36.6%) match CDC figures, and the clinical gradients hold. The signal a real analyst would find in NHANES is present in the synthetic data — which is the whole point of calibration. Prototype your query here, and the relationships carry over.

One honest caveat. Generation only reproduces the relationships it models. Here age and BMI are drawn nearly independently (r ≈ 0.01), whereas real cohorts show a mild age–BMI correlation — so a study that hinges on that specific interaction should validate against real data first. Calibrated synthetic data is for realistic structure and privacy-safe iteration, not for discovering a correlation nobody built in. We cover that line in what is synthetic data →

§ · Reproduce it

Generation is deterministic, so this command yields the exact 5,000 rows analyzed above:

# with the person CLI
person-cli -n 5000 -s 42 -f csv > persons.csv

# or pull the same volume from the API — see the Quickstart
#   POST /v1/datasets/person  { "clientId": "...", "count": 5000, "seed": 42 }

New to the API? Start with the developer Quickstart →

Developer Quickstart →Person Profile generator →NHANES-calibrated data →What is synthetic data →