📊 statistics

Coarsened data in small area estimation: a Bayesian two-part model for mapping smoking behaviour

This paper proposes a Bayesian two-part unit-level Small Area Estimation framework that explicitly models coarsening mechanisms to improve the accuracy and reliability of smoking prevalence and intensity estimates for Italian regions and age groups, demonstrating through simulations and empirical application that ignoring such data limitations leads to biased results.

Original authors: Aldo Gardini, Lorenzo Mori

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aldo Gardini, Lorenzo Mori

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a perfect photo of a crowded room to count how many people are smoking and how much they are smoking. But there's a problem: the camera is a bit blurry, and the people in the room are also a bit fuzzy about their own habits.

This paper is about fixing that blurry photo to get a clear picture of smoking habits across different regions and age groups in Italy. Here is how the authors did it, explained simply:

The Problem: The "Fuzzy" Survey

The researchers used a big national survey (the EHIS) to ask people, "Do you smoke?" and "How many cigarettes do you smoke a day?"

But the data they got back was messy in two specific ways:

The "Rounding" Effect: People aren't good at math when they are tired or just guessing. If they smoke 12 cigarettes, they might say "10." If they smoke 17, they might say "20." It's like trying to measure a table with a ruler that only has marks for 5, 10, and 15 inches. You end up with a pile of answers at those "nice" numbers, and the real numbers in between get lost.
The "Cutoff" Effect: The survey had a rule: "If you smoke more than 20, just say 20." So, someone smoking 25 and someone smoking 40 both get recorded as "20." This hides the truth about heavy smokers.

The Double Trouble

Usually, statisticians have two tools to fix bad data:

Tool A (Small Area Estimation): This is like using a telescope. If you don't have enough people in a small town to get a good answer, you look at the neighboring towns and the whole country to "borrow strength" and make a smart guess for the small town.
Tool B (Fixing the Blur): This is like using photo-editing software to sharpen the image and guess what the real numbers were behind the rounding.

The problem is that most statisticians use Tool A but forget Tool B. They try to sharpen the image after they've already tried to guess the small town numbers, or they ignore the blur entirely. The authors say, "If you ignore the blur, your guess for the small town will be wrong, and you won't even know how wrong it is."

The Solution: A "Two-Part" Detective Kit

The authors built a new statistical model that acts like a two-part detective kit. They split the problem into two separate cases:

Case 1: The "Yes/No" Detective (Do they smoke?)
First, they figure out who smokes at all. This is a simple "Yes" or "No" question. Since people are usually honest about whether they smoke or not, this part is straightforward. They use the "telescope" method (borrowing data from neighbors) to get a stable answer for every region and age group.

Case 2: The "How Much" Detective (How much do they smoke?)
This is the hard part. Here, they have to deal with the "fuzzy" numbers (the rounding and the 20-cigarette cutoff).

The "Ghost" Variable: They imagine a "ghost" number that represents the real amount of cigarettes a person smokes. This ghost number is smooth and continuous (like 12.4 or 17.8).
The "Heaping" Mechanism: They built a rulebook to explain how the real ghost number turns into the fuzzy survey number. They assume people have different "rounding styles": some round to the nearest whole number, some to the nearest 5, and some to the nearest 10.
The "Mixture" Model: They realized that smokers aren't all the same. Some are light smokers, and some are heavy smokers. So, they didn't just use one average curve; they used a "mixture" of two different curves (like mixing two different colors of paint) to capture the fact that there are two distinct groups of smokers.

What They Found

The authors ran a simulation (a practice run with fake data) to see what happens if you ignore the "fuzzy" rounding.

The Result: If you ignore the rounding, your estimates are like a map with the wrong landmarks. You might think there are fewer heavy smokers than there really are, and your "confidence intervals" (your safety net) are too small, making you feel more sure than you should be.
The Fix: Their new model, which accounts for the rounding and the two types of smokers, gave much more accurate maps. It correctly identified where the heavy smokers were and gave a realistic range of uncertainty.

The Real-World Picture (Italy)

When they applied this to real data from Italy, they found some interesting patterns:

Age Matters: Younger people smoke less often, but if they do smoke, they might smoke heavily. The 50–64 age group had the highest intensity of smoking.
Geography Matters:
- Southern Italy: Had fewer smokers overall, but those who did smoke tended to smoke more heavily.
- Northern Italy: Had more smokers overall, but they tended to smoke less heavily on average.
The "Campania" Example: In the Campania region, fewer young people smoke compared to others, but the ones who do smoke are very heavy smokers.

The Bottom Line

This paper teaches us that when we ask people to count things like cigarettes, they will round their answers. If we want to understand health trends in small towns or specific groups, we can't just take the numbers at face value. We need a model that understands how people round their numbers and how they group them. By doing this, we get a clearer, more honest picture of public health, which helps leaders make better decisions on where to focus their resources.

Technical Summary: Coarsened Data in Small Area Estimation: A Bayesian Two-Part Model for Mapping Smoking Behaviour

Problem Statement
Public health monitoring frequently relies on self-reported survey data, which are inherently subject to measurement errors such as recall bias, rounding, and systematic digit preference (heaping). These coarsening mechanisms reduce data resolution and can introduce significant bias into domain-level estimates. Furthermore, when estimating indicators for small geographic areas or specific socio-demographic subgroups, sample sizes are often insufficient for reliable direct estimation. This paper addresses the dual challenge of limited sample sizes and data coarsening in the context of estimating smoking indicators (prevalence, intensity, and heavy smoking proportion) across Italian regions and age groups using the 2019 European Health Interview Survey (EHIS). Standard Small Area Estimation (SAE) methods, while effective for borrowing strength across domains, typically treat reported values as error-free, thereby failing to account for the distortion caused by heaping and top-coding (e.g., grouping all values above 20 cigarettes).

Methodology
The authors propose a Bayesian unit-level SAE framework designed specifically for semi-continuous, coarsened responses. The model is structured in two distinct parts:

Part One (Prevalence): A logistic mixed-effects model estimates the probability of an individual being a daily smoker ( $W$ ). This component uses a standard logistic regression with domain-specific random effects to handle the binary nature of the data.
Part Two (Intensity): A flexible mixture model estimates the average daily cigarette consumption ( $Z$ $Z$ ) among daily smokers.
- Latent Structure: The model posits that the observed integer count ( $Z^*$ ) is a coarsened realization of an underlying latent continuous variable $Z$ .
- Coarsening Mechanism: An explicit latent variable $G$ is introduced to model the heaping behavior. $G$ represents the rounding regime (nearest integer, nearest multiple of 5, or nearest multiple of 10). The probability of $G$ depends on the latent intensity $Z$ via a proportional odds model, allowing the coarsening mechanism to be non-ignorable (i.e., dependent on the value of $Z$ ).
- Distributional Assumption: The latent intensity $Z$ is modeled using a two-component Lognormal Mixture (LNM). This "Mixture of Experts" approach allows component-specific intercepts and covariate effects, capturing heterogeneity in consumption patterns that a single Lognormal distribution cannot.
- Top-Coding: The model explicitly accounts for right-censoring where values $\ge 20$ are recorded as a single category (21 in the survey).
- Inference: The model is fitted using Hamiltonian Monte Carlo (HMC) via Stan. The authors derive sufficient conditions on the prior distributions of variance components to ensure the existence of posterior moments for the target indicators, addressing a known technical issue with Lognormal likelihoods.

Key Contributions

Integration of Coarsening into SAE: The paper bridges a gap in the literature by embedding an explicit coarsening mechanism within a unit-level SAE framework. Unlike previous works that focus on individual-level outcomes or ignore heaping in domain estimation, this approach jointly estimates the small area parameters and the coarsening process.
Two-Part Mixture Specification: The model accommodates the semi-continuous nature of the data (mass at zero) and the heterogeneity of positive values through a two-component Lognormal mixture, coupled with a covariate-dependent mixing probability.
Theoretical Guarantees: The authors provide formal conditions (Propositions 1 and 2) regarding prior specifications that guarantee the finiteness of posterior moments for the target indicators, a critical requirement for valid Bayesian inference in this context.
Handling Non-Ignorable Coarsening: By modeling the dependence of the rounding mechanism on the latent intensity, the approach corrects for biases that arise when heaping is informative.

Results

Simulation Study: A model-based simulation study compared the proposed coarsening-aware mixture model (LNM-C) against alternatives: a standard Lognormal model (LN), a Lognormal model with coarsening (LN-C), and a mixture model ignoring coarsening (LNM).
- When coarsening depends on the latent value (non-ignorable), ignoring it leads to biased point estimates, inflated Root Mean Square Error (RMSE), and severe undercoverage of credible intervals, particularly for the proportion of heavy smokers ( $HS_d$ ).
- The LNM-C model consistently achieved near-nominal coverage (approx. 90%) and lower bias/RMSE across scenarios, demonstrating robustness even when heaping patterns were complex.
Empirical Application (Italy): Applied to 84 region-age domains:
- Model Fit: Posterior predictive checks confirmed that the LNM-C model closely tracked the empirical distribution of reported cigarette counts, including the heaping at multiples of 5 and 10 and the top-coding at 20, whereas models ignoring coarsening showed significant deviations.
- Estimates: The analysis revealed distinct age and regional patterns. Smoking prevalence peaked in the 25–49 age group, while smoking intensity and the proportion of heavy smokers increased with age, peaking in the 50–64 group.
- Regional Dynamics: While prevalence showed no strong spatial clustering, smoking intensity was higher in Southern Italy and lower in the North-East. The model provided stable estimates for domains with small sample sizes where direct estimators were unreliable.

Significance and Claims
The paper claims that ignoring data coarsening in small area estimation can yield biased and unstable domain estimates with poor interval coverage, particularly for indicators sensitive to the tail of the distribution (e.g., heavy smoking). By explicitly modeling the rounding and top-coding mechanisms within a Bayesian two-part SAE framework, the proposed approach improves accuracy and achieves near-nominal coverage. The empirical application demonstrates that this methodology provides a detailed and reliable picture of smoking patterns across region–age domains, thereby supporting more targeted public health policies. The authors emphasize that while the model is non-trivial, the release of the Stan code facilitates its adaptation for other applications involving heaped or top-coded self-reported data. The work does not claim to solve all measurement error issues but specifically targets the bias introduced by known coarsening mechanisms in the context of small area inference.