📊 statistics

Dynamic Mini Max Design and Sequential HB Inference for Repeated Surveys

This paper introduces a Dynamic Mini-Max (DMM) framework for repeated surveys that jointly optimizes sample size and wave overlap to reduce costs by approximately 6.3% while ensuring superior precision for both population levels and movements compared to classical designs, as demonstrated through Australian Census data and simulations.

Original authors: Siu-Ming Tam

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Siu-Ming Tam

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 the manager of a massive, ongoing survey, like a national census or a monthly employment check. Your job is to ask thousands of people questions to get a snapshot of the country's economy. You have two main goals:

The Snapshot: How many people are employed right now?
The Movie: How has employment changed since last month?

The paper introduces a new, smarter way to run these surveys called Dynamic Mini-Max (DMM). Think of it as a "Goldilocks" strategy that finds the perfect balance between asking enough people to get accurate answers and not wasting money or annoying too many people.

Here is how the paper explains this new system, using simple analogies:

1. The Old Way vs. The New Way

The Old Way (Classical Design):
Imagine you are taking a photo of a crowd every month. To get a good picture, you hire a fixed number of photographers (a fixed sample size). To see how the crowd moved, you try to keep some of the same photographers from last month to take the next photo.

The Problem: This method is rigid. It often hires too many people (wasting money) or doesn't keep enough of the same people to accurately track movement. Also, the old math used to calculate "how much things changed" often ignores the fact that the crowd itself is changing randomly, leading to confidence intervals that are too narrow (overconfident).

The New Way (Dynamic Mini-Max):
This is a smart, adaptive system. It doesn't just pick a number of people; it calculates the exact number needed and the exact number of people to keep from the previous month.

The "Mini-Max" Name: Think of this as a tightrope walker. The walker wants to minimize the "Max" risk (error) while minimizing the "Cost" (money). The system constantly adjusts the rope to ensure the error never gets too big, but the cost stays as low as possible.

2. The Two-Part Engine

The DMM framework runs on two main gears working together:

Gear A: The Smart Planner (Dynamic Mini-Max Design)

Before the survey starts, this planner looks at the budget and the rules. It asks: "If I interview 40,000 people instead of 42,000, and I keep 90% of the same people from last month, will I still get a good answer?"

The Result: In the paper's test, this planner found that they could cut the sample size by about 4% (saving roughly 6% in total costs) while still meeting all the accuracy rules. It's like realizing you don't need to fill the whole stadium to get a good view of the game; you just need the right seats.

Gear B: The Time Traveler (Sequential Hierarchical Bayes Update)

This is the "secret sauce" for tracking changes over time.

The Analogy: Imagine you are trying to predict the weather. The old way looks only at today's temperature. The new way looks at today's temperature plus a "memory" of what the weather was like yesterday, last week, and the month before.
How it works: The system uses a "memory model" (called an AR(1) model) to carry information forward. It doesn't treat each month as a fresh start. Instead, it says, "We know what the population looked like last month; let's use that knowledge to help us understand this month."
The Benefit: This allows the system to be much more accurate about changes (movements) than the old method.

3. The Big Win: Tracking Movement

The paper highlights a major difference in how the two methods handle "Movement" (changes over time).

The Old Method's Blind Spot: The old method calculates uncertainty based only on how many people were interviewed. It assumes the population is a static painting. If the population is actually a moving movie, the old method gets confused and underestimates the risk. In the paper's test, the old method only got the "change" right 82% to 96% of the time.
The New Method's Success: The DMM method accounts for both the "sampling noise" (interviewing people) and the "population noise" (the fact that the population itself is changing). In the test, this new method got the "change" right 100% of the time.
The Metaphor: If the old method is like trying to measure the speed of a car by looking at a single photo, the new method is like watching a video. It understands that the car is moving, so it calculates the speed correctly.

4. The "Burden" and "Budget" Rules

The system is also very practical. It respects two real-world limits:

Respondent Burden: You can't keep asking the same people the same questions forever, or they will get annoyed and quit. The system sets a "cap" (e.g., keep no more than 90% of the same people) to ensure fresh voices are heard.
Budget: It has a strict spending limit. It automatically shifts money from hiring new people (expensive) to keeping old people (cheaper) as long as the accuracy rules are met.

Summary of Results

In a simulation using Australian data:

Cost Savings: The new method saved about 6.3% in fieldwork costs compared to the standard method.
Accuracy: It was just as good at measuring the current state (levels) as the old method.
Superiority in Change: It was vastly superior at measuring changes (movements), achieving perfect coverage in the simulation where the old method failed to reach the 95% confidence target.

In a nutshell: The paper proposes a smarter, more flexible way to run repeated surveys. It uses a "memory" of past data to shrink the sample size (saving money) while simultaneously making the estimates of change over time much more reliable. It's like upgrading from a static map to a live GPS that knows the traffic is moving.

Technical Summary: Dynamic Mini–Max Design and Sequential HB Inference for Repeated Surveys

Problem Statement
National Statistical Offices (NSOs) conducting repeated surveys face a dual challenge: producing reliable estimates of population levels at each wave and accurately estimating movements (changes) between successive waves. While cross-sectional designs focus on levels, repeated surveys require simultaneous precision for both levels and movements. The precision of movement estimates depends not only on sample size but also on the degree of sample overlap between waves. Existing literature often treats these problems separately or relies on static allocation and retrospective estimation. Specifically, classical rotation designs (e.g., Patterson, 1950; Rao and Graham, 1964) optimize overlap for a fixed sample size without incorporating model-based precision, while composite estimation methods do not address prospective design. Furthermore, classical confidence intervals for movements typically account only for sampling variance, ignoring the model variance component arising from temporal dynamics, which can lead to under-coverage when population parameters evolve stochastically.

Methodology
The paper proposes a Dynamic Mini-Max (DMM) framework that integrates a Dynamic Mini-Max Design with a Sequential Hierarchical Bayes Update (SHBU). This framework extends the static Mini-Max approach (Tam, 2026a) to a repeated-survey setting using a three-component Hierarchical Bayes (HB) state-space model at the stratum level:

Sampling Model: Accounts for complex survey designs (clustering) via effective sample sizes. It uses a Binomial likelihood for binary variables (on the logit scale) and a Fay–Herriot normal model for continuous variables.
Linking Model: Borrows strength across strata within a wave using auxiliary covariates and random effects.
Propagation Model: An AR(1) process that describes temporal dynamics, serving as an informative prior for the current wave based on the state of the previous wave.

The DMM Design Component
The design phase jointly optimizes the sample size ( $n_{h,t}$ ) and the number of overlapping units ( $m_{h,t}$ ) for each stratum $h$ at wave $t$ . The objective is to minimize total fieldwork cost subject to:

Simultaneous precision constraints for levels (Coefficient of Variation, CV) and movements (Standard Error, SE).
Respondent burden limits (maximum overlap fraction $\rho_{max}$ ).
A total fieldwork budget.

Crucially, the movement SE constraint is decomposed into a sampling variance component (dependent on design variables) and a model variance component (dependent on the AR(1) and linking model parameters, estimated from previous waves). The optimization treats $(n_{h,t}, m_{h,t})$ as joint decision variables, unlike classical methods where sample size is often fixed.

The SHBU Component
The Sequential Hierarchical Bayes Update propagates posterior uncertainty from wave $t-1$ to wave $t$ .

Mechanism: Instead of running independent MCMC chains for each wave, the SHBU uses the posterior draws from wave $t-1$ to condition the prior for wave $t$ . Specifically, for each draw $b$ from the wave $t-1$ posterior, the stratum states are treated as fixed conditioning values for the AR(1) prior in wave $t$ .
Estimation: The framework produces coherent joint estimates for levels and movements. The movement estimator is defined as the difference between paired posterior draws ( $\hat{\theta}^{(b)}_{d,t} - \hat{\theta}^{(b)}_{d,t-1}$ ), preserving the posterior covariance induced by sample overlap and temporal dependence.
Intervals: The paper defines Calibrated Bayes Intervals (CBIs) for both levels and movements. These intervals explicitly incorporate both the sampling variance (from the design) and the model variance (from the HB propagation), ensuring valid coverage in a setting where the population evolves over time.

Key Contributions

Joint Optimization: The paper introduces a framework that simultaneously optimizes sample size and overlap to meet precision targets for both levels and movements, a problem distinct from static cross-sectional allocation.
Coherent Inference: It provides a unified inferential framework where level and movement uncertainties are jointly quantified. The movement CBI correctly accounts for the model variance component ( $\hat{V}_{mod}$ ), which classical design-based intervals ignore.
Sequential Updating: The SHBU allows for automatic sequential updating without the need for ad hoc composite estimator chaining or truncation of historical data, as the full posterior history is encoded in the AR(1) prior.
Small Area Estimation: By leveraging the HB linking model, the framework enables reliable estimates for small domains (sparse strata) that would otherwise require prohibitively large sample sizes under a purely design-based approach.

Results
The framework was evaluated using a four-wave simulation based on 2021 Australian Census microdata (wave 1) and simulated data for waves 2–4, calibrated to historical Australian Bureau of Statistics Labour Force Survey dynamics.

Cost Efficiency: Starting from the same 5% proportional allocation ( $n_A = 42,018$ ) used in a classical design, the DMM reduced the required sample size to $n^* = 40,251$ (a 4.2% reduction) while meeting all precision constraints. This resulted in a fieldwork cost saving of approximately 6.3%.
Level Coverage: Both designs achieved comparable level coverage and point accuracy (Maximum Absolute Relative Error ratios between 0.844 and 1.263).
Movement Coverage: A significant divergence was observed. The DMM achieved 100% movement coverage across all 27 domain-variable cells. In contrast, the classical design achieved only 82%–96% coverage (87.5%–95.0% nationally). The classical intervals under-covered because they failed to account for the model variance component inherent in the stochastic evolution of the population.
Sensitivity: The results indicate that the validity of the DMM level inference is sensitive to the correct calibration of the AR(1) propagation coefficient; misspecification can degrade coverage.

Significance
The paper claims that the DMM framework offers a superior approach for repeated surveys by addressing the limitations of static design and classical inference in a dynamic environment. Its primary significance lies in providing coherent joint uncertainty quantification for levels and movements, ensuring that confidence intervals for movements are valid even when population parameters evolve stochastically. By jointly optimizing design variables and utilizing sequential Bayesian updating, the framework achieves cost savings and maintains rigorous precision standards without the operational intractability of extending composite estimator chains over many waves. The paper concludes that while the framework relies on the assumption of a stationary AR(1) process, the availability of historical data allows for empirical calibration and model selection to mitigate misspecification risks.