Using Bayesian Evidence Synthesis to estimate the number of sex workers in the United Kingdom

This study employs Bayesian Evidence Synthesis to integrate historical data with recent UK population and clinic attendance records, producing a robust estimate of approximately 84,000 sex workers in the United Kingdom (95% credible interval: 49,000–130,000) to address critical data gaps for targeted public health support.

The Gist

The Big Picture: Counting the Unseen

Imagine you are trying to count how many people are hiding in a giant, dark forest. You can't see them all because they are trying to stay hidden, or they are afraid to come out, or they are moving around too much. This is the challenge the researchers faced: estimating the number of sex workers in the UK.

Because sex work is often stigmatized or legally complex, many workers don't show up in official government records (like census data) or standard surveys. Previous attempts to count them were like guessing the number of fish in a lake by looking at just one bucket of water; they gave a single number but didn't tell you how wrong they might be.

This paper's goal was to create a more reliable map of this hidden population, not just by guessing, but by combining all the old maps we already have and admitting exactly how uncertain we are about the details.

The Method: The "Master Chef" of Data

The researchers used a technique called Bayesian Evidence Synthesis. Think of this like a master chef trying to make the perfect soup.

1. The Ingredients (Old Studies): Over the last 30 years, different researchers have tried to count sex workers. Some used data from clinics, some from online ads, and some from surveys of support services. Each of these studies is an "ingredient." Some ingredients are fresh and high-quality; others are a bit old or were measured with a shaky hand.
2. The Recipe (The Model): Instead of just picking the "best" study and ignoring the rest, the researchers built a mathematical recipe (a computer model) that mixes all these ingredients together.
3. The Secret Sauce (Uncertainty): The special part of their recipe is how they handle the "shaky hands." They didn't just take the numbers from the old studies as absolute facts. Instead, they treated every old number as a "best guess" with a margin of error. They fed these "guesses" into their model, allowing the uncertainty to flow through the whole process.

The Analogy: Imagine you are trying to guess the weight of a mystery box.

• Old Way: You ask five friends. One says 10kg, one says 50kg, one says 20kg. You just average them to get 26.6kg and say, "That's the weight."
• This Paper's Way: You ask the five friends, but you also ask, "How sure are you?" One friend says, "I'm guessing 10kg, but I could be off by 5kg." Another says, "I'm guessing 50kg, but I'm pretty sure it's between 40 and 60." The model then combines these guesses and their confidence levels to give you a final answer that says, "The weight is likely around 26kg, but it could realistically be anywhere between 15kg and 40kg."

The Results: The Final Count

After mixing all the data and running the complex math, the model produced a new estimate:

• The Number: There are approximately 84,000 sex workers in the UK.
• The "Maybe" Range: Because the data is tricky, the researchers are 95% confident the real number is somewhere between 49,000 and 130,000.
• The Context: This represents about 0.12% of the entire UK population.

To put this in perspective, the paper notes this group is smaller than the population of men who have sex with men (about 1% of the population) but potentially larger than the population of people who inject drugs.

Why This Matters (According to the Paper)

The paper argues that having a number with a clear "margin of error" is better than having a single, precise-looking number that might be wrong.

• Better Planning: If health services know the population is likely between 50k and 130k, they can plan resources (like vaccines, testing, or support services) that are robust enough to handle the higher end of that range, rather than underestimating the need.
• Reducing Inequality: The paper states that sex workers often face health inequalities and barriers to care. By having a better estimate, health officials can better understand the scale of the problem and design targeted support to reduce these gaps.

The "Fine Print" (Limitations)

The authors are very honest about the flaws in their "soup":

• No Perfect Data: They had to rely on old studies that didn't have their own error bars. They had to invent a system to judge which old studies were "better" based on how they were done.
• Time Travel: The data spans nearly 25 years. The world of sex work has changed (especially with the internet), so the model assumes the population hasn't changed drastically in structure, which might not be true.
• No Direct Contact: The model didn't talk directly to sex workers; it only analyzed existing reports. The authors admit that future work needs to involve sex workers directly to get better data.

Summary

This paper didn't go out and count sex workers one by one. Instead, it acted as a data detective, taking all the fragmented, imperfect clues from the last 30 years, combining them with a sophisticated mathematical method, and producing a single, well-reasoned estimate: 84,000 people, with a clear warning label that the real number could be significantly higher or lower. The goal is to use this number to help health services support this community more effectively.

Technical Summary

Technical Summary: Bayesian Evidence Synthesis for Estimating the UK Sex Worker Population

Problem Statement
Sex workers in the United Kingdom face significant health inequalities and barriers to accessing healthcare, often exacerbated by stigma, discrimination, and legal contexts. A critical obstacle to providing targeted, evidence-based public health support is the lack of reliable data regarding the size of the sex worker population. Existing estimates, derived from various studies between 1999 and 2017, vary widely (ranging from ~35,000 to ~105,000) and, crucially, lack quantified uncertainty. Previous studies often treated their estimates as point values without confidence intervals, failing to account for the inherent uncertainty in the parameters and methodologies used to derive them. This absence of robust, uncertainty-quantified data hinders effective resource allocation, outbreak response planning, and the design of tailored health interventions.

Methodology
The authors employ a Bayesian Evidence Synthesis framework to integrate seven historic estimation studies (spanning 1999–2017) with recent national population data and Genito-Urinary Medicine Clinic Attendance Data (GUMCAD). The core methodological innovation is the treatment of historical studies not as "black boxes" but as mechanistic processes with uncertain parameters.

1. Hierarchical Model Structure:

   • Level 1 (Study-Specific Estimates): Instead of accepting reported point estimates, the model reconstructs each study's estimate from its documented parameters (e.g., service contacts, multipliers, ratios). Each parameter is assigned a prior distribution, allowing the inherent uncertainty of the original methodology to propagate through the model. The observed study counts are modeled as noisy log-normal realizations of these mechanistic estimates.
   • Level 2 (Hierarchical Pooling): Study-specific estimates are soft-constrained to a study-specific proportion ($\theta_i$) scaled by the current UK population.
   • Level 3 (Population-Level Parameters): The study-specific proportions are drawn from a common Beta distribution governed by the overall population proportion ($p$) and a between-study heterogeneity parameter ($\tau$).

2. Quality Weighting System:
   To address the lack of standard errors in historical studies, the authors developed a systematic quality weighting system based on WHO guidelines. This system assigns weights (0–100) based on:

   • Data Collection Method (67% weight): Evaluated on population coverage (sectors, genders, ages, geography) and temporal relevance. Coverage is prioritized over recency.
   • Method-Specific Criteria (33% weight): Evaluated on the execution of the specific method (e.g., independence of sources in multiplier methods, outreach in enumeration).
     These weights determine the variance in the likelihood function; higher-quality studies receive lower variance (higher precision), while lower-quality studies receive higher variance.

3. Likelihood and Priors:

   • Likelihood: Weighted log-normal distributions are used to compare observed study counts with model-derived estimates. This choice respects the positivity constraint of population sizes and accounts for the multiplicative nature of estimation errors.
   • Priors: The overall population proportion ($p$) uses a Beta(12, 9988) prior, centering the estimate around 0.12% based on previous research but allowing data influence. The heterogeneity parameter ($\tau$) uses a log-normal prior to allow for between-study variation.

Key Results

• Population Estimate: The model estimates a total of 84,000 sex workers in the United Kingdom.
• Uncertainty: The 95% credible interval is 49,000 to 130,000, representing 0.121% of the current UK population (approx. 69.3 million). The median estimate is 81,000, indicating a slightly right-skewed posterior distribution.
• Model Adjustments: The synthesis made principled adjustments to individual studies. For instance, the Cusick et al. (2009) estimate was substantially adjusted upward (+31%) to align with the prior-implied estimate, while Import IO (2017) and Pitcher (2015) were adjusted modestly downward. The GUMCAD data, derived from clinical counts, showed the widest credible interval due to the compounding uncertainty of the scaling factors used.
• Robustness: Sensitivity analyses (leave-one-out and uniform weighting) confirmed that the results are robust. Removing any single study shifted the estimate by less than 2.3%, and changing the weighting scheme resulted in a negligible 0.2% difference. The model successfully reproduced the pattern of observed estimates without overfitting.

Significance and Claims
The paper claims three primary contributions:

1. Revised Estimate: It provides a comprehensive, synthesized estimate that leverages collective information from nearly two decades of research, reducing the influence of any single study's limitations.
2. Quantified Uncertainty: Unlike previous point estimates, this approach rigorously propagates uncertainty from study-specific parameters and between-study heterogeneity to the final result. This provides policymakers with a credible interval essential for robust planning and risk assessment.
3. Extensible Framework: The hierarchical Bayesian model offers a replicable and extensible framework. It can incorporate new evidence (e.g., from surveillance systems or digital platforms) without requiring a complete re-analysis, maintaining methodological consistency.

Limitations and Future Directions
The authors acknowledge several limitations:

• Data Constraints: The underlying studies lack standard errors, necessitating the use of a subjective quality weighting system.
• Temporal Stability: The model assumes temporal stability across studies spanning 25 years, potentially missing shifts due to technological changes (e.g., online platforms) or legal developments.
• Methodological Inheritance: The approach still relies on multiplier methods and extrapolations, inheriting the coverage limitations of the original studies.
• Lived Experience: The study did not involve co-production with sex worker-led organizations.

The paper concludes that while this estimate provides a crucial national baseline, future work must prioritize co-produced, community-led methodologies to improve accuracy, inclusivity, and the safety of data collection. The authors emphasize that population size estimation should serve to reduce health inequalities and inform targeted support, while strictly guarding against the repurposing of data for enforcement or surveillance.