An Office for National Statistics (ONS) release published on 1 October 2026 estimated that 974 deaths of people experiencing homelessness were registered in England and Wales in 2025. The ONS gave a 95% confidence interval of 908 to 1,041. Its estimate for 2024 was 978, and the office described the two years as similar. The four-death difference should not be read as a firm change in direction: these are model-based estimates and the confidence interval makes their uncertainty visible. The figures refer to the year in which a death was registered, not necessarily the year in which it occurred. Registration can take place later, particularly when a coroner investigates a death.

The published total is not simply a count of records directly identified. The ONS identified 834 deaths in registration data for 2025 and used capture-recapture modelling to estimate a total of 974. The model therefore added an estimated 140 deaths beyond those found directly. Capture-recapture examines overlap between different ways of identifying cases to estimate cases that may not appear in any one source. The office used three groups of evidence: residence or no-fixed-abode information, coroner information and place-of-death information. It first searched complete death-registration records using multiple strategies, then had analysts review cases with ambiguous or conflicting evidence. In a second stage, it fitted a model for each registration year at the England and Wales level.

The ONS sets out important limits on interpretation. It selected the Mth Chao lower-bound estimator, which it describes as robust but conservative. The resulting estimates should therefore be treated as lower bounds, and the true total may be higher. The evidence sources come from the same administrative system and may not be fully independent. Homelessness may not be known to the person supplying information, the registrar or the coroner, or may not be entered in the relevant fields. The model is intended to account for some deaths not directly identified, but it does not turn the estimate into a complete count. The release’s definition is based on death-registration information indicating that a person was sleeping rough or using emergency or temporary accommodation at or around the time of death. It does not align directly with the definitions used in statutory homelessness or rough-sleeping statistics.

Over the longer series, estimated deaths rose from 550 in 2013 to 974 in 2025, an increase of 77.1% according to the ONS. Directly identified deaths rose from 415 to 834 over the same period. These series describe different things: identified deaths are cases found in the available registration records, while estimated deaths include the additional cases inferred by the model. The ONS also said it had improved identification for registrations from 2013 to 2021, resulting in more deaths being identified for those years than in its previous release. The modelling method did not change, but the improvements in record searches affected earlier figures. That note matters when comparing this release with older versions of the series; they should not be treated as if every historical figure had been produced from an unchanged identification process.

The 2025 estimate included 833 male deaths and 141 female deaths, representing 85.5% and 14.5% of the estimated total respectively. Among English regions, London had the highest estimated number at 151, followed by the North West at 146 and the South East at 144; the estimate for Wales was 63. Regional and subgroup figures are allocated from the England and Wales model estimate according to each area’s share of identified deaths; the ONS did not fit a separate model for every region. It therefore advises caution with small-area figures and annual changes. The crude rate of 21.3 deaths per million people is also not a death rate among the population experiencing homelessness: its denominator is the general population aged 15 to 74, and the rate is not age-standardised. These statistics establish neither a cause nor a causal relationship. They need to be read with their registration definition, estimation method and uncertainty in view.