Background People experiencing homelessness have high rates of emergency admissions and poor health outcomes, but homelessness is not routinely recorded in hospital data. We aimed to estimate the number of hospital admissions for people experiencing homelessness in England.

Methods We developed a novel homelessness phenotype in national Hospital Episodes Statistics Admitted Patient Care data. This included people recorded with: ‘no fixed abode’ (with exceptions), registration with a homeless-exclusive general practice (GP), and the International Classification of Disease (ICD-10) code for homelessness. We used three-source capture-recapture Poisson regression methods with Bayesian model averaging to estimate the number of admissions for this population. This method is used to estimate the size of populations where a direct count is not possible, such as in these data where homelessness is not routinely recorded. It uses information from the overlap between the three defined homelessness codes to estimate the total number of admissions. We then calculated an inflation factor which is the ratio between the estimated total and the observed number of admissions. We computed point estimates of total admissions and inflation factors with 95% confidence intervals (95% CI) for three non-consecutive years, 2013/14, 2015/16, and 2017/18, using Stata 17.

Results The number of observed admissions with at least one code for homelessness was 27,124 in 2013/14, 31,933 in 2015/16, and 34,790 in 2017/18. We estimated that there were 5.07 times more admissions [95% CI 4.71 -5.42] among people experiencing homelessness in 2017/18 than were observed in the dataset (176,342 total admissions [95% CI 164,031 – 188,654]). The estimated total number of admissions decreased over time, while the observed number increased. Thus, the corresponding inflation factors also decreased over time [6.66 [6.09 – 7.23] in 2013/14 and 5.75 [5.30 – 6.19] in 2015/16).

Conclusion Our study suggests there were over five times as many hospital admissions for people experiencing homelessness in 2017/18 than we could observe directly. The main strengths of this study are its size and robust analytical approach. We have tested and adjusted for the underlying capture-recapture assumptions, but residual bias and confounding may remain. Hospitals may consider applying our methods to estimate the size of their local population experiencing homelessness, while also aiming to standardise coding. The ICD-10 code is the most valid and reliable as it requires the attending clinician to ‘diagnose’ homelessness. National implementation of the ICD-10 code for homelessness is needed to inform better service planning and delivery.