Estimated numbers of homeless and homeless mentally ill people in north east Westminster by using capture-recapture analysisBMJ 1994; 308 doi: https://doi.org/10.1136/bmj.308.6920.27 (Published 01 January 1994) Cite this as: BMJ 1994;308:27
- N Fisher,
- S W Turner,
- R Pugh,
- C Taylor
- Pathfinder Community and Specialist Mental Health Services, Springfield Hospital, London SW1 7DJ Camden and Islington Community Health Services NHS Trust, Hampstead Road, London NW12 2LT National Addiction Centre, Institute of Psychaitry, London SE5 8AF.
- Accepted 1 October 1993
Objectives: To use routinely collected data to provide a reliable estimate of the size and psychiatric morbidity of the homeless population of a given geographical area by using capture-recapture analysis.
Design: A multiple sample, log-linear capture-recapture method was applied to a defined area of central London during 6 months. The method calculates the total homeless population from the sum of the population actually observed and an estimate of the unobserved population. Data were collected from local agencies used by homeless people.
Subjects: Homeless people in north east West-minster residing in bed and breakfast accommodation and hostels or sleeping rough who had contacted statutory or voluntary agencies in the area.
Results: 2150 contacts by 1640 homeless people were recorded. The estimated unobserved population was 3293, giving a total homeless population for the period of around 5000 (SD 1250). Mental health problems were significantly less prominent in the unobserved compared with the observed population (23% (754) v 40% (627), P<0.0001). For both groups the prevalence varied greatly with age and sex.
Conclusions: Capture-recapture techniques can overcome problems of ascertainment in estimating populations of homeless and homeless mentally ill people. Prevalences of mental illness derived from surveys that do not correct for ascertainment are likely to be falsely inflated while at the same time underestimating the total size of the homeless mentally ill population. Population estimates derived from capture-recapture techniques may usefully provide a good basis for including homeless populations in capitation calculations for allocating funds within health services.
Homelessness is associated with ill health
Current estimates of the size of homeless populations are unreliable because of problems in identifying all homeless people for a given time period
Capture-recapture methods overcome problems of ascertainment by calculating the size of the unobserved population.
This study estimated the population of unobserved homeless to be twice the size of the observed population
Mental health problems are significantly less common in the unobserved or hidden homeless compared with those who are easily surveyed.
Homelessness is associated with considerable ill health1,2 and increased use of acute general hospitals.3,4 The planning of health services led by needs for homeless people is hampered by a lack of reliable data on the size of homeless populations.5 For example, the 1991 census failed to identify any people sleeping rough in Birmingham on census night.6 The prominence of mental health problems among the homeless in inner city areas has generated increasing interest. As much as half of the homeless population may have some form of mental disorder, although estimates of the extent and severity of these disorders varies widely.*RF 7-9*
Working for Patients gives each health authority responsibility to determine the size and morbidity of both its resident and homeless populations.10,11 The size of homeless populations is difficult to estimate because of problems in identifying all homeless people in a given period. This is a result of the relative rarity and elusiveness of homeless people compared with resident populations.12 Capture-recapture techniques are increasingly being used to overcome these problems of ascertainment13,14 and have been recommended for use with homeless populations.15
Capture-recapture methods derive their name from censuses of wildlife populations. In medical practice the technique allows the number of cases in a defined population to be estimated by using two or more sources of cases. These could include records from hospital or general practice or any other points of contact. Taken alone each of these sources may undercount the actual number of cases, as indeed would a simple aggregate of the sources that excluded duplicate cases (cases identified at more than one source). Capture-recapture techniques use information provided by duplicate cases to allow the number of people not identified at any of the sources to be calculated.14 An estimate of the total number of cases is thus the sum of the calculated unobserved population and the observed population. In its simplest form the capture-recapture technique makes a range of assumptions about the enumerated population being homogeneous and closed to arrival or departure of cases and the sources being independent of each other. The use of multiple sources and log-linear modelling applied to this technique has been recommended when these assumptions are violated.*RF 13-16*
We report a pilot application of this technique which used several sources to estimate the size of the homeless population of a fixed geographical area over a limited time. We proposed that data sources could be derived from existing lists of homeless people, such as booking in sheets in hostels. In addition, we expected that these lists would contain enough data to describe the population in terms of demography and prevalence of mental health problems.
Subjects and methods
The study area was north east Westminster, an area of central London with a large homeless population. The sources providing data for the capture-recapture analysis were derived from services used by homeless people within this area. Statutory and voluntary agencies that agreed to participate included hospitals, local authority social services, a primary health care centre for homeless people, hostels, an outreach service for the homeless mentally ill, the probation service, and the registrar of deaths for Westminster.
Homelessness was defined to include those placed in emergency bed and breakfast or rented accommodation in the private sector; residents of night shelters, traditional hostels, and short stay hostels; those sleeping rough or on the streets; those in precarious accommodation (for example, squats); and those in non-residential institutions with no other permanent accommodation (for example, hospitals). When ambiguity about homelessness existed recorded addresses were checked against lists of addresses of hostels and bed and breakfast accommodation used to place homeless people within the study area in a manner similar to that used by Victor et al.17
All data were obtained by examining records held at each source. Name, date of birth, and sex were recorded to identify duplicate cases. In addition, if the information was available, type of homelessness and presence of any mental health problems (including those related to use of drugs and alcohol) were recorded. No subjects were interviewed. Data were collected at the end of two periods of three months: July to September 1991 and October to December 1991. The number of people presenting to each source, rather than the number of contacts, was recorded. Strict confidentiality was preserved. The data collected were not used for purposes other than this study.
Duplicate cases were identified by an algorithm based on name and date of birth which allowed for various name spellings. We then examined these matches and judged whether they were obvious alternative spellings of the same name or were likely to be different names.
The heterogeneity of the homeless population was acknowledged by stratifying the sample by three factors: sex, presence of mental health problems, and age (under 30 or 30 and over). This gave a total of eight population subgroups (see table II). The total population of homeless people was thus the sum of these subgroups. Violation of assumptions relating to dependencies between sources and population subgroups was managed by log-linear modelling by using the GLIM statistical package18 (see appendix). Three possible types of dependencies were incorporated into the log-linear model estimation procedure (see appendix). The first was dependencies between subpopulation groups - for example, the association of age with an increased likelihood of mental health problems. The second was dependencies between subpopulations and sources - for example, the presence of mental problems increasing the probability of presentation at a particular source. The last was dependencies between sources - that is, presentation at one source increasing the likelihood of presentation at additional sources. The model reported was the one which produced the most acceptable fit to the observed counts.
The observed sample
During the two sampling periods a total of 2150 contacts by 1640 homeless people were recorded, with 393 contacts at more than one source. No deaths of homeless people were identified from death certificates; this source was thus excluded from further analyses. There were no differences between the two sampling periods (representing summer and winter) in numbers of contacts to the remaining agencies (table I). As a result the two periods were aggregated in all further analyses.
Of the 1640 people, 1337 (82%) were men and 303 (18%) women (table II). Of the men, 896 (70%) were aged 30 and over compared with only 104 (41%) women X(sup2)=102.3, df=1, P<0.0001). A total of 627 (40%) people were identified as having mental health problems. For those aged 30 and over the proportions of men and women with mental health problems were similar (399(44%) men, and 49(47%) women; X(sup2)=0.2, df=1, P>0.5); in the under 30s, however, there was a significant difference (138 (34%) men compared with 41 (23%) women; X(sup2)=7.1, df=1, P<0.01). The pattern of use of services was uneven. A person was more likely to contact more than one agency if he or she was 30 and over (X(sup2)=5.0, df=1, P<0.025) or had a mental health problem (X(sup2)=70.7, df=1, P<0.0001) (table III).
We found strong interdependencies between population subgroups and sources. The older subgroup was more likely to go to hostels while the younger was more likely to present to the probation service; women were more likely to go to social services, hostels, and hospitals. Of necessity the presence of mental health problems was associated with contact with the homeless mental health team. The above heterogeneity accounted for almost all dependencies between sources apart from, firstly, social services and the homeless mental health team and, secondly, hostels and the homeless mental health team. The nature of these dependencies reflects the known use of services and gives the model validity.
The unobserved homeless population was calculated as 3293 (table III), giving a total estimated population of around 5000 (SD 1250). This indicates that a simple survey would have ascertained barely one third of the total homeless population. The demographic characteristics and the relative distribution of mental health problems were similar for the estimated unobserved population and the observed sample (tables II and IV). There were, however, significantly fewer mental health problems in the unobserved compared with the observed population (23% (754) v 40% (627), X(sup2)=145.8, df=1, P<0.0001).
The results of the capture-recapture analysis indicate, firstly, that the homeless population of north east Westminster is substantially larger than previously estimated, and, secondly, that there are important differences between the observed and unobserved populations, particularly in the prevalence of mental health problems.
The completion of this pilot project suggests that log-linear capture-recapture analysis can be applied to existing datasets to estimate the size and nature of the homeless population for a given geographical area. Questions of validity and reliability cannot be conclusively answered as no gold standard for the size of this population exists and there are no similar studies for comparison. With this in mind the results of the study should be taken in the context of several methodological limitations.
The population was not closed. People could have moved into or out of the area and become or stopped being homeless. Potentially this inward or outward flow could have been determined by fitting more complex log-linear models18; this, however, was beyond the remit of this pilot project. The estimate is therefore dependent on the time period of the study, with the expectation that the total would increase over time as new cases arrive. Thus the figure of 5000 should be taken to indicate the number of people who were homeless in north east Westminster for at least part of the six months studied. This population estimate remains valuable for planning health services as all the members of that population would have a valid claim to use the local health care services.
People whose age was unknown did not present at more than one source. Consequently it was not possible to complete any capture-recapture analysis on this group. Clearly it is improbable that no people from this group were unobserved. The total unobserved population was about twice the size of the observed population. If a similar relation applied to the group of unknown age then about 100 people could be added to the unobserved total.
Data on mental health problems relied on the use of information recorded at each source and cross referencing between sources. For some sources we assumed that the quality of this information was good (for example, hospitals); for other sources the information was likely to be less robust. Problems associated with drugs and alcohol were recorded separately from other mental health problems, but the numbers were too small to allow for a discrete capture-recapture analysis.
Identification of all homeless people presenting to the study sources was likely to be incomplete. This was most obvious at the registry of deaths, where no deaths of homeless people were identified from death certificates. This was an improbable finding2 and may have resulted from a practice of using the last known address rather than the current situation in completion of death certificates. Some duplicate cases may have been missed if people used aliases when presenting at different sources.
Comparison with other studies
A recent report places this study in context. Black et al estimated that within what was Bloomsbury Health Authority, which encompasses north east Westminster, there were 3467 homeless people.3 The area covered by this study was about half the size of Bloomsbury Health Authority, yet 1640 homeless people were identified and the capture-recapture analysis estimated that there were a further 3293 homeless people that were unobserved during the study period.
Clearly from both the observed and estimated unobserved populations the homeless population is heterogeneous both demographically and in distribution of mental health problems. The prevalence of mental health problems in the observed population was similar to that reported previously.8 Strikingly, for all age and sex categories the prevalence in the estimated unobserved population was significantly lower (table III) and in total was barely half that of the observed population. This indicates that homeless people with mental health problems are disproportionately prominent. Consequently prevalences derived from surveys that do not correct for ascertainment are likely to be falsely inflated while at the same time underestimating the total size of the homeless mentally ill population.
Capture-recapture techniques can overcome problems of ascertainment in the estimation of homeless and homeless mentally ill populations and is an advance on other current methods of estimation. Population estimates derived from capture-recapture techniques may usefully provide a good basis for including homeless populations in capitation calculations for allocating funding in health services.
Nigel Fisher was supported as a research worker in the department of psychiatry, Middlesex Hospital, by a grant from the special trustees of the Middlesex Hospital. The authors are grateful to all the agencies who cooperated during the course of this study.
Details of the statistical analysis
An estimate of the total population (N(subT)) in a simple two source analysis uses information on the relative frequency of duplicate cases (N) and cases in only the first sample (N(sub1)) or the second (N(sub2)) in relation to the number of unobserved cases where N(subT)x(N(sub1)=N(sub2))/N. The assumption that the captures are independently made is crucial in allowing this relation to be determined. In multiple sources recapture analysis the increased information on different multiple captures allows this assumption to be tested and allows models which incorporate heterogeneity among subgroups in the population. These models are analysed with standard log-linear modelling programs,18 in this case GLIM3.77.19
The model allowed for all interactions between subgroups. First order interactions were included for dependencies between subgroups and individual sources and for dependencies between sources. The related higher order interactions were not included. Owing to the special position of the homeless mental health team in relation to mental illness the basic model was extended to include all interdependencies between this source and all combinations of subgroups.
Interactions in the fitted model (apart from those with the homeless mental health team) were removed as being unnecessary if the effect was less than four times its standard error. This was a harsh criterion deliberately used to avoid “overfitting” in such a large model, which would have been likely to inflate the overall population estimate. The retained interactions between sources and subgroups (see text) were strong (X(sup2)=243.6, df=8, P less than 0.0001); the between source dependencies, which were present even after allowing for this subgroup heterogeneity of capture, were by contrast weak and although significant (X(sup2)=25.9, df=2, P less than 0.001) were excluded from the final, conservative model, again to avoid any overfitting. Had they been retained in the estimation procedure the overall population estimate would have been 5% higher but of similar structure. The final model scaled deviance was acceptably close to its degrees of freedom (X(sup2)=338.7, df=481), suggesting a good but not overly close fit and a total population estimate of 3293.