The capture-recapture method using two sources enabled estimation that about 6000 homeless people died in France from January 2008 to December 2010, i.e. around 2000 deaths per year, with a large confidence interval. The number of homeless deaths was greatly underestimated by the two sources, CMDR and the CépiDc, taken separately. The CépiDc could not probably enhance the completeness of the collection of the homeless deaths since the only solution would be to introduce a specific location for the homeless status on the death certificate. On the contrary, in view of the missing data in the CMDR database, the CMDR has planned to develop its network to increase the completeness and standardization of its database. The estimation could be more accurate by enhancing the matching rates and therefore by improving the quality of the CMDR data. The age and the day of death constituted necessary information to avoid duplicates. The CMDR has planned to implement strategies to improve the quality of these data, particularly regarding age and date of death. One solution to improve the matching would be to provide a unique identifier but the medical section of the death certificate is anonymous and confidential.
The only recent estimate of the homeless population of France was made by the INSEE, which reported that there were 144,000 homeless people in 2012 . In 2001, the first national survey of the homeless had estimated there were 86,500 homeless people in France. In the period of our study (2008–2010), the size of the homeless population was unknown. The structures and size of homeless population are constantly evolving. In particular, in recent years, homeless families become the fastest growing segment of the homeless population in Paris area . Thus, estimates are only valid for the dates they are conceived for. In addition, the definition of the homeless population in the present study was not the same as for the INSEE study: it did not include people sleeping rough and not using institutional facilities, or homeless people living in towns with a population of less than 20,000. For this reason, the mortality rate of the homeless population could not be estimated in this study or compared with the international literature.
Considering the small number of deaths common to both sources, the estimates of the number of deaths are imprecise and should be considered with caution. However, the low data completeness of each source is plausible, first because the information network of the CMDR is highly dependent on local initiatives and does not cover the whole of France homogeneously, and secondly because medical death certifiers rarely consider homelessness to be relevant information to be reported with the causes of death.
Some of the general conditions that validate the use of a capture-recapture method were not completely fulfilled by this study. First, the definition of homeless people was not the same for the two sources. While the homeless deaths reported in CMDR included all forms of homelessness (defined by INSEE), the homeless deaths reported in the CépiDc database are those in which the physician considered that homelessness played a role in the death. The CépiDc homeless death definition is likely to be a subcategory of the overall definition of homeless death. However, the proportion of deaths with ICD code 'Z59.0' (homelessness) that was not retrieved in the CMDR database was unexpectedly large (89%). In addition, we supposed that homelessness plays a role in the majority of cases even if it is not the underlying cause of death. Homelessness is associated with high increased risks of all-cause mortality . As such, in most of the cases the medical certifier should declare homelessness on the death certificate. However the low occurrence of homeless reporting on death certificate is likely to be attributable to the focus that medical certifier may put on medical rather than social conditions. Therefore, the definitions are not necessarily so different, and being caught by each of the two sources may be seen as independent random events.
This study showed some heterogeneity in the capture, specifically regarding the areas in which deaths occurred. For this reason, the estimates were stratified by the geographic area, but the results remained unchanged. The dependence between the two sources was not quantitatively evaluated but is likely to be low. However, a positive dependence would imply that the total number of homeless deaths was underestimated. The condition requiring perfect matching could not be met in this study. The matching algorithm was implemented in order to minimize the number of false matches by reducing the number of possible multiple matches between the CMDR and the entire database. Excluding the 391 unmatched records from the calculation is equivalent to assuming that the proportion of death certificates with the 'homeless' code in the CépiDc record was the same as that for the matched records, which seems the most reasonable option.
This study represented the first part of a broader project whose objective was to describe the mortality among the homeless population in France. The next step is to describe the deaths characteristics (age, location, season and causes of deaths) and to compare to the mortality in French population, in order to implement efficient strategy, at national level, provided by policymakers and public health professionals serving this population, and not only through local initiatives.