Strengths and limitations of this study
The study had the strength of being a large population-based study with high data quality. First of all, the personal identification number from the Danish Civil Registration Register ensured reliable linkage of information. The validity of the Danish Cancer Register is high compared with other disease registers23 and the National Discharge Register concerning the administrative data has been validated finding agreement of 98.5%.18 The information about GP services was retrieved from the National Health Insurance data, which is used for remuneration of GPs and validated successively by the health authorities.
Our results should, however, be interpreted with caution. First of all, as the study is retrospective it potentially suffers from the typical weaknesses of such studies: reversal of causality, misclassification of exposure, and selection bias of cases and/or controls. Although the design of the study is retrospective, we are, by using the dates in the registers, sure that the exposure of interest, home visits by GP, and contacts with district nurse, occurs before the outcome, place of death. Furthermore, by analysing all cancer deaths in a well-defined region and timeframe we are sure that our study is not biased from differential selection of either cases or controls on the basis of their exposure status.
The most ideal design would be a prospective study but this would be difficult to conduct without risk of substantial bias. Prediction of a patient death can not be precise,24 that is, entry to a prospective study will introduce selection bias. Furthermore, it would be unethical to preclude a GP home visit if requested by the patient.
Attrition due to deterioration of the condition, drop-outs, and small sample size25,26 is a well-known problem in palliative care research, and can hamper conduction of prospective randomised clinical trials. The major problem induced by the retrospective data collection is the left censoring, that is, restricting the analysis to the last 3 months. We had to introduce left censoring (date of death minus 3 months) to get good data quality but as a consequence we cannot allow for any events prior to the 3 months before death in our study. We did, however, find similar results for those diagnosed less than 3 months before death and those who were given a diagnosis prior to the 3 months before death.
The patient's preference for a home death is a known predictor of dying at home,27 but we do not have information recorded on this in our data. As such, we cannot rule out whether the strong association between GPs paying a home visit and dying at home is a marker of such a decision being made either by the GP and the patients jointly, or by the patient alone to achieve death at home. In spite of this, we consider our results interesting, as the established dose–response relationship would seem to confirm the usefulness of GPs' visits for implementing such a decision.
In relation to care from other sources, the existence of an informal carer is also a known predictor of dying at home.8 We tried to account for this by including proxy variables (marital status, children, sex, and age) in the analyses and found that the association between GP home visits and place of death was unchanged.
The most important unobserved patient characteristic (apart from the preference for place of death and the care from other sources discussed above) is the patient's general need of care. We tried to account for this by adjusting for number of ambulatory contacts, type of cancer, sex, and age in the analyses. Given the size of the association, we find it unlikely that the strong association can be entirely explained by a patient's general need of care. Furthermore, the need for symptom management and with that, the need of acute care interventions by GPs including out-of-hours consultations, increases in the last weeks of life. We found, however, that the association persisted eliminating all services the last 2 weeks in the adjusted regression and in the matched case-control study, matching according to the period before the last admission of the patient.
We investigated whether the association of home visits with home death was a result of different exposure time, that is, time not in hospital. However, stratifying on time at home we found a significant adjusted and unadjusted association. Furthermore, our analysis identified the rate of home visits as a predictor of place of death which showed a strong dose–response relationship with respect to place of death. In the latter analysis the amount of time spent at home in the last 3 months is implicitly accounted for, since the rate is the number of home visits relative to time spent at home. The interpretation is, therefore, that among patients spending similar amounts of time at home, those with more home visits have a higher chance of dying at home.
Findings regarding home visits by community nurses seemed to show the same tendency as home visits by a GP, but this result is entirely based on an urban population and can only be extrapolated to the entire population with caution.
