We respond to Shah et al’s letter of 2nd May 2013 as we share their concern why similar data sources should produce different conclusions about morbidity and mortality after bereavement. We have re-extracted and analysed data from The Health Improvement Network primary care database following their comments about our unexposed sample (Shah et al. letter dated 21st March 2013). In summary, we extracted a new cohort of unexposed people where we no longer make it a requirement for the cohabitee to remain alive throughout the follow-up. In instances where the cohabitee died during the study period, data were censored at the point of death so that the unexposed were not in the study once they were bereaved.

These analyses (tables 1-3 and figure 1: http://www.plosone.org/at...) produced similar results to those we reported in our original paper. In our new data there was an average of 5.2 unexposed for every exposed person (82,371 versus 15,748). Table 1 shows that demographic and social variables are similar for the exposed and unexposed. In Table 2, we observe, as reported previously, that the exposed were prescribed significantly more psychotropic drugs in the six months before the cancer death/index date, even after adjustment for age, gender, prescribing index, smoking status, heavy alcohol intake and Townsend deprivation quintile. GP consultations were also higher in the exposed than unexposed group in the six months before the cancer death/index date. Lastly, in table 3 we confirm our original finding of lower all-cause mortality throughout follow-up in the exposed (bereaved) than unexposed. This persisted after adjustment for demographic and other factors. We also examined all-cause mortality for the first 18 months following the cancer death (or index date in unexposed), a period which we consider to be most likely related to the bereavement. After adjustment, all-cause mortality was 11% lower in people bereaved following a cancer death than in people not so bereaved (table 3). We also found that new courses of hypnotics and antidepressants were more frequently prescribed in the exposed than unexposed and that this held after adjustment. There were, however, fewer consultations with the GP over the course of follow up in the exposed than the unexposed (adjusted IRR 0.93; 95% CI 0.92, 0.93). This was the one finding which varied from our original analysis, when we reported a significant excess in consultation rates for the exposed.

As stated in our initial response to Shah et al., we believe the differences in our results are mainly attributable to two factors. First, we are still not convinced that Shah et al. have been able to account sufficiently for a strong non-linear relationship between age and mortality using a standard Cox regression model. Second, we remain concerned that their analysis is subject to immortal time bias. This would be the case if they have included denominator time before the first death of both the deceased and the surviving partner, as it results in a bias towards decreased mortality among the non-bereaved. For example, let us say that Shah et al. had a couple who were in the database for one year before the death of the first individual and the bereaved then died after another year (Figure 1). The result of such analysis would then conclude that the risk of death was doubled following bereavement when in fact there was no difference. In this example, the year before the subject who is later bereaved should not be counted in the denominator as during that time he or she is effectively ‘immortal’. We ask Shah et al. to clarify how they have dealt with time in both individuals of the couple before the death.

Data from the meta-analyses by Moon et al. (2011) and cited by Shah et al in their recent communication revealed that “the widowhood effect” was stronger in the first six months after bereavement than beyond it. Nearly half of the 12 studies that followed participants for longer than six months produced equivocal estimates of effect. Furthermore, Moon et al, reported that excess mortality following bereavement in older people (>65 years) was only significant in men (RR 1.23). These findings further reinforce our belief that risk of increased morbidity and mortality after bereavement remains a topic for further investigation.