We thank Shah et al. for raising awareness of the discrepancy between our two studies on bereavement (1,2). It is always worrying when studies try to answer similar questions using a similar data source, but reach different conclusions. However, this is not the first time this has happened with analyses of data from primary care databases. Often this relates to different study designs answering slightly different aspects of the same question or different ways of handling important confounding factors (3,4).

In this case age is probably the strongest predictor of mortality. The association is likely to be nonlinear; most couples in their sixties may expect to live together for another 20 years. However, only a few then live for another 20 years, whether they are bereaved or not. Age is also likely to be strongly associated with bereavement and therefore a potential critical confounder of the association between bereavement and mortality.

Shah et al. chose to use a study design whereby “bereavement was modelled as a time-dependent variable so that after bereavement, the status of the surviving partner was changed to bereaved and the impact of different periods after bereavement on mortality could be examined.” (2) They identified 171,720 couples and followed them up for nearly 4 years. While some detailed information is given about the age distribution at the start of the follow-up (81% of the women and 74% of the men were below 75 year, Table 1 of their paper) no direct information is available on the age distribution at the start of the bereavement. However, we can deduce from the information on the age at all deaths (45% were below 75 years in Table 2 of their paper) that the bereaved individuals were substantially older (see figure, based on data from table 1 and 2 from Shah et al.). They did include age in their regression models, but no further information was provided on how the models were specified. The question is whether the slightly increased mortality, found by Shah et al., in their bereaved sample, could be due to an incorrect specification of age. The major imbalance in the age distribution between the two cohorts also raises the question about the precision and accuracy of the estimates in the tails of the distribution. For example, there may only be a few non-bereaved left for comparison in the 85+ age group and likewise this study design only identified a few bereaved in their fifties (see figure: http://www.plosone.org/at...).

Fig: Data plotted from those given in Shah et al. (Table 1 and 2). Age at baseline and age of those who died. No information is given in the paper about age at bereavement.

Further, we wonder if Shah et al. (2) have introduced an immortal time bias in the way they have designed their study (5). It appears that the survivor first contributes time to the non-bereaved sample and then to the bereaved sample. However, by definition the survivor is immortal in the non-bereaved period. Hence, including the time before the death of their spouse would inflate the survival time of the non-bereaved cohort dramatically. This may simply be an omission in the description of the study design, but we would welcome Shah et al.’s clarification of this point.

In contrast, we used a relatively simple design to compare survival in individuals whose partner died from cancer to survival in individuals whose partner was still alive. We stratified our comparison sample such that the exposed (bereaved) and unexposed (non-bereaved) were of similar age and sex. In addition, we adjusted for age and sex in our analyses to account for any remaining effects of age. The method of stratification is a non-parametric approach that accounts for the potentially non-linear effects of age. Further, this optimises the statistical power of the study by providing approximately equal proportions (1:5) of bereaved and non-bereaved at each age group and sex.

We now address some of the more specific comments from Shah et al:
As highlighted above, we stratified our comparison cohort so that they were of similar ages and sex to the bereaved sample. Hence, a simple comparison with death rates in the general population is no longer valid.
Although we noted in our discussion that, because of changed domestic circumstances, bereaved people might be more likely to leave the practice, there is no evidence that this means they have higher mortality than those who stayed. We would also like to clarify that we identified couples by their family number at the start of the follow up, not the end.

We found that the consultation rate and prescription of psychotropic medication was higher in the bereaved in the 6 months before, and in the follow-up period after, death. This is to be expected because of the medical arrangements involved in assisting a dying partner and the psychological strain that occurs. To conclude that this is an indication they must have had higher mortality is conjecture. One could as easily suggest that it also indicated better primary care which had a role in reducing mortality.

Finally, we were puzzled that Shah et al. find the published literature on increased morbidity and mortality after bereavement so convincing. If this were the case, why did they do their study? We undertook this piece of research because we did not find the evidence convincing and stated this in our introduction (1).

1. King M, Vasanthan M, Petersen I, Jones L, Marston L & Nazareth I. Mortality and Medical Care after bereavement: a general practice cohort study. PLoS ONE 2013 e52561.
2. Shah SM, Carey IM, Harris T, DeWilde S, Victor C, Cook DG. Do good health and material circumstances protect older people from the increase in mortality after bereavement. Am J Epidemiol. 2012;176(8):689-698.
3. de Vries F, de Vries C, Cooper C, Leufkens B, van Staa T-P. Reanalysis of two studies with contrasting results on the association between statin use and fracture risk: the General Practice Research Database. International Journal of Epidemiology 2006;35:1301-1308.
4. Osborn D P J, Limburg H, Walters K, Petersen I, King M, Green J, Watson J, Nazareth I. Relative incidence of common cancers in people with severe mental illness. Cohort study in the United Kingdom THIN primary care database. Schizophrenia Research 2012 epub ahead of print.
5. Suissa S. Immortal time bias in pharmacoepidemiology. American Journal of Epidemiology. 2008;167:492-499.

Michael King*
Irene Petersen#
Louise Jones*
Louise Marston#
Irwin Nazareth#

*Unit of Mental Health Sciences
Faculty of Brain Sciences
University College London Medical School
Second Floor
Charles Bell House
67-73 Riding House Street
London W1W 7EH
UK

# Research Department of Primary Care and Population Sciences
Institute of Epidemiology and Health Care
University College London Medical School
Royal Free Campus
Rowland Hill Street
London NW3 2PF
UK

Author for correspondence: Michael King
michael.king@ucl.ac.uk