In the movies and in certain kinds of romantic literature, we sometimes come across a deathbed scene in which a dying person holds onto life until some special event has occurred. For example, a mother might stave off death until her long-absent son returns from the wars. Do such feats of will occur in real life as well as in fiction?

The author, David P. Phillips, looked at selected data and concluded that famous people were often able to postpone their deaths until after a birthday, in that there were fewer than expected deaths (a "death dip") in the months preceding the birthmonth and more than expected deaths in the birthmonth and succeeding months. Because the collection of essays in which his paper appeared is so well-known, Phillips' essay attracted considerable attention and one of the authors uses it as an example in his widely used statistics textbook.[2]

Phillips has subsequently reported finding a death dip among Jewish males before Passover[3] and among elderly Chinese Women before the Harvest Moon Festival,[4] though there were considerable opportunities in these studies for a selective use of data. The Jewish study used data for Californians with "common" Jewish surnames that had at least ten listings in the central Los Angeles telephone directory, and did not find a death dip among nonwhite Jews or female Jews. No reason was given for excluding Jews whose names did not appear at least ten times in this particular telephone directory. In addition, Phillips excluded people named Ash or Bach because he didn't consider these names sufficiently Jewish, but included people named Asher or Brody as Jewish enough for his purposes. The Chinese Harvest Moon study reports results for women who were at least 75 years old when they died. No reason is given for the exclusion of women who died between the ages of 65 and 70 or 70 and 75. Interestingly, Phillips reports that there was no statistically significant death dip for women under the age of 75, and interprets this absence as support for his assertion that the holiday is most important for older women. While the holiday may be important for older women, the fact that there were different results for those under 75 and over 75 indicates that his results depend on whether the dividing line is drawn at 65, 70, 75, or some other number.

One of the present authors (Smith) has had four students write term papers attempting to replicate Phillips' findings with completely different data sets. None found a statistically significant death dip. One of these papers looked at 400 randomly selected persons in the Biography Almanac (which lists more than 20,000 internationally famous newsmakers);[5] there were slightly fewer deaths in the month preceding the birthmonth than in the month after (30 versus 32), but the P-value was 0.58. A study of 372 deceased persons in the 1972 issue of Who Was Who in the USSR found slightly more deaths in the month preceding the birthmonth than in the month following the birthmonth (32 versus 31) and even more deaths two and three months before the birthmonth, with the result almost statistically significant at the 5 percent level (P-value = 0.056), a result that directly contradicts Phillips' conclusion![6] A study of 630 obituaries from the Los Angeles Times found slightly fewer deaths in the month preceding the birthmonth than in the month after (50 versus 53), but many more deaths in the second month before the birthday than in the second month after (63 versus 50); the P-value was 0.59.[7] The fourth study examined 264 obituaries from four newspapers and compared the death date to the birthdate (rather than the birthmonth), and found slightly fewer deaths in the month preceding the birthdate than in the month after (18 versus 20); but the P-value was 0.30.[8] Figure 1 shows the results of combining the first three studies (which all followed Phillips in using the birthmonth as the reference point). The deaths near the birthmonth were almost exactly equal to the expected values for the null hypothesis that there is no relationship between the deathmonth and birthmonth. While there were slightly fewer deaths than expected during the month preceding the birthmonth (112 versus 116.83), there were far more than expected during the second and third months preceding the birthmonth (134 and 131, respectively). A chi-square test gives a P-value of 0.500. If there is no relationship between the deathmonth and birthmonth, deviations as large or larger than those shown in Figure 1 can be expected exactly 50 percent of the time.

Puzzled by the fact that four independent attempts to confirm Phillips' theory failed, we decided to reexamine the data that launched Phillips' career. One perplexing aspect of his analysis, given his opening description of the dramatic deathbed scene, is that he lumped together all deaths that occur during the birthmonth, not distinguishing those that occurred before the birthday from those that occurred afterward. Instead he separated the deaths into these twelve monthly categories: birthmonth, one month before birthmonth, one month after birthmonth, and so on.

Phillips interprets those deaths that occurred during the deathmonth as evidence that these people were able to postpone death until the celebration of their birthdays. In fact, when we reexamined his data, we found that of the 26 people who died during their birthmonths, 13 died before their birthdays, 1 died on his birthday, and 12 died after their birthdays! These 26 people who died close to their birthdays were not at all successful in postponing death.

Given the alleged importance of the birthday event, a more natural set of twelve categories is one month preceding the birthday, one month following the birthday, and so on. And this is what we used. We counted the 30 days preceding the birthday as one month preceding the birthday, 31 to 60 days preceding the birthday as two months preceding the birthday, and so on. The birthday itself and the subsequent 29 days were counted as one month following the birthday. The next 30 days were two months after the birthday. To account for the fact that a year has either 365 or 366 days, rather than 360, the six intervals farthest from the birthday have 31 days rather than 30.

Table 1 shows our data for the deceased people in Four Hundred Notable Americans,[9] the first work analyzed by Phillips. Phillips gives data for 348 people, reporting that, "The total number of deaths is less than 400 because (1) some of those in the source volume have not yet died; (2) for some of those in the volume, the month of birth and/or death is not known. By using more recent reference works, we were able to find complete birth and death information for 386 persons in Four Hundred Notable Americans. Table 1 shows the results for these 386 people. The number of deaths during the month preceding the birthday is only slightly smaller than expected, and is larger than the number of deaths for two and three months preceding the birthday. A chi-square test gives a P-value of 0.418, far from statistical significance at the 5 percent level (and far from the 0.025 value reported by Phillips).

We next turned to three additional samples analyzed by Phillips: persons listed in both Who Was Who in America[10] and in a U.S. appendix to Royalty, Peerage and Aristocracy of the World.[11] We followed Phillips in excluding famous people whose surnames do not appear in this appendix, although there is no persuasive reason for doing so. Phillips justifies his focus on famous people by arguing that their birthdays may be celebrated publicly with considerable attention. In the United States, at least, fame is not noticeably diminished by the absence of one's name from a list of aristocratic families. Indeed, it could well be argued that those persons listed in Who Was Who solely because of their surname are the ones who should be excluded. Those who have made it into Who Was Who without the benefit of an aristocratic surname are more likely to be well-known persons whose birthdays are celebrated publicly, than are those who have nothing more going for them than their surname. Phillips' decision to exclude nonaristocratic newsmakers is but one of many seemingly arbitrary decisions that could have been made after looking at the data, rather than before. The fact that Phillips excluded those who did not have aristocratic surnames from his Who Was Who samples, but not from his 400 Notable Americans data, is certainly questionable. In collecting these data, we were also struck by the very large number of published compilations of notable people; a determined data miner could surely find a subset of some collection that confirms almost any theory.

Phillips' three Who Was Who samples are based on every name in the 1951-60 edition, every name in the 1943-1950 edition, and every other name in the 1897-1942 edition. In each case, the surnames had to be listed in the Royalty appendix, and persons who appear in more than one volume or in Four Hundred Notable Americans were not included a second time. Phillips reports that each sample showed a death dip in the month preceding the birthmonth. We did not find this death dip.

Our samples are somewhat larger than Phillips' samples. Phillips reports that his 1943-1950 sample excludes those who died during World War II and that his 1897-1942 sample excludes those who died during the either World War. We did the same, although the argument for doing so is unpersuasive. It is reasonable to exclude combat deaths; however, Who Was Who does not identify whether the person died in combat and it seems unlikely that a soldier who died in combat had accomplished enough in an abbreviated life to be listed in Who Was Who. Most of those excluded by Phillips are almost surely not combat deaths. A practical problem with Phillips' report is that the time period spanning each World War is not clearly defined, which is evidently the reason for the modest differences in our sample sizes.

It is also puzzling that Phillips did not exclude those who died during these world wars from his first sample (Four Hundred Notable Americans) and that he did not exclude from any of his samples those who died during time periods spanned by other wars--including the Revolutionary War, Civil War, Spanish-American War, and Korean War. Instead of arbitrarily excluding people who might have died in combat in two selected wars, it seems more reasonable to use all the available data. If some of these notable people did die in combat, there is no reason to suppose that the timing of these deaths would reverse an otherwise systematic relationship between deathday and birthday. At most, it would dilute the relationship somewhat.

Table 2 shows that in two of the three Who Was Who samples, the number of deaths in the month preceding the birthday was larger than expected--the opposite of Phillips' death dip theory--though none of the results are statistically significant at the 5 percent level. The only sample to show a death dip has a P-value of 0.99. Table 3 shows the combined results for the three Who Was Who samples and for all four samples. There are a relatively large number of deaths in the month preceding the birthday and the P-value of 0.051 for the four combined samples is very close to statistical significance at the 5 percent level. No doubt, with a little effort we could have made it statistically significant by finding reasons for excluding some people from these data.

We did find, as did Phillips, a relatively large number of deaths in the months following the birthday. One explanation of this rise in deaths in the months preceding and following the birthdate is that, rather than postponing death until after the birthday, the anxiety associated with this milestone and the excesses associated with its celebration are sometimes fatal. Another explanation is that Phillips' results are just a fluke created by a selective use of data.