Medium-term mortality of Dutch professional soccer players.
Koning, Ruud H. ; Amelink, Remko
Introduction
Professional soccer players are expected to be in better health
than the average member of their age cohort, because their profession
demands a high level of physical fitness. Moreover, during their active
career, their health is monitored on a regular basis by physicians
(usually employed by the teams for whom they play). However, whenever a
(former) well-known player dies, such an event is publicised widely.
Three examples are the deaths of Daniel Jarque, Marc-Vivien Foe, and
David di Tommaso. Daniel Jarque died on 8 August 2009. He was a player
for Espanyol, a team in the Primera Division (the highest Spanish soccer
league). Marc-Vivien Foe died from a fatal cardiac arrest in 2003 whilst
playing a game for the Cameroon national team. He was playing for
Manchester City at that time. In the Netherlands, FC Utrecht player
David di Tommaso died unexpectedly of a cardiac arrest in 2005.
Perceived mortality of soccer players may be distorted.
In this article, we examine medium-term mortality of professional
soccer players in The Netherlands. We consider a sample of all players
active in the highest professional soccer league in The Netherlands
(Eredivisie) in three seasons 1970/1, 1971/2, and 1972/3, and examine
their status on 1 January 2009. More specifically, we address two main
questions. First, are mortality rates of these soccer players
significantly lower or higher than the ones in the Dutch population in
general? Second, we examine whether there is any heterogeneity of
mortality rates between teams. Even though it may be interesting to
hypothesise on the causes of differences in mortality rates, if any, we
are not able to attribute any deviations from expected mortality to
causal factors. Causal explanation requires far more detailed data that
are not presently available. We develop a methodology that can be
generalised to other groups of athletes.
This study is related to and complements a number of other
articles. Teramoto and Bungum (2010) survey fourteen articles on
mortality and longevity of elite athletes. The common findings in the
studies reviewed are that elite endurance athletes and mixed-sports
athletes (soccer players belong to this category) survive longer than
the general population. The likely primary cause of this effect is lower
cardiovascular disease mortality. However, results on power athletes are
mixed.
Sarna, Sahi, Koskenvuo, and Kaprio (1993) collected data on
athletes representing Finland in elite international contests, and match
athletes with healthy non-athletes of the same age and region using the
Finnish Defence Forces conscription database. They find that Finnish
team members have higher life expectancy that is mainly explained by
decreased risk of cardiovascular mortality.
In the case of soccer, particular attention has been given to
amyotrophic lateral sclerosis (ALS) as a cause of mortality (Belli and
Vanacore 2005; Taioli 2007). Belli and Vanacore estimate standardised
proportionate mortality ratios for a number of causes of death for
24,000 Italian soccer players active in the period 1960-1996. They find
that the mortality ratios adhere substantially to expected mortality
with the exception of mortality for diseases of the nervous system. In
particular, ALS is more prevalent than expected, and this is possibly
related to the use of dietary supplements and drugs.
Mortality among elite athletes as professional soccer players may
differ from mortality in the general population for a number of reasons.
First, athletes are healthy because they are a self-selected group of
the population. Second, during their active career, their health status
is monitored closely, and they have access to high quality medical care.
Third, the soccer players we consider in this article may be relatively
well-off after their career, enabling them to have better access to
medical care after their career. Finally, we mention that being
successful and having a high social status may also contribute to lower
mortality outcomes (as has been shown for Academy Award winners by
Redelmeier and Singh (2001)). On the other hand, mortality could be
higher due to physical strain during the active career, possible use of
performance enhancing drugs, or a celebrity life-style after (or
possibly even during) the active career. Another reason for higher
mortality could be depression or other mental health related disorders
as a result of poor adjustment to an alternative career or lifestyle.
Given the growth in the size of the soccer industry, different mortality
is potentially an important issue.
Due to the long time span between the active career and the moment
of investigation, we need to allow for changes in mortality rates over
time. Mortality rates have decreased for all ages, mainly due to
improvements of medical knowledge and living conditions. In the analysis
below, we take these changes into account.
Data and Methods
In this section, we present our research design in three steps.
First, we discuss compilation of the dataset of professional soccer
players. Using this dataset, we estimate expected mortality over time.
Second, since these mortality rates change over time, we discuss how we
allow for these rates to change over time. Third, we indicate how we
compare expected mortality (based on the adjusted mortality rates) with
observed mortality (as measured in the dataset with players).
We begin by describing how we collected the basic dataset of soccer
players. We focus on all players who have played in the highest level of
Dutch professional soccer ('Eredivisie') in one of the three
seasons 1970-71, 1971-72, and 1972-73. For all players, we assess on 1
January 2009 whether he is living or dead (and the date of death). We
choose players from these seasons for the following reasons. First,
enough time should have elapsed between the active career and the moment
of measurement of survival. If the moment of measurement is too soon
after the active career, individual survival will be censored for too
many players, making it hard if not impossible to distinguish between
expected survival and observed survival. By 1 January 2009, 28 out of
371 players had died, survival in our sample is censored for 92.5 per
cent of all observations. By considering three seasons instead of one
season (for example, 1970-71), the number of observations is increased
from 240 to 371. We use three consecutive seasons, so that all players
have access to the same general state of medical knowledge, and have
approximately similar stocks of health.
Lists of almost all players who were active in these three seasons
were provided by the Koninklijke Nederlandse Voetbal Bond (the Dutch
soccer association), and Infostrada (a private company specialising in
collecting and publishing sports data). The list of Infostrada contained
names of players who have played at least one match in one of the three
seasons, and we use that list. Contract players with no match
appearances are not representative of the population of top level
professional soccer players and are not included. The list also
contained the date of birth of each player, this information was checked
against other sources, such as club websites and books. No discrepancies
were found. To complete the data set, we collected information on
mortality of each player using a variety of sources, websites of the
teams being one of them. Finally, each team was sent a list with the
names of the players, their dates of birth and dates of death if
applicable, with a request to check these dates. We only use information
that has been validated explicitly by the teams. Teams that did not
respond, or were unable to check the dates of birth and death, are not
included in the analysis.
Table 1 presents the raw counts of deceased players by team.
Thirteen out of 23 teams that played in the highest league in at least
one of the three seasons responded. Some teams played only one or two
seasons in the highest league, as for example Vitesse. The number of
players in each pair of columns is not unique, so the last pair of
columns is not the sum of the earlier columns. In total, 28 out of 371
players have died between their active career and 1 January 2009 (7.5
per cent).
In the second step, we need to estimate mortality in a group with
this age composition. To do so, we need age and calendar time specific
mortality rates. The Actuarial Association publishes mortality tables in
The Netherlands for five year periods (for example, 2000-05), separately
for men and women. Since we consider professional soccer players, we use
the tables for men only. In the remainder, p(x, t) denotes the
probability that a male aged x survives one more year, in calendar year
t. The corresponding mortality rate is q(x, t) [equivalent to] 1 - p(x,
t). Survival probabilities have, in general, increased markedly over
time (see for example figures A-1 and A-2 in the Appendix). The one year
mortality rate of a 25-year old male has decreased from q(25, 1970) =
0.0008582 in 1970 to q(25, 2010) = 0.0006914 in 2010. Decreased
mortality rates translate into increased residual expected lifetimes:
for a 25-year old male residual lifetime has increased from 47.6 years
in 1970 to 52.4 years in 2010. We smooth mortality rates over time using
splines (Harrell 2001; Currie, Durban, and Eilers 2004) so as to avoid
sudden discrete jumps in the one year mortality rates. If we were to
base our analysis on constant mortality rates of 1971 (that is, we would
not allow for improvements of medical science over time), we would
underestimate survival and we would be biased towards concluding that
mortality among soccer players is lower than mortality in the general
population. In the last section, we discuss the sensitivity of our
results to this approach.
Now that we have a portfolio of 371 soccer players of different
ages, and one year mortality rates, we derive the probability
distribution of the number of survivors at any moment in (calendar) time
by simulation. We denote the number of players of age x at calendar time
t by N(x, t), so that the total number of players alive at calendar time
t is N(t) [equivalent to] [[summation].sub.x] N(x, t). In our model, we
assume that all players are born on 1 January. Other than Dudink (1994),
we do not reject the hypothesis that birthdates of players are
uniformity distributed throughout the year. Let the probability that a
player aged x alive at t survives for s years be denoted by p(x, t, t +
s) (so that one-year survival probability is p(x, t) [equivalent to]
p(x, t, t + 1). This s-year survival probability is the product of
one-year survival probabilities, taking aging and time dependence into
account:
p(x, t, t + s) = p(x, t, t + 1) x p(x + 1, t + 1, t + 2) x ... x
p(x + s - 1, t + s - 1, t + s) (1)
We fix [t.sub.0], the starting point of the analysis to be 1
January 1971. The number of surviving players N(x + s, [t.sub.0] + s)
follows a binomial distribution with parameters N(x, [t.sub.0]) and p(x,
[t.sub.0], [t.sub.]+ s). The total number of players who survive at
[t.sub.0] + s is
N([t.sub.0] + s) = [summation over X + S] N(x + s, [t.sub.0] + s),
s = 1, 2, ..., 38.
with the summation taking place over all relevant values of x + s.
Since we consider three seasons, in our calculations we allow for an
increase of our portfolio of players at times [t.sub.0] + 1 and
[t.sub.0] + 2 (when the new players of 1971-72 and 1972-73 are added to
the portfolio). The distribution of N([t.sub.0] + s) is not available in
closed form, therefore we approximate it by simulating the individual
terms in the sum in equation (2). This distribution gives survival over
time if the pool of soccer players were to have similar mortality to the
general population. We refer to this distribution as population
survival.
As opposed to Belli and Vanacore (2005), Sarna et al. (1993), and
Taioli (2007), we do not have access to administrative data. For this
reason, we are unable to perform a case-control study, and we do not
know the exact cause of death. While our dataset is less rich in this
respect, our approach is flexible and can easily be extended to other
settings.
[FIGURE 1 OMITTED]
Results
Mortality per team is given in Table 1. We have been able to obtain
complete information on 371 unique players, 28 of whom died before 1
January 2009. The first question is how observed mortality compares to
expected mortality, using the approach discussed in the previous
section. First, in Figure 1 we graph median survival over time and its
associated 80 per cent and 90 per cent confidence bands. The confidence
bands are obtained by linear interpolation. The solid decreasing line in
the center of the bands shows the expected development of the number of
players in the portfolio, if the general population mortality rates were
to apply to them. As discussed in the previous section, this curve does
reflect decreasing mortality rates over time. The dashed line indicates
observed survival, and after 2001 observed survival is higher than the
80 per cent upper confidence limit. By 1 January 2009, observed survival
is noticeably higher than the 90 per cent upper confidence limit. We
have calculated the exact p-value in 2009 to be 0.003. In Table 2 we
give observed values for observed survival n(t) for the period 2000-09.
Also, we show how population survival is distributed around this point.
The column p-value equals Pr(N(t) [greater than or equal to] n(t)), the
lowest level of significance that would give a rejection of the null
hypothesis in year t. Note that this is a one-sided p-value, the
hypothesis being tested being that observed survival exceeds population
survival. As of 2006 the p-value is smaller than the level of
significance a = 0.05. Observed survival is significantly higher than
survival in the general population, in other words, mortality among
soccer players is lower than the mortality in the general population. In
the Appendix, we give similar tables for survival of players from each
of the three seasons, and these tables lead to the same conclusion.
The second question is whether mortality varies by team. The
players represented thirteen different teams. In total, we have
information on 35 team-years out of 54 possible teams-years (eighteen
teams times three seasons). The aggregate mortality rate is 7.5 per cent
over 37 years (so the average mortality rate is 0.27 per cent annually).
The number of deaths per team seems to vary, but the number of players
at risk differs by team as well. To assess whether or not this variation
by team is systematic, we calculate Pearson's [chi
square]-statistic, which is 16.2 in this case. Under the null hypothesis
of equal mortality between teams, we calculate the p-value as 0.12, by
permutating all deaths and survivals over the teams. We prefer a
permutation test to an asymptotic test because of the limited number of
observations and small mortality rate. The hypothesis of equal mortality
between teams cannot be rejected.
Discussion
In this article, we have shown that the medium-term mortality in a
sample of professional soccer players active in the highest league of
Dutch soccer (1970-71 until 1972-73) is significantly lower than
mortality in the general population (p = 0.003). We do take into account
that mortality rates decrease over time. The difference between expected
survival and survival in the portfolio of soccer players has been
increasing since 2006. Also, we have shown that mortality does not vary
significantly by team.
In our analysis, we smoothed mortality rates over time. We
performed the same test as presented in the last column of Table 1 under
two alternative models. First, we did not smooth mortality rates, but we
used the five-year mortality rates as published by the Actuarial
Association. That is, we allowed for mortality rates that change over
time, but in a discrete, discontinuous fashion. The p-value in 2009 is
in that case 0.003 as well. The second alternative scenario is to ignore
changes in mortality and assume that the mortality rates of 1971 apply
to the whole 1970-2009 period. In that case, the p-value in 2009 is
0.000, smaller than 0.003. The difference between this approach and our
approach is perhaps best illustrated by looking at survival by 1 January
2000. Allowing for non-constant mortality rates, the p-value is 0.195,
under the incorrect assumption of constant mortality rates, it is p =
0.004. In other words, expected survival is underestimated by assuming
mortality rates that are constant over time.
As a final remark, we want to point out that the methodology
developed in this article can be easily applied to other sports to
assess development of mortality patterns over time. Data requirements of
the approach are modest. In particular, this approach can be used to
shed more light on the inconsistent results found for athletes
participating in anaerobic (power) sports (see for example, in Teramoto
and Bungum 2010).
Acknowledgements
We thank Infostrada, KNVB, the Actuarial Institute, and Henk Grim
for helping compiling the datasets. We also thank the football teams
that were able to check our data. This research is part of the Research
program Passion, Practice and Profit. The program ran from 2007 to 2010
and was financed by the Dutch Ministry of Health, Welfare and Sport
(VWS). It was coordinated by the W. J. H. Mulier Institute for Social
Science Sports research. In the program researchers of the institute
joined forces with researchers from four associated universities
(Tilburg University, University of Amsterdam, Utrecht University and
University of Groningen).
Appendix
In figures A-1 and A-2, we give two examples how one-year survival
probabilities change over time, for two different ages, x = 24 and x =
65. The dots indicate the probabilities as published by the Actuarial
Association. As discussed in the text, the Actuarial Association
publishes tables in five year intervals. The smooth curve is the spline
we used to smooth the survival rates. Smoothing was obtained by
estimating cubic B-splines with two knots, one at 1981 and one knot at
1995. These spline functions have a continuous first and second
derivative. By using such smooth functions, we implicitly assume that
survival (or its counterpart, mortality) changes gradually over time.
[FIGURE A-1 OMITTED]
[FIGURE A-2 OMITTED]
[FIGURE A-3 OMITTED]
Table A-1: p-values of null-hypothesis, t = 2009, players 1970/71.
Portfolio season 1970/71
Pr(N(t) Pr(N(t) Pr(N(t)
Year n([t.sub.0]) n(t) < n(t)) = n(t)) > n(t)) p-value
2000 240 228 0.709 0.090 0.201 0.291
2001 240 228 0.814 0.066 0.120 0.186
2002 240 228 0.893 0.043 0.064 0.107
2003 240 227 0.909 0.037 0.055 0.091
2004 240 227 0.956 0.020 0.025 0.044
2005 240 225 0.942 0.024 0.034 0.058
2006 240 225 0.976 0.011 0.013 0.024
2007 240 223 0.971 0.012 0.016 0.029
2008 240 223 0.990 0.005 0.005 0.010
2009 240 222 0.994 0.003 0.003 0.006
Table A-2: p-values of null-hypothesis, t = 2009, players 1971/72.
Portfolio season 1971/72
Pr(N(t) Pr(N(t) Pr(N(t)
Year n([t.sub.0]) n(t) <n(t)) = n(t)) > n(t)) p-value
2000 235 225 0.812 0.072 0.116 0.188
2001 235 224 0.820 0.067 0.113 0.180
2002 235 224 0.894 0.044 0.062 0.106
2003 235 223 0.906 0.039 0.055 0.094
2004 235 221 0.873 0.047 0.081 0.127
2005 235 220 0.895 0.039 0.066 0.105
2006 235 220 0.950 0.020 0.030 0.050
2007 235 219 0.963 0.016 0.021 0.037
2008 235 219 0.986 0.007 0.007 0.014
2009 235 219 0.995 0.002 0.002 0.005
Table A-3: p-values of null-hypothesis, t = 2009, players 1972/73.
Portfolio season 1972/73
Pr(N(t) Pr(N(t) Pr(N(t)
Year n([t.sub.0]) n(t) < n(t)) = n(t)) > n(t)) p-value
2000 221 211 0.663 0.106 0.231 0.337
2001 221 210 0.665 0.102 0.234 0.335
2002 221 210 0.771 0.080 0.149 0.229
2003 221 209 0.783 0.074 0.143 0.217
2004 221 208 0.800 0.068 0.132 0.200
2005 221 206 0.747 0.076 0.178 0.253
2006 221 206 0.848 0.053 0.100 0.152
2007 221 205 0.875 0.044 0.081 0.125
2008 221 205 0.937 0.025 0.038 0.063
2009 221 204 0.954 0.019 0.027 0.046
[alpha] = 0.05
[alpha] = 0.10
[alpha] > 0.10
References
Belli, S. and Vanacore, N. (2005) 'Proportionate mortality of
Italian soccer players: Is amyotrophic lateral sclerosis an occupational
disease?, European Journal of Epidemiology, 20(3), pp. 237-242.
Currie, I., Durban, M. and Eilers, P. (2004) 'Smoothing and
forecasting mortality rates, Statistical Modelling, 4(4), pp. 279-298.
Dudink, A. (1994) 'Birth date and sporting success, Nature,
368, p. 592.
Harrell, Jr, F. E. (2001) Regression Modelling Strategies,
Springer, New York.
Redelmeier, D. A. and Singh, S. M. (2001) 'Survival in Academy
Award-winning actors and actresses, Annals of Internal Medicine,
134(10), pp. 955-962.
Sarna, S., Sahi, T., Koskenvuo, M. and Kaprio, J. (1993)
'Increased life expectancy of world class male athletes, Medicine
and Science in Sports and Exercise, 25(2), pp. 237-244.
Taioli, E. (2007) All causes mortality in male professional soccer
players, European Journal of Public Health, 17(6), pp. 600-604.
Teramoto, M. and Bungum, T. J. (2010) 'Mortality and longevity
of elite athletes, Journal of Science and Medicine in Sport, 13, pp.
410-416.
Ruud H. Koning *
Remko Amelink *
* University of Groningen, The Netherlands
Professor Ruud H. Koning is professor of Sports Economics at the
University of Groningen, The Netherlands. His research fields include
economics and econometrics of sports, applied microeconometrics and
applied actuarial science. He is head of the Department of Economics,
Econometrics, and Finance of the Faculty of Economics and Business.
Besides, he is member of the supervisory board of Algemeen Belang
(funeral insurer), and avid soccer fan. He can be reached at
r.h.koning@rug.nl.
Remko Amelink a senior actuarial analyst at ABN AMRO Insurance.
Currently, he is following the Executive Master of Actuarial Science
program at TiasNimbas (Tilburg, The Netherlands) to become a certified
actuary. His email address is amelinkr@hotmail.com.
Table 1: Number of players in portfolio. The total number of
unique players of each club can be found in column 'Total'. Some
players have played for multiple clubs in the dataset. Therefore,
the total number in the dataset is smaller than the sum of the
unique players of every club.
Club 70-71 71-72
Total Number of Total Number of
number of deceased number of deceased
players players players players
Ajax 15 3 18 0
AZ 23 1 0 0
FC Groningen 0 0 22 4
FC Twente 22 2 22 2
FC Utrecht 26 3 21 3
FC Volendam 24 0 26 0
Feyenoord 21 2 22 1
Haarlem 24 4 0 0
NAC Breda 21 0 19 0
NEC 22 0 24 0
PSV 17 0 19 0
Sparta 25 3 20 3
Vitesse 0 0 22 2
Total 240 18 235 16
Club 72-73 Overall
Total Number of Total Number of
number of deceased number of deceased
players players players players
Ajax 19 0 23 3
AZ 22 2 32 2
FC Groningen 22 5 26 5
FC Twente 17 2 25 2
FC Utrecht 22 2 34 4
FC Volendam 0 0 31 0
Feyenoord 21 0 32 2
Haarlem 18 2 32 4
NAC Breda 20 1 30 1
NEC 22 0 38 0
PSV 18 0 26 0
Sparta 20 3 30 3
Vitesse 0 0 22 2
Total 221 17 371 28
Table 2: Development of complete portfolio over time,
with p-values by year.
Whole portfolio
Pr(N(t) < Pr(N(t) Pr(N(t) >
Year n([t.sub.0]) n(t) n(t)) = n(t)) n(t)) p-value
2000 371 354 0.805 0.059 0.136 0.195
2001 371 353 0.849 0.048 0.103 0.151
2002 371 353 0.927 0.027 0.046 0.073
2003 371 351 0.925 0.026 0.048 0.075
2004 371 349 0.928 0.024 0.047 0.072
2005 371 347 0.936 0.022 0.043 0.064
2006 371 347 0.977 0.009 0.014 0.023
2007 371 345 0.982 0.007 0.011 0.018
2008 371 345 0.995 0.002 0.003 0.005
2009 371 343 0.997 0.001 0.002 0.003
[alpha] = 0.05
[alpha] = 0.10
[alpha] > 0.10
