,mW'ZTZ.

,,~

" Jll
ELSEVIER S~IENCE Mechanisms of Ageing and Development
IRELAND 74 (1994) 15-33

The analysis of survival (mortality) data:

Fitting Gompertz, Weibull, and logistic
functions

David L. Wilson
Department of Biology, University of Miami, P.O. Box 249118, Coral Gables,
FL 33124, USA

(Received 16 July 1993; revision received 27 November 1993; accepted 20 December 1993)

Abstract

Survival functions are fitted to survival data from several large populations. The Gompertz
survival function corresponds to exponential mortality rate increases with time. The Weibull
survival function corresponds to mortality rates that increase as a power function of time. A
two-parameter, logistic survival function is introduced, and corresponds to mortality rates
that increase, and then decrease, with time. A three-parameter logistic-mortality function also
is examined. It reflects mortality rates that rise, and then plateau, with age. Data are from
published studies of medflies, Drosophila, house flies, flour beetles, and humans. Some survi-
val data are better fit by a logistic survival function than by the more traditionally used
Gompertz or Weibull functions. Gompertz, Weibull, or logistic survival functions often fit the
survival of 95+% of a population, and the 'tails' of the survival curves usually appear to fall
between the values predicted by the three functions. For some populations, such 'tails' appear
to be too complex to be fit well by any simple function. Survival data for males and females
in some populations are best fit by different functions. Populations of 100 or more are needed
to distinguish among the functions. When testing effects of environmental or genetic manipul-
ations on survival, it has been common to determine the changes in parameter values for a
given function, such as Gompertz. It may be equally important to determine whether the best-
fit function has changed as well.

Key words." Aging; Mortality; Gompertz

004%6374/94/$07.00 © 1994 Elsevier Science Ireland Ltd. All rights reserved.

SSDI 0047-6374(93)01420-D
16 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

1. Introduction

Two functions have been widely used by biologists to fit survival and mortality
data [1,2]. The most popular is termed Gompertz [3], and specifies an exponential
rise in mortality rate with age of a population. A Gompertz function appears to 'fit'
survival data for a variety of populations [2]. The other function is Weibull [4],
which describes mortality rate as a power function of time. While it has been
recognized that deviations from Gompertz mortality occur [5,6], recent studies [7,8]
indicate that some populations exhibit quite dramatic departures from Gompertz
kinetics.
If different populations are best fit by different survival functions, then a new tool
for the analysis of survival data is possible. Not only different species, but also
populations that differ because of genetic or environmental manipulations, could
best be fit by different survival functions. To develop such a tool, several functions,
which span the variety of survival curves, are needed.
Here, several previously published sets of survival data from large populations are
examined. Gompertz, Weibull, and a two-parameter, logistic survival function are
fit to the data by determining parameter values giving the least error. The 'fits' of
the three functions to the data then are compared. The fits of the functions to data
from males and females of the same species also are compared. In addition, the fits
of several functions having more than two parameters are examined. An estimate is
made of the minimum population size needed to determine which function best fits
the survival of a population.

2. Methods

2.1. Functions

Mortality rates and survival are related to one another by the equation:

ds/dt = - m s

where s = fraction of a population surviving, m = mortality rate, and ds/dt = rate of

change of s with time.
This equation was used to determine the survival function corresponding to a
mortality function, or visa versa. The following functions were studied:

A. Two-parameter functions.
1. Gompertz (exponential) mortality function:

m = A e Gt

where A and G are parameters to be fitted, t = time, and e = exponential =

2.71828...
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 17

The corresponding Gompertz survival function:

eO]
where exp = exponential.

2. Weibull (power) mortality function:

m = at ~z-1)

where a and g are parameters to be fitted.

The corresponding Weibull survival function:

s=exp [- (g) tg]

3. Mortality function corresponding to logistic survival:

w t w-I
m-
v TM + t w

where v and w are parameters to be fitted.

The logistic survival function:

1
S~
l+(t) w

B. Functions with more than two parameters

1. Logistic mortality function (3 parameters):
ab
m=
a + ( b - a ) e -kt

where a, b, and k are parameters (a = minimum mortality rate, at t = 0, and b =

maximum mortality rate).
The corresponding survival function:

s = e x p [ - b t - ( b / k ) ln a + ( b - a ) e-kt

2. Gompertz-Makeham (Comfort, 1979) mortality function (3 parameters):

m = A e ct + Mo

where A, G, and M0 are parameters.

18 D.L. Wilson/Mech Ageing Dev. 74 (1994) 15-33

The corresponding Gompertz-Makeham survival function:

s = exp [-Mot+ ( A ) (1- eat) ]

2.2. Sources of data and population sizes

The following sources of data were used for the analyses. House flies consisted of
populations of 4000 males and 4000 females [9] (1 day's data in the life table was
missing in the published paper, but the values were easily inferred from the data on
the previous and following days). Medflies consisted of 1.2 million males and females
[7], and separate tables for 598,118 males and 605,528 females were from Carey [10].
Corrections to several typographical errors in the Carey data were made and con-
firmed with Carey (pers. commun.). Drosophila data were from two studies: Curts-
inger et al. [81, who studied a population of 5751 males from an inbred line (Block
IV; actual life-table data obtained from Curtsinger, pers. commun.); and Pearl and
Parker [11], who studied a population of 1415 females and 1407 males from line 107
('wild type') and 524 females and 456 males from a 'pure, vestigial' line. Flour beetles
consisted of 854 females and 893 males from Pearl et al. [12], with data taken from
day 50 onwards. Human data was from two sources: U.S. Census data from 1988
[13] (cross-sectional study) with data taken from age 10 onwards; and a longitudinal
study of 4404 males and females beginning at age 100 from Thatcher [14].
For both flour beetles and humans, early deaths, which did not appear to be
senescence-related, were eliminated from the analysis by only taking data from 50
days and older in the beetles and 10 years and older for the U.S. Census data on
humans. In both cases, the early drop in survival was followed by a plateau before
significant mortality rate increases from senescence.

2.3. Data analysis

In all cases, the survival functions (rather than mortality functions) and survival
data (time interval, number of animals alive at start of interval, and number of ani-
mals dying during the interval) were used to determine the parameter values, as the
survival functions tend to give more reliable values for the parameters than do the
corresponding mortality functions [15].
To speed data analysis, a C-language computer program was developed that, from
survival data, determines 'best fit' parameter values for the Gompertz survival func-
tion, Weibull survival function, and logistic survival function. The program is IBM ®
PC compatible, and uses a simplex algorithm [16,17] to do non-linear, least-squares
analyses. The simplex algorithm requires initial estimates for parameter values, from
which to begin the iteration. (Proper choice of initial parameter values reduces the
chance of getting trapped in a local minimum.) For both Gompertz and Weibull, lin-
ear regression analyses were performed on the log-transform of the mortality func-
tions to obtain initial parameter values. For the logistic survival function, the
median life span, as determined by the Weibull non-linear analysis, was used as the
D.L. Wilson/Mech. Ageing Dev. 74 (1994) 15-33 19

initial value for parameter v, and the initial value for parameter w was arbitrarily
set at 4. Parameter values that minimize error (error = sum of the squares of the
differences between data and calculated function values) were determined by the
program. The program is available from the author.
Other functions were 'fit' using SigmaPlot ®, a commercially available curve-fitting
and plotting program, which uses a Marquardt-Levenberg algorithm [18] to
minimize the same error function. When tested on the same data and function, the
simplex and Marquardt-Levenberg methods gave nearly identical parameter values.
The error values given in this paper are:

~ SSR
n-2

where SSR = the sum over all time intervals of the squares of the differences between
the calculated values of survival and the actual values, and n = number of time inter-
vals (data points).
In most cases, an un-weighted analysis was performed. This can contribute to a
better 'fit' for the tails of the curves, but does not take into account the expected
scatter in data, which will be greater on days when fewer animals die. The analysis
was repeated for some of the cases using an error function that was weighted by the
square root of the number of animals that died in each time period. In this case a
weighted error function also was calculated, and consisted of the square root of the
weighted sum of the squares divided by the sums of the weights. In all cases exam-
ined, the weighting produced small changes in the values of the parameters, and did
not change the functions that 'best-fit' the data.
Other C-language programs were developed to randomly select individuals from
several of the populations, and to determine survival and mortality values from the
equations, given parameter values.

3. Results

3.1. Fits of 2-parameterfunctions

In general, once the three survival functions, Gompertz, Weibull, and logistic sur-
vival, are fit to a data set, the Gompertz will show the most dramatic deotine in
numbers at the 'tail' of the survival curve, with Weibull declining more gradually,
and logistic survival declining even more gradually. The reason for this can be seen
if the three survival functions are examined. Gompertz survival declines as the ex-
ponential of an exponential of time; Weibull survival declines as an exponential of
a power function of time; and logistic survival gives a decline that is a reciprocakof
time to a power. (The same argument for the corresponding mortality functions is
as follows: for Gompertz, the mortality rate climbs exponentially with time; for
Weibull, the mortality rate climbs as a power function of time, and for the mortality
function corresponding to logistic survival, the mortality rate at first increases with
time, but then decreases.) j
20 D.L. Wilson/Mech. Ageing Dev. 74 (1994) 15-33

An analysis of the data from a variety of insects and for humans shows that dif-
ferent populations appear to be better fit by different functions, with Gompertz giv-
ing the best fit to survival data from some populations, Weibull giving the best fit
to other data, and logistic survival giving the best fit to the survival data from yet
other populations. Some populations exhibit survival curves that fall between pairs
of the functions. Also, for some populations, males and females appear to follow dif-
ferent curves. These results are presented below.

A. Medflies
Survival data from female medflies [10] are better fit by the logistic survival func-
tion than by either Gompertz or Weibull (Fig. 1 and Table 1). While a logistic curve
gives the best fit to the data, Fig. 1 also shows that the tail of the survival data
deviates from logistic survival: the last fraction of the surviving medflies dies at a
greater rate than predicted by the best-fit logistic survival parameters. The mortality
rates for the part of the data that is well fit by the logistic curve first increase with
age, and then decrease (Fig. 1), in agreement with Carey et al. [7].
An alternative logistic survival function,

l
S--
1 + ae bt

where a and b are parameters, did not fit the medfly data as well as the logistic survi-
val function given in the methods section (data not shown).
Male medflies [10] were better fit by a Weibull function (Table 1), with data at
the tail of the curve falling between Weibull and logistic survival (data not shown).

Bo Flour beetles
In contrast to the female medfly population, female flour beetles are best fit by
a Gompertz function (Fig. 2 and Table 2). The males are best fit by a Weibull func-
tion, with the tail of the survival curve falling between Weibull and Logistic (Fig.
2 and Table 2). Thus, in this species, and under the environmental conditions used
in the study by Pearl et al. [12], the survival data of males and females is fit best by
different functions. In fact, while male flour beetles have a significantly shorter mean
life span, they have a longer maximum life span, reflecting the differences in the
shapes of the curves (Fig. 2). For those beetles surviving 50 days, males have an ad-
ditional mean life expectancy of 143 days, while females have 175 days. Yet, the life
span of the longest lived males is about 100 days greater than the females. It is clear
that males and females of the same species can follow different survival curves.

C. House flies
House flies also can show different survival curves for males and females (Fig. 3
and Table 3). In this case, males are best fit by logistic survival while females exhibit
Gompertz kinetics. As was the case with female medflies, the last few percent of male
house flies show higher mortality rates than predicted by logistic survival (Fig. 3).
Female house flies also appear to deviate at the end of the curve, but in this case
they show lower mortality than that predicted by Gompertz (Fig. 3), and survival
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 21

FEMALE MEDFLIES
10
~ . ' ' 0.2 ' , ' , ' , '

Logistic s

08
" ~"'.~ I ..... Gompertz
",'~ ~ ......... Weibull j
c

!
o 0.6

',~~ 30 40 50 60 70
"~ 0.4

02

O0
10 20 30 40 50 60 70

O1

o
:~ 001
u
2
LOGISTIC S
'~ 0001
Ii ", ~-.,,. LOGISTIC M

0.0001 GOMPERTZ IIl~ ' • " " " ~

~L \~ W E I B U L L " ~
', %
0.00001 I I , I I I ",, I L
20 40 60 80 ~ O0 120 140

0.20 , ~ ,

015

010
2

005

000
5 10 15 20 25 50 55 40

Time (doys)

Fig. !. Survival curves and a mortality curve for female medfiies. Data (from Carey [10]) are shown as
open circles. The fit of four functions is shown. For each function, parameter values were adjusted to give
the best fit to the data. The four functions are the logistic-survival function, the Gompertz survival func-
tion, the Weibull survival function, and the logistic-mortality function. All functions are given in the
Methods section. The logistic mortality function has three parameters, the other functions have two par-
ameters. A re-plot of the data with a log scale for survival allows a viewing of the fit to data at the tail
of the curve. The mortality rates given by the logistic survival function for the first 38 days also is shown.
For those days, the logistic survival function gives a good fit to the data.
22 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

Table I
Medflies ( A n a l y s i s o f d a t a f r o m C a r e y [10l)

Function P a r a m e t e r values Error

Females
Gompertz s A = 0.00987; G = 0.1181 0.0261
Weibull s a = 0.00130; g = 2.468 0.0155
Logistic s v = 17.74; w = 3.875 0.00434
Logistic m a = 0.0006476; b = 0.01163; k = 0.04387 0.00208
Males
Gompertz s A = 0.00601; G = 0.122 0.0154
Weibull s a = 0.000325; g = 2.833 0.00543
Logistic s v = 21.02; w = 4.406 0.0119
Females (weighted analysis)
Gompertz s A = 0.00975; G = 0.122 0.0481
Weibull s a = 0.00110; g = 2.54 0.0289
Logistic s v = 17.73; w = 3.853 0.00747

falls between Gompertz and Weibull predictions. The male house flies show a rise
and then a decline in mortality rate with age (Fig. 3), as expected from logistic
survival.

D. Drosophila
Three inbred strains of Drosophila that have been the subject of large-cohort survi-
val studies were examined. These were wild-type and vestigial lines from Pearl and

F h l : i J ~ , ~ B E ! IL E i
1 O0

0 75

"2
o
"d

~ C bO

t~

C, 2S

0 ()O I , . . . . . . . . . - . . . . . . . . .

Time (days)

Fig. 2. Survival curves f o r m a l e a n d female f l o u r beetles. D a t a ( f r o m Pearl et al. [12]) a r e s h o w n as o p e n

circles f o r the m a l e s a n d closed circles f o r the females. O n l y the f u n c t i o n t h a t best fit e a c h d a t a set is
s h o w n . F o r males, the W e i b u l l survival f u n c t i o n is s h o w n . F o r females, the G o m p e r t z survival f u n c t i o n
is s h o w n . T h e inset s h o w s s o m e detail o f the tail o f the curves.
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 23

Table 2
F l o u r Beetles (Analysis o f d a t a f r o m Pearl et al. [121)

Function P a r a m e t e r values Error

Females
Gompertz s A = 0.00130; G = 0.0105 0.0104
Weibull s a = 0.0000133; g = 2.266 0.0329
Logistic s v = 167.6; w = 3.427 0.0601
Males
Gompertz s A = 0.00218; G = 0.0109 0.0288
Weibull s a = 0.0000732; g = 2.016 0.00915
Logistic s v = 127.98; w = 3.21 0.0196
(weighted analysis)
Gompertz s A =0.00211; G = 0.0113 0.0329
Weibull s a = 0.0000682; g = 2.033 0.0106
Logistic s v = 128.12; w = 2.033 0.0183

Parker [11] and an inbred line from a more recent study of Curtsinger et al. [8]. The
wild-type lines from Pearl and Parker [11] and the males from Curtsinger et al. [8]
all were best fit by Gompertz (Table 4), but, as with house flies, deviated from
Gompertz at the tails of the survival curves (Fig. 4). The data from Curtsinger et al.
[8] show some scatter that probably was caused by the feeding schedule of the ani-
mals [8]. The Pearl and Parker [11] data were not given for each day, so the form
of their data presentation could have masked similar scatter.
The vestigial line of Pearl shows different features. Females exhibit survival that
is best fit by a Weibull function, while survival of vestigial males is roughly midway
between Weibull and logistic survival, slightly favoring logistic (Fig. 5 and Table 4).
Both the vestigial males and vestigial females differ from wild type, which followed
Gompertz.

E. Humans
Fig. 6 shows cross-sectional data on humans from 1988 U.S. census data [13]. The
analysis confirms what has been known for some time [1], that humans exhibit survi-
val that is well fit by Gompertz, from age 10 until at least 85 years of age (Table
5). The population size for the cross-sectional study was 246 million, and was based
on the resident population in the United States in 1988 (based on 1980 census data)
[13].
An analysis of Thatcher's [14] data on humans over the age of 100 indicates that
mortality rates continue to increase, but at a lower rate, at very old ages (Fig. 6).
Gompertz continues to give a better fit to the data, but the mortality rate doubling
time (MRDT) has increased from about 8 years to about 16 years (Table 5, using
the fact that MRDT = (In 2)/G, where G is the Gompertz exponential parameter).
Thus, human survival, while best fit by Gompertz, appears to exhibit a change in
the Gompertz exponential parameter at very advanced ages. It should be noted,
however, that the Thatcher data combines males and females.
 24 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

FEMALE HOUSE FLIES

i i
0 75

i O 5O
o,
001
0 25

0 O 0 1 0 ' ~ i , ---060001 ~1 i b i ,'

Time(doys) 10 20

Tim~
30

(doys)
40 50 6u

MALE HOUSE FLIES

1 O0 ~ -

075

i 0 50 °
~ 901
O25
" 0
" 0

iho
~ ", 0000 ~
0oo
10 20 30 40 ......
50
o ool
20 30 40
Time (doys) Time (doys)
MALE HOUSE FLIES

015

#.
olo

005

/
J
t, i~ I

Time (doys)

Fig. 3. Survival curves and a mortality curve for house flies. Data (from Rockstein and Lieberman [9])
are shown as open circles. Both linear and log survival axes are shown, to allow a full view of the fit at
high and low survival levels. For the male house flies, a curve showing mortality rate for the first 30 days,
as determined by the logistic survival function, is given. The best-fit curve from the logistic survival func-
tion is shown as a solid line; the best-fit curve from the Gompertz survival function is shown as a long-
dashed line; and best-fit curve from the Weibull survival function is shown as a short-dashed line.

3.2. Fits of functions with more than two parameters

The very large cohort study of medflies [7,10] was subjected to further analysis
using functions with more than two parameters. Several such functions were tested,
and one gave somewhat better fits to the tail of the survival curve than did the two-
parameter functions used above.
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 25

Table 3
H o u s e flies (analysis o f d a t a f r o m Rockstein a n d L i e b e r m a n [9])

Function P a r a m e t e r values Error

Females
Gompertz s A = 0.00573; G = 0.0802 0.0079
Weibuli s a = 0.000028; g = 2.615 0.0217
Logistic s v = 27.873; w = 4.047 0.048
Males
Gompcrtz s A = 0.00679; G = 0.1778 0.0374
Weibull s a = 0.000515; g = 2.978 0.0220
Logistic s v = 15.745; w = 4.595 0.0062

Several three- and four-parameter logistic survival functions that were tested gave
only very minor improvements in the fit of the function to the medfly data. Even
with weighting (using reciprocal of the value of survival fraction as the weighting
variable), the tested functions showed only a minor improvement in fit (data not
shown).
Similarly, the Gompertz-Makeham function, with its third parameter, did not
show improvement over the two-parameter Gompertz function.
Table 4
Drosophila (analysis o f d a t a f r o m C u r t s i n g e r et al. [8], a n d Pearl a n d P a r k e r [1 I])

Function P a r a m e t e r values Error

Males ( Cartsimger et al. [8])

Gompertz s A = 0.00658; G = 0.101 0.0290
Weibull s a = 0.000412; g = 2.663 0.0387
Logistic s v = 22.79; w = 4.021 0.0623
Males ( Curtsf~ger et al. [8])(wdghted aawlysis)
Gompertz s A = 0.00689; G = 0.0987 0.0330
Weibull s a = 0.000510; g = 2.587 0.0452
Logistic s v = 22.83; w = 3.670 0.0660
LINE 107 (wild type; Pearl and Parker [11])
Females
Gompertz s A = 0.00178; G = 0.0646 0.0126
Weibull s a = 0.0000052; g = 3.34 0.0294
Logistic s v = 48.20; w = 5.057 0.0562
Males
Gompertz s A = 0.00157; G = 0.0731 0.0154
Weibull s a = 0.00000323; g = 3.52 0.0224
Logistic s v = 45.87; w = 5.344 0.0471
VESTIGIAL (Pearl and Parker [11])
Females
Gompertz s A = 0.0176; G = 0.0657 0.0262
Weibull s a = 0.00520; g -- 1.884 0.00698
Logistic s v = 18.06; w = 2.994 0.0301
Males
Gompertz s A = 0.0170; G = 0.1285 0.0417
Weibull s a = 0.00433; g = 2.243 0.0207
Logistic s v = 13.28; w = 3.533 0.0164
 26 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

MALE DROSOPHILA (CURTSINGER ET AL) MALE DROSOPHILA (CURTSINGER ET AL)

1O0 1

0 75 01

i 050
©Q ",,,
i O25
F

000
10 20 50 40 5,3,
Time (days} Time (days)

MALE DROSOPHILA (LINE 107; PEARL & PARKER) FEMALE DROSOPHILA (LINE 107; PEARL & PARKER
r ,, ~ j , r , 2 L _ 2 ~ ! , ,_ - . ~ _ _ ~ .... ,

o8 ~', o8

o,~ ~.,

: me (days) Time (doy~)

Fig. 4. Survival curves for Drosophila. In all cases data are shown as open circles. Survival for male Droso-
phila from Curtsinger et al. [8] is shown in linear and log form. The males significantly deviate from
Gompertz only very late in the life span. Survival for males and females from 'wild-type' (line 107) Droso-
phila from Pearl and Parker [11] are also shown. Curves from the three survival functions are shown as
in Fig. 3.

A logistic mortality function (see Methods), with three parameters, did give a bet-
ter fit to the 'tail' of the female medfly survival data, but that function also deviated
from the data at later times (see Fig. 1 and Table 1). The logistic mortality function
describes a mortality rate that increases to a plateau with time.
Of course, the presence of three parameters in the logistic mortality equation gives

FEMALE DROSOPHILA (VESTIGIAL; PEARL & PARKER MALE DROSOPHILA (VESTIGIAL; PEARl & pARKER
10

08 C5 I "k

~°~
i o¢

0?

9 c - b ,:
o If] 2(] ~L) 4(I ~) ,i
Time (doys} T me (days}

Fig. 5. Survival curves for vestigial Drosophila. Data (from Pearl and Parker, [I II) are shown as open
circles. Male and female vestigial Drosophila appear to follow different survival curves from the wild-type
Drosophila. Curves from the three survival functions are shown as in Fig. 3.
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 27

HUMANS (1988 U.S. CENSUS DATA

100
HUMANS I00+ (THATCHER)
~o . . . .

~" 0.75
.2
0
"~" 050 06
- - Logi~ X ~ O4
..... Gompertz ~,~.
m
09
025 .......... W e i b u l l ",~ x
............ L o g , s t i c m ',~,~ O2
0 000
20 30 40 ~o 60 7o 8o 90 ioo Ho
lOO 102 104 ~06 108
Time (years) T i ~ (yeats)

Fig. 6. Survival curves for humans. Data (from 1988 U.S. Census [13] and from Thatcher [14]) shown
as open circles. Four survival functions are shown for the U.S. Census data, as indicated. The Gompertz
curve and the logistic mortality curve are almost coincident. The best-fit curve from the Gompertz survi-
val function is shown with the Thatcher data.

that function enhanced flexibility, as can be seen in its ability to give a good fit to
the Gompertz-like, human survival (Fig. 6 and Table 5), which is quite a different
shape of curve from that of the female medflies.

3.3. Size of populations required to determine best-fit function

A determination was made of the number of organisms needed to distinguish

among survival functions. Three populations, female medflies [10], Drosophila wild-
type females [11], and Drosophila vestigial females [11], were tested for the number
of organisms needed to determine which survival function, Gompertz, Weibull, or
logistic survival, gave the best fit to the population.
For the medflies, a set of tests with a sample size of 25 clearly indicated that this
was too small a size to distinguish among the functions. Results were not consistent.
This confirmed experiments that the author has done on a number of samples of 25
nematodes (C elegans), which sometimes were fit best by one function and
sometimes by another.

Table 5
Humans (analysis of data from 1988 U.S. Census ages 10-85 [13]; and Thatcher [14], age 100+)

Function Parameter values Error

Males and Females (U.S. Cen-~s)
Gompertz s A = 0.0000843; G = 0.0831 0.00455
Weibull s a=4.1 x I0-1°;g=5.25 0.0211
Logistic s v = 77.99; w = 7.48 0.0238
Logistic m a = 0.000084; b = 39 600; k = 0.0832 0.00441
Males and Females (from Thatcher, [14])
Gompertz s A = 0.498; G = 0.0410 0.0031
Weibull s a = 0.5301; g = 1.051 0.0044
28 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

Samples of 100 then were tested. One hundred individuals were selected at random
from the populations, and the function giving the best fit to the subset of 100 in-
dividuals was determined. This was repeated seven to ten times for each of the three
populations. Table 6 shows that in all but two cases, the function giving the best fit
to the sample of 100 was the same as the function best fitting the whole population.

3.4. Importance of separating male~female data

In the appendix to a recent paper by Juckett and Rosenberg [19], an analysis of

the medfly data of Carey et al. [7] used a log-log graphing of age versus mortality
rate. They identified several slopes, with sharp breakpoints, and suggested that the
medfly population was composed of several sub-populations dying from different
processes. Fig. 7A reproduces their analysis, which was performed on the entire
medfly population, with males and females combined. The straight line indicates one
of the slopes covering roughly days 14-30 that Juckett and Rosenberg [19] had iden-
tified. As can be seen in Fig. 7B, which plots the male and female data separately,
that particular slope appears to be an artifact of combining the mortality rates of
the two sexes. The female curve 'breaks' at about 14 days and the male curve 'breaks'
at about 30 days. Neither curve taken alone shows a separate slope that exists only
for the period from day 14 to 30.

3.5. Some properties of the logistic survival function

Logistic functions have not been as widely used as Gompertz or Weibull func-
tions, but the introduction here of the logistic survival function has two justifica-
tions. First, it fit some of the survival curves better than the Gompertz or Weibull
functions. Second, the two parameters from this particular logistic function have
some useful characteristics. The parameter v is an estimate of the median life span
for the population, as can be seen in Fig. 8A, which shows a family of curves with
the value of v held constant, while w varies.
The parameter w is related to the rate of decline, or slope of decline, of a popula-
tion size with age. It has been popular recently to speak of the extent to which a
survival curve is rectangular [20], so a measure of such rectangularization can be
useful. Fig. 8A indicates that, with v held constant, as w increases, so does the rec-

Table 6
Number of times that a function is the best fit for random samples of 100 from various populations

Population Best fit for entire population Best fit for samples of I00
Gompertz Weibull Logistic S
Dros. w.t.f. Gompertz 8 2 0
Dros. vest. f. Weibull 0 7 0
Medfly f. Logistic S 0 0 7
Dros., Drosophila; w.t., wild-type; f., female; vest., vestigial; S = survival.
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 29

A 1

o Females plus Males

0
N 0.1
0
T

2 e
O
C~ o
e

o
o
~d 0.01 o
o
0 o
o
o

0.001 I
10 O0
B 1 I

0.1 oo°
2 Females 0 00
Males 0 00
OO o •

0.01
O

0.001 I
10 00

Age (days)

Fig. 7. Log-log plots of mortality rate as a function of age for medflies. Survival data was from the life
tables of Carey et al. [7] and Carey [10]. Mortality rate was calculated as Hazard, which is the deaths
per unit time divided by the average number of individuals alive during the time interval. Fig. 8A shows
the mortality rate for males and females together. The straight line in A shows one of three lines (that
exhibiting the middle slope) identified by Juckett and Rosenberg [19]. Fig. 8B shows the data for mort-
ality of males and females, separately. The slope of the data in A corresponding to the straight line can
be seen to be an artifact of merging the data from males and females.

t a n g u l a r i z a t i o n o f the survival curve. H o w e v e r , Fig. 8B s h o w s that, if v varies, keep-

ing w c o n s t a n t d o e s not give a c o n s t a n t rate o f decline. Instead, as s h o w n in Fig.
8C, one m u s t keep the ratio o f w over v a c o n s t a n t in order to gain an estimate o f
the rectangularization o f the survival curve.
30 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

Logistic S u r v i v a l v = 25 Logistic Survival v = 5--40

w : I 25-10 w= 50
1.00

• , • O w 1,25
=
- - 1 00;
(t,i I , 0 v = 5
O v = 10
• w=2.5
i .... 4 r
• '.', ~ ",','=5 "v v v : 20
d~
0.75 o • ~, • w = 10 (375 • v = 4u

o ;V,
-6
>
\oT',?,
%
"~ 0 5 0 ~:b..o 0 50

(',"o"'o~. 1
v

0.25 ' v, • 0"'°-'~0~! 0.25

i
, v, -,..., \\i
"'Lv .... . "v
_ i 9, o •v v " • i
0.000 101 210 v..,~. - ~ ~ 000
30 40 50 (')C 1C 2 (]' 3,, 40 5C ;) b
Time Time

Logistic Survival v = 5-40

w == v/4
1 O0

X~O '~ v,.,. " " 0V = 10

' • v 20 =

• v ,¢ v v = 30
075 • v = 40

\
"6
>
co
050

0.25
\
0 O0
~0 20 30 40 50 50
T;me

Fig. 8. Properties of the logistic survival function. Plots of the function s = 1/[1 + (t/v)Wl are shown. Fig.
8A shows a set of curves with v held constant, which gives identical median life spans = v. Fig. 8B shows
a set of curves with w held constant, which gives varying slopes to the decline in survival. Fig. 8C shows
a set of curves with w/v held constant, which exhibit a rather constant 'rectangularization.'

4. Discussion

Three f u n c t i o n s , each w i t h t w o parameters, were fit to survival data f r o m a variety

o f p o p u l a t i o n s . T h e f u n c t i o n s are logistic survival, and the survival forms o f the
G o m p e r t z and W e i b u l l functions. T h e first 95% or m o r e o f i n d i v i d u a l s d y i n g in each
o f the p o p u l a t i o n s were well fit by one o f the three f u n c t i o n s , and the 'tails' o f the
survival curves generally h a d v a l u e s that fell b e t w e e n t w o o f the fitted functions.
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 31

The two-parameter, logistic-survival function,

s = 1/[1 + (t/v) w]

is introduced here, and some survival data were better fit by it than by the more
traditional Gompertz or Weibull functions. The parameter v gives the median life
expectancy for the population (when t = v, s = 0.5), and the ratio of the two parame-
ters, w/v, gives a measure of the rectangularization of the survival curve.
The populations that were best fit by logistic survival had higher mortality than
predicted by the logistic survival function after first 95-99% of the population has
died. So, the 'tails' of such survival data tend to fall between the predictions of logis-
tic survival and Weibull. Mortality rates for such populations did remain lower than
predicted by either Gompertz or Weibull functions, and clearly exhibit a rise fol-
lowed by a decline with age.
One can get somewhat better fits to the 'tail' of the data curves by using functions
with more than two parameters, such as a 3-parameter, logistic-mortality function,
but it appears that no simple function will fit the complicated tail of a population
such as the medflies of Carey et al. [7]. The 3-parameter logistic-mortality function
exhibits a good ability to 'fit' a variety of survival curves. However, functions with
larger numbers of parameters should be used with caution, because of the flexibility
that additional parameters alone can provide for adjusting to the possible shapes of
a curve. By the adjustment of parameter values, the three-parameter, logistic-
mortality function can be made to fit curves that range from Gompertz-like (human)
to those close to logistic survival (female medflies). That function reflects a mortality
rate that increases with age to a plateau value.
Males and females sometimes are best fit by different functions. This was the case
for the flour beetles, house flies, and medflies. Thus, mixing data from males and
females can mask differences in survival characteristics. Also shown was the danger
of using mixed-sex data for analyzing mortality rates and causes. The fact that males
and females die by different curves suggests that different kinetics underlie the deaths
of the two groups, and hints of the possibility of different underlying causes of death
as well. Thus, this analysis points to some interesting species meriting further study
on the physiological bases for mortality increases with age.
When comparing the effects of environmental or genetic manipulations on survi-
val of a population, investigators frequently will compare the best-fit parameters for
a single function, such as Gompertz. The change in the values of the parameters then
is used to give insight into mechanisms of aging. The results of this paper suggest
that it also can be important to test the fit of survival data to different functions.
As indicated by a comparison of wild-type and vestigial Drosophila, the shapes of
survival curves can change more dramatically than can be accounted for by modifi-
cation of parameter values for a single function. Thus, genetic or environmental
manipulations may change the function that best fits the data as well as changing
the best-fit parameter values for a given function. The analysis of more than one
function would appear to be necessary if the manipulations produce more dramatic
changes in the shape of the survival curve. It is quite possible that populations can
32 D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33

be altered, by environment or genetics, to produce senescence or deaths in a popula-

tion by a different mechanism, perhaps by eliminating or adding a major cause of
mortality.
For the populations studied, random sampling experiments indicate that 100
organisms often are enough to determine which function (Gompertz, Weibull, or
logistic survival) fits the survival data best, although larger numbers will enhance the
accuracy of the determination. These random sampling tests also suggest that the
differences in 'fit' of the three functions are significant, at least for the populations
studied, since the fit is reproducibly best with one function.
Of course, there is no known reason why a given population should be fit by any
one curve, or why a curve fitting the survival of a population for one part of its life
span should necessarily fit another part of the life span. Earlier research has used
mortality curves to explore the nature of the underlying causes of senescence, or has
attempted to use mortality data to propose multiple causes of senescence [19,21,22].
If there are a variety of causes for the increase in the rate of mortality with age, which
would seem to be the case in many organisms, then what is being seen in the survival
data is the sum of deaths from a number of causes. More complex survival curves
are therefore expected. Indeed, some of the survival data appear to fall between the
fitted functions, rather than following a particular function (see female house flies
and male vestigial Drosophila). Furthermore, one use of the analysis proposed here
is to identify deviations in mortality rates at particular ages (such as late ages in the
medflies) that stand in need of explanation.
Lestienne [22] has suggested that organisms exhibiting Gompertz mortality could
not have such mortality attributed to a slowing of the recovery machinery that
restores an organism's homeostasis after a challenge or disease. He indicates that
such a cause of death would exhibit a constant rate of mortality. It is interesting to
note that the logistic mortality function (three parameter) does show an increase in
mortality to a constant rate of mortality at later ages, and does model the survival
of some populations.
It is interesting to note that the data analyzed here confirm the work of Carey et
al. [7] and Curtsinger et al. [8] that some populations show a slowing of mortality
rates at advanced ages. Male houseflies certainly exhibit not only a slowing, but
apparently a decline in mortality rates at advanced ages. Analysis of data from a
population of grain beetles [23] showed that survival of males at 29.1 °C was best fit
by a logistic survival function (data not shown), so they too show a slowing, or
decline, in mortality rates at advanced ages.
Data from populations of 100 or more individuals appear to give reproducible
curves that can be fitted by the Gompertz, Weibull, and logistic survival functions.
For some populations, even using the best-fit (least-squares error) values for
Gompertz and Weibull, the functions do not fit the survival data well, while logistic
survival or logistic mortality functions do fit well. This can be seen, for female
medflies, in Fig. 1. Notice that the failure to fit by the Gompertz or Weibull func-
tions is not just at the tails of the curves. The top of the survival curve is fit no better
(Fig. 1A), and it is clear that the data follow a differently shaped curve. More
generally, the fits of survival data by different functions can be used as a tool for
D.L. Wilson~Mech. Ageing Dev. 74 (1994) 15-33 33

the analysis of populations, and of changes in populations due to environmental or

genetic modifications.

5. Acknowledgements

This research relied upon the data gathering efforts of others, which the author
gratefully acknowledges. I thank James Curtsinger for providing his Drosophila
survival data.

6. References

! A. Comfort, The Biology of Senescence, Elsevier, New York, 1979.

2 C.E. Finch, Longevity, Senescence, and the Genome, University of Chicago Press, Chicago, 1990.
3 B. Gompertz, On the nature of the function expressive of the law of human mortality and on a new
mode of determining the value of life contingencies. Phil. Transact. R. Soc. (Lond), 115 (1825)
513-585.
4 W. Weibull, A statistical distribution function of wide applicability. J. Appl. Mech., 18 (1951)
293-297.
5 A. Economos, A non-Gompertzian paradigm for mortality kinetics of metazoan animals and failure
kinetics of manufactured products. Age, 2 (1979) 74-76.
6 M. Witten, A return to time, cells, systems, and aging V. Further thoughts on Gompertzian survival
dynamics - - the Geriatric years. Mech. Ageing Dev., 46 (1988) 175-200.
7 J.R. Carey, P. Liedo, D. Orozco and J.W. Vaupel, Slowing of mortality rates at older ages in large
medfly cohorts. Science, 258 (1992) 457-461.
8 J.W. Curtsinger, H.H. Fukui, D.R. Townsend and J.W. Vaupel, Demography of genotypes: failure
of the limited life-span paradigm in Drosophila melanogaster. Science, 258 (1992) 461-463.
9 M. Rockstein and H.M. Lieberman, A life table for the common house fly, Musca domestica. Geron-
tologia, 3 (1959) 23-36.
10 J.R. Carey, Applied Demography for Biologists, Oxford University Press, New York, 1993.
11 R. Pearl and S.L. Parker, Experimental studies on the duration of life. IX. New life tables for Droso-
phila. Am. Nat., 58 (1924) 71-82.
12 R. Pearl, T. Park and J.R. Miner, Experimental studies on the duration of life. XVI. Life tables for
the flour beetle Tribolium confusum Duval. Am. Nat., 75 (1941) 5-19.
13 U.S. Census, National Center for Health Statistics, Department of Health and Human Services,
Vital Statistics of the United States, 1988, 1Iol. H--Mortality Part A, Hyattsville, 1991.
14 A.R. Thatcher, The span of life. J. Inst. Actuaries, 109 (1982) 346-349.
15 D.L. Wilson, A comparison of methods for estimating mortality parameters from survival data.
Mech. Ageing Dev., 66 (1993) 269-281.
16 J.A. Nedler and R. Mead, A simplex method for function minimization. Comp. J., 7 (1965) 308-313.
17 M.S. Caceci and W.P. Cacheris, Fitting curves to data: the simplex algorithm is the answer. Byte,
9 (1984) 340-361.
18 D.W. Marquardt, An algorithm for least-squares estimation of nonlinear parameters. J. Soc. lnd
Appl. Math., 11 (1963) 431-441.
19 D.A. Juckett and B. Rosenberg, Comparison of the Gompertz and Weibull functions as descriptors
for human mortality distributions and their intersections. Mech. Ageing Dev., 69 (1993) 1-31.
20 J.F. Fries and L.M. Crapo, Vitality and Aging, W.H. Freeman & Co., New York, 1981.
21 B.L. Strehler and A.S. Mildvan, General theory of mortality and aging. Science, 132 (1960) 14-21.
22 R. Lestienne, On the thermodynamical and biological interpretation of the Gompertzian mortality
rate distribution. Mech. Ageing Dev., 42 (1988) 197-214.
23 L.C. Birch, Experimental background to the study of the distribution and abundance of insects. Eco-
logy, 34 (1953) 698-711.