American Journal of Epidemiology Vol. 162, No.

8
Copyright 2005 by the Johns Hopkins Bloomberg School of Public Health DOI: 10.1093/aje/kwi272
All rights reserved; printed in U.S.A. Advance Access publication August 31, 2005

Original Contribution

Variation in the Seasonal Diagnosis of Acute Lymphoblastic Leukemia: Evidence

from Singapore, the United States, and Sweden

Fei Gao1, Per Nordin2, Ingela Krantz2, Kee-Seng Chia3, and David Machin1,4
1
Division of Clinical Trials and Epidemiological Sciences, National Cancer Centre, Singapore.
2
Skaraborgsinstitutet, Skovde, Sweden.
3
Centre for Molecular Epidemiology, National University of Singapore, Singapore.
4
Medical Statistics Unit, School of Health and Related Research, University of Shefeld, Shefeld, United Kingdom.

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

Received for publication December 29, 2004; accepted for publication May 13, 2005.

This study investigated, by summing data over successive years, the evidence for the seasonal diagnosis of
acute lymphoblastic leukemia. To do so, the authors estimated the dates of peak diagnosis over a range of
geographic locations including Singapore (19681999), Hawaii and mainland United States (19731999), and
western Sweden (19771994) at latitudes of 1.16
N to 58.24
N. In contrast to other studies, the authors used case-
by-case information on dates, gender, and age rather than grouped data for analysis. The seasonal pattern was
estimated by tting a von Mises distribution to the data from each location. No seasonal pattern was found in
Singapore, which is close to the equator and does not have marked climatic changes. Likewise, seasonality was
not demonstrated in Hawaii or mainland United States despite a 26.18
range of latitudes. In contrast, a signicant
peak (early January) was observed for western Sweden that appeared strongest for males (December 22, 95%
condence interval: November 16, January 16) and those less than age 20 years (January 14, 95% condence
interval: December 8, March 27). Thus, despite a wide geographic range of localities, there is little evidence of any
seasonality in the diagnosis of acute lymphoblastic leukemia in most populations studied and no strong evidence of
any inuence of climate (as expressed by latitude).

leukemia, lymphoblastic, acute; seasons

Abbreviations: ALL, acute lymphoblastic leukemia; SEER, Surveillance, Epidemiology, and End Results.

Many infectious and some chronic diseases have charac-
teristic seasonal onsets (1), and, in certain instances, these
seasonal rhythms can be related to etiologic precipitating
factors (2). Thus, it has been speculated that acute lympho-
blastic leukemia (ALL) in children is virus related (3) and
may be the consequence of an abnormal response to com-
mon infections occurring in early life (4). If infections are
involved, then some seasonal pattern of onset might be ex-
pected to be associated with them (5). In general, demon-
stration of seasonal variation in the onset of leukemia may
provide insight into potential risk factors (6).
The first published studies on the seasonality of leukemia
appear to be from Belgium (7, 8), which reported a November
February peak in acute leukemia. Since then, numerous re-
ports (we identified more than 30 in a systematic literature
search) from 15 countries have specifically investigated the
seasonality of ALL.
While many of these studies have identified an obvi-
ous seasonal pattern, others have not found a significant
seasonal variation. For example, for England and Wales
combined, a summer peak was noted in two age groups
(019 and 2044 years) in studies from the early 1960s

Correspondence to Fei Gao, Division of Clinical Trials and Epidemiological Sciences, National Cancer Centre, 11 Hospital Drive, Singapore
169610 (e-mail: ctegfe@nccs.com.sg).

753 Am J Epidemiol 2005;162:753763

754 Gao et al.
(9, 10), while one study some 30 years later showed no such
pattern (11).
Inconsistency between results may itself be informative,
perhaps reflecting various levels of between-population het-
erogeneity and different patterns of seasonality induced by
possible causative agents. It is also possible that seasonality
may be more pronounced within subtypes of leukemia or,
for example, a particular locality. There was little evidence
of seasonality in a national data set from Great Britain
(England, Scotland, and Wales), but seasonality (summer:
MayOctober) was evident in one regional data set from the
West Midlands, England (12). The authors concluded that
further work on seasonality needs more sophisticated anal-
ysis, controlling for broad geographical heterogeneity (12,
p. 678). However, to our knowledge, no formal synthesis of
published reports has been attempted to date. As a conse-
quence, the etiology of leukemia in this respect, in both
children and adults, still remains essentially undetermined.
One study, reporting some 20 years ago, examined in
particular the influence of latitude on presentation of ALL
for US subjects aged 019 years by using monthly data from
the Surveillance, Epidemiology, and End Results (SEER)
Program and from an independent survey of a 57-county
study area in the eastern half of Nebraska (13). The authors
examined the seasonal variation in ALL by using periodic
regression of monthly rates; they noted up to three peaks
(April, August, and December) in registries north of 40
N
and the same, but in different months (February, July, and
October), for the southern counterparts. They commented
that the observed peaks in monthly ALL risk coincide with
seasonal elevations in the rates of allergenic and infectious
diseases, elements of which are capable of promoting lym-
phocytic proliferation and transformation (13, p. 915).
The SEER Program records, and makes available to re-
searchers, individualized data on each case from 11 registries
in the United States over a range of latitudes from 21.18
N
to 41.36
N. We have the same detail for cases of ALL from
part of western Sweden (latitude 58.24
N) and the whole of
Singapore (latitude 1.16
N). In these two registries, the ac-
tual day is recorded and not just the month, as with SEER.
The purposes of our study were to search for a seasonal
component in the presentation of ALL in the registry data
from these three countries, to determine the influence of
latitude (if any), and to investigate the role of subject age
and gender.
MATERIALS AND METHODS
For all registries, we restricted our analysis to cases of
ALL (International Classification of Diseases, Eighth Re-
vision, code 204.00 or International Classification of Dis-
eases, Ninth Revision, code 204.0) (14, 15). For the US
locations, the state capitals were used to define latitude
and longitude.
Data
Singapore (latitude 1.16
N, longitude 103.51
E). We used
Singapore Cancer Registry data that included all cases of
cancer occurring in citizens and permanent residents of the
island (population 3.5 million) over the period 19681999.
Disease classification follows the International Classifica-
tion of Diseases, Ninth Revision. In addition, age, gender,
ethnic group, date of birth, place of birth, and date of di-
agnosis are available on an individual case basis. A total of
941 ALL cases were registered for that time period. Two
cases were excluded from our analysis of age (for one, only
the year was recorded; for the other, the diagnosis was re-
corded a few days earlier than the birth).
United States (latitude 21.18
N47.36
N, longitude 72.41
W
157.52
W). The SEER Program of the National Cancer
Institute records data from 11 cancer registries over a wide
geographic area in the United States, and we used the data
for the time period 19731999. The SEER locations in-
cluded the island of Hawaii and the following mainland
areas: metropolitan Atlanta, Georgia; Connecticut; metro-
politan Detroit, Michigan; Iowa; New Mexico; the San
Francisco-Oakland Standard Metropolitan Statistical Area,
Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012
California; Seattle, Washington; Utah; and Los Angeles and
San Jose-Monterey, California. Seattle contributed data for
the years 19741999, metropolitan Atlanta for 19751999,
and Los Angeles and San Jose-Monterey for 19921999
only.
SEER collects patient-specific information on tumor site,
histology, gender, age, ethnicity, and date (month and year
only) of diagnosis from all residents diagnosed with cancer
in collaborating states or localities. Data for 9,158 ALL
patients were collected, but 38 cases were excluded from
our seasonality analysis because their month of diagnosis
was not available.
Western Sweden (latitude 58.24
N, longitude 13.50
E).
Individual dates of each arrival at the hospital for any dis-
ease were available for part of the West Gotaland Region in
western Sweden from 1977 to 1994 (the population of the
catchment area was 270,000 in 1995), and we used these
data. Disease classification follows a Swedish version of the
International Classification of Diseases, Eighth Revision
before 1987 and a Swedish version of the International
Classification of Diseases, Ninth Revision thereafter. For
our purposes, the date of admission to the hospital was
considered the date of diagnosis. In addition, age and gender
are available on an individual-case basis.
A total of 81 cases were diagnosed with ALL, and we re-
viewed their diagnoses noted at each visit. As a consequence,
two cases were excluded because, on most occasions, their
diagnosis was recorded as lymphosarcoma or reticulosar-
coma but only once as ALL. Corrections to the date of pre-
sentation regarding ALL were made for seven cases.
Statistical analysis
Data from each calendar year were first standardized to
365 days and were then converted to an angle between 0

and 360
. We illustrate these data graphically in a rose di-
agram format (figure 1), where each petal is of a standard
month or 360/12 30
. The segments are ordered from
January to December (clockwise) starting from due north.
In these diagrams, the square root of monthly totals is used
to preserve equal areas for each unit of frequency, as in
a conventional histogram.
Am J Epidemiol 2005;162:753763
 Seasonal Variation in ALL 755

TABLE 1. Latitude and longitude of the four locations and information on number, age, and gender of
acute lymphoblastic leukemia cases obtained from cancer registries in Singapore (19681999), Hawaii
(19731999), mainland United States (19731999), and western Sweden (19771994)

United States
Characteristic Singapore Sweden
Hawaii Mainland

Latitude 1.16
N 21.18
N 33.45
N47.36
N 58.24
N

Longitude 103.51
E 157.52
W 72.41
W122.21
W 13.50
E
No. of cases 939 396 9,185 79
Age in years
Mean (range) 16.99 (0.0988.32) 21.50 (0102) 22.50 (0102) 18.62 (190)
Gender
No. of males (%) 548 (58.4) 221 (55.8) 5,296 (57.7) 49 (62.0)

To examine the data for the presence of distinct annual Sweden). The latitudes range from 1.16
N (Singapore) to

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

peaks, histograms were also constructed (figure 2) in which
the summated year representing the 12 months was repeated
(once) to emphasize the circular nature of the calendar. The
associated von Mises distribution was added. This distribu-
tion assumes a single peak for seasonality within a year and
is described by the parameters l and j, from which the peak
magnitude, R, can be derived. All peak dates, l, are pre-
sented in the format (month, day) of a 365-day year. Al-
though this format may suggest spurious precision, it is used
for clarity and to facilitate comparisons. The inverse of j,
which indicates the amount of variation about the peak date,
is also displayed in all tables in this paper. Small values of R
(small j) would usually indicate the absence of seasonality,
whereas those close to the maximum of 1 (large j) would
indicate a very sharp peak in (here) the number of cases. In
this study, all j were small, so that R
j/2; therefore, only j
is shown in the tables. Departures from a uniform distribu-
tion of cases over the year were tested by the Mardia v2
statistic with 2 degrees of freedom (16). Details of the von
Mises distribution and the Mardia test statistic are given in
the Appendix.
The bootstrap technique (17) was used to obtain a 95
percent confidence interval for the peak date by using
2,000 bootstrap samples of the same size as the number of
patients under consideration. Because the bootstrap method
of calculation for confidence intervals requires individual
dates, and SEER does not provide precise dates of diagnosis,
for calculation purposes (95 percent confidence interval
only) we randomly assigned to each case a day from their
corresponding month of diagnosis. In this simulation, the
monthly total, fg, for each year was assumed to be distrib-
uted uniformly over the bin size of 30
, and a random sam-
ple of size fg was drawn from that bin. All of the peaks and
the corresponding confidence intervals are displayed in
a forest plot. We comment on the format of a confidence
interval in this context when describing figures 35.
RESULTS
Table 1 summarizes the geographic details of the four
locations (Singapore, Hawaii, mainland United States, and
58.24
N (Sweden) and the longitudes from 157.52
W
(Hawaii) to 103.51
E (Singapore). On a per-population ba-
sis, the numbers of cases of ALL were very similar across
regions, although the variation in the annual number was
considerable, and there was a suggestion of rising numbers
in Singapore (data not shown). The average age at presen-
tation was similar among the four location groups, and more
males than females (approximately 1.5:1) were diagnosed.
The numbers of cases presenting at each location on a cu-
mulated monthly basis over the years for which data were
used are given in figure 1. Although no single peak seemed
to occur in Singapore or in either Hawaii or mainland United
States, for western Sweden a peak was suggested over the
interval between 300
and 0
45
or, equivalently, mid-
November to the end of February.
To some extent, the rose diagram format obscures the
shape of the distribution over the year; therefore, we plotted
data for Singapore, Hawaii, and Sweden again in a (double-
year) histogram format (figure 2). Superimposed on these
plots are the corresponding von Mises distributions. It is
clear from this figure that the von Mises distribution reason-
ably describes the seasonal pattern for these cases and sug-
gests that the strength of the peak is increasing from
Singapore to Sweden. However, a more detailed examina-
tion of the values of j in all of the intermediate-latitude
registries (table 2) was far from suggestive of a general
pattern. Furthermore, there was no clear evidence of any
systematic change in the estimated date of peak diagnosis
over the changing latitudes.
The estimated peak diagnosis is given for each location in
table 2. For the total of ALL cases, the peak at August 14 in
Singapore was not strong (R 0.040, 95 percent confidence
interval: May 1, October 29). In addition, the peak at May
15 in the United States was not strong (R 0.031, 95 per-
cent confidence interval: April 19, June 8). In contrast, a sta-
tistically significant peak was apparent in early January (p
0.02), with a moderate amplitude of R 0.223 in western
Sweden. However, the sample (N 79) was small and the
95 percent confidence interval (November 17, February 10)
rather wide.
The 13 estimated peak dates ranged over 10 months of the
year from January 3 in western Sweden to October 23 in San

Am J Epidemiol 2005;162:753763
756 Gao et al.

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

FIGURE 1. Rose diagram of the seasonal distribution of acute lymphoblastic leukemia cases in Singapore (19681999), Hawaii (19731999),
mainland United States (19731999) (metropolitan Atlanta, Georgia; Los Angeles, California; New Mexico; San Jose-Monterey, California; the San
Francisco-Oakland Standard Metropolitan Statistical Area, California; Utah; Iowa; Connecticut; metropolitan Detroit, Michigan; and Seattle,
Washington), and western Sweden (19771994). Nov, November; Feb, February; Aug, August.

Jose-Monterey, southern California (figure 3). Ten of the 11
peak dates in the United States were located on the arc
between Singapore and Sweden, suggesting a weak latitude
effect. Only the date for San Jose-Monterey, California, was
found within the opposite arc. However, a plot of the esti-
mated peaks, with corresponding 95 percent confidence in-
tervals, by increasing latitude showed no clear, systematic
pattern. Furthermore, all but one of the corresponding esti-
mates of j were less than 0.12 (small), implying little evi-
dence of seasonality in most registries except for western
Sweden (j 0.458). The size of the symbol d represent-
ing peak date of diagnosis is proportional to j, except that

Am J Epidemiol 2005;162:753763
 Seasonal Variation in ALL 757

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

FIGURE 2. Repeated histogram of corrected monthly numbers of acute lymphoblastic leukemia cases, with the corresponding tted von Mises
distributions, in Singapore (19681999), Hawaii (19731999), and western Sweden (19771994). l, peak date; j, estimated concentration
parameter of the von Mises distribution; J, January; F, February; M, March; A, April; M, May; J, June; J, July; A, August; S, September; O, October;
N, November; D, December.

those peak dates with j < 0.1 are labeled , and the
length of the vertical lines, one for each location, represents
the 95 percent confidence interval for the peak date. Because
the confidence intervals may wrap around December 31
and January 1, different symbols are used to indicate the
lower (D) and upper (=) confidence limits, which in-
dicate the direction to go along the arc from that point to find
the peak.
For the places from Singapore to San Francisco, at a lat-
itude of <40
Nthe cutpoint used previously for compar-
ison purposes (13)the estimated peaks ranged from April
to October (table 2, figure 3). For the locations at a latitude
of 40
N, from Utah to western Sweden, the estimated
peaks ranged from January to February (winter) except in
Seattle, where the latitude is 47.36
N and the estimated peak
was July 19 (summer).

Am J Epidemiol 2005;162:753763
758 Gao et al.

TABLE 2. Circular analysis of acute lymphoblastic leukemia cases from cancer registries in Singapore (19681999), Hawaii (1973
1999), mainland United States (19731999), and western Sweden (19771994), ordered by latitude

95% condence interval Mardia v2

Location Latitude Longitude No. j* Peak l
* Peak date p value
(bootstrap) statistic

Singapore 1.16
N 103.51
E 939 0.080 222.61 August 14 May 1, October 29 2.993 0.224
Hawaii 21.18
N 157.52
W 393 0.112 92.98 April 5 December 21, July 20 2.457 0.293
Mainland United States
Metropolitan Atlanta,
Georgia 33.45
N 84.23
W 570 0.024 168.75 June 21 November 7, October 3 0.158 0.924
Los Angeles, California 34.03
N 118.15
W 1,363 0.070 118.34 April 30 January 21, July 17 3.298 0.192
New Mexico 35.05
N 106.39
W 587 0.093 172.48 June 24 January 30, August 30 2.546 0.280
San Jose-Monterey,
California 36.78
N 121.54
W 320 0.004 291.33 October 23 July 12, June 10 0.003 0.999
San Francisco-Oakland,
California 37.48
N 122.21
W 1,236 0.064 218.75 August 10 May 29, November 16 2.493 0.288
Utah 40.46
N 111.53
W 693 0.039 45.22 February 15 September 3, July 13 0.517 0.772
Iowa 41.40
N 91.32
W 1,082 0.038 44.33 February 14 August 23, July 7 0.788 0.674

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

Connecticut 41.46
N 72.41
W 1,063 0.028 49.70 February 20 September 27, July 31 0.429 0.807
Metropolitan Detroit,
Michigan 42.20
N 83.03
W 1,173 0.054 4.04 January 5 August 16, May 3 1.719 0.423
Seattle, Washington 47.36
N 122.20
W 1,063 0.047 196.43 July 19 March 8, December 7 1.164 0.559
Western Sweden 58.24
N 13.50
E 79 0.458 2.78 January 3 November 17, February 10 7.875 0.020

* j, estimated concentration parameter of the von Mises distribution, which indicates the amount of variation; l
, estimated season peak or
mean direction.

In the broad age categories of 019 and 20 years, there

FIGURE 3. Peak dates (95% condence intervals) of acute
lymphoblastic leukemia in Singapore (19681999), Hawaii (1973
1999), mainland United States (19731999), and western Sweden
(19771994), ordered by latitude (
N). Jan, January; Mar, March; Jul,
July; Sep, September; Nov, November. d, peak date of diagnosis; ,
peak date with j < 0.1; D, lower limit of the 95% condence interval;
=, upper limit of the 95% condence interval. The size of d
is proportional to j, except that those peak dates with j < 0.1
are labeled , and the length of the vertical lines, one for
each location, represents the 95% condence interval for the peak
date.
appeared to be no trend across the latitudes for children or
adults (table 3, figure 4). Although the values of j tended to
increase once the subjects, within each registry, were di-
vided into these two age groups, suggesting a difference
between the younger and older cases in the corresponding
peak dates of diagnosis, no systematic pattern emerged of
one age group experiencing a peak earlier in the year than
another.
A more detailed examination using six age groups (02,
39, 1019, 2049, 5069, and 70 years; details not
shown) failed to identify any subgroup for which there
was a clear seasonal peak. In addition, we found no evidence
of any systematic trend across the ages in date of peak
diagnosis.
There were no systematic patterns within or between gen-
ders across the registries (table 4, figure 5). Furthermore, no
association between peak date of diagnosis and longitude
was observed.
DISCUSSION
Many studies have investigated the seasonal patterns in
the presentation of ALL, but (to our knowledge) none has
used individual case-by-case methods in their approach to
analysis, instead grouping the data into monthly counts.
Such an approach will be less statistically sensitive and
may therefore fail to identify patterns that may be present.

Am J Epidemiol 2005;162:753763
 Seasonal Variation in ALL 759

TABLE 3. Circular analysis of acute lymphoblastic leukemia cases by age in Singapore (19681999), Hawaii (19731999), mainland
United States (19731999), and western Sweden (19771994), ordered by latitude

95% condence interval Mardia v2

Location Age (years) No. j* Peak l
* Peak date p value
(bootstrap) statistic

Singapore 019 684 0.111 201.56 July 24 May 30, October 20 4.188 0.123
20 253 0.109 303.45 November 4 July 15, March 17 1.506 0.471
Hawaii 019 263 0.147 117.36 April 29 February 8, July 28 2.836 0.242
20 130 0.140 31.88 February 2 September 17, June 16 1.270 0.530
Mainland United States
Metropolitan Atlanta, Georgia 019 358 0.091 9.25 January 10 October 12, May 24 1.492 0.474
20 212 0.216 183.32 July 5 April 23, September 4 4.898 0.086
Los Angeles, California 019 907 0.100 131.92 May 14 March 9, July 24 4.494 0.106
20 456 0.049 46.72 February 17 September 11, July 13 0.548 0.760
New Mexico 019 376 0.132 162.34 June 14 March 25, August 22 3.260 0.196
20 211 0.050 228.45 August 20 March 5, February 4 0.263 0.877
San Jose-Monterey, California 019 226 0.079 22.67 January 23 September 8, May 15 0.697 0.706

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

20 94 0.192 207.19 July 30 April 13, December 18 1.715 0.424
San Francisco-Oakland,
California 019 761 0.093 211.14 August 3 May 18, October 28 3.269 0.195
20 475 0.027 266.22 September 27 April 20, February 27 0.169 0.919
Utah 019 499 0.070 85.94 March 29 November 21, August 8 1.212 0.546
20 194 0.117 316.32 November 17 June 11, February 13 1.319 0.517
Iowa 019 674 0.074 42.33 February 12 October 1, May 28 1.861 0.394
20 408 0.022 212.97 August 4 March 19, February 7 0.096 0.953
Connecticut 019 653 0.048 42.37 February 12 September 14, July 20 0.766 0.682
20 410 0.010 155.79 June 7 November 4, October 5 0.021 0.989
Metropolitan Detroit, Michigan 019 768 0.089 25.51 January 26 October 28, April 20 3.045 0.218
20 405 0.062 273.77 October 5 May 20, February 21 0.772 0.680
Seattle, Washington 019 696 0.066 256.61 September 18 June 1, January 22 1.523 0.467
20 367 0.131 140.37 May 23 March 16, October 9 3.146 0.208
Western Sweden 019 63 0.458 13.56 January 14 December 8, March 27 6.278 0.043
20 16 0.606 328.23 November 29 September 7, February 12 2.692 0.260

* j, estimated concentration parameter of the von Mises distribution, which indicates the amount of variation; l
, estimated season peak or
mean direction.

Our approach with individual data enables the von Mises
distribution to be used to describe the variation over the
calendar year, and estimates of the corresponding parame-
ters allow the peak date to be identified, the strength of the
peak to be determined (over a range from 0 to 1), and con-
fidence intervals to be calculated. If there is little or no
evidence of a peak, one is nevertheless estimated; however,
the corresponding confidence interval will cover the major-
ity of the year, reflecting the uncertainty associated with
such an estimate.
For this study, we were able to obtain individualized data
from western Sweden, the Singapore Cancer Registry, and
the SEER database of 11 registries in the United States for
a total of 10,599 cases over the years 19681999.
Singapore (latitude 1.16
N) is very close to the equator;
therefore, if there was indeed an influence of season on the
date of peak diagnosis of ALL, it should be weak in this
country because there are no marked seasonal weather pat-
terns. In contrast, western Sweden (latitude 58.24
N) has
seasons that differ markedly from winter, with few daylight
hours and temperatures below 0
C, to summer, with almost
continual daylight and temperatures ranging from 25
C to
30
C. Thus, one might expect to identify a seasonal pattern
for ALL if one exists, and western Sweden was indeed
the only location of the 13 studied in which a statistically
significant peak (January 3) was identified. However, the
sample size was small (79 cases in total), so there is a danger
of a spurious finding especially because there are regions
covered by the US registries where the climate (but not day
length) is similar to that of western Sweden, but no peaks
were established there. For the US registries as a whole, and
despite a wide range of latitudes extending from Hawaii to
Seattle (a difference of 26.18
), there is little evidence to
suggest a climatic effect.
Apart from the Swedish data just noted, in which the peak
was confined to young males, there was little other evidence

Am J Epidemiol 2005;162:753763
760 Gao et al.
FIGURE 4. Estimated peak dates (95% condence intervals) of
presentation of acute lymphoblastic leukemia in children and adults in
Singapore (19681999), Hawaii (19731999), mainland United
States (19731999), and western Sweden (19771994), ordered by
latitude (
N). Jan, January; Mar, March; Jul, July; Sep, September;
Nov, November. d, peak date of diagnosis; , peak date with j < 0.1;
D, lower limit of the 95% condence interval; =, upper limit of the 95%
condence interval. The size of d is proportional to j, except that
those peak dates with j < 0.1 are labeled , and the length of the
vertical lines, one for each location, represents the 95% condence
interval for the peak date.
to suggest that, within specific age groups, strong peaks are
present or that there is any systematic trend across ages
regarding diagnosis, even when age was categorized into
groups ranging from 02 years to the elderly. There was
also little evidence of any gender effect in any location or
any suggestion of a consistent effect across locations.
Our failure to find any such influences may have resulted
from not studying registries in more regions of the world,
weaknesses in the analytical approach, and (nonrecorded)
etiologic factors playing a major role and obscuring any
latitude, longitude, gender, and age effects. Thus, we are
aware that we were not able to perform an individual-case,
systematic overview of the ALL studies (we identified more
than 30) that have been conducted. Also of concern is the
loss of statistical sensitivity because the SEER data provide
only the month and not the day of diagnosis. Were the
precise days available, then angular regression models
(18) could be used to investigate the associations rather than
the somewhat descriptive approach that we had to use. Fur-
ther work is also required to confirm (or otherwise) the
utility of the von Mises distribution as an adequate descrip-
tion of seasonality; once more, only precise dates would
allow a thorough investigation.
For some diseases, seasonal variation in known or un-
known precipitating factors will depend on climate and
a range of population characteristics (19), and these in turn
will induce seasonal patterns in the disease itself. However,
apart from gender and age, unavoidably all of the possible
Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012
precipitating factors are expressed merely through the lati-
tude (and longitudes) of the individual registry locations.
These clusters will clearly obscure the influence of, for ex-
ample, case-specific socioeconomic status and some envi-
ronmental factors (20, 21) that have been linked to the risk
of developing ALL.
Our findings are consistent with a report concluding that
there was no seasonality in the United States across all ages
(22) but not with a later reanalysis of these same data (13).
This reanalysis identified up to three peaks in each of the
nine locations studied for subjects aged 019 years. Our
view is that this result more likely reflects the statistical
methodology applied to data that are essentially random
(but grouped into 12 bins), so that a trigonometric model
with many parameters will mirror the vagaries of such
a toothy distribution and not reflect the smoothed (and more
likely) pattern of a uniform distribution over the year. We
could not substantiate the claimed summer peak in the di-
agnosis of ALL in children (23).
In contrast, for western Sweden, the significant peak in
winter (early January) is close to the peak of onset reported
for Capetown, South Africa (winter: JuneAugust) (24) and
the peak of symptoms occurring in Shiraz, Iran (winter:
October) (5). However, this information differs from the
summer peaks reported for onset in England and Wales (9,
10) and diagnosis in England (6). These contrasting results
for ALL may be due to the different statistical approaches
that have been used, the different age groups chosen for
analysis and reporting, or the different date of onset of
ALL considered, although the delay between clinical symp-
toms in children and diagnosis is not likely to be great (4).
The seasonality observed in western Sweden (if it could
be firmly established) may, in any event, be related more to
precipitating than to etiologic factors. For example, eventual
cases might first present as a result of lowered immunity to
infections; these may be common in winter or summer, de-
pending on their type and their ALL being detected as a re-
sult. The overall health care system in Sweden is ranked
highly (25) and provides relatively open access to care. Con-
sequently, the peak of moderate magnitude in January may
reflect post-Christmas and New Year festivities delaying
self-referral and not the presence of an etiologic determinant.
Am J Epidemiol 2005;162:753763
 Seasonal Variation in ALL 761

TABLE 4. Circular analysis of acute lymphoblastic leukemia cases by gender from Singapore (19681999), Hawaii (19731999),
mainland United States (19731999), and western Sweden (19771994), ordered by latitude

95% condence interval Mardia v2

Location Gender No. j* Peak l
* Peak date p value
(bootstrap) statistic

Singapore Male 548 0.092 237.08 August 29 May 25, November 30 2.328 0.312
Female 391 0.074 196.79 July 19 February 20, December 4 1.073 0.585
Hawaii Male 220 0.108 85.40 March 28 November 14, August 31 1.284 0.526
Female 173 0.120 101.73 April 14 November 26, August 24 1.229 0.541
Mainland United States
Metropolitan Atlanta, Georgia Male 339 0.007 277.84 October 9 July 30, July 10 0.008 0.996
Female 231 0.062 160.14 June 12 November 26, October 18 0.445 0.801
Los Angeles, California Male 767 0.052 128.23 May 11 December 13, September 29 1.051 0.591
Female 596 0.094 111.21 April 23 January 13, August 18 2.605 0.272
New Mexico Male 340 0.159 180.07 July 2 April 25, September 3 4.291 0.117
Female 247 0.029 91.72 April 3 October 27, September 23 0.105 0.949
San Jose-Monterey, California Male 200 0.059 118.51 May 1 October 19, September 6 0.346 0.841

Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012

Female 120 0.110 297.74 October 29 June 26, March 23 0.723 0.697
San Francisco-Oakland,
California Male 723 0.075 229.17 August 21 May 8, November 15 2.019 0.365
Female 513 0.053 197.70 July 20 February 19, December 15 0.721 0.698
Utah Male 375 0.012 348.76 December 20 August 8, July 16 0.028 0.986
Female 318 0.077 54.22 February 24 September 17, June 24 0.942 0.625
Iowa Male 644 0.055 102.52 April 14 October 27, September 6 0.991 0.609
Female 438 0.086 350.83 December 22 July 22, May 7 1.628 0.443
Connecticut Male 585 0.013 6.657 January 7 September 14, August 22 0.047 0.977
Female 478 0.053 61.24 March 4 September 6, June 30 0.668 0.716
Metropolitan Detroit, Michigan Male 669 0.050 338.26 December 9 July 29, May 14 0.835 0.659
Female 504 0.072 27.51 January 28 September 5, July 7 1.315 0.518
Seattle, Washington Male 638 0.032 210.99 August 2 February 2, December 26 0.331 0.848
Female 425 0.071 186.59 July 9 March 21, December 21 1.083 0.582
Western Sweden Male 49 0.742 349.42 December 22 November 16, January 16 11.871 0.003
Female 30 0.274 77.94 March 22 November 19, August 8 1.107 0.575

* j, estimated concentration parameter of the von Mises distribution, which indicates the amount of variation; l
, estimated season peak or
mean direction.

Despite investigation of seasonal patterns in the presen-
tation of ALL, as indicated by the date of diagnosis, over
a range of latitudes from 1.16
N to 58.24
N, there is no
clear message with respect to their interpretation. Neverthe-
less, patterns may have emerged if the date of first symp-
tom had been recorded and studied, because, for example,
a significant seasonal variation for Hodgkins disease has
been reported in this date but not in the date of diagnosis
(26). Thus, this and other studies (27, 28) suggest that date
of first symptom versus date of diagnosis more closely re-
flects the event that precipitates the clinical onset of disease.
However, the induction period of ALL is sufficiently short,
so a similar seasonal pattern should be observed for date of
first symptom and date of diagnosis (27). Although the in-
terval is indeed short for the majority of pediatric ALLs, it
has been claimed that ALL can be clinically silent for
months or even years in some cases (4). Such late cases
are likely to dilute seasonal patterns.
A similar lack of consistent findings has been reported
for other forms of leukemia with respect to a peak in the
date of diagnosis. For example, a significant summer peak
was reported for ALL but not for acute myeloid leukemia
in the United States (23), while a significant autumn peak
(November) for ALL but (bimodal) peaks for acute mye-
loid leukemia in winter and spring were noted in Shiraz,
Iran (5). In contrast, no clear evidence of seasonality
for ALL, acute myeloid leukemia, and chronic myeloid
leukemia was reported from England and Wales (11), but
none of these studies have been analyzed on a case-by-case
basis.
Given the small numbers of cases from western Sweden,
no firm evidence from the United States, and the expected
absence in Singapore, any suggestions for peak seasonality
of diagnosis could all be ascribed to chance variations. Like-
wise, the corresponding degree of seasonality reported in
other studies may be enhanced (or obscured) by local

Am J Epidemiol 2005;162:753763
762 Gao et al.
FIGURE 5. Gender-specic peak dates (95% condence intervals) for
acute lymphoblastic leukemia cases in Singapore (19681999), Hawaii
(19731999), mainland United States (19731999), and western
Sweden (19771994), ordered by latitude (
N). Jan, January; Mar,
March; Jul, July; Sep, September; Nov, November. d, peak date of
diagnosis; , peak date with j < 0.1; D, lower limit of the 95%
condence interval; =, upper limit of the 95% condence interval. The
size of d is proportional to j, except that those peak dates with j < 0.1
are labeled , and the length of the vertical lines, one for each
location, represents the 95% condence interval for the peak date.
referral characteristicseven leading to a false indication of
an underlying climatic component.
ACKNOWLEDGMENTS
The authors gratefully acknowledge the Singapore
Cancer Registry for allowing access to the data.
Conflict of interest: none declared.
REFERENCES
1. Linet MS. The leukemias: epidemiologic aspects. New York,
NY: Oxford University Press, 1985:185222.
2. Douglas S, Cortina-Borja M, Cartwright R. Seasonal variation
in the incidence of Hodgkins disease. Br J Haematol 1998;
103:65362.
3. MacMahon B. Is acute lymphoblastic leukemia in children
virus-related? Am J Epidemiol 1992;136:91624.
4. Greaves M. A natural history for pediatric acute leukaemia.
Blood 1993;82:104351.
5. Karimi M, Yarmohammadi H. Seasonal variations in the onset
of childhood leukemia/lymphoma: April 1996 to March 2000,
Shiraz, Iran. Hematol Oncol 2003;21:515.
6. Badrinath P, Day NE, Stockton D. Seasonality in the diagno-
sis of acute lymphocytic leukaemia. Br J Cancer 1997;75:
171113.
7. Lambin P, Gerard MJ. Seasonal variation in the frequency of
acute leukemia. (In French). Sang 1934;8:7302.
8. Allan TM. Seasonal onset of acute leukaemia. (Letter). Br Med
Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012
J 1964;2:630.
9. Lee JAH. Seasonal variation in the clinical onset of leukaemia
in young people. Br Med J 1962;1:17378.
10. Lee JAH. Seasonal variation in leukaemia incidence. (Letter).
Br Med J 1963;2:623.
11. Douglas S, Cortina-Borja M, Cartwright R. A quest for sea-
sonality in presentation of leukaemia and non-Hodgkins
lymphoma. Leuk Lymphoma 1999;32:52332.
12. Gilman EA, Sorahan T, Lancashire RJ, et al. Seasonality in the
presentation of acute lymphoid leukaemia. Br J Cancer 1998;
77:6778.
13. Harris RE, Harrell FE Jr, Patil KD, et al. The seasonal risk of
pediatric/juvenile acute lymphocytic leukemia in the United
States. J Chronic Dis 1987;40:91523.
14. World Health Organization. International classification of
diseases. Manual of the international statistical classification
of diseases, injuries, and causes of death. Eighth Revision.
Geneva, Switzerland: World Health Organization, 1967.
15. World Health Organization. International classification of
diseases. Manual of the international statistical classification
of diseases, injuries, and causes of death. Ninth Revision.
Geneva, Switzerland: World Health Organization, 1977.
16. Mardia KV. Statistics of directional data. New York, NY:
Academic Press, 1972.
17. Fisher NI. Statistical analysis of circular data. Cambridge,
United Kingdom: Cambridge University Press, 1993.
18. Gao F, Seah SKL, Foster PJ, et al. Angular regression and the
detection of the seasonal onset of disease. J Cancer Epidemiol
Prev 2002;7:2935.
19. Little J, Elwood JM. Seasonal variation. In: Elwood JM, Little J,
Elwood JH, eds. Epidemiology and control of neural tube de-
fects. Oxford, United Kingdom: Oxford University Press, 1992.
20. Stiller CA, Parkin DM. Geographic and ethnic variations in
the incidence of childhood cancer. Br Med Bull 1996;52:
682703.
21. Hjalmars U, Gustafsson G. Higher risk for acute childhood lym-
phoblastic leukaemia in Swedish population centres 197394.
Swedish Child Leukaemia Group. Br J Cancer 1999;79:303.
22. Walker AM, van Noord PA. No seasonality in the diagnosis of
acute leukemia in the United States. J Natl Cancer Inst 1982;
69:12837.
23. Ross JA, Severson RK, Swensen AR, et al. Seasonal variations
in the diagnosis of childhood cancer in the United States. Br J
Cancer 1999;81:54953.
24. Lanzkowsky P. Variation in leukaemia incidence. (Letter).
Br Med J 1964;1:910.
Am J Epidemiol 2005;162:753763

25. World Health Organization. The world health report 2000.
Health systems: improving performance. Geneva, Switzerland:
World Health Organization, 2000:1515.
26. Westerbeek RM, Blair V, Eden OB, et al. Seasonal variations
in the onset of childhood leukaemia and lymphoma. Br J
Cancer 1998;78:11924.
27. Higgins CD, dos-Santos-Silva I, Stiller CA, et al. Season of
birth and diagnosis of children with leukaemia: an analysis of
over 15 000 UK cases occurring from 195395. Br J Cancer
2001;84:40612.
28. Gao F, Machin D, Khoo KS, et al. Seasonal variation in
breast cancer diagnosis in Singapore. Br J Cancer 2001;84:
11857.
APPENDIX
The von Mises Distribution
The von Mises distribution with a single peak at l has the
following probability density function:
1 N
Mh expfj cosh  lg;
2pI0 j
PN 1

j2s
where I0 j s0 s!2 2 and j is positive. R, the stronger the peak identified at l.
Here, h represents the angle, over a range of 0360
(or
02p radians), that can be equivalent to the date of diagnosis
within a year. For example, if this date is February 28, 2004,
then h 360 3 (31 28)/366 58.03
or 1.0123 radians; if
it is February 28, 2005, then h 360 3 (31 28)/365
58.19
or 1.0151 radians. After converting each date to an
angular day for the N subjects concerned, thePpeak l is
N
PN as follows: First, calculate C i1 coshi ,
estimated
S i1 sinhi and note whether each is less than or greater
than 0. Next, calculate l0 arctan (S/C), which gives a value
Am J Epidemiol 2005;162:753763
Seasonal Variation in ALL 763
in degrees (or radians) of between 90
and 90
. Finally, the
estimated peak angle l is equal to 1) l0 itself, if S > 0 and
C > 0; 2) l0 180
, if C < 0 irrespective of the value of S;
or 3) l0 360
, if S < 0 and C > 0. Once l is calculated in
degrees, it can be converted to a date in a standard year of
365 days.
The second parameter of the von Mises distribution j is
termed the concentration parameter and relates to the inverse
of the variance of the distribution. The algebraic expression
for the estimate of j from the data is complex. However,
a very good approximation is given by the following:
8
< 2R R3 5R5 =6 R < 0:53
j 0:4 1:39R 0:43=1  R 0:53  R  0:85;
:
1=R3  4R2 3R R > 0:85
where
Downloaded from http://aje.oxfordjournals.org/ by guest on November 28, 2012
p
C 2 S2
R :
R is the estimated magnitude of the peak at the date
identified and takes values of 01. The larger the value of
It can be shown that when j is large, suggesting a strong
peak, the shape of the von Mises distribution is close to that of
a Normal distribution,pwith
the mean at l and a standard
deviation equal to 1= j: In contrast, for small j, the von
Mises distribution tends to the uniform distribution and is
spread evenly over the whole 360
. This situation is clearly
indicative of no peak in date of diagnosis being present.
A formal test of the null hypothesis, j 0, is made by
referring the value of M 2NR2 to the chi-squared distri-
bution with 2 degrees of freedom.