Variation and Information

in White Blood Cell

Differential Counts
Svend Juul, M.D.,
Joseph S. Pliskin, Ph.D.,
and Harvey V Fineberg, M.D., Ph.D.

The magnitude and sources of variation in the white blood cell (WBC) count and dif-
ferential count affect their information content and clinical value. This study
describes components of variation in the WBC count and differential, estimates the
magnitude of each component, and uses computer simulations to compare the infor-
mation conveyed by the total WBC count and by the WBC differential count,
expressed as the number of each type of cell and as the percentage of each cell type.
Biologic variation is much greater than statistical sampling variation in differen-
tial WBC counts, even when a relatively small number of cells is classified. The com-
monly reported neutrophil percentage is inferior both to the neutrophil count and to
the total WBC count in its ability to detect neutrophilia and to detect recovery from
elevated levels. This conclusion holds for single as well as for sequential WBC differ-
ential determinations and regardless of the level of test result at which the clinician
considers disease to be present. The total WBC count and the neutrophil count differ
little in performance, so a neutrophilic patient’s return to normal levels can safely be
detected and monitored by relying on the less expensive total WBC count and for-
going repeated requests for differential counts. (Med Decis Making 4:69-80, 1984)
Introduction
The total white blood cell (WBC) count and the WBC differential count are
frequently-used clinical tests. The information conveyed by a single differ-
ential count or by a series of counts depends on the origin and magnitude of
variation in reported results. Since a differential count is more labor-

This research was supported in part by grants from the Robert Wood Johnson Foundation
and the National Center for Health Services Research (HS 03314-03).
From the Institute of Social Medicine, Umversity of Aarhus, Denmark; the Department of
Industrial Engineering and Management and University Center for Health Sciences, Ben-
Gurion University of the Negev, Beer-Sheva, Israel; and the Institute for Health Research, A
joint program of the Harvard Community Health Plan and Harvard University, Harvard
School of Public Health, Boston. Please address requests for reprints to Dr. Fmeberg, Institute
for Health Research, Harvard School of Pubhc Health, 677 Huntington Avenue, Boston, Mas-
sachusetts 02115, USA.
70

intensive and expensive than a total WBC count, it would be of practical

value in specified clinical conditions to know what amount of additional
clinical information is conveyed by a differential count as compared to a
total WBC count.
Typical uses of the WBC differential count are to detect infection and to
monitor the effects of treatment. Though the WBC count-in particular the
neutrophil count-tends to increase during bacterial infection, the clinical
usefulness of the WBC count as an aid to diagnosis may be controversial
even in a common problem like appendicitis [1,2]. Such disagreement may

be based on differences in the performance of the test in different popula-

tions and settings, on differences in expectations for test performance, and
on differences in the form for expressing results.
The WBC differential count is typically performed in two stages. First
the total concentration of white blood cells is determined as cells/mm3.
Then the differential distribution is obtained by estimating the relative fre-
quency of each cell type in a random sample of cells. The proportion of
each cell type can be multiplied by the total count to obtain an estimate of
the differential concentration (cells/mm3). In most laboratories the differ-
ential distribution is reported as the percentage of each cell type, rather than
as separate cell counts.
This paper describes five principal components of variation in the WBC
differential, assesses statistical sources of variation, and compares the mag-
nitude of statistical variation to that of interindividual and intraindividual
biologic variation. Computer simulations examine the abilities of the total
white blood cell count, the neutrophil percentage, and the neutrophil count
to discriminate among patients with and without bacterial infection. In par-
ticular, this paper assesses the extent to which the type of test (total WBC
count or differential count) and the form of reporting the differential (as a
percentage or as a count) affect the amount of information conveyed by the
test.

Sources of Variation
Table 1 summarizes five sources of variability in WBC differentials: inter-
individual variation (among individuals), intraindividual variation (within
an individual), statistical sampling, inconsistent cell classification, and
inconsistent sample preparation. The table shows how these sources affect
the cell counts from different types of specimens.
The source of greatest variability is interindividual variation, i.e., varia-
tion in the WBC differential counts among different, apparently healthy
individuals [3]. The distribution around the population mean value for a
cell type reflects the interindividual variability in that cell type. Typically,
clinical results of WBC differentials are interpreted in relation to a reference
interval based on samples drawn from a population of healthy individuals.
Though some studies have shown differences in average cell counts by age,
sex, and racial characteristics, these differences are much smaller than the
71
72

interindividual differences in a healthy population [3] and are of little con-

sequence in evaluating the test results in single patients.
The second largest source of variability is physiological variability within
a single patient, or intraindividual variation [3]. In the short term (hours to

months) this variability is partly diurnal, and it is also related to activity,

diet, medication, and other stimuli [4]. The intraindividual variation is of
greatest concern in the interpretation of a series of examinations in one indi-
vidual.
The third source of variability is statistical sampling [5], which comes
from the two independent steps in performing a differential: estimating the
total number of white blood cells per mm3and estimating the proportion of
each cell type. From these the concentration of each cell type can be
obtained by multiplying the total WBC count by the proportions.
Figure 1 illustrates these three components of variation in the distribution
of a cell count in a healthy individual. The population mean is represented
by point A, around which there is considerable interindividual variation.
The average value for one member of the population is represented by B,
around which there is some intraindividual variation. C represents the
actual concentration of that cell type in that individual at a given point in
time. Due to statistical sampling variation there is some variation in the
recorded value d that estimates the actual concentration C.

Figure 1. Sources of variation in the determination of white blood cell concentration.

A Population mean
=

B = Individual mean
C = Actual individual value at a given time
d Measured value at the same point in time
=

The area under each curve equals unity. The curves may be viewed as having differ-
ent vertical scales.
 73

The final two sources of variation, inconsistent sample preparation and

inconsistent cell classification, bear on the evaluation and comparison of
various technologies to perform differential counts. The term analytic vari-
ation includes these two sources, as well as the variation due to statistical
sampling. The magnitude of variation due to inconsistent cell classification
has been examined by Bacus [6] and by Winkel et al. [7]. Inconsistent cell
classification typically contributes very little to observed variation, and for
this reason inconsistent sample preparation and cell classification will not
be separately quantified in the subsequent analysis.

Measures of Variation. The probability distribution of a variable x may

be characterized by its mean value X, and by its standard deviation SD(x).
In ratio scales (scales with a natural zero, e.g., length, counts) variation in x
is sometimes expressed as the coefficient of variation, CV(x) = SD(x)lX. CV
is a dimensionless measure, and particularly useful when the SD varies
directly with the mean. CV is often used in hematology to describe variabil-
ity in cell counts.
In biologic measurements a logarithmic transformation frequently
improves the correspondence of an empirical distribution to a normal distri-
bution. The standard deviation of a log transformed distribution is denoted
SD(log x). Like CV, SD(log x) is dimensionless. SD(log x) can be approxi-
mately calculated from CV(x) by the formula
SD(log x) = log [CV(x) + (CV(X)2 + 1)112].
A justification for this approximation is available from the authors.
When two independent sources of variation are combined, the combined
variation can be calculated as CV(x) CV(xl)2 + CV(X2)2 , and correspond-
=

ingly for the log transformed distribution:

[SD(log X)]2 = [SD(log xl)]2 + [SD(log X2)]2.
This applies, for example, when combining the total white blood cell count
with the differential distribution to calculate the differential count.

REFERENCE VALUES IN HEALTHY INDIVIDUALS. Medical textbooks pro-

vide different ranges for the levels of white blood cells in healthy individuals
[8-11]. One inconsistency in the figures given by one text is that the sum of
the lower ranges of all cell types is higher than the lower range given for the
total WBC count [9]. A key study by Orfanakis et al. [12] demonstrated that
the counts of individual cell types in 291 healthy subjects were log normally
distributed. Orfanakis’ main findings are summarized in the first three
columns of Table 2. The fourth column shows the geometric mean, i.e., the
retransformed value of the mean in a logarithmic transformation. The geo-
metric means are close to the median values reported in the second column,
supporting the assumption of a log normal distribution. The standard
 74

Table 2. Reference Values and Ranges of White Blood CeHs

______
[Counts of cells(x) per mm 31

&dquo;
From Orfanakis et al. [12].
b Geometric mean = antilog [log x2 5% +log X97 s%)~2].
‘SD(log x)=(log x9~s%-log X2 5oy,,)/3.92. For eosinophils and basophils, SD(log x)=
(log xso% -log x2 5 %) /1.96.

deviation of the log normal distribution is shown in the fifth column of the
table. In a normal distribution the central 95 percent of the distribution is
encompassed within 1.96 standard deviations on either side of the mean,
yielding the figures in the fifth column.
The following sections on sampling variation and biologic variation
describe the breakdown of the total variation of the last column into its sta-
tistical and biological components. The latter will be further dissected into
interindividual and intraindividual components.

SAMPLING VARIATION. Both the total WBC count and the determination
of the proportion of each cell type involve sampling variation. We calculate
the variation from these two sources separately.

Sampling Variation in the Total WBC. Most contemporary total WBC

counts are performed by machines such as the Coulter S counter. Variation
in the total WBC count depends on statistical properties of sampling and on
machine imprecision. Typically, 10 000 cells are counted, giving rise to a
Poisson variation corresponding to a CV of 0.01. The analytic variation of a
total WBC count performed on a Coulter S counter has been empirically
estimated to correspond to a CV of 0.04 [13], indicating some degree of
machine imprecision. We will use 0.04 to represent the coefficient for varia-
tion attributed to the total WBC count.

Statistical Sampling Variation in the Differential Count. Sampling varia-

tion in the count for each cell type is due in part to statistical variation in the
differential classification of cells, and in part to variation in the total WBC
count.
The differential count is calculated by multiplying the total WBC count
by the proportion of the cell type of interest. If k of n cells are identified as a
given cell type, the estimated proportion of that cell type is p kln. The =
 75

standard deviation of the estimate is SD(p)=[p(I-p)lnl 1/2 , based on

the binomial distribution. The corresponding CV is given by CV(P)=
[(1- p)/np]l/2.
As can be seen from the above formula, the coefficient of variation for
the proportion of cell types depends on both the proportion of the cell type
and the number of cells counted. For example, if 200 cells are counted (for
the differential) and the observed proportion is p = 0.2,

CV(p) _ [0.8/(0.2)(200)] 1/2 =

0.141.

The coefficient of variation of the concentration of a cell type CV(x) incor-

porates the sampling variation in the total WBC (CV = 0.04):
[CV(x)]Z - [CV(p)]2 + [0.04]2 =
0.0216;
hence CV(x) = 0.147.
Table 3 presents for each cell type the assumed proportion p, the coeffi-
cient of statistical sampling variation of the proportion CV(p), the co-
efficient of total sampling variation CV(x), and the standard deviation of
the log normal distribution SD(log x) in a classification of 200 cells. These
calculations yield estimates of the sampling variation of the cell counts
reported by Orfanakis et al. [12].
BIOLOGIC VARIATION. The sampling component of variation can be sub-
tracted from total variation to yield an estimate of biologic variation. Bio-
logic variation includes two components, interindividual variation and
intraindividual variation. Table 4 summarizes these components of varia-
tion for each of the five major types of white blood cells. Biologic variation
is much larger than sampling variation for neutrophils and lymphocytes;
these cell types are the most important clinically and typically constitute
more than 90 percent of the circulating WBC population.

Information Value of the WBC Differential:

Computer Simulation of Changes During Infection
We examined the abilities of WBC counts and WBC differential counts, as
simulated by a computer, to identify patients with bacterial infection and to
monitor recovery from infection. We compared the information conveyed
by the WBC differential when expressed as proportions and as cell counts,
and we compared these expressions with the information conveyed by the
total WBC count. We proceeded in three steps to simulate results in a popu-
lation of 1000 individuals without infection (healthy). Then we proceeded in
a similar way to estimate results in a population of 1000 individuals with
infection (sick).
As a first step we simulated the average counts for each individual in the
healthy population, using the mean counts of each cell type in the second
column of Table 2 and the interindividual variation in the fourth column of
76

Table 3. Estimated Statistical Sampling Variation of White Blood Cells,

by Classification of 200 Cells

a
Proportions according to the relative distribution of median values in Table 2, second column.
bCV(p) = [(1- p)/200pI1l2
c
[CV(x)l2= [CV(p)]2 + [O.04¡2. CV(x) = Coefficient of variation for the cell
~ SD(Iog x) log [CVM + (CV (X)2 + 1)1121
=

Table 4. As a second step we simulated the concentration of each cell type in

an individual at a point in time by applying the intraindividual variation of
that cell type (Table 4, column 5) to the individual average values obtained
in the first step. Finally, we simulated in each patient the classification of
100 randomly selected cells, assuming no errors in cell classification.
We followed the same approach to simulate cell counts in the population
with infection. We considered the five major white blood cell types (neutro-
phils, lymphocytes, monocytes, eosinophils, and basophils) as coming from
five independent cell populations. (This assumption would not apply to the
subgrouping of neutrophils according to maturity, i.e., band versus seg-
mented forms.) We assumed the mean neutrophil count in a sick individual
was double the level for that individual when healthy. We further assumed
that the levels of nonneutrophil cells do not change with infection.

Table 4. Components of Biologic Variation in White Blood Cells

a
From Table 2, column 5.
b From Table 3, column 4.
c (Col. 3)2 = (Col. 1)2 - (Col. 2)2
(Col. 3)2 = (Col. 4)2 + (Col. 5)2.
Allocations to the last two columns are proportional to the interindividual and intra-
individual variation found by Statland et al. [3].
 77

To judge whether these assumptions are reasonable, we examined the

records of 56 patients at Mt. Auburn Hospital, Cambridge, Massachusetts,
with primary or secondary diagnosis of gram-negative sepsis. The review
confirmed that an increase in the neutrophil count to twice the usual level or
more is a typical occurrence during that type of infection. The assumption
that other cell types are unaffected during infection was not directly con-
firmed, as lymphocytes tended to decrease and monocytes to increase dur-
ing gram-negative sepsis. These changes were, however, much smaller than
the change in the neutrophil count, and they tended to balance each other,
so that the total count of nonneutrophil cells remained approximately

unchanged during infection. To avoid unnecessary complexity of the com-

puter model we did not introduce changes in lymphocyte and monocyte
levels at the time of infection.
The analysis addresses two types of clinical problems: the ability of a
single white blood cell differential to distinguish sick from healthy individ-
uals, and the ability of repeated WBC differentials to identify individuals
with changing clinical conditions.

Figure 2. ROC curves showing the abilities of the neutrophil count, neutrophil per-
centage, and total WBC count to distinguish neutrophilic from nonneutrophilic indi-
viduals.
78

The neutrophil count, total WBC count, and neutrophil percent in

patients with and without infection were used for calculating points on
ROC curves. The ROC curves in Figure 2 describe the ability of the neutro-
phil count, the total white blood cell count, and the neutrophil percentage
to distinguish between sick and healthy individuals under the conditions
described above. The location of curves nearer to the upper left corner indi-
cates superior performance. The area under the ROC curve was used as an
index of test performance, an area of 1.0 indicating a perfect test and an
area of 0.5 indicating a test without informative value. The area was com-

puted by the method described by Swets [14].

The location of ROC curves in Figure 2 indicates that the neutrophil per-
centage is inferior for distinguishing healthy from sick individuals, as com-
pared to the neutrophil count, which in turn is slightly better than the total
WBC count. This conclusion holds regardless of the level of test result at
which the clinician considers disease to be present. The areas under the three
curves are listed in the first row of Table 5, confirming the visual impression
from Figure 2.
We examined the ability of the white blood cell differential to indicate
recovery from disease, by simulating results for two hypothetical popula-
tions, each of 1000 individuals. For each individual in the &dquo;recovery&dquo; popu-
lation the assumed neutrophil level shifted from double the assumed healthy
level back to the healthy level. In the &dquo;no recovery&dquo; population each individ-
ual’s average neutrophil count was simulated to remain double the healthy
level for that individual. The underlying intraindividual variation and the
analytic sampling variation remained in effect.
Changes from one determination to a second can be examined by means
of the ratio or by means of the difference between the determinations. We
examined the abilities of the ratio and the difference to distinguish members
of the recovery group from the no recovery group by constructing ROC
curves for the neutrophil count, total WBC count, and neutrophil percent-

age.
The test performances are summarized in the second and third rows of
Table 5. The ratio between determinations was superior to the difference
between determinations. The neutrophil count and the total white blood cell
count were superior to the neutrophil percentage. Successive WBC counts
within an individual can discriminate disease better than a single determina-
tion because successive counts in an individual are not influenced by inter-
individual variation.
We also simulated a succession of five determinations in the recovery and
no recovery populations, the recovery population being subject to a gradual
reduction in neutrophil level. In both groups cell counts were subject to
intraindividual variation. We calculated the regression coefficient and the
correlation coefficient for the changes in neutrophil count, total white
blood cell count, and neutrophil percentage over time, and used these
 79

Table 5. Performance (Area under the ROC curve)

of the White Blood Cell
Differential

results to construct the ROC curves summarized in the last rows of Table 5.
The correlation coefficient seems to be a slightly better predictor of
recovery than the regression coefficient for sequential neutrophil and total
WBC counts.

Conclusions
Compared to biologic sources, statistical sampling makes a relatively small
contribution to variation in the white blood cell differential count, even
when only 100 cells are classified. The origin and magnitude of variation
determine the potential performance of any technology to count and clas-
sify white blood cells. The relatively small statistical component of variation
limits the potential added precision of counting large numbers of cells, as
done, for example, by automated differential counters.
Based on simulated determinations of white blood cell counts under stan-
dard conditions, we conclude that the neutrophil percentage is inferior to
the total WBC count and to the neutrophil count in ability to identify indi-
viduals with elevated neutrophil concentrations. The neutrophil percentage
is also inferior in ability to identify individuals in whom the neutrophil level
changes over time. Laboratory reports of the WBC differential count
should prominently show the numerical counts of each cell type, perhaps
indicating percentages in parentheses. The total WBC count and the neutro-
phil count based upon a classification of 1000 cells differ little in perform-
ance. This conclusion is consistent with the findings of Brecher et al., who
analyzed 415 sequences of WBC and differential counts in hospital patients
80

and found only four in which the neutrophil count had changed without a
concomitant change of comparable magnitude in the total WBC count [15].
The clinician who wishes to detect and monitor a neutrophilic patient’s
return to normal levels can safely rely on the total WBC count and forgo
repeated requests for differential counts.

References
1. Raftery AT: The value of the leukocyte count in the diagnosis of acute appendi-
citis. Br J Surg 63:143-144, 1976
2. Bolton JP, Craven ER, Croft RJ, Menzies-Gow N: An assessment of the value
of the white cell count in the management of suspected acute appendicitis. Br J
Surg 62:906-908, 1975
3. Statland BE, Winkel P: Physiological variability of leukocytes in healthy sub-
jects. In, Koepke JA, ed: Differential Leukocyte Counting. Aspen, Colo: CAP
Conference, 1977, pp 23-37
4. Boggs DR: Response of blood leukocytes to common stimuli. In, Koepke JA,
ed: Differential Leukocyte Counting. Aspen, Colo: CAP Conference, 1977, pp
47-52
5. Rumke CL: The statistically expected variability in differential leukocyte count-
ing. In, Koepke JA, ed: Differential Leukocyte Counting. Aspen, Colo: CAP
Conference, 1977, pp 39-45
6. Bacus JW: The observer error in peripheral blood cell classification. Am J Clin
Pathol 59:223-230, 1973
7. Winkel P, Statland BE, Harris SC, et al: A study of variation of concentration
values of leukocyte type 2. Analytical considerations. Manual vs. automated cell
counting. Advances in automated analyses. Technicon International Congress.
Tarrytown, New York: Mediad Incorporated, 1977
8. Rumke CL, Bezemer PD, Kuik DJ: Normal values and least significant differ-
ences for differential leukocyte counts. J Chronic Dis 28:661-668, 1975
9. Cecil RL, Loeb RF: Textbook of Medicine (13th ed). Philadelphia: Saunders,
1971
10. Harrison TR: Principles of Internal Medicine (9th ed). New York: McGraw-
Hill, 1980
11. Wintrobe MM: Clinical Haematology (7th ed). Philadelphia: Lea and Febiger,
1974
12. Orfanakis NG, Ostlund RE, Bishop CR, Athens JW: Normal blood leukocyte
concentration values. Am J Clin Pathol 53:647-651, 1970
13. England JM: Total and differential leukocyte count. Br J Haematol 33:1-7, 1976
14. Swets JA: ROC analysis applied to the evaluation of medical imaging tech-
niques. Invest Radiol 14:109-121, 1979
15. Brecher G, Anderson RE, McMullen PD: When to do diffs. How often should
differentials be repeated? Blood Cells 6:431-439, 1980