Statistical methods have great value in medicine

and should be of interest to practicing physicans

Statistics in the
Practice of Medicine
Robert G. Hoffmann, PhD, Gainesville, Fla.

PHYSICIANS and statisticians first step is to establish clinical criteria for normal¬
PRACTICING
may not to have many mutual interests,
seem
but perhaps they have more in common than is
ity and admit subjects to the study only if they
meet the clinical criteria.
apparent at first. The purpose of this article is to How many normal subjects would be required?
describe a few statistical tools and to illustrate how There is no way to answer this question without
these tools may be useful in the practice of first specifying the accuracy to which the normal
medicine. limits are wanted, and an estimate of normal sub¬
As a starting point, consider
a fairly frequently ject variability is also needed. Information is avail¬
encountered medical problem, diabetes mellitus. able about some tests in the literature, so the liter¬
ature data will be used. The results of testing 60
2

According to some physicians, they believe that

diabetics are best maintained at slightly elevated clinically normal subjects are shown in Table 1.
blood-glucose levels, but others feel that they Table 1 was formed simply by setting up the
should be maintained within limits of clinical nor- glucose scale shown in column 1; tallying the glu¬
mality. Which group is right or are they both cose values from the list supplied by the laboratory
right? And what is clinical normality and how is it (column 2); and adding up the tally marks as
determined? Suppose we consider the question of shown in column 3. The cumulative frequency
clinical normality first, partly because it is rela- shown in column 4 is obtained by successively add¬
tively easy to discuss and partly because it is a ing numbers of subjects beginning at the top of
portion of the problem of maintaining diabetics the table. The cumulative per cent in column 5 is
in control.
Table 1.—Frequency Distribution of 60 Normal Subjects
Normal Values by Blood Glucose
According to one textbook,1 the normal range (l)
Blood (3)
(4)
Cumu¬
(5)
Cumu¬
for level of blood glucose after fasting (which is Glucose,
Ml
(2)
Tally
Subjects,
No.
lative lative
%
Mg/100 Frequency
just called glucose after this) is 70-110 mg/100 ml of 70-74 1 1 1.7
whole blood. This may not be true for a given 75-70 7 8 13.3
80-84 IÍ I 11 19 31.7
laboratory, because methods for determining glu¬ 85-89 i mu mu i 16 35 58.3
cose concentrations differ among various labora¬ 90-94 I Hill II 12 47 78.3
tories. Suppose we decide to see whether this 95-99 4 51 85.0
100-104 4 55 91.7
range is reasonable for a specific laboratory. The 105-109 2 57 95.0
110-114 2 59 98.4
115-119 1 60 100.0
From the J. Hillis Miller Health Center, University of Florida. Total Subjects: 60

Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
 the cumulative frequency converted to a per cent
of total subjects. For the moment ignore these
cumulative values.
The discussion of normal values was started par¬
tially as a means of illustrating statistical tech¬
niques and how they may be useful in the practice
of medicine. With the information displayed in the
first three columns of Table 1 we are ready to
begin discussion of analogies between medical and
statistical problems and approaches to them.
Apart from blood glucose level, how is a diag¬
nosis of diabetes made? From the clinical point of
view, it is a pattern of responses made by a patient,
with the pattern consisting of certain signs and
symptoms exhibited by the patient. Statisticians
look for "response patterns" too, but they look for
them in groups of measurements such as the glu¬
cose tests shown in Table 1. The pattern is shown
by the numbers of subjects having given glucose
values. Note, for example, that the "most popular"
glucose value is in the range 85-89 mg, with 16
subjects. Values on either side of this range are
increasingly less popular as we go farther away
from it on either side.
The pattern exhibited by the normal glucose sub¬
jects is most easily seen in chart form; Fig 1 ex¬
hibits the data graphically. For a given value, the
number of subjects is shown by the bar height
above the value. Patterns such as this are called
"frequency distributions" because the 60 subjects
have been distributed according to their glucose
values, and the bar heights correspond to the
frequency.
These laboratory results may now be compared
with the textbook range mentioned previously, 70-
110 mg. Even a gross appraisal of the situation
reveals that differences exist between the data and
the textbook range. None of the values for the
normal subjects reached the lower limit of the
textbook range, but several exceeded it. It looks
as though it would be best to establish a range
from the data. To help do this some statistical tools
will be used, but before they are discussed, con¬
sider again what is meant when we say "normal
range."
With unlimited time and facilities millions of
normal subjects could be tested for glucose. The
result should be a distribution similar to that shown
in Fig 1. But it is not practical to test millions of
normal subjects, so we must make best use of the
information we have—the 60 subjects. To do this,
an assumption must be made. The hypothetical
millions of normal subjects would produce a
smooth, symmetrical distribution curve similar to
the one seen in Fig 1. The only way to test the
reasonableness of this assumption is to actually
test several hundred normal subjects and see what
the distribution looks like; but apparently no one
has done this.
The smooth curve shown in Fig 1 is actually a
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
mathematical expression which is usually referred
to as the "normal distribution." As far as medicine
is concerned, this is an unhappy name because it
is only a mathematical expression. Happily, how¬
ever, it has been named too after a famous mathe¬
matician by the name of Gauss, so it will be re¬
ferred to as the gaussian distribution.
The gaussian distribution is of great importance
in statistics and the applications being discussed,
but in an article only a few of its features can
be described. The gaussian distribution shown in
Fig 1 was fitted (by certain mathematical pro¬
cedures) to the 60 normal blood sugar values.
There are two reasons why a mathematical
expression is sought to describe the glucose dis¬
tributions; first, to summarize the 60 values in a
convenient form and, second, to use this summary
to estimate what would be obtained with a much
larger series of measurements. Bemember, what is
sought is a clinically normal range using the infor¬
mation obtained in testing 60 normal subjects.
One feature of the gaussian distribution is that
it never touches the bottom of the chart ( abscissa )
on which it is plotted. From the practical point of
view, this means it would admit clinically normal
subjects with very high and very low glucose
values, provided a of subjects was
large number
examined. This that to use the gaussian
means
distribution to establish the normal range, its tails
must be chopped off. The two chopping points will
define the normal range. The problem at the mo¬
ment is how to define and then determine these
two points. To solve it, knowledge of a few mathe¬
matical properties of the gaussian distribution is
needed.
The distribution is a bell-shaped curve ( as shown
in Fig 1) which is determined by two quantities
the estimates of which are computed from the
measurements. These are the mean and the stand¬
ard deviation. Here is a little about these two
quantities.
Symbolic How Computed
Expression Add up all values and di-
Mean _ 2.x -vide by number of values.
Estimate
_ n
Subtract mean from each
each differ-
_value, square
Standard /2(x "x)a ence, sum the resulting
Deviation s = V n
—
1 squares, divide by one less
Estimate —
than the number of values,
and, finally, take the square
root.
From the practical point of view, the mean locates
the curve. If, for example, 100 mg is added to each
of the 60 glucose values, the curve would just slide
up the glucose scale by 100 mg.
The standard deviation describes how "fat" or
"skinny" is a given normal distribution. Note that
if all the values were the same as the mean, the
standard deviation would be zero.
It is not obvious from what has been said up to
 X=89.8 MG/IOOML
S= 9.6 MG/IOOML
TEXTBOOK RANGE.
60-110 MG
SICK" SUBJECTS
X = 10
"I-1-T
70 75 80 85
BLOOD GLUCOSE MG/IOOML ARBITRARY SCALE

Fig .—Frequency distribution of 60 clinically normal subjects by

1 Fig 3.—Two theoretical gaussian distributions summed to simulate
fasting blood glucose. The bell-shaped curve is a gaussian distribu¬ a composite distribution. The two gaussian components represent

tion fitted to the normal subjects' values. "clinically normal" and "sick" subjects.

(PLOTTED ON NORMAL PROBABILITY PAPER I

THEORETICAL
GAUSSIAN
DISTRIBUTION

95% LIMITS FOR

NORMAL RANGE _

-1-1-1-1-1—
70 75 85 90 95 100 105
BLOOD GLUCOSE MG/100 ML ARBITRARY SCALE
(END POINTS OF INTERVALS)

Fig 2.—Cumulative frequency distribution of 60 clinically normal Fig 4.—Two theoretical distributions and their sum, representing a
subjects by fasting blood glucose, plotted on normal probability composite, on probability paper. One component represents "clin¬
paper. The straight line is a theoretical gaussian distribution. ically normal," the other "sick" subjects.

this point, but it is possible to plot a gaussian dis¬ ping points.We are now ready to define the nor¬
tribution so that it has a mean of zero and a mal range from the 60 normal subjects.
standard deviation of unity. These are called unit Normal limits are arbitrarily defined to be the
gaussian curves, and the unit of measurement is points which enclose 95% of the values obtained
the standard deviation. This is a very important by testing clinically normal subjects. The mean
point because when a standard deviation is com¬ and standard deviation for the 60 normal subjects
puted from a set of measurements, the units are are:
the same as the measurements. Another feature of Mean = x = 89.8 ]
the gaussian curve is that about 95% of the area , , lmg/100 ml of blood
enclosed by the curve is contained within the Standard
range defined by both adding and subtracting two
Deviation = s = 9.6 J
standard deviations to and from the mean. The two Upper
limit: x + 2s = 89.8 + 19.2 = 109
standard deviations can be used to chop off the mg/100 ml of blood
tails of the gaussian curve, although any multiples Lower
of standard deviations could be taken as the chop- limit: x 2s = 89.8 19.2 71
— —

Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
 The computed limits are fairly close to the text¬
book values.
What has just been done may seem like a great
deal of trouble to obtain normal limits, and the
assumptions could certainly be questioned. There
is this about what has been done, however, that is
well worth keeping in mind: Given the same set of
60 glucose measurements, anyone, anywhere, any¬
time would obtain the same normal range if he
made the same assumptions and went through the
same set of computations. Assumptions must be
made unless definite information exists, and in a
problem of this kind, we will always be faced with
some uncertainty.
One practical difficulty with obtaining normal
limits as has just been done, is the fact that a
group of clinically normal subjects had to be ob¬
tained and tested. To do this for large numbers of
subjects is not possible for most laboratories. What
is about to be described is a method for obtaining
normal value information without testing any se¬
lected group of clinically normal subjects. We will
simply capitalize on the fact that most of the pa¬
tient specimens sent to laboratories for analysis are
normal with respect to many tests. To do this a
simple statistical technique is needed which is now
described.
The gaussian, or any other, distribution can be BLOOD GLUCOSE MG/IOOML
plotted in cumulative form. Befer again to Table 1 Fig 5. —Cumufotive distribution of 449 tests of fasting blood glucose.
for the glucose values tabulated in cumulative These values represent routine tests of patients' specimens. They are
form. In graphic form, the curve is "S" shaped, plotted on normal probability paper and a gaussian curve is eye-
rising to a maximum of the total number of meas¬ fitted to the "clinically normal" component.
urements. For a given glucose value the curve
shows the number of subjects having that value
or a lower one.
The glucose tests shown in Fig 1 were plotted on
ordinary graph paper, using an arithmetic scale for 90

the glucose values. The next illustration, Fig 2, 70 GAUSSIAN ESTIMATE OF THE
shows the glucose data plotted on a special graph 60-
CLINICALLY NORMAL COMPONENT

paper called normal (gaussian) probability paper. 50

The cumulative form of the distribution is used.
10
Note that the glucose scale is unchanged from the
previous plot (Fig 1) but the number of subjects 30"

(shown as apercentage of the total) along the 20"

side is not arithmetic scale. This special graph

an 10-

paper serves the useful purpose of "straightening

out" a cumulative gaussian distribution. It forms a 15 35 55 75 95 115 135 155 175 195 215 235 255 275 295 115

straight line. This feature allows a given set of BLOOD GLUCOSE MG/100 ML

data to be plotted on probability paper and a nor¬
mal distribution "fitted" to it. A ruler is used to
draw a straight line through the points. This is
called an "eye fit." The points around the 50%
point (side scale) are given the greatest weight.
This is all that is needed in the way of statistical
techniques solve the problem of obtaining nor¬
to
mal values from clinical laboratory data.
Before actual laboratory data are considered,
however, charts which have been prepared to show
how mixtures of gaussian distributions look when
plotted on probability paper should be considered.
Fig 6.—Frequency distribution of 449 tests of fasting blood glucose.
These values represent routine tests of patients' specimens. A gaus¬
sian curve has been fitted to the "clinically normal" component.
Laboratory data consists of specimens from healthy
as well sick persons. Gaussian distributions can
as
be used describe the values from sick as well as
to
healthy persons, but, starting with the values as
they come from the laboratories, the distributions
are mixed.
Fig 3 shows two gaussian distributions plotted
on regular graph paper computed to simulate a

Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019

25 45 65 85 105 125 145 165 185 205 225 245 265
BLOOD GLUCOSE MG/IOOML
Fig 7.—Frequency distributions of 500 tests of fasting blood glucose.
One distribution represents laboratory work in a general hospital
and the other represents work done in a nearby private laboratory.
mixed or composite distribution. The two gaussian
distributions are also shown separately in the figure.
The one on the left represents clinically normal
subjects and the one on the right, the sick subjects.
They were computed so that out of every group of
ten subjects, eight were normal and two were sick.
The standard deviation for the sick subjects was
also chosen to be five times the magnitude of the
standard deviation of the normal subjects.
Fig 4 shows these distributions plotted on prob¬
ability paper. This figure deserves careful study.
Note that the composite distribution can be de¬
scribed by not one, but two straight lines. Each
of these two lines corresponds to one of the two
distributions shown in Fig 3. Note how closely the
normal component of the composite corresponds to
the normals plotted separately. Probability paper
has separated the two components for us.
The separation is not perfect because the slope
of the line which would be used to estimate the
normals is not quite as steep as it should be. This
means that a graphic estimate of the standard
deviation, on which the limits of clinical normality
are based, will be a little too broad, but if the
example is reasonable, the bias is not very large.
The bias in the sick component is much greater,
but estimating this component is not part of the
immediate goal.
This statistical technique may now be put to
work on a real set of data. Fig 5 shows a probabil¬
ity plot of 449 fasting blood-glucose values from
the laboratory.* A straight line has been fitted by
eye to the component which represents clinically
normal subjects. At the 2.5% and 97.5% points (side
scale) horizontal lines have been projected to the
line representing the clinically normal components.
At their intersections, lines have been dropped to
the glucose scale at the bottom of the chart. These
mark the clinically normal range, which is 55-120
mg. Previously ( Fig 4 ) it was shown that the nor¬
mal range would be a little too broad, so it might
be better to take the 5% and 95% points as the
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
chopping points. If this were done, then the limits
would be 60-115 mg, a range that coincides fairly
closely to the one quoted from the textbook. A
mathematically more sophisticated method exists 3
for dissecting mixed distributions, but it is a little
too intricate to discuss here.
The blood glucose data are plotted in the form
of a regular frequency distribution in Fig 6. The
clinically normal component is shown as a gaussian
distribution. The standard deviation for the normal
component was estimated from the cumulative dis¬
tribution in Fig 5. Note that the gaussian estimate
is a little too wide but not grossly so. The height of
the normal component was chosen so as to coincide
with the height of the distribution of laboratory
values.
The discussion of uses of statistics in the prac¬
tice of medicine was introduced by considering a
question concerning the treatment of diabetics.
One portion of the problem involved normal val¬
ues. We have seen that with the aid of some simple
statistical techniques a basis for normal limits can
be defined reasonably exactly. A method has been
suggested for obtaining normal limits without test¬
ing any special group of clinically normal subjects.
This statistical technique can be used for obtain¬
ing any normal values in medicine where a group
of measurements are available and the mathemat¬
ical assumptions are reasonable.
The blood glucose data used in the illustrations
were selected from a large body of laboratory data
involving many tests and from several laboratories.
These data were chosen so that reasonable agree¬
ment was obtained when it was compared with
textbook data. (In contrast to the normal range of
60-115 mg obtained from the laboratory data illus¬
trated, data from another laboratory yielded a
range of 88-135 mg. The same statistical procedure
was used for both sets of data.) Anyone who uses
these techniques will not necessarily find such good
agreement. If disagreement is substantial, which is
right—the textbook or the estimates using the statis¬
tical techniques? Perhaps it would be more appro¬
priate to ask, "Which is the best normal range?" No
range is likely to be wholly right or wrong. With
regard to the normal range based on current clinical
values, this feature might be considered: The esti¬
mate is based on the local situation, and this is im¬
portant. The estimates are based on values as they
are being reported locally, whether they agree with
textbooks or not.
There is another point about the normal range
problem which might be considered. Begardless
of the source of the range, it is assumed that the
laboratory testing procedure is stable. At present,
the stability of many laboratory procedures could
be improved. Pathologists are at work on the prob¬
lem by means of their quality control programs,
but much remains to be done. More will be said
 120-
100-
—
PRIVATE LABORATORY
80-
-HOSPITAL LABORATORY
60-
40-
20-
"i-1-
100 MO 120 130 140 150 160
NORMAL QUOTIENT VALUES
Fig 8.—Frequency distributions of 500 tests by fasting blood glu¬
cose. One distribution represents values from a general hospital
laboratory and the other, values from a private laboratory. The
measurement scale (mg/100 ml of blood) has been transformed to
"normal quotient values."
about quality control programs later on in this
manuscript.
How do we know whether two groups of meas¬
urements differ from one another? The normal
ranges just discussed differed somewhat from one
another, but all were based on the examination
of limited numbers of measurements. Problems
where limited numbers of measurements are avail¬
able are those where statistical techniques can be
helpful, so they will be discussed next.
For example, if some patient specimens were
tested in a private laboratory, it would be pertinent
to inquire as to whether the private laboratory's
results coincide with those obtained in a nearby
hospital's laboratory. For given laboratories, a good
means of obtaining information is to send aliquots
of a "standard sample" to both laboratories and
compare the reported results. All of this takes extra
time and work, however, so the clinical values will
again be used as a source of information. From the
statistical point of view, this will be an example
of comparisons among groups.
Comparisons Among Groups—I
The problem is to determine whether a hospital
laboratory is reporting values comparable to those
reported by a private laboratory. To obtain infor¬
mation pertinent to the problem, 500 consecutively
determined blood glucose values obtained after
fasting were copied from the records of the hos¬
pital and the private laboratory. In the hospital
laboratory, the glucose oxidase method was in use
and in the private laboratory the Folin-Wu method
was being used. Frequency distributions were tabu¬
lated, and the results are shown in Fig 7.
At this point in the preparation of the paper,
work was stopped. The glucose distributions are
so obviously different that no statistical methods
are needed to detect the difference, as might have
been expected from the different testing methods.
The private laboratory is located across the street
from the hospital, so a single group of physicians
will send patients to either laboratory. Surely the
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
Table 2—Blood Urea Nitrogen Test Results (Mg/100 Ml)
First Period Second Period
15 14
10 17
12
12
16 16
20 22
13 13
18 18
22 11
14 15
21 16
physicians were aware that glucose results were
different for the two laboratories, but the differ¬
ence did not help the physicians in the care of
their patients. Was there something that could be
done to help the physicians without disturbing the
laboratories to any great extent? A method was
found, which is about to be described. This is not
exactly conventional statistics, so we digress from
our description of comparisons among groups.
What is about to be done is to change the units of
measurement, which for glucose are milligrams per
100 ml of blood, to some new units of measure¬
ment. These new units are arbitrarily called "nor¬
mal quotient values."
Normal Quotient Values
Before the proposed units of measurement are
described a few words about the medical measure¬
ment problem seem to be in order. There are cer¬
tainly hundreds, if not thousands, of different kinds
of measurements made in medicine.4 Pediatricians
are concerned with growth rates; radiologists are
concerned with organ sizes; gynecologists are con¬
cerned with pelvic measurements; and most every¬
one is concerned with clinical laboratory measure¬
ments. Many kinds of measurement scales are in
use because of the diverse nature of the things that
are measured.
What is proposed here is a single measurement
scale that could be substituted for many of the ex¬
isting measurement scales. The proposed scale is
based on the standard deviation of clinically nor¬
mal subjects. In other words, it is based on the
biological rock of clinical normality. Transforming
any measurement scale to one based on the stand¬
ard deviation of normal subjects is all that is re¬
quired. First, determine the standard deviation of
the normal subjects as described earlier and, sec¬
ond, divide each original scale value by the stand¬
ard deviation just computed. The units of measure¬
ment are now standard deviations. This is a
somewhat awkwardly narrow scale because of the
large number of subjects which are included over
a range of one standard deviation, so each new
Table 3.—Blood Urea Nitrogen Averages and Variances
for Two Different Periods
First Period Second Period
Average. 16.1 15.4
Variance 16.3 10.3
.
 Table 4.—Null Hypothesis Tests for Averages and
Variances of Blood Urea Nitrogen for Two Periods
First Second
Period Period
Average . 16.1 15.4
Variance . 16.3 10.3
scale value is now multiplied by 5. Only
step is needed in the transformation.
To the new scale values just computed add (or
subtract) a constant determined so that the mid¬
point of the range of clinically normal subjects is
100. The number 100 was chosen because it is easy
to remember, has three significant figures in it,
and contains no decimal points. This completes
the transformation, and the resulting units are
called "normal quotient values." In symbolic nota¬
tion, this is all that has been done:
(5)x
——+ k
y =
where y is the normal quotient value;
x is a value on the original measurement scale;
a is the standard deviation of normal subjects;
k is a constant chosen so that the mid-point of the
range of normal quotient units is 100.
In actual practice, the scale transformation need
be computed only once. In clinical laboratories no
additional work would be required to report, in
normal quotient units, results of testing patients'
specimens, after the initial transformation had been
computed. Of course a transformation would be
required for each test procedure.
The results of transforming the two glucose dis¬
tributions shown in Fig 7 are seen in Fig 8. Note
how similar are the two distributions now. The
mid-point of clinically normal subjects is 100 units;
the clinically normal range is 90-110 units (± 2
standard deviations) and subjects with values out¬
side the normal range can be judged in terms of
deviations of healthy subjects.
Although the distributions seen in Fig 8 are for
blood glucose, similar computations could be made
for any reasonably large group of measurements.
The resulting distributions might not be the same,
but the mid-point of the normal subjects would be
100. The range of clinical normality would also be
90-110 normal quotient units. If the private and
hospital laboratories were reporting their results in
normal quotient units, the physicians' problems
about differences between laboratories would no
longer exist.
So much for normal quotient units for the pres¬
ent. Other potential uses for them will be men¬
tioned later in connection with a few remarks
about electronic computers. For the moment, how¬
ever, a return to the problem of comparisons
among groups is in order.
Comparisons Among Groups—II
Consider another laboratory problem which
arises when a change in testing procedure is made.
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
HEALTHY SUBJECTS
Null FALSE NEGATIVE TESTS
Hypothesis
Probability FALSE POSITIVE TESTS
0.56
0.25 SICK SUBJECTS
one more
MEASUREMENT
SCALE
UPPER LIMIT OF
NORMAL RANGE
Fig 9.—Simulated distributions of "healthy" and "sick" subjects.
An area of uncertainty exists where the two distributions overlap
each other.
The data presented in Table 2 came from a hospital
laboratory when the blood urea nitrogen procedure
was changed from the urease-NessIer method (first
period) to the Autoanalyzer (second period). The
question the laboratory director wants answered is
this: Did the biochemical procedure change affect
the blood urea nitrogen test results? We will as¬
sume that the two sets of 10 measurements are
random samples of normal values from the two
periods.
The laboratory director's question, "Did the bio¬
chemical procedure change affect the test results?"
cannot be answered without being more specific.
The type of change that may have occurred must
be specified. There are two things which could
have happened. First is a change in level of test
results as reflected by the averages for the two
periods. Second is a change in the spread or scatter
of the test results as reflected in their variance. The
variance is the square of the standard deviation.
The averages and variances are presented in Table
3. Neither the averages nor the variances are the
same for the two periods. Have real changes oc¬
curred, or are the differences just due to the par¬
ticular subjects who happened to be tested during
the two periods? This is a typical elementary sta¬
tistics problem. Information pertinent to the ques¬
tion is obtained by performing what are called
"tests for statistical significance." This is what
they mean.
TEXTBOOK NORMAL RANGE 8-16 MG
COMPUTED NORMAL RANGE 5-25MG
—I-1-1-1-1-1-1-1-1
25- 35- 45- 55- 65- 75- 85- 95- 105-
B.U.N. MG/IOOML
Fig 10.—Frequency distribution of 500 tests by blood urea nitrogen
value. These represent results of testing patients' specimens. A
normal range obtained from a textbook and a computed normal
range are shown.

PERCENT OF
SUBJECTS 50
WHO ARE SICK
MEASUREMENT SCALE
-UPPER LIMIT OF NORMAL RANGE
Fig 11.—A hypothetical "illness likelihood curve." Height of the
curve Indicates for each point on the measurement scale the per¬
centage of subjects who are sick.
We hypothesize that there was no change in
BUN values when the testing procedure changed.
This is called the "null hypothesis," and here is
what is done to see whether it is true. We compute
a probability that the null hypothesis is true. If the
resulting probability is very small then grave
doubts exist that the null hypothesis is true and we
proceed to reject it; otherwise we accept it. This
is something like betting on a horse race. Before
a bet is placed we would want to know what the
odds are that a given horse would win. Most peo¬
ple would not want to bet much money on a horse
where the odds were 20 or 100 to 1 against his
winning. The "horse" in our laboratory problem is
the null hypothesis, which says that no change
occurred in BUN levels with the change in testing
procedure. I do not know how odds are determined
for horse races, but in statistics they are based on
the observed measurements.
There are separate null hypothesis tests for the
averages and the variances. For details of how they
are computed, reference should be made to a book
on statistics. The F test, named after the late Sir
Ronald Fisher, will be used for the variances and
Student's t test for the averages. Results of the
computations yield the probabilities shown in
Table 4. The averages and variances are included
for easy references. The null hypothesis probabil¬
ities shown in Table 4 are about 1 in 2 for the
averages and 1 in 4 for the variances. If this were
a horse race, the odds of winning would be good
with the null hypothesis. In other words, the prob¬
abilities provide evidence that the BUN values
did not change when the change from a urease-
Nessler method to the Autoanalyzer was made. The
results apply of course only to this one laboratory.
Perhaps this is not a fair way of illustrating sta¬
tistical methods, but far more laboratory data were
collected than the two groups of 10 measurements
just described. In fact about 150 BUN values for
each of two periods were obtained. A plot of their
frequency distributions (not shown here) reveals
nothing which could be interpreted as a change in
levels. The point here is that the test for statistically
significant differences is confirmed when larger
numbers of values are available.
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
From the practicing physician's point of view,
he would be interested to know that no change oc¬
curred in BUN levels with the change to the Auto-
analyzer. If the laboratory was reporting normal
quotient values, no change at all would be ex¬
pected.
We have discussed the simplest kind of com¬
parisons among groups, because there were only
two groups. Statisticians must be prepared to com¬
pute probabilities where any number of groups
and any numbers of subjects are involved. Anyone
who is interested in pursuing the matter further
should do some reading in "The Design of Experi¬
ments" chapters in statistics books. Statistical prob¬
lems involving groups and the null hypothesis
just described are of greatest interest in medical
research rather than practice. There are, however,
many kinds of "groups" problems. The one des¬
cribed very briefly below is probably of greatest
interest to practicing physicians. Physicians would
call it a problem in diagnosis and statisticians
would call it a discrimination problem.
Discrimination Among Groups
In earlier discussions of normal limits, distribu¬
tions of values from clinical laboratories were re¬
garded as composite distributions. A composite
distribution is shown again in Fig 9. For the
moment, ignore the vertical lines enclosing areas
identified by a, b, and c. The problem now is to de¬
termine the effectiveness of whatever is being meas¬
ured, to separate the healthy from the sick subjects.
It is obvious from study of Fig 9 that the point on
the measurement scale which will misclassify the
fewest subjects, ie, the smallest number of false
positive and negative subjects, is the point where
the two distributions cross each other. This is
shown in Fig 9 as the upper limit of the normal
range. An assumption is made, however, that the
consequences of falsely misclassifying (positive or
negative) a patient are equal. If this is not true,
then a different point must be chosen as the upper
limit of the clinically normal range. It should also
be apparent that the problem is complex from
many points of view. For example, the relative
numbers of healthy and sick subjects must be con¬
sidered and their numbers will depend on local
circumstances. Perhaps the complexity of the prob¬
lem is in part responsible for the remarkably little
work done on it. To my knowledge, essentially
one paper has been written on the subject.5 Refer¬
ence is made to the simpler types of discrimination
problems. Attention has and is being given to the
more complex problems. The problems and meth¬
ods of approach to them are being described under
the general headings of "Uses of Computers in
Medicine" and "Discriminant Functions."
No matter how complex the problem, however,
it is worth noting that obtaining information about
some local problems is very easy. Consider a speci-

X =
22.6
S 3.6
=
-i-1-r i-r-1-r
14- 16- 18 24- 26- 28- 30-
NUMBER OF PLUS TESTS
Fig 1 2.—Frequency distribution of 33 "number plus" points by num¬
ber of plus tests. Each number plus point represents, out of 50 con¬
secutive routine BUN tests, the number having 15 mg/100 ml or
more of BUN.
fie example. Fig 10 shows
a distribution of blood
urea nitrogen for 2-months' period. This is a
a an outline of them is given here.
distribution of values resulting from testing pa¬
tients' specimens. In fact, some of these values
were considered earlier in connection with the
discussion of comparisons among groups. The text¬
book range of clinical normality is shown on the
figure as well as a normal range computed with the
aid of normal probability paper. Which normal
range is best, ie, misclassifies the fewest patients?
It should not be much work to review a few pa¬
tients' charts and see what happened in the areas
where the two ranges do not coincide. At least
some information would be obtained at no great
amount of effort. There is nothing either to keep
from establishing more than the two groups, normal
and sick.
The normal range could be dispensed with al¬
together and what might be called an "Illness Like¬
lihood Curve" could be computed. The idea is a
very simple one. Looking again at Fig 9 and noting
the areas of the curves identified as a, b, and c,
one can see that if a patient was reported to have
a test value in area a, he obviously would be
healthy as far as this test is concerned. Patients
with values falling in area b, however would be
mostly healthy, but some of them would be sick.
In area c, the situation has changed completely
and all patients with test values this high or higher
would be sick. All that is done is to compute the
percentage of sick patients at each point on the
measurement scale. In area a this is zero; for area b
it is about 10%; and for area c it is 100%. This
resulting curve will thus indicate an "illness likeli¬
hood" in terms of percentage; hence the name. Fig
11 shows an hypothetical illness likelihood curve.
For any point on the measurement scale the likeli¬
hood of a subject's being sick can be seen at a
glance. The point where the healthy and sick
curves (Fig 9) cross each other is the point where
a subject would have a 50% chance of
being sick.
Many things would influence a given illness
likelihood curve, the most important of which is the
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
relative proportions of healthy and sick subjects.
It would be easy, however, to keep track of these
proportions for many measurements. In the next
section, a method is described for doing this as
well as for doing other things at the same time.
The usefulness of "illness likelihood" curves re¬
mains to be tested, however, so the discussion is
continued in terms of normal limits.
Regardless of how normal limits are determined,
it is assumed that the laboratory testing procedure
is stable. This means that a blood specimen which
has 18 mg of urea nitrogen in it is always reported
from the laboratory as containing approximately 18
mg. The control of laboratory accuracy is a difficult
and ever-present problem. Statistical methods can
contribute toward a solution of this problem. A
more complete description of some methods has
been accepted for publication elsewhere,6 so only
Control of Laboratory Accuracy
If limits of clinical normality are to have mean¬
ing, the laboratory testing procedure must be
stable. On the other hand, if the laboratory testing
procedure is not stable, then the limits will shift.
This is a clue toward use of the clinical values for
the purpose of laboratory control.
Note the vertical line drawn through the BUN
distribution shown in Fig 10. Approximately 70% of
all BUN tests have values above this line, ie, are
15 mg or more. If the laboratory testing procedure
is stable, approximately 70% of the values will be
15 mg or more in consecutive groups of laboratory
tests. This is the basis of the control method.
In order to easily identify tests, the values
falling to the right of the vertical line are arbi¬
trarily called "plus tests." The vertical line is de¬
termined by the highest point on the distribution
(in statistical language called the mode). The con¬
trol procedure consists of simply counting the
number of "plus tests" in consecutive groups of 50
tests and plotting the "number plus" on a control
chart. For details reference should be made to the
paper about the method.
Physicians might be a bit suspicious of a control
procedure that is based on testing patients' speci¬
mens. How do we know whether a group of speci¬
mens might not be sent to the laboratory at one
time and all of them have high nitrogen concen¬
trations? A partial answer to this question is seen
by study of Fig 10. For example, only about one
specimen in five has value of 30 mg of BUN
a
or more, so it is very unlikely that large numbers
of high values will come in at the same time. Prob¬
ably more convincing evidence is found in Fig 12.
Fig 12 shows a frequency distribution of 33
number plus points. Each unit in the distribution
is the number of tests out of a group of 50 that
had a value of 15 mg of BUN or more. It thus
represents 1,650 consecutive BUN tests. Note that
 it suggests the familiar bell-shaped gaussian dis¬
tribution. The laboratory testing procedure as re¬
flected by this distribution was reasonably stable
over the period represented by these tests. Large
numbers of high-valued specimens are not re¬
ceived at the same time and the proportion of
healthy and sick subjects is reasonably stable.
Most practicing physicians are not directly
cerned with operating clinical laboratories, but
they are concerned about laboratory accuracy.
They want information from the laboratory about
normal values and they want assurance that the
testing procedures are stable. In other words, they
want information. Both producer and consumer of
information should agree on how the information
is produced and organized if misunderstandings
are to be avoided. This brings up a major point of
this paper: The greatest value of statistics to the
practicing physician is organized information. Sta¬
tistics offers the organizing methods, but the in¬
formation must come from medicine itself.
Summary and Comment
A few features of some statistical methods have
been described with emphasis on applications of
interest to practicing physicians. The manner in
which frequency distributions are formed, how
theoretical distributions can be used as a base for
limits of clinical normality, how measurements are
used to separate healthy from sick patients, and
how information pertaining to the control of clin¬
ical laboratory measurements is easily obtained
from the results of testing patient specimens are
all outlined. These are only a few examples of
potential uses of statistics in the practice of medi¬
cine. Other examples appear in my statistics text.7
Common to all the above are essentially three
things :
1. Statistical services must be available.
2. Information about specific problems must
be collected.
3. Communication channels for the informa¬
tion must be established; otherwise it will
be wasted.
For some problems, little time or effort is required
to gather and organize information. The clinical
laboratory should be able to provide distributions
of laboratory test values, and the medical records
department can perform a similar function using
information obtained from patients' charts. Refer¬
ence here is made to problems of local interest, for
example those affecting a single hospital staff or a
county medical society. The applications, however,
are not so limited, as will be discussed presently.
Large-scale applications, however, suggest mechan¬
ical means of handling information because of ex¬
tensive amount of work involved.
Computers
be helpful for some applications suggested here.
For example, a computer program is being de¬
vised to do the following:
Downloaded From: https://jamanetwork.com/ by a The Ohio State University Health Sciences Library User on 04/06/2019
Fed into the machine will be 500 or 1,000 consecutively
determined laboratory measurements and the date on which
each test was performed. First it will check the measure¬
ments for a shift in the laboratory testing procedure and it
will note the results of its check. Second, it will tabulate a
frequency distribution of the values. If a shift has occurred
in the testing procedure, correction will be made for the
shift before the distribution is tabulated. Third, it will sepa¬
rate the "clinically normal component" from the remainder
con¬ of the distribution. The separation will include computing
the means and standard deviations of all the resulting com¬
ponent distributions. Fourth, it will compute a "normal
quotient scale." Finally it will compute an "illness likeli¬
hood curve."
All of this will be done automatically. Similar
kinds of things could be done for other medical
measurements.
The above uses of statistics are described in
terms of local problems. There is no reason why
some problems be studied on a state or
cannot
nation-wide basis. A few details would have to be
omitted, but other information would be obtained.
For example, if distributions of growth rates of chil¬
dren were submitted from each of 1,000 pediatri¬
cians scattered over the country, it soon would
really be known how fast children are growing.
Many of the statistical methods outlined here
would be useful for this and other similar prob¬
lems. Information that is now lying fallow in rec¬
ords would be organized and put to use.
It should be clear by this time that properly used
statistical methods can provide a set of aids to the
practicing physician. These aids would be abstracts
of information organized by statistical methods
which would allow the physician in a sense to
share the experiences of hundreds or thousands of
others. Organized information is certainly not new
to medicine. A hospital chart is very carefully or¬
ganized, but it describes only a single patient. The
kinds of things discussed here involve only a single
aspect of many patients. Each kind of information
has its own uses and limitations. Each should com¬
plement the other.
The values in Fig 5 were furnished by John B. Henry, MD, Director
o£ the Clinical Laboratory, J. Hillis Miller Health Center. An Auto-
analyzer was used to test the specimens, which represent routine clini¬
cal work.
References
1. Cecil, R.L., and Loeb, R.F.: A Textbook of Medicine,
10th Ed, Philadelphia: W. B. Saunders Co., 1959.
2. Lozner, E.L., et al: Intravenous Glucose Tolerance
Tests, J Clin Invest 20:507, 1941.
3. Hald, A.: Statistical Theory with Engineering Applica-
tions, New York: John Wiley and Sons, Inc., 1952.
4. Sunderman, F.W., and Boerner, F.: Normal Values in
Clinical Medicine, Philadelphia: W. B. Saunders Co., 1950.
5. Dunn, J.E., Jr., and Greenhouse, S.W.: Cancer Diag-
nostic Tests, Principles and Criteria for Development and
Evaluation, Federal Security Agency, Public Health Service,
National Cancer Institute, Publication No. 9, Washington,
D.C.: US Government Printing Office, 1950.
6. Hoffmann, R.G., and Waid, M.E.: The Number Plus
can
Method of Quality Control of Laboratory Accuracy, Amer
J Clin Path, to be published.
7. Hoffmann, R.G.: Statistics for Medical Students,
Springfield, Ill.: Charles C Thomas, Publisher, 1963.