SALARY SURVEYS

Compensation
Measures of Central
Tendency,Location,and
Dispersion in Salary
Survey Research
Robert M. Halley
Compensation Consultant Proper application of
Swanson Consulting
salary data requires a
clear understanding of
the methods and
measures used.

t
o obtain the maximum utility from on the method of data selection. In a probability
salary survey results, it is essential sample, the data elements are selected in a ran-
to know the mathematical attrib- dom fashion from a sample population such that
utes and proper applications of each item has an equal chance of inclusion and
the various statistical measures each sample of a particular size also has an equal
used to summarize survey find- chance of selection. For this reason, a probability
ings. This article outlines the sample permits generalizations about the larger
use and interpretation of the de- population universe from which it is drawn. In
scriptive statistics commonly addition, the accuracy of an estimate from a
found in commercial salary probability sample may be specified within de-
survey reports and suggests some finable limits. Finally, a probability sample per-
general guidelines for their mits the use of both descriptive1 and inferential
implementation. statistics.2 An example of a probability sample
is the Washington State Legislature Salary Survey
(see Exhibit 1).
Probability and In a nonprobability sample, the data elements
Nonprobability Sampling are selected in a nonrandom or biased manner.
Sample surveys may be classified as either Factors other than chance such as subjective
probability or nonprobability samples depending judgment, self-selection, or availability adversely

Keywords: statistics; salary survey; compensation; measures of central tendency; location and dispersion;
salary research
DOI: 10.1177/0886368704268598

© 2004 Sage Publications

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
SEPTEMBER/OCTOBER 2004 39
 SALARY SURVEYS
Compensation

EXHIBIT 1
Survey Sampling Techniques

Feature Probability Sample Nonprobability

Selection method Samples selected randomly, namely, every No random or chance selection procedure
element in the sample population has an applied, for example, self-selection, judgment
equal chance of inclusion or availability, and so on
Major types Simple random sample, stratified random Convenience sample (most salary surveys),
sample, cluster sample, systematic sample judgment sample, quota sample
Utility of results Permits statistical inference, namely, Does not permit statistical inference, results
accurate conclusions about a population may not be generalized beyond the actual
based on information gathered from a sample group, and valid management
sample judgments can be made if the sample is
representative

affect the probability of inclusion of survey ele- nized in a frequency distribution that arrays pay
ments. This means that some sample units have a rates (by category) in terms of their frequency of
high probability of selection, whereas others have occurrence. A frequency distribution is described
virtually no chance of inclusion (with no way of by the following three principle features: (a) its
determining the probability of either their inclu- shape or form, (b) its cluster around the central
sion or exclusion). For this reason, the results of a value, and (c) its degree of variability.
nonprobability sample may not be generalized Descriptive statistics are employed to describe
beyond the actual sample group to a larger popu- various aspects of frequency distributions. Data
lation universe. The precision of a nonprobability are condensed and summarized via descriptive
sample estimate cannot be determined. Al- statistics to better understand their meaning and
though a nonprobability sample allows the use of to prepare them for further analysis. The intent of
descriptive statistics to summarize the important descriptive measures is to represent an entire
features of a sample distribution, it does not per- data set with a single number. Because the result-
mit the use of inferential statistical analysis (see ing data reduction causes a loss of information,
Exhibit 1). several different types of statistics may be neces-
The majority of commercial salary surveys are sary to fully represent the various features of a
nonprobability convenience surveys. Data ele- distribution of salary rates. Once pay distribu-
ments are included because of the “voluntary re- tions have been fully described, various compar-
sponse” of survey subscribers who are willing and isons can be made between them.
able to provide input. Even if participation in a
convenience survey is extensive, its results may Measures of Central Tendency
be generalized only to the immediate sample it- Measures of central tendency are used to sum-
self and no farther. However, if a convenience marize (with one or a few statistics) a distribution
sample is representative, it is just as useful as a of pay rates in terms of its center. They represent
random sample for making management deci- the pay rate that is typical, expected, or represen-
sions, although statistical inferences may not be tative in a set of salary data. Because the center of
made from the data. The Management and Pro- a distribution may be defined in several ways
fessional Salary Survey published by Milliman (e.g., balance point, geographical center, or high-
USA is an example of a convenience survey (see est point), there are several different measures of
Exhibit 1). central tendency, including the mean (simple,
weighted, or trimmed), median, and the mode.
There is no one right way to measure central
Measures of Central Tendency tendency. Depending on the nature of a set of
and Dispersion data (e.g., its quality, quantity, distribution, dis-
The results of salary surveys are reduced and persion, and representativeness), certain mea-
summarized by various measures of central ten- sures of central tendency (and location) are more
dency, location, and/or dispersion. Data ob- suitable than others. The task of the compensa-
tained from salary surveys are commonly orga- tion professional is to evaluate distributions of

40 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS

Compensation
EXHIBIT 2
Measures of Central Tendency

Representation of Amount of Affected by

Statistical Measure Distribution Information Use Extreme Values
Simple mean (i.e.,
unweighted mean) The balance point or fulcrum of All of it Yes
the distribution; average score
deviations from the mean sum
of zero
Weighted mean The balance point or fulcrum of All of it Yes
the distribution
Trimmed mean The balance point or fulcrum of Nearly all data in Very little
the distribution distribution once
the extreme cases
are removed
Median The midpoint of the distribution Very little No
Mode The highest point in a distribution; It depends; based on Only if the mode is an
most common score or scores most frequent score extreme value
or scores

EXHIBIT 3
Measures of Dispersion

Statistical Measure Amount of Information Used Affected by Extreme Values

Range Just extreme values Yes
Interquartile range Middle half of the data No
Standard deviation All of it Yes

pay and to choose the best measure of central rate represents a set of salary data. These mea-
tendency to describe a set of pay rates. However, sures help determine the goodness, quality, or re-
once a measure is chosen, it should be consis- liability of generalizations made from the pay
tently maintained throughout and should not be data. The greater the dispersion of the pay data,
deviated from without a good reason and without the less reliable to average salary reported (see
noting it (see Exhibit 2). Exhibit 3).
Among the major measures of central tenden-
cy, the mean is the most stable, followed by the
median, and lastly the mode. Simple Mean
Definition. “The simple mean is the sum of
Measures of Dispersion scores in a distribution divided by the number of
A salary rate assumes meaning only when it is scores.”3
contrasted with other pay rates or other statistics.
Calculation. The simple mean is calculated by
To more fully describe a distribution of pay rates
totaling the average salary paid in each organiza-
or to more accurately interpret a pay rate, addi-
tion and then dividing by the total number of or-
tional information is required concerning the
ganizations (see Exhibit 4).
dispersion of salary rates about the measure of
central tendency. Mathematical characteristics. The simple (or
Measures of dispersion indicate how scattered arithmetic) mean “gives equal weight to the
or spread out the salary data are around a mea- salary paid by each organization regardless of the
sure of central tendency. The purpose of this type number of incumbents.”4 In other words, salary
of measure is to assess how well an average pay data from a company with a single incumbent

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

SEPTEMBER/OCTOBER 2004 41
 SALARY SURVEYS
Compensation

act value of every pay rate in a data set. Thus, a

EXHIBIT 4
change in any pay rate will alter the value of the
Calculating the Simple Mean mean.
Company A $22,000 The mean permits an entire distribution of pay
Company B $26,000 rates to be characterized with a single score and
Company C $30,000 facilitates the meaningful interpretation and un-
Company D $20,000 derstanding of group differences.
Company E $30,000 The simple mean is particularly sensitive to
5 $128,000 extreme pay rates when they are not distributed
n=5 evenly on either side of it. Unlike the median or
mode, neither of which is impacted by the ex-
Simple Mean + $128,000 = $25,600 treme pay rates, the simple mean is pulled in
5
the direction of the skew (extreme rate) (see Ex-
hibit 6).
When the simple mean is greater in value than
the median, the distribution is said to be posi-
EXHIBIT 5 tively skewed, and when the simple mean is less
The Mean as the Balance than the median, it is said to be negatively
Point of a Distribution
skewed.
Scores about the Mean If a distribution of pay rates is markedly
(5) skewed,5 the simple mean will provide an erro-
(4)—-(5)—-(6) neous estimate of central tendency because it is
(2)—-(3)—-(4)—-(5)—-(6)—-(7)—-(8) not a central score in this context. Schoeninger
Mean Score = 5 and Insko reported that “when a distribution is
| markedly skewed the mean (income) is likely to
Score Deviations about the Mean be greater than that obtained by 70 to 80 percent
(0) of the persons in the community.”6
(–1)—-(0)—-(+1) When a distribution of pay rates is mildly
(–3)—-(–2)—-(–1)—-(0)—-(+1)—-(+2)—-(+3) skewed, either the mean or median is an appro-
| priate measure of central tendency. If a distribu-
Mean Deviation = 0
tion of pay rates is symmetrical (normally distrib-
(Deviations from the mean sum to zero) uted), the mean, the median, and the mode will
be identical. Under such conditions however, the
mean should be employed because it contains
have the same influence in determining the sim- more information than either the median or
ple mean as salary data from a company with 20 mode and thus is more reliable (see Exhibit 7).
incumbents in the same job. For this reason, the The simple mean is also a more stable and re-
simple mean is sometimes referred as the “com- liable measure of central tendency than either
pany mean.” the median or mode because it is least subject to
In a distribution of pay rates, the simple mean sampling variation. Furthermore, because the
represents the balance point or center of gravity simple mean varies less from sample to sample, it
because the sum of the deviations of all pay rates normally provides a closer estimate of the popu-
from the mean pay rate is zero. lation mean than the sample median or mode to
Note: The concept of the mean can be illus- the population median or mode. It is for this rea-
trated by the use of a weight and scale analogy. If son that the simple mean is used in inferential
each pay rate in a set of rates were treated as a statistics whenever possible (see Exhibit 8).
separate weight (based on its numerical value) The simple mean is the most versatile and
and arrayed in rank order along a measurement commonly used measure of central tendency.
scale continuum, the fulcrum (or balance point) Due to its mathematical properties, the simple
of that distribution would constitute the simple mean plays an important part in several more ad-
mean (see Exhibit 5). vanced statistical operations such as the standard
Next to its statistical cousin the weighted deviation, variance, correlation, and regression.
mean, the simple mean is the most sensitive in- In pay distributions, the simple mean and the
dex of central tendency because it reflects the ex- median are seldom equal because salaries tend to

42 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS

Compensation
EXHIBIT 6 EXHIBIT 7
The Mean Is Drawn in the The Normal Distribution
Direction of Extreme Pay Rates

Frequency
Frequency

Mode Mean

Pay Rate
Median Mean }
Median} All Equal
The mean is drawn in the direction of skewedness (or toward the tail). Mode }

amount rapidly and spread upward toward the by extreme pay rates, and (e) plays an important
top of the range due to the impact of pay-for- role in higher statistical applications such as
performance and seniority systems as well as standard deviation, variance, correlation, and re-
the performance of additional (yet unidentified) gression as well as in inferential statistics.
duties. The principle difference between the simple
Research has shown that the mean of a distri- mean and the weighted mean is that the latter as-
bution of pay rates may exceed the median (or signs equal weight to the pay rate of every em-
midpoint of a distribution) by 3% to 5% because ployee (rather than each employer) represented
of the presence of very high rates of pay. Richard in the survey sample. For this reason, the weight-
Henderson warns, “If an organization . . . uses the ed mean could be thought of as the “employee
mean as the value setting pay practices it may be mean” (see Exhibit 10).
overpaying by 3 to 4 percent.”7 Because the pay data of each company are
Finally, the simple mean is less volatile than weighted by the number of incumbents in the
the weighted mean and thus better suited for job, the data from companies with many incum-
making year-to-year comparisons. bents will have greater influence on the weighted
mean than the data from employers with few
incumbents.
Weighted Mean The weighted mean is the most sensitive mea-
Definition. “The weighted mean is the sum of sure of central tendency to extreme pay rates in a
the mean of each group multiplied by its respec- data set. For instance, when the mean pay rate
tive weight (the n in each group), divided by the with the greatest weight (number of cases) is sit-
sum of the weights (total n).”8 uated on either the high or low side of a distribu-
tion of means, then it will pull the weighted mean
Calculation. The weighted mean is calculated in that direction and away from the simple (un-
by totaling the pay rate of each employee and weighted) mean.
then dividing by the total number of incumbents When the survey sample participants are high-
or by weighting the mean pay rate of each com- ly representative of the market for a job, the
pany surveyed by the number of pay rates that weighted mean (which reflects the pay rate of
went into its calculation in the first place (see Ex- each job incumbent) will represent the market
hibit 9). value of the position. In effect, the number of in-
Mathematical characteristics. The weighted cumbents represents the number of potential
mean possesses most of the same properties as openings for that job.
the simple mean, including that it (a) serves as Note: Because they share the same mathemat-
the balance point or center of gravity in a data set, ical properties, the simple mean and weighted
(b) reflects all the pay rates in a distribution of mean may be combined to reflect the influence
salaries, (c) is impacted by the exact numerical of both employers and employees on the market
value of each pay rate in a data set, (d) is affected value of a job.

SEPTEMBER/OCTOBER 2004 43
Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS
Compensation

EXHIBIT 8
Applications of the Simple Mean

Best Applications Poor Applications and Cautions

When the intent is to reflect every bit of information in When the distribution of pay rates is markedly skewed (in
a set of pay rates which case the mean is drawn in the direction of the
skew and does not provide a central pay rate)
When the distribution of pay rates is symmetrical in When the differential between the mean and median is
shape rather than skewed greater than 5%
When the sample size is small and where the dispersion When the distribution of pay rates is bimodal (in which
of rates of pay is within rational limits case the two modes are often reported instead of the
mean because they indicate where the bulk of the
distribution is located)
When computing year-to-year change amounts
When “uncertainty exists as to what companies or
employee data are represented in the sample”19
When “the sample is incomplete and not fully
representative of the industry (e.g., one or more
major comparator companies declined participation)”20
When the objective is to determine the common pay
practice among companies
When the intent is to use the data in higher statistical
applications (like regression or correlation) and/or
with inferential statistics

Mathematical characteristics. This measure of

EXHIBIT 9
central tendency excludes the extreme pay rates
Calculating the Weighted Mean in a distribution while utilizing the preponder-
Company A 10 x $15,000 = $150,000 ance of the remaining salary data. The trimmed
Company B 20 x $12,000 = $240,000 mean provides a balance or compromise be-
Company C 15 x $14,000 = $210,000 tween the median and the mean. The trimmed
3 45 x $41,000 = $600,000 mean can also take the form of either the simple
mean or the weighted mean.
Weight Mean = $600,000 = $13,333
Common versions of the trimmed mean
45
include
Simple Mean = $41,000 = $13,667
3 a. excluding the highest and the lowest pay
rates and computing the mean of the remaining
pay data,
Trimmed Mean
b. eliminating the top 5% and the bottom 5%
Definition. “The trimmed mean of a set of data of pay rates and computing the mean of the mid-
is the mean of the middle part of the data, after dle 90% of the pay data, and
trimming off the extreme pay rates.”9
c. removing the top 10% and the bottom 10%
Calculation. The trimmed mean is calculated of pay rates and computing the mean of the mid-
by taking the same number of pay rates from the dle 80% of the pay data.
top and bottom of a set of salaries and calculating
the simple mean. The simple mean is calculated Note: Trimming more than 20% of the pay
by totaling the average salary paid in each organi- rates is generally not advised. You might as well
zation and then dividing by the total number of just use the median salary rate, which trims all
organizations (see Exhibit 11). but one or two pay rates (see Exhibit 12).

44 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS

Compensation
EXHIBIT 10
Applications of the Weighted Mean

Best Applications Poor Applications and Cautions

When the intent is to utilize every bit of the information When there is uncertainty about what companies and
employee data are contained in the sample
When the pay distribution is symmetrical (normal) in When the distribution is markedly skewed (in which case
shape rather than skewed (see Exhibit 6) the weighted mean is drawn in the direction of the skew
and does not provide a central pay rate)
When there is confidence that the survey sample When the distribution contains only a small number of
represents an organization’s major competitors and cases
by extension the applicant pool from which it draws
employees
When employers responding to the survey “represent When the differential between the mean and median
at least 20% of the total non-agricultural work force pay rates is greater than 5%
in the labor market”21
When a job is “common to only a few employers or When the distribution of pay rates is bimodal (in which
to only one category of employer (and) in your case the two modes are often reported instead of the
judgment, you are including a sufficient number mean because they indicate where the bulk of the
of these employers to represent at least a quarter of distribution is located)
the total employment in that job”22

Median
EXHIBIT 11
Calculating the Trimmed Mean Definition. “The median is the value of the
middle item when the data are arranged in as-
$40,000 cending or descending order.”10
$39,000
$39,000 Calculation. The median is calculated by first
$39,000 ranking a set of pay rates. If there are an odd
$37,000
number of rates, then the median is the middle
$36,000
$34,000 pay rate (see Exhibit 13).
$30,000
Mathematical characteristics. The median is a
$28,000
$26,000 measure of location that indicates the geographi-
$25,000 cal center of a distribution of pay rates. It is the
“middle data point in an ordered array of data.”11
Simple Mean = $33,909
The median cleaves a data set in two so that half
N = 11 of the pay rates are less than (or equal to) its val-
Top and bottom pay rates trimmed ue and half the pay rates are greater than (or
equal to) its value. It is the value at the 50th per-
($40,000)
$39,000 centile, the fifth decile, and the second quartile
$39,000 (see Exhibit 14).
$39,000 Although the mean takes into account the ac-
$37,000 tual value of each pay rate in its calculation, the
$36,000
median (once pay rates have been rank ordered)
$34,000
$30,000 treats each pay rate equally regardless of its actu-
$28,000 al value. The value of the median depends on the
$26,000 order among pay rates rather than their numeri-
($25,000) cal value.
Trimmed Mean = $34,222 The computation of the median involves a
simple count of pay rates with each rate counting
N=9
as only one case (irrespective of its value). When

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

SEPTEMBER/OCTOBER 2004 45
 SALARY SURVEYS
Compensation

EXHIBIT 12
Applications of the Trimmed Mean

Best Applications Poor Applications and Cautions

When there is no rational explanation for extreme When business competitors are located at either extreme
pay rates of the distribution of salary rates
When attempting to use most of the pay information When the distribution of pay rates is bimodal (in which
while simultaneously eliminating the impact of just case the two modes are often reported instead of the
a few of the extreme pay rates (outliers) mean because they indicate where the bulk of the
distribution is located) (see Exhibit 17)
When a compromise between the mean and the
median pay rates is sought

EXHIBIT 13 EXHIBIT 14
Calculating the Median The Relation of the Median to the Mean

Median = (n + 1)
2
Raw Data – Ordered Data
Frequency

32 39 (7)
37 39 (6)
50% 50%
38 38 (5) of of
36 37 (4) Pay Pay
Rates Rates
39 37 (3)
39 36 (2)
Rates
37 32 (1)
MEDIAN Pay Rate
n=7
If balance Lever falls to
Median = 7 + 1 or 4th pay rate point here. right.
2
If balance Lever
when the salary data are order. point here. balances.
The median is 37.
MEAN
When there is an even number of salary rates, the
median is the middle of the pay rates.
Less than
Raw Data – Ordered Data
More than
50% of the pay { } 50% of the
rates. pay rates.
13 14 (8)
10 13 (7)
14 13 (6)
7 12 (5) there is an odd number of pay rates, it is simply
8 10 (4) the middle rate. When there is an even number of
7 8 (3)
12 7 (2) pay rates, it is the simple average of two center-
13 7 (1) most rates. When more than two scores share the
middle value, calculation of the median becomes
n=8
a somewhat more complicated proposition12 (see
Median = 8 + 1 or 4.5th pay rate footnote 16).
The median is easier to obtain than the mean.
2
In fact, the median can even be computed when
when rates are ordered. The median is 11 some of the pay rates are missing from the top or
(10 + 12) = 11 bottom of a distribution.
The median provides the “typical” pay rate in a
2
distribution of pay rates. It indicates the pay rate

46 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS

Compensation
EXHIBIT 15
Applications of the Median

Best Applications Poor Applications and Cautions

When the objective is purely descriptive and there is For use in regression analysis because that statistic
no intention to use this measure of central tendency assumes a normal distribution
in subsequent statistic analysis
When there are inexact data at the extremes of the For use with inferential statistics
distribution
When the complete elimination of the impact of When the distribution of pay rates is bimodal (in which
extreme pay rates is desired and the location of the case the two modes are often reported instead of the
middle pay rate is sought median because they indicate where the bulk of the
distribution is located) (see Exhibit 17)
When the distribution is markedly skewed, for example,
the differential between the mean and the median is
greater than 5%
When confidence in the sampling and data analysis
is lacking and knowledge of the “typical” pay rate for
a job is sought
When the pay rate of the middle person in a range earns
in other companies for comparable work is desired

of the employee in the middle of a distribution of

EXHIBIT 16
pay rates.
Calculating the Mode
Because the median uses only the value(s) in
the center of a distribution of pay rates, it is in- $30,000
sensitive to extreme pay rates. Research has ($29,000)
shown that the median is typically 3% to 4% less ($29,000)
than the mean. For this reason, it is ideal for use ($29,000)
$27,000
in describing data sets that are highly skewed (see
$26,000
Exhibit 15). $24,000
Unlike its statistical relative the mean, the me- $24,000
dian cannot generate an overall median for com- $20,000
bined data. For example, if the middle value of $18,000
$16,000
several categories of data were known, the overall
$15,000
middle value could not be generated by combin-
ing them. In addition, the median is not suitable The mode is $29,000
for use in higher statistical applications such as
the correlation and regression or inferential sta-
tistics. Finally, the median should not be com-
of pay data. When displayed graphically, the
bined with the mean to form a new measure of
mode constitutes the highest point (or points) in
central tendency.
a frequency distribution of pay data.
The mode is the simplest and the easiest mea-
sure of central tendency to obtain. Once pay data
Mode has been grouped, it is found by inspection of a
Definition. “The mode is defined as the most data set rather than calculation.
frequent score in a distribution.”13 A distribution of pay rates may have more than
one mode. When a distribution has two or more
Calculation. Once the pay rates are grouped,
pay rates with the highest (yet equal) values, the
the mode is obtained by inspection of the data set
distribution is classified as bimodal, trimodal, or
(see Exhibit 16).
multimodal.
Mathematical characteristics. The mode iden- In a bimodal distribution, the mode differs
tifies the most common pay rate (or rates) in a set greatly from both the mean and the median. In

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

SEPTEMBER/OCTOBER 2004 47
 SALARY SURVEYS
Compensation

may constitute the mode of a pay distribution

EXHIBIT 17
without constituting a majority of the rates.
Bimodal Distribution Ironically, because the mode is the most prob-
able pay rate in a distribution, if the modal pay
rate were used to predict each pay rate in a distri-
Frequency

bution of pay rates it would make fewer errors

than any other rate. Whereas the mean and medi-
Pay an may generate values that do not reflect any
Rate real pay rates, the mode represents an actual pay
rate.
Mode 1 Mean Mode 2 Pay distributions with identical modes (modal
Median
pay rates) can possess radically different forms,
whereas pay distributions can have similar forms
this circumstance, the mean and the median do a yet radically different modes.
better job of reporting the middle of the distribu- In a normal (or symmetrical) distribution with
tion, whereas the modes indicate where the bulk a single mode at the center, the mean median
of the pay lies. If a distribution (of pay rates) is bi- and mode will all be identical (see Exhibit 7).
modal, the location of the two modes is typically However, because the mode typically uses only a
reported. When all pay rates in a data set have the small portion of the pay data in a distribution, it
same frequency, the distribution is said to have is rarely chosen as the sole descriptive measure to
no mode (see Exhibit 17). represent an entire data set where the mean or
The mode is seldom used as a measure of cen- median are available. Finally, the mode does not
tral tendency with pay data (even when a distrib- enter into any of the higher mathematical rela-
ution is markedly skewed) unless the peak of the tionships that are integral to inferential statistics
frequency distribution is extremely high. (as does the mean).
The mode is most useful when there is a large
number of pay rates and the data have been
grouped. However, when there are only a small Percentile
number of pay rates or the data are ungrouped,
Definition. “A percentile is the pay rate that a
the measure is not very informative because the
certain percentage of the salary data is less than.
likelihood of each pay rate being different from
(Note: the mathematically rigorous definition for
every other pay rate is greater (see Exhibit 18).
percentile is that it is the pay rate which a given
“The mode is not necessarily a measure of
percentage of salary data is less than or equal
central tendency.”14 Because the mode is unaf-
to).”15
fected by pay rates outside the modal interval,
there are situations where the mode is anything Calculation. The percentile is calculated by (a)
but the typical pay rate. The mode provides only sorting the pay data in order, (b) numbering them
a rough indication of location because it ignores (from low to high), and (c) numbering all repeat
a considerable portion of a data set. A pay rate values. The percentile is derived from the formu-

EXHIBIT 18
Applications of the Mode

Best Applications Poor Applications and Cautions

When there are a large number of pay rates and the data The mode is not very informative in small groups of pay
have been grouped data; in groups of 10 pay rates or less it may occur that
no two rates are alike
When the peak of the distribution of pay rates is
unusually high
When conducting trend pattern analysis (e.g., planned
pay increases) in common practice
When a quick, rough estimate of central tendency
is sought

48 Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

 SALARY SURVEYS

Compensation
EXHIBIT 19 EXHIBIT 20
Calculating the Percentile Applications of the Percentile

Raw Data Rank Order Best Poor Applications

Applications and Cautions
451 9
450 8 When a comparison of When the sample size
422 7 a firm’s pay philosophy is small
417 6 or practice against the
414 5 market for a job is
410 4 desired (e.g., some
397 3 companies target base
360 2 pay to the 60th)
332 1
When a picture of the
n=9 dispersion of a set of
pay data is sought
Percentile = p(n + 1) (e.g., the interquartile
70th percentile = (.70)(9 + 1) range)

= (.70)(10)
= 7th data point
= 422 EXHIBIT 21
Calculating the Range

la p(n + 1), where p = the percentage of the per- For example, if the highest pay rate in a distribution of
centile and n = the number of pay rates. If the p(n salaries is $68,000 and the lowest pay rate $56,000,
+ 1) observations fall between two pay rates, then then the range would be $12,000 ($68,000 – $56,000 =
calculate a value proportionally between them $12,000)
using linear interpolation. Note: The p(n +1) for-
mula presented is the most widely used. Not all
texts use (n + 1). This can make a difference with formed into percentile ranks. For the percentile
small samples (see Exhibit 19). rank of a pay rate to be meaningful, it must also
be made in relation to some reference group (e.g.,
Mathematical characteristics. The percentile is
the market or major competitors).
a measure of position that identifies the location
Percentiles permit comparisons only at the or-
of a pay rate in a distribution of rates. The per-
dinal level. This means that such comparisons in-
centile specifies the percentage of rates that are
dicate only that one pay rate is larger or smaller
less than (or equal to) a given pay rate in a distri-
than another rate but does not specify by how
bution (e.g., the 70th percentile is that point at
much. An appropriate analogy for ordinal-level
which 70% of the pay data fall under when rated
comparisons is a horse race (where the horses
on size). It also indicates the position of pay rates
finish in first, second, and third place).
that are larger than a proportion of rates.
For a percentile to make sense, there must be a
The percentile is not necessarily a measure of
minimum number of two data points lying out-
central tendency or typicality but rather more like
side the percentile. 16
the median. In fact, the 50th percentile of a distri-
bution is the median. However, unlike the medi-
an, the percentile may reflect all areas of a distri-
Range
bution (of pay rates), not merely in its middle
section. Definition. “The range is the (actual numeri-
The percentile is often used to compare an or- cal) difference between the largest and smallest
ganization’s pay philosophy against the market. pay rates in a distribution.”17
For example, an organization might target its
Calculation. To calculate the range, you sub-
base pay to the 60th percentile of the market rate
tract the lowest pay rate in a distribution from the
(see Exhibit 20).
highest pay rate (see Exhibit 21).
A pay rate has little meaning or utility in isola-
tion. To provide a comprehensible basis for inter- Mathematic characteristics. The range is a
preting and comparing pay rates, they are trans- measure of dispersion that indicates the spread

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

SEPTEMBER/OCTOBER 2004 49
 SALARY SURVEYS
Compensation

EXHIBIT 22
Applications of the Range

Best Applications Poor Applications and Cautions

When a quick and rough indication of pay dispersion When there is a very large sample (because there is a
is sought greater probability of including extreme pay rates and
thus generating an inordinately wide range)
When the complete span of pay data reported is desired Whenever it is feasible to use the interquartile range
When small data sets of 10 pay rates or less are When knowledge of the variability of the bulk of the
employed between the extreme scores is sought

or variability of a distribution of pay rates. It is the

EXHIBIT 23
simplest, crudest, and most straightforward in-
Calculating the Interquartile Range
dex of market pay variability.
The utility of the range is limited because it re- 651 11
flects only the two most extreme pay rates in a 498 10
distribution (which are themselves often rare or 475 9
unusual happenstance). Furthermore, the range 451 8
430 7
ignores the remainder of the pay data in distribu-
415 6
tion and reveals nothing about the variability of 407 5
pay rates between the extremes. For example, 380 4
when a distribution of pay rates has an extremely 346 3
small or extreme large pay rate, the range will be 310 2
295 1
wide even though the bulk of pay data evidence
only moderate dispersion (see Exhibit 22). Percentile = p(n + 1)
The range for large samples is wider than for n = 11
small samples because the former has a greater 75th percentile = (.75)(11 + 1)
probability (i.e., a larger number of cases) of in- = (.75)(12)
cluding extreme values than the latter. In addi- = 9th data point
= 475
tion, the range is an unstable measure of disper-
sion that varies considerably from one sample to 25th percentile = (.25)(11 + 1)
the next. Such instability is particularly true of = (.25)(12)
= 3rd data point
small samples. To overcome the limitations of the
= 346
range the interquartile range is often used.
Interquartile range = 75th percentile – 25th
percentile
Interquartile Range (IQR) 475 – 346 = 129

Definition. “The inter-quartile range is the dif-

ference between the 75th and the 25th per-
centiles. It is the range of the middle 50% of the The IQR reports pay data in the middle portion
salary data.”18 of a distribution situated between the 25th and
75th percentiles. The width of the interquartile
Calculation. The IQR is obtained by first calcu-
range indicates the degree of dispersion among
lating the salary rates at the 75th and 25th per-
pay rates (i.e., how tightly or loosely clustered pay
centiles and then subtracting the pay rate at the
rates are around one another).
25th from the 75th percentile pay rate (see Exhib-
The IQR provides a more meaningful measure
it 23).
of pay variability than the range because it is un-
Mathematical characteristics. The IQR is a affected by extreme pay rates at the ends of the
measure of dispersion that indicates the degree distribution. When the interquartile range is ex-
of market pay variability (see Exhibit 24). tremely wide, it raises doubts about the quality of

50 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015
 SALARY SURVEYS

Compensation
EXHIBIT 24
Applications of the Interquartile Range

Best Applications Poor Applications and Cautions

When overcoming the instability of the range as a When the sample size is small
measure of dispersion is desired
When a distribution is markedly skewed When the precise interpretation of a pay rate within a
distribution is sought
When an alternative method of showing market pay When the desire is to take advantage of all the pay data
variability is desired
When a comparison of the variability of the middle 50% When the intention is to enter into any of the higher
of the data among several distributions are required mathematical relations that are integral to inferential
statistics

survey matches (i.e., whether survey participants in the behavioral sciences. St. Paul, MN: West
were matching the same job). Publishing Company. (p. 60). Elifson, K. W.
The quartiles (Q1, Q3) in conjunction with the (1982). Fundamentals of social statistics.
median provide an indication of the center, Reading, MA: Addison-Wesley. (p. 95).
spread, and shape of a distribution of pay rates. 4. Daniels, L. (Ed.). (1994). Puget Sound regional
When either Q1 or Q3 is farther away from the salary survey summary report. Seattle, WA:
median than the other quartile, the distribution Milliman & Robertson, Inc. (p. 18). Clark, L. J.
is said to be skewed in that direction. (1993). Business statistics I. Piscataway, NJ:
The quartiles typically provide a more stable Research & Education Association. (p. 20).
measure of pay rate variation than the range be- 5. Rees, D. G. (1989). Essential statistics. Lon-
cause (being less impacted by extreme rates) they don: Chapman & Hall. (pp. 27, 42). Note: To
vary less from sample to sample. The quartiles are determine the skewness of a distribution of
also a more informative measure of dispersion pay rates, use the following equation. Marked
with markedly skewed data than the range be- skewness means that the measure of skew-
cause they portray the unequal spread of pay ness is greater than +1 or less than –1, a rough
data in a distribution. guide.
The IQR, similar to the median, does not enter
into any of the higher mathematical relationships (Sample mean – sample median)
that are basic to inferential statistics. Sample standard deviation

Moore, D. S., & McCabe, G. P. (1993). Intro-

Notes duction to the practice of statistics. New York:
1. Vogt, W. P. (1999). Dictionary of statistics and W. H. Freeman. (p. 34). Note: To identify sus-
methodology. Thousand Oaks, CA: Sage. pected outliers, single out values falling at
(“Descriptive statistics: Procedures for sum- least 1.5 x interquartile range below the first
marizing, organizing, graphing, and, in gen- quartile or above the third quartile.
eral, describing quantitative information,” 6. Schoeninger, D. W., & Insko, C. A. (1971). In-
p. 79). troductory statistics for the behavioral sci-
2. Vogt (1999). (“Inferential statistics: Statistics ences. Boston: Allyn & Bacon. (p. 35).
that allow one to draw conclusions or infer- 7. Henderson, R. I. (1989). Compensation man-
ences from data. Usually this means coming agement: Rewarding performance (5th ed.).
to conclusions [such as estimates, general- Englewood Cliffs, NJ: Prentice Hall. (p. 262).
izations, decisions, or predictions] about a 8. Elifson (1982, p. 107).
population on the basis of data describing a 9. Davis, J. H., Gluckman, D. M., & Beatty, J. R.
sample,” p. 149). (1990). Quantitative methods. Scottsdale, AZ:
3. Williams, F. (1968). Reasoning with statistics. American Compensation Association. (p.
New York: Holt, Rinehart & Winston. (p. 32). R2.27).
Pagano, R. R. (1990). Understanding statistics 10. Davis et al. (1990, p. R2.26).

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015

SEPTEMBER/OCTOBER 2004 51
 SALARY SURVEYS
Compensation

11. Weiss, N. (1989). Elementary statistics. Read- used. Compare p and (1 – p) with the value of
ing, MA: Addison-Wesley. (p. 59). n/(n + 1). If p or (1 – p), is larger than n/(n +
12. Schmidt, M. J. (1974). Understanding and us- 1), the formula will not work, as it will gener-
ing statistics: Basic concepts. Lexington, MA: ate a value outside the range of the data.”
D. C. Health. (“When more than two scores (Davis et al., 1990, p. R2.33)
share the middle value, and there are rela-
tively few observations involved, computing Number of Cases Needed to
Ensure Two Data Points
the median becomes a slightly more compli- Outside of the Percentile(s) Percentile(s)
cated matter. Consider the distribution: 2, 3,
4 50th
3, 4, 7, 7, 7, 7, 7, 8, 9. The middle score is 7, but
four other scores also share the value. In such 8 25th, 75th
cases, one has to determine where the medi- 20 10th, 90th
an lies between the upper and lower true lim-
its of the value 7, that is, where exactly the 17. Herzberg, P. A. (1983). Principles of statistics.
median falls between 6.5 and 7.5. Here, the New York: John Wiley. (p. 60).
middle score stands second among five 18. Davis et al. (1990, p. R2.34).
scores that are equal. The median is this two- 19. Bjorndal, J. A., & Ison, L. K. (1991). Mastering
fifths the difference between 6.5 and 7.5 Market Data. Scottsdale, Arizona: American
above the lower true limit of 6.5. The median Compensation Association. (p. 8).
is 6.5 + 0.4 = 6.9,” p. 78.) 20. McMahon, J. R., & Hand, J. S. (1991). Measur-
13. Pagano (1990, p. 67). ing the Marketplace. Scottsdale, Arizona:
14. Davis et al. (1990, p. R2.28). American Compensation Association. (pp.
15. Davis et al. (1990, p. R2.31). 14-15).
16. “In order to determine whether there are a 21. Lutz, Carl F. How to develop, conduct, and use
sufficient number of cases to compute a sta- a pay/fringe benefit survey. Crete, Illinois: Ab-
ble percentile for very small n, very small p, bott, Langer & Associates (p. 38).
or very large p the following formula can be 22. Lutz, Carl F. (pp. 37-38).

Robert M. Halley, CCP/CBP, is a compensation consultant associated with Swanson Consulting, a labor re-
lations consulting firm, in Seattle, Washington. He has more than 15 years of experience in the field of
compensation and has previously served as manager of compensation services for Associated Grocers, Inc.
and compensation administrator for the Port of Seattle. He formerly edited and published the Compara-
tive Salary Survey Report, a major regional salary survey. Originally trained as a behavioral scientist, he
holds advanced degrees in the social sciences and education. His professional specialties include salary
survey research, compensation program development, and salary administration training.

52 COMPENSATION & BENEFITS REVIEW

Downloaded from cbr.sagepub.com at EASTERN KENTUCKY UNIV on May 26, 2015