You are on page 1of 8

American Marketing Association

Modeling and Forecasting Sales Data by Time Series Analysis


Author(s): S. G. Kapoor, P. Madhok and S. M. Wu
Source: Journal of Marketing Research, Vol. 18, No. 1 (Feb., 1981), pp. 94-100
Published by: American Marketing Association
Stable URL: http://www.jstor.org/stable/3151318
Accessed: 28-10-2015 10:12 UTC

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/3151318?seq=1&cid=pdf-reference#references_tab_contents

You may need to log in to JSTOR to access the linked references.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

American Marketing Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of Marketing
Research.

http://www.jstor.org

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
- y |~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

S. G. KAPOOR, P. MADHOK, and S. M. WU*

Time series modeling technique is used to model a series of sales data


in which seasonality causes distinct spike peaks. The analysis of actual sales
data shows that the seasonality in the data can be approximated by a
deterministic function and the stochastic component is a sixth-order autoregres-
sive moving average model. Use of the combined deterministic and stochastic
models to derive the minimum mean squared forecast yields reliable results.

Modeling and Forecasting Sales Data by


Time Series Analysis

Obtaining accurate sales forecasts for the selection tic difference operator of the form (1 - Bs). However,
of proper marketing strategy is a problem of growing the use of this operator may be ineffective or even
concern. Most of the prevailing forecasting methods inappropriate in certain situations (Wu 1977), especial-
make use of static techniques like regression and ly those cases in which the seasonality in the data
correlation analysis (Pindyck and Rubenfeld 1976). is not homogeneous.
These techniques ignore the dynamics or memory of The purpose of our article is to illustrate the use
the process involved and treat the data, which are of an alternative modeling approach to time series
mostly in the form of time series, as being independent. analysis in an actual business situation wherein the
Serial or autocorrelation, generally present in such nonstochastic structure in the data is captured by the
data, then distorts the analysis, masking the true nature use of deterministic functions. The alternative ap-
of the interrelationships and inflating the forecast proach is different from the Box-Jenkins method both
errors. A popular technique which does not require in model philosophy and in modeling strategy (Wu
the assumption of independent data is exponential 1977).
smoothing (Brown 1962). However, it is an ad hoc The time series data on the sale of a particular
technique and involves the problem of selecting, often line of consumer products are analyzed. A brief
arbitrarily, an appropriate weight factor. overview of the alternative modeling procedure for
Time series analysis provides an ideal solution to the nonstationary time series is given first; the applica-
the problem of serially correlated data. Some work tion of the technique in forecasting product demand
in marketing research using time series analysis has is then described and discussed.
been reported recently (Helmer and Johansson 1977;
Mabert and Radcliffe 1974; Moriarty and Adams 1979; MODELING PROCEDURE FOR THE
Tiao and Pack 1975). The advantage of time series NONSTA TIONAR Y TIME SERIES
modeling over other econometric methods of obtaining Consider a scalar time series, Y, t E Z (Z is the
forecasts has been demonstrated by Geurts and Ibra- set of integers). It can be shown that any covariance
him (1975). Most recent work in time series modeling stationary time series can be decomposed linearly into
has followed the Box-Jenkins (1970) method whereby two parts, one deterministic and the other stochastic,
seasonality in data is accounted for by using a stochas- i.e.:
(1) Y,= (t, ) + Xt, t E Z
*S. G. Kapoor is Assistant Professor, Department of Mechanical where ,u(t, *) is a deterministic function for the mean
and Industrial Engineering, University of Illinois at Urbana- with parameter vector 1. For a stationary process,
Champaign. P. Madhok is Research Engineer, Honeywell, Inc. S. ,u(t, i) is a constant whereas for a nonstationary
M. Wu is Professor, Department of Mechanical Engineering, Uni-
versity of Wisconsin-Madison.
process it is a function of time. Xt is a stochastic
series with zero mean and can be further represented
94

Journal of Marketing Research


Vol. XVIII (February 1981), 94-100

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
MODELINGAND FORECASTINGSALESDATA 95

as an autoregressive moving average model (ARMA) Figure 1


as follows. FLOWCHART
OF MODELINGPROCEDURE

(2) X, = X,_,X- Oeka_, + a,, t E Z


X,I k-I

where n and m are the autoregressive and moving


average orders, respectively, and (, and ok are the
autoregressive and moving average parameters. The
a, is a discrete white noise process of strength c2
(i.e., E(a,a,+k) =
kaa2 t, k E Z where 8k is the
Kronecker delta function).
Modeling Steps
Let us reconsider equation 1. Assume that the
function p,(t, *) is known but the parameters, 4, are
not. One of the objectives of the modeling process
is to obtain the estimates of these parameter values.
For a nonstationary time series the modeling proceeds
in three stages. In the first stage, the initial values
of the parameters of the deterministic function, ,I(t, 4),
are obtained by the method of least squares. The
second stage determines the order (n, m) and the initial
values of 4)'s and O's for the ARMA model of the
stochastic component, X,. Finally, in the third stage,
a simulataneous estimation of the model is carried
out by using the initial values obtained from the first
two stages. The three stages are detailed hereafter.
Stage 1. Assume that the output data Y,are available
for N time periods. The initial values of 4, denoted
by 4i, are obtained by fitting the tentative model
Y, = I(t, ,) + ,, t= 1, 2..., N
where X, are the residuals. We obtain the parameters,
4i, by minimizing the sum of squares of the X,.
Stage 2. In this stage, we fit an ARMA model to
the residuals X, obtained from Stage 1. The adequate Stage 3. The last stage of the modeling precedure
consists of fitting the simultaneous models
ARMA model is obtained by systematically fitting
higher order models (Wu 1977). For example, in the Y = (t,(I)+ X,
case of scalar models, we successively fit ARMA (2n, n m
2n - 1) models. At each stage, the conditional least
X, = ,X,_,- Oka,_ + a,
squares estimates of 4), 42, ...f (), , 01, 029... 9 0n-, -i lk-I
are obtained by minimizing the sum of squares of
residuals. The adequacy of the model is checked by where the a, are residuals. The initial values, 4,
statistical tests such as an F-test (Wu 1977) or by obtained from Stage 1 and i,, 0k from Stage 2 are
the plots of the residuals, Q-statistic (Box and Jenkins used to start the nonlinear regression routine for the
estimation of the parameters which are obtained by
1970), etc.
Let us write the model obtained for the residuals minimizing the sum of squares, S = N a2 A final
X, as check for the adequacy of the model is then undertak-
en. Figure 1 is a flowchart depicting the three stages
ni~ m
of modeling.
Xt E Xt-, E kda,-k + a,.
i-\1 k-I DATA DESCRIPTION
The tildes on the 4('s and O's indicate that the parame- Figures 2a and 3a are the plots of the two series
ters obtained are only initial values and will be used of data used in the analysis. Series 1 is the total market
to obtain the final estimates in Stage 3. The tildes demand for all products of both the firm and the
on the X, and a,, likewise, represent the fact that rest of the competitors. Series 2 is the total demand
the X, and d, themselves are only residuals from Stage for the firm's products. The firm has a market share
I and Stage 2, respectively. of about 60%. For each series, there is a total of

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
96 JOURNALOF MARKETINGRESEARCH,FEBRUARY1981

Figure 2 a dummy variable which takes a value 1 for the periods


SALESSERIES1 corresponding to August peak and 0 for other periods.
Likewise, C2d2 approximates the December peak,
where d2 is another dummy variable accounting for
'1 rI
I
I
)(DATA
I
FOR
I
PERIODS
w
I
periods corresponding to December peaks. Therefore,
72,73,74
the deterministic function ,i(t, u) is of the following
form.
(/)
I- x(t, ,) = a + C1di + C2d2 where i = [a, Cl, C2]; a
is the average value of the sales data for all periods
excludingAugust and December.
After the deterministic function, pL(t,4),was decided
upon, the model was fitted in the following sequence.
0 20 40 60 80 100 In Stage 1, the initial values, v = [&, C,, C2 ], were
(0) ACTUAL DATA ALONG WITH FORCASTS obtained for the parameters of the deterministic func-
AND THEIR
i- 95% CONFIDENCE LIMITS
tion, u(t,4*), by minimizing the sum of squares of
the residuals X,. The results for series 1 and 2 are
150 1 | , | Il given in Table 1. The total sum of squares of residuals
DECEMBER PEAK
from Stage 1 is 3455 for series 1 and 1073 for series
2. The deterministic function estimated in Stage 1
( AUGUST PEAK
100I
/
for series I and 2 is shown in Figures 2b and 3b,
respectively. The residuals, X,, from Stage 1, i.e.,
*3
MEAN LEVEL [Y, - px(t, *)], for series 1 and series 2 are shown
/
in Figures 2c and 3c.
o50 . .

Figure 3
0 20 40 60 80 100
SALESSERIES2
(b) DETERMINISTIC FUNCTION (STAGE I)

An -Irv ~~I . -_ -
I
-
I
- v

2
2.
Z o
-2' o . 80 100 ,
00,1 20 40 60
->
(c) RESIDUALS FROM STAGE I

71 observations (periods) taken at intervals of every


four weeks covering a period from February 1971
(0) ACTUAL DATA ALONG WITH FORECASTS
to August 1976. AND THEIR 95% CONFIDENCE LIMITS
As is apparent from the plots, both series have
nonstationary components characterized by sharp
peaks which occur with regularity during the months
of August and December. In the trade, the minor
peak in August is referred to as the "back to school ,)
peak." The major peak in December which is due (0
z
to the surge in demand during the Christmas shopping
season is known as the "Christmas peak."
MODELING THE DA TA
Before modeling the data, one must ascertain the
actual form of the deterministic function, ,i(t, *). The
(b) DETEMINISTIC FUNCTION (STAGE I)
function must account for both peaks and also for on.
Av I II aI r I
)
the mean level of sales for the periods other than
the peak periods of August and December. The August
peak in the data is approximated by C, d, where C,
is the average value of the August sales for five years
z
_ rtI
L
"0 20 40 60
I.
80 100

from 1971 to 1976 represented in the data and d, is (C) RESIDUALS FROM STAGE I

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
MODELINGAND FORECASTINGSALESDATA 97

Table 1
MODELSFOR DATA

Series 1 Series 2
Stage 2 Stage 3 Stage 2 Stage 3
Parameters Stage I A RMA(6,5) simultaneous Stage I A RMA (6,5) simultaneous
a 31.4 31.9 20.8 21.2
cl 13.9 13.5 6.5 5.87
C2 76.0 70.6 44.4 42.8
I0, 1.5109 1.6412 0.4969 0.5697
4)2 -0.7329 -0.9525 -0.1166 -0.1784
c43 -0.8245 -0.6638 -0.1384 -0.1092
4h4 1.0971 1.1778 0.5455 0.5510
5^ -0.4049 -0.6014 -0.6673 -0.7248
4)6 -0.2080 -0.0985 -0.2638 -0.2085
01 1.4124 1.4204 0.5002 0.5179
02 -0.9050 -0.8716 -0.3243 -0.3191
03 -0.7445 -0.7345 -0.1354 -0.1320
04 1.2616 1.2655 0.6549 0.6539
05 -0.7770 -0.7551 -0.8189 -0.8287
Residual sum
of squares 3455 2197 2120 1073 796 788
Q-value' 14.2 13.6 21.5 19.2
'95% x2 value (table) = 28.9.

Next, ARMA models were fitted to the residuals, Another reason for not using the difference operator
X,, sequentially by using a (2n, 2n - 1) strategy to is that it is a stochastic operator; that is, it assumes
determine the order of the models and the initial values a relationship to the previous year's sales. However,
of the parameters to be used in Stage 3. The details the seasonality could be more deterministic than sto-
of the (2n, 2n - 1) modeling strategy are given by chastic. The occurrence of peaks every December
Wu (1977). For both series, ARMA(6,5) models were is simply explained by the fact that Christmas is in
tentatively taken as adequate models as indicated by December and does not necessarily imply a relation-
the statistical tests including the F-test and the Q-sta- ship to the sales of the previous years. In other words,
tistic; the parameters are listed in Table 1. (The order the occurrence of peaks is only a function of time
of the adequate ARMA model was confirmed in the and may not be related causally to previous peaks.
final estimation.) The function Ix(t,*) used to account for the August
Finally, in Stage 3, the combined deterministic- and December peaks recognizes the deterministic
stochastic models were estimated (Table 1). As a finmal nature and the inhomogeneity of the seasonality in
check for adequacy, the residual autocorrelations for the data. However, one may question whether the
30 lags were tested to determine the presence of ARMA(6,5) models for the stochastic component for
recognizable patterns or correlation. The adequate both series capture the 13-period cycle adequately.
models were then used to derive the forecasts. To answer this question, the autoregressive polyno-
mials of both models were decomposed into their
MODEL INTERPRETATION A ND monomial factors (Table 2).
JUSTIFICA TION The nth-order autoregressive polynomial can be
Figures 2a and 3a suggest an annual cycle in the factored as follows.
data because the peaks in the data occur exactly after
every 13 periods. (Each period is four weeks.) One
(1- ,B)k X, = (1 - XkB)
might therefore be led to conclude that the difference k-l k=1
operator (1 - B13) could be used to represent this
cycle. However, one of the reasons for avoiding its where B is the backshift operator defined as BkX,
use is that the seasonality, in this case, is not homoge- = X,_k, k = 0, + 1, +2, ..., and kare termed the roots
neous. In other words, sales during the Christmas or eigenvalues of the autoregressive operator. Each
season may appear to be related to sales during the pair of complex roots gives rise to different frequen-
last Christmas season but sales are more dependent cies. The absolute values of the roots indicate the
on their immediate past than on sales a year ago. extent to which the frequencies are damped. An
The operator (1 - B ) does not reflect this inhomo- absolute value of one is the upper limit for stationary
geneity and imputes a dependence only on last year's processes and represents a completely undamped
sales for all 13 periods. frequency.

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
98 JOURNALOF MARKETINGRESEARCH,FEBRUARY1981

Table 2
ROOTSOF AUTOREGRESSIVE FOR SERIES1 AND 2
POLYNOMIALS

Series I Series 2
(6,5)a (6,5)a
Absolute Frequency Time Absolute Frequency Time
Roots value cycles/period period value cycles/period period
Xi,, 2 0.86 ? i 0.46 0.98 0.078 12.73 0.87 ? i 0.46 0.99 0.078 12.81
X3, X4 0.46 ? i 0.85 0.96 0.171 5.85 -0.02 ? i 0.99 0.99 0.252 3.96
X5 -0.86 0.86 - -0.91 0.91 - -
h6 -0.13 0.13 -0.22 0.22
aARMA order.

From Table 2, we see that the roots for series 1 available subsequently and are indicated in Figure 4
include a pair of complex roots i, K2= 0.86 ? i 0.46 by crosses (X). These were also compared with their
with an absolute value of 0.98 corresponding to a forecasted values (Table 3).
time period of 12.73. Similarly, series 2 has a pair For period t = 72, the sales predicted at period t = 71
of roots Xi, k2 with an absolute value of 0.99 corre- differed from the actual demand by 21% for series
sponding to a time period of 12.81. We see, therefore, 1 and 14% for series 2. This discrepancy could be
that the annual cycle is strongly represented in both attributed to the fact that the firm conducted a major
models even though we have models of order lower promotion campaign during period 72 with discounts
than 13. In fact, by not explicitly considering a term and sales incentives, the effect of which could not
for dependence on last year's data, we put more weight be reflected in the past data. For the next periods,
on recent observations in deriving the forecasts. This the forecasts proved very accurate. For series 1, the
consideration was dictated by the nature of the market. two steps ahead forecast made at period t = 71, i.e.,
Whether the peak in this December will be higher for period t = 73, with its 95% confidence limit is
depends more on the current strength of the economy 33.0 ? 11.0 units. As actual demand was 31 units,
as revealed in the trends from the recent periods of the error is 6%. Similarly, for series 2, the two steps
September, October, and November than on last ahead forecast is 21.4 ? 6.5 and the actual demand
December's sales. was 21 units, a difference of only 2%. Note that the
confidence limits of the forecast for both series are
FORECA STS
very wide. This is to be expected because of the
After identifying the models for series 1 and 2, nonstationary nature of the data and the relatively
we obtained the forecasts by using the usual rules short length of the time series record (71 periods)
of conditional expectation (Box and Jenkins 1970). available to model the data.
The results of the forecasts made at period t = 71 It is important to mention that the effect of promo-
for the next 12 periods and their 95% confidence limits tions and sales can be factored in the estimation of
using the normally independently distributed forecast the model parameters. However, in the present case,
errors are shown in Figures 2 and 3. For clarity, the promotions and sales incentives were not identified
last few observations and the forecasts for series 2 before modeling and thus contributed to past error.
are reproduced on a large scale in Figure 4. The hatched The forecasts made by the proposed method were
area covers the region between the upper and lower compared with those obtained by the Box-Jenkins and
95% confidence limits. Actual observations for the generalized least squares methods. For illustration,
next three periods, i.e., for t = 72, 73, and 74, become the models were fitted to series 2. Using the Box-Jen-

Table 3
ACTUAL DEMAND COMPARED WITH FORECASTED VALUES

Series 2
Series I Proposed method Box-Jenkins
Actual Forecasted Percentage Actual Forecasted Percentage Forecasted Percentage
Period demand values error demand values error values error
72 42 33.2 + 10.7a -21 27 23.1 ? 6.5a -14 20 ? 6.9 -26
73 31 33.0 + 11.0 +6 21 21.4 + 6.5 +2 24 + 7.1 +14
74 40 37.3 + 11.4 -7 29 24.9 +?6.6 -14 26 +?7.2 -11
a95% confidence limits.

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
SALESDATA
AND FORECASTING
MODELING 99

Figure 4
FORECASTSAND THEIR95% CONFIDENCELIMITSFOR SERIES2

X DATA FOR PERIODS 72, 73, AND 74


0 FORECASTS

Q)
I-
Z

67 72 77 82
PERIODS

kins identification, estimation, and diagonistic check- was 21% and 28%, respectively, in contrast to a
ing procedure, we found the adequate model for series difference of 14% from the proposed method. For
2 to be t = 73, the two methods gave approximately the same
percentage error in the predicted values.
(1 - .31B)( - B3)x,= a,.
CONCLUSION
The model was used to obtain the forecasts and their
95% confidence limits for the periods t = 72, 73, and A deterministic function of time trend in conjunction
74; the results are given in Table 3. The percentage with a stochastic model is used to fit the series in
errors for the periods t = 72 and 73 are significantly which the seasonality is inhomogeneous and of a
larger (26% and 14%) than the percentage errors by deterministic nature. Two series of real market demand
the proposed method (14% and 2%). For period t = data are analyzed to demonstrate the applicability of
74, the percentage errors by the two methods are the modeling procedure. For those series in which
close to each other. Furthermore, the forecasts ob- a few sharp peaks appearing at regular intervals depict
tained by the Box-Jenkins method have wider confi- the nonstationary components, the deterministic
dence limits (6.9-7.2) than the forecasts obtained by component is approximated by a function involving
the proposed method (6.5-6.6). the dummy variables accounting for peaks and the
The generalized least squares method was also used stochastic component is represented by the ARMA
to predict the sales forecasts for periods t = 72, 73, models. Comparison of the minimum mean squared
and 74. The predicted values for three periods were error forecasts obtained from these models with
found to be 21.3. The difference between the actual subsequently available observations for the next three
and predicted values for periods t = 72 and t = 74 periods shows that the actual values are well within

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions
100 JOURNALOF MARKETINGRESEARCH,FEBRUARY1981

the 95% probability limits of the forecasts. Our use of the term "nonstationary" when the
Clearly time series analysis is widely applicable in nature of a series of data appears to be dependent
marketing research. An important advantage is the on time origin follows the terminology in stochastic
ability to predict and explain both short- and long-term processes, systems analysis, and control theory. The
effects. However, the care must be taken in modeling models for such a series of data therefore need to
economic time series as the simple strategies of dif- include functions which depend on time origin.
ferencing data by prevailing techniques may not yield
reliable forecasts. Our time series modeling procedure REFERENCES
is different from the Box-Jenkins method in two main Box, G. E. P. and G. M. Jenkins (1970), Time Series Analysis,
aspects. First, it avoids the tedious stage of identifying Forecasting, and Control. San Francisco: Holden-Day.
orders n and m of the ARMA(n, m) model from the Brown, R. G. (1962), Smoothing, Forecasting, and Prediction
examination of plots of the estimated autocorrelation of Finite Time Series. Englewood Cliffs, New Jersey:
and partial autocorrelation functions as recommended Prentice Hall, Inc.
in Box-Jenkins procedure and provides a step-by-step Geurts, M. D. and I. B. Ibrahim (1975), "Comparing the
Box-Jenkins Approach with the Exponentially Smoothed
modeling strategy which can be easily executed on Forecasting Model with an Application to Hawaii
a digital computer to obtain the statistically adequate Tourists," Journal of Marketing Research, 12 (May),
model irrespective of its order. 182-7.
The second difference is in the modeling of nonsta- Helmer, R. M. and J. K. Johansson (1977), "An Exposition
tionary time series. Box and Jenkins use the word of the Box-Jenkins Transfer Function Analysis with an
"nonstationary" for discrete ARMA models with one Application to the Advertising-Sales Relationship," Jour-
or more roots with absolute value one, e.g. random nal of Marketing Research, 14 (May), 227-39.
walk, EWMA, etc. However, the probabilistic proper- Mabert, V. A. and R. C. Radcliffe (1974), "A Forecasting
ties are still independent of the origin, although theo- Methodology as Applied to Financial Time Series," Ac-
counting Review (January), 61-75.
retically their variance is infinite. As a matter of fact Moriarty, M. and A. Adams (1979), "Issues in Sales Territory
in nonstationary cases, where trends and seasonality
Modeling and Forecasing Using Box-Jenkins Analysis,"
are dominant in the data, the sample autocorrelation Journal of Marketing Research, 16 (May), 221-32.
functions fail to damp out quickly. Thus, tentative Pindyck, R. S. and D. L. Rubenfeld (1976), Econometric
identification of the model from their plots is almost Models and Economic Forecasts. New York, McGraw-
impossible. That is the reason for applying differencing Hill Book Company.
or seasonality operators (which in turn must be guessed Slutsky, E. (1927), "The Summation of Random Causes
from the data or sample autocorrelation functions, as the Source of Cyclic Processes" (Russian with English
etc.). The danger of such an indiscriminate operating summary), Problems of Economic Conditions (Institute
or smoothing of the data merely to simplify analysis of Economic Conjecture, Moscow), 3, revised English
ed. in Econometrika, 5, 105 (1937).
has been pointed by Slutsky [1927]. Such an operation
Tiao, G. C. and D. J. Pack (1975), "Modeling the Consump-
itself may introduce spurious trends and periods in tion of Frozen Concentrated Orange Juice: A Case Study
the resultant series which are not present in the data. of Time Series Analysis," Academic Economic Papers,
The final fitted model, though statistically adequate 3 (February).
and apparently parsimonious, may give a completely Wu, S. M. (1977), "Dynamic Data System-A New Model-
distorted picture of the structure of the original series. ing Approach," A SME Transactions, Series B, 99, 708-14.

ERRATUM

The article "Measuring Personal Values: An Evaluation of Alternative Methods" by Thomas J. Reynolds
and James P. Jolly in the November 1980 JMR contains an error in the formula for transforming Taus. The
correct formula is:

TaU,=/2arcsin concordant pairs


= total - 1/2:tl(t, - 1) Vtotal -
2 arcsin 1/2t2(t2 - 1)

This content downloaded from 143.167.2.135 on Wed, 28 Oct 2015 10:12:09 UTC
All use subject to JSTOR Terms and Conditions

You might also like