Beruflich Dokumente
Kultur Dokumente
Applied Economics
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/raec20
To cite this article: Brian W. Gould (1997) Consumer promotion and purchase timing: the case of cheese, Applied
Economics, 29:4, 445-457, DOI: 10.1080/000368497326949
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy, completeness, or suitability
for any purpose of the Content. Any opinions and views expressed in this publication are the opinions
and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of
the Content should not be relied upon and should be independently verified with primary sources of
information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands,
costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or
indirectly in connection with, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or
systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution
in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at
http://www.tandfonline.com/page/terms-and-conditions
Applied Economics, 1997, 29, 445Ð 457
53706 , USA
I. INTRODUCTION and Forker, 1986; Sun et al., 1995; Ward and Dixon, 1989;
Ward and McDonald, 1986).1 Little work has been done at
The US dairy industry has responded to decreasing per the household level to analyse the impact of dairy product
capita dairy product consumption by adopting industry- promotion or to analyse the impacts of non-advertising
wide advertising and promotion campaigns. Examples of promotion e orts. With proposed changes in US dairy
these programmes include the 1995 $52 million Fluid Milk policy to a more market oriented system, the use of both
Processor Education Programme and the cheese, butter and advertising and non-advertising promotion e orts to main-
other dairy product promotion e orts of the National Dairy tain industry revenues can be expected to increase. Under-
Promotion and Research Board. Recent evaluations of sev- standing household impacts is important to understanding
eral dairy advertising campaigns have indicated a positive the potential bene® ts of such e orts.
net e ect of such programmes (Forker and Kinnucan, 1991; Household level analyses of the impacts of commodity
Cheese Reporter, 1995; Sun et al., 1995). For example, Kaiser promotion programmes typically hypothesize impacts on
and Roberte (1995) evaluated generic ¯ uid milk advertising both quantity and timing of product purchases (Neslin et al.,
in New York City over the January 1986Ð December 1992 1985; Gupta, 1988; 1991). Figure 1 provides one representa-
period and found a signi® cant impact on whole and low fat tion of the relationship between product promotion, quanti-
milk demand. Larson (1992) argues for continued expansion ty purchased and the dynamics of these purchases. With
of such generic agricultural commodity promotion pro- commodity promotion e ort, there may be a direct positive
grammes given the positive secondary bene® ts of increased impact on quantity purchased and a negative impact on the
e ectiveness of associated branded promotion e orts and length of time between purchases. With less time between
protection of market share from processed substitute prod- purchases, this implies that household stocks may be larger.
ucts. With larger household stocks, the lower the amount pur-
Previous evaluations of the e ect of promotion e orts on chased. The net e ect of promotion, therefore depends on
US dairy product demand have tended to focus on generic the direct purchase impacts and indirect stock e ects. The
advertising programmes at the aggregate commodity or possible con¯ ict of the direct and indirect quantity impacts
market level (Blaylock and Blisard, 1988; 1990; Kinnucan implies that unless shorter interpurchase time is recognized,
1
Two studies that have examined the e ect of coupon use on a speci® c food commodity are those of Ward and Davis (1978a) and of Lee
and Brown (1985). They examine the e ect of coupon use on frozen concentrated orange juice demand. Both studies using monthly
household level purchase data found signi® cant e ects of coupon redemption on quantity of juice purchased. Neither study investigate the
e ect of such promotion on purchase timing.
0003Ð 6846 Ó 1997 Routledge 445
446 B. W . Gould
I I . D E S C R I P TI O N O F T H E MO D EL O F
I N T ER P U R C H A S E T I M E
1 2
n i+ 1
plicit objective undertaken by retailers so as to counter-act L L i (I | l ) = ln
anticipated marketing e orts of competitors. Alternatively,
Õ H (t ij )r S(t ij )
ij
(2)
j= 1
without such a predatory situation, purchase acceleration
= n i ln(l ) - l (t is + t ic ) (i = 1, ¼ , N)
may result in consumers stockpiling the commodity that
they would have purchased regardless of the promotion where I is the vector of interpurchase time, t ij the interpur-
(Neslin et al., 1985). chase time for the ith individual on the jth event, n i the
In the present analysis and as a ® rst step in investigating number of events (purchase occasions), t is the sum of n i in-
the relationships shown in Fig. 1, we focus on the dynamics terpurchase times and N households. To account for the
of the consumer purchase process. In particular, we examine possibility that the last observation may be right censored
the e ect of coupon-based price deals on interpurchase (i.e. no purchase has occurred by the last time period of
times. We apply econometric models of duration to a observation for the ith individual), r ij is the Kronecker delta
frequently purchased food commodity, cheese. We conduct equal to 1 if there is a purchase on the jth event, 0 if there is
this analysis by using a data set which consists of a right censoring, and t ic censored time, which is zero if the ith
household panel observed over 170 consecutive weeks observation is not censored (Gupta, 1991, p. 3; Yamaguchi,
from March 1991 to June 1994. Besides purchase quantity 1991).
and price, information with respect to coupon use and From (2), if at time t ij a purchase occurs, the contribution
household demographic characteristics are available. to the likelihood function of each purchase occasion is the
The models presented below build upon each other in probability density function of the occurrence of the event at
terms of assumed distribution of interpurchase time, time t ij , f (t ij ), the product of H (t ij ) and S(t ij ) from (1). If an
e ect of previous purchases, role of demographic charac- observation is right censored, the contribution to the overall
teristics and e ect of unobserved interpurchase time likelihood function is the probability of not having the event
heterogeneity. occur between 0 and t i , S(t ij ) (Yamaguchi, 1991, p. 11).2
2
It is important to include observations that are right censored. If we omit observations that are right censored or assume an arbitrarily
large duration value for these censored observations biased distribution parameter estimates may be obtained. For these right censored
observations, information concerning survival up to t i are accounted for without making assumptions about the timing of the event’s
future occurrence (Yamaguchi, 1991; Tuma and Hannan, 1979).
Consumer promotion and purchase timing 447
As shown by the hazard rate in (2), one shortcoming of ith consumer becomes:
e
assuming an exponentially distributed interpurchase time is
`
that a consumer’s purchase probability for the next period is L i (I) = L i (I | l )C (l )dl (6)
una ected by the length of time since the last purchase 0
(Gupta, 1991; Chat® eld and Goodhardt, 1973; Jeuland et al.,
1980). To overcome this shortcoming many researchers where L (I | l ) is the conditional likelihood shown in (4), C (l )
have adopted the gamma distribution and in particular the is the gamma distribution:
Erlang-2 form in analyses of non-durable purchase duration
times (Herniter, 1971; Chat® eld and Goodhardt, 1973; a
C (l ) = (a l )( v ± 1) (± a l )
e (7)
Zufryden, 1978; Jeuland et al., 1980; Gupta, 1988).3 T (v )
When interpurchase times are distributed according to
Erlang-2, the density function, survior function, and hazard and T (v ) is the gamma function (McDonald and Butler,
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
rates are: 1987, p. 232; Wagner and Taudes, 1986).4 The resulting
log-likelihood function with the assumption of interpur-
f (t) = l 2 te(
± l t)
chase time being distributed Erlang-2 with gamma hetero-
S (t) = (1 + l t)e(
± l t) (3) geneity is:
2 2 ni ± 1
f (t) l t
H (t ) = = L L i (t | l ) = v ln(a ) + ln t ip + + ln(v + j)
S(t ) (1 + l t) j= 0
1 2
where l is the distribution location parameter (Gupta, (2n i + v )t ic
1991). Thus from (3), the hazard rate is now dependent on + ln 1 + - (2n i + v )ln (t is + t ic + a ) (8)
t is + t ic
interpurchase time. The associated likelihood function used
to estimate l is: (i = 1, ¼ ,N)
ni
L i (I | l ) = Õ f (t ij )S (t ic ) Although (6) allows l to have a distribution, the e ect of
j= 1 exogenous factors on the distribution of interpurchase time
3 4
ni has still not been incorporated. As an alternative to (6) and
= Õ l 2
t ij e(
± l t ij )
(1 + l t ic )e( ± l t ic )
(7), market and household characteristics can impact inter-
j= 1
purchase time via the following:
= l 2 ni
t ipe(
± l ( tis + t ic ))
(1 + l t ic ) (i = 1, ¼ , N) (4)
l i (t) = l 0 e( b Xit)
(9)
where t ip the product of interpurchase times. The associated
log-likelihood function is: where X it are time dependent explanatory variables and
b and l 0 are parameters to be estimated. Using (9) along
L L i (I | l ) = 2n i ln(l ) + ln (t ip ) - l (t is + t ic )
with (5) results in interpurchase time being the result of
+ ln (1 + l t ic ) (i = 1, ¼ , N) (5) a nonhomogeneous Poisson process (Gupta, 1991). Assum-
ing l 0 time invariant, the survivor and density functions are:
The above model assumes that all consumers have the
e e
same hazard rate function given the constant l distribution
1 2 1 2
t t
parameter. In reality di erential purchase rates exist across S(t) = 1 + l (t ) dt exp - l (t ) dt
0 0
consumers due to household and market characteristics
(Gupta, 1988; Helsen and Schmittlein, 1992; 1993; Ward and = (1 + D (t ))exp( - D (t)) (10)
1e e
Davis, 1978a; Neslin et al., 1985). A method that has been
2 1 2
used to allow for heterogeneity in hazard rates is to assume t t
3
We can represent the gamma distribution ( fG ) as:
±
l rtr e± 1 l t
fG (t) =
r- 1
When the distribution parameter r takes only integer values this distribution is referred to as an Erlangian distribution. When r is set equal
to 2, this is referred to as an Erlang-2 distribution (Cox and Lewis, 1966).
4
Other methods to include heterogeneity can be found in McDonald and Butler (1987) and Butler and Worrall (1991).
448 B. W . Gould
where purchase time but allows for di erences in these time-
k± 1 dependent covariances (X it ) across households.
D (t) = l (1) d 1 + + l (w) + l (k)(t - d1 - g (k - 2))
w= 2
and d1 is the proportion of the ® rst week included in the III. DE SCRIPTION OF THE CONSUMER
® rst purchase, k is the current week of purchase (with the P A N EL A N D E X P LA N A T O R Y
week of last purchase set to 1), g equal to 1 if k equals 1 V A RI A B L E S
and l ( ´) de® ned by (9) (Gupta, 1991, p. 6). The resulting
log-likelihood function can be shown to be: We apply the theoretical models represented by (5), (8), (11),
and (13) to an analysis of US cheese purchase. The purchase
L L i (I | l 0 ) = 2n i ln l 0 + ln G p
i + ln V p
i - l 0 (V s
i + V c
i) data used are obtained from a March 1991Ð June 1994 US
+ ln (1 + l 0V
c
i) (11) consumer panel maintained by Nielsen Marketing Research
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
1 2
j± 1
= G + + G +G - d i1 - - 2) ] if no purchases had occurred during the previous week be-
V ij i1 d i1 iw ij [t ij g (j
w= 2 cause of not purchasing during a given week or the result of
being away from home due to vacation, business trip, etc. For
1 2
± 1
jc
V c
i = G i1 d i1 + + G iw +G ij c [t ic - d i1 - g ( jc - 2) ] this analysis we include households that reported continu-
w= 2
ously over 170 weeks. This does not imply that households in
The above model assumes no unobserved heterogeneity the panel purchased each week but during weeks where
in the hazard rate function given that l 0 is time and house- cheese was not purchased for at-home consumption, NMR
hold invariant. Unobserved heterogeneity in consumers’ was given this information. Given the size of the household
hazard rates can be accounted for by allowing l 0 to be panel we randomly selected households from the continuous
distributed across the population. If we assume l 0 to be panel. In order to avoid extremely long interpurchase times,
distributed according to the gamma distribution shown in we include households that have more than two purchase
(7), the likelihood function which allows for both unex- occasions each year over the 170 week period. Purchase
plained heterogeneity and time-varying covariates can be opportunities and occasions are de® ned on a weekly basis.
represented as: We apply our econometric model to four di erent
e
v p p cheeses: all, processed, natural cheddar, and cottage cheese.5
` a G iV
L i (I) = = i
L i (t | l ) g(l )dl Table 1 provides an overview of the samples’ cheese pur-
0
0 0 0
V s
i + V c
i +a (2 ni+ v )
chase characteristics. For all cheese, over 18% of the pur-
3 43 4
(2n i + v ) V c (2 n i ± 1 ) chase weeks occurred with some type of cents-o coupon.
1+ Õ + j)
i
3 (v (12) This is similar to that observed for processed cheese. Only
V s
i + V c
i +a j=0
3% of cottage cheese purchase occasions involve the use of
with log-likelihood: some type of coupon. In contrast to our initial hypotheses,
2 ni ± 1 we see reduced `coupon week’ interpurchase times only for
L L i (I | l 0 ) = v ln (a ) + ln G p
i + ln V p
i + + ln (v + j) all cheese. This result may be deceiving given that there are
j=0
other market and demographic characteristics impacting
1 2
(2n i + v ) V c
interpurchase time as well as coupon use. The duration
ln 1 + - (2n i + + + a ) (13)
i s c
v )ln(V V
V s
i + V c
i +a i i
model estimated here isolates the impact of each of these
variables on purchase timing.
(i = 1, ¼ , N)
With this formulation, di erences in interpurchase times
Purchase characteristics a¤ ecting timing
are accounted for by two components: one which allows for
heterogeneity in base hazard rate which is time invariant For interpurchase time models represented by (11) and
and a second component which assumes homogeneity of the (13) we incorporate household and market variables as
e ect of changes in time-dependent covariates on inter- distribution shifters. Table 2 provides the de® nition of these
5
Given that our analysis covers 170 weeks of potential purchases, an examination of aggregate US cheese supply and demand
characteristics revealed modest annual growth rates in supply and per capita consumption. For example, over the study period both the
US supply and per capita consumption of cheese increased between 1Ð 3% annually over the four-year period 1991Ð 1994 (Putnam and
Allshouse, 1995).
Consumer promotion and purchase timing 449
Table 1. Cheese purchase characteristics
Table 2. DeÞ nition and mean values of demographic and cheese purchase characteristics
Type of cheese
Purchase characteristics
PURCHÐ L AGa Amount purchased on last purchase occasion (1b) - 1.2 1.2 1.0 1.5
COUPÐ V AL UEb Value of coupons used per occasion ($/1b) + 0.80 0.74 1.02 0.47
POS Ð CHANGEa Ratio of change in price to reference price when - - 0.156 - 0.112 - 0.064 - 0.077
there is a price increase (# )
NEGÐ CHANGEa Ratio of change in price to reference price when + 0.116 0.088 0.056 0.062
there is a price decrease (# )
SUMMERa Dummy variable for June, July and August (0/1) + 0.279 0.263 0.241 0.279
HOL IDAY a Dummy variable for November and December (0/1) + 0.127 0.149 0.162 0.127
Household characteristics
INV Ð HHSIZEc Inverse of household size (1/# of members) - 0.477 0.466 0.448 0.491
BL ACKc Dummy variable = 1 if meal planner is Black (0/1) ? 0.055 0.049 0.062 0.011
HISPANICc Dummy variable = 1 if meal planner is ? 0.033 0.061 0.025 0.026
Hispanic (0/1)
PER < 5c Per cent of household members less than - 0.034 0.055 0.050 0.044
6 years (%)
PER6Ð 13c Per cent of household members between - 0.065 0.075 0.081 0.059
6 and 13 years (%)
PER14Ð 18c Per cent of household members between - 0.046 0.047 0.042 0.036
14 and 18 years (%)
POV Ð RA TIOc Ratio of household income to poverty ? 3.43 3.35 3.45 3.31
threshold income (# )
SUBURBc Dummy variable = 1 if household resides ? 0.120 0.131 0.167 0.140
in suburb (0/1)
RURAL c Dummy variable = 1 if household resides ? 0.064 0.086 0.040 0.083
in rural area (0/1)
The `Expected sign’ column refers to hypothesized coe cient signs for variables in Equations 11 and 13.
a
Indicates mean taken over all purchase occassions.
b
Indicates mean taken over all purchase occassions where there is a coupon used.
c
Indicates mean calculated over the number of households consuming each cheese.
variables along with sample means and expected direction made last period minus consumption. Jain and Vilcassim
of the impact of changes in these variables on purchase (1991) outline several empirical problems with estimating
hazard rate. From Figure 1, beginning period inventory is household inventory from a data set that does not explicitly
hypothesized to have a direct impact on interpurchase time collect such information. In order to avoid potential biases
(Neslin et al., 1985; Gupta, 1988, 1991). Beginning house- in making assumption concerning initial inventories and
hold inventory equals last periods inventory plus purchases estimating household consumption rates, we follow Jain
450 B. W . Gould
and Vilcassim (1991) by including lagged volume purchases, represents the value of these coupons. Ward and Davis
PURCHÐ LAG, as an explanatory variable (Helsen and (1978b), using a similar variable, found that coupon redemp-
Schmittlein, 1992; 1993). We hypothesize that interpurchase tion rates are determined, in part, by coupon face value. If
time would increase (and hazard rate would decrease), the coupon redemption induces a stockpiling type of behaviour
greater the amount purchased on the last occasion. Thus we in which consumers purchase sooner than usual than we
expect a negative coe cient for this variable in the empirical would expect a positive coe cient for this variable in (11)
applications of (11) and (13) (Jain and Vilcassim, 1991). and (13).
As suggested by Gupta (1988; 1991) we would expect that Previous analyses of cheese demand have shown that
a price drop may cause the consumer to purchase earlier during the holiday period, cheese demand increases (Sun et
than usual (increase the rate of purchase). To examine this al., 1995; Blaylock and Smallwood, 1983; 1986). The vari-
e ect we include the variables POSÐ CHANGE and ables SUMMER and HOL IDAY are used to account for
NEGÐ CHANGE. The ® rst variable represents the relative hypothesized di erential hazard rates during the summer
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
change in product price since the last purchase when there is months, June, July and August, and over the Thanksgiving
a shelf price increase. NEGÐ CHANGE is similarly de® ned and Christmas holiday period, respectively. We hypothesize
except for a price decrease. Following, Rajendran and Tellis increased probabilities of purchase during these times,
(1994), we could have used previous purchase price as shorter interpurchase times and therefore positive coe -
a point of comparison. Instead, we use a measure of the cients for these variables.
consumers `reference price’. For the present analysis we
de® ne reference price as:
Household characteristics a¤ ecting purchase timing
RefÐ Pricec º 0.571 * Pricec ± 1 + 0.286 * Pricec ± 2
Demographic characteristics such as household size, com-
+ 0.143 * Pricec ± 3 (14) position, ethnicity and income have been shown to be im-
where c refers to purchase occasion, and Price is per pound portant determinants of cheese demand (Blaylock and
shelf price. The use of the above declining weights approx- Blisard, 1983; 1986; Heien and Wessells, 1990). In the pres-
imates a geometric function with a common ratio of 0.5 ent analysis the variable INV Ð HHSIZE represents the in-
(Rajendran and Tellis, 1994, p. 27). verse of the number of resident household members.6 One
From (14), POS Ð CHANGE and NEGÐ CHANGE are de- would expect that the greater the household size, the more
® ned as: quickly household inventories will be depleted, ceteris
paribus, implying a shorter interpurchase time and an in-
POSÐ CHANGE creased hazard rate. Thus we would expect a negative co-
5
e cient for this variable (Jain and Vilcassim, 1991).
RefÐ Pricec - Pricec
if (RefÐ Pricec - Pricec) < 0 We control for household composition by including the
º RefÐ Pricec variables PER < 5, PERÐ 13, and PER14Ð 18. Since our data
0 otherwise includes birth month and year for each household member,
(15) we update these variables on a monthly basis. It is unclear
NEGÐ CHANGE as to the impact of having young children in the household
5
RefÐ Pricec - Pricec on overall cheese purchases. Young people eat less com-
if (RefÐ Pricec - Pricec) > 0 pared to adults but many adults face dietary restrictions on
º RefÐ Pricec
the amount of fat and cholesterol that can be consumed.
0 otherwise Three separate composition variables are included so as to
With a price increase relative to the reference price, allow for di erential impacts of children on household inter-
we would expect an increase in interpurchase time, de- purchase time as children age.
creased purchase probability and therefore a negative The characterization of the ethnicity of each household is
POSÐ CHANGE coe cient. Conversely, a positive coe - based on characteristics of the main meal planner which was
cient is hypothesized for NEGÐ CHANGE. Using these two assumed to be the female head, if present. Previous Tobit-
variables we can test the hypothesis that the impacts of based analysis of US household cheese consumption have
a price increase or decrease are symmetric but of the op- found that non-white households had a lower probability
posite sign. and consumed less than white households (Blaylock
The variable used to re¯ ect the e ect of coupon redemp- and Smallwood, 1983; 1986) The variables BL ACK and
tion on interpurchase time is COUPÐ V AL UE which HISPANIC are used to capture di erences across ethnic
6
Following industry standards, cottage cheese is not included in the all cheese category. A detailed listing of cheeses in each category can be
obtained from the author upon request.
Consumer promotion and purchase timing 451
groups. If this decrease in demand is re¯ ected in both the I V . C H A R A C T E R I Z A TI O N O F T H E
number of purchases and purchases per occasion, we would D I S T R I B U T I O N O F I N T E R PU R C H A S E
expect negative coe cients for these variables in (11) T I M ES
and (13).
Household pre-tax income in the data set is reported in 16 Sixteen models were run: four cheeses under four distribu-
categories ranging from less than $5000 to more than tion assumptions of interpurchase time and gamma hetero-
$100 000. To convert these categorical data to a continuous geneity. Estimation was undertaken using the Maximum
form, we assumed the midpoint of each category to be Likelihood module within the GAUSS software system.
household income. For households with income above From the above discussion, the Erlang-2 model (5) is not
$100 000 an income of $150 000 was assumed. To control for nested within the Erlang-2 with heterogeneity. As such the
household size, composition and income, the variable use of a likelihood ratio test is not appropriate for evaluat-
POV Ð RA T IO is the ratio of household pre-tax income to ing the performance of one model over the other. In order to
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
poverty threshold income as de® ned by the Bureau of Cen- evaluate heterogeneity versus no-heterogeneity assumption
sus (Department of Commerce, 1995). Poverty income is we adopt the method suggested by Schwarz (1978) and
used by the Bureau of Census to estimate the number of reviewed by Rust and Schmittlein (1985). The test statistic
individuals and families in poverty. Poverty threshold in- developed by Schwarz is:
come levels are dependent on number and age distribution
z º ln(L L ) - 1
2 (ln N )P* (15)
of household members. We are unsure of the e ect of this
variable on interpurchase time. For low income households, where P* is the number of estimated parameters. In com-
the ability to purchase large amounts per purchase occasion paring non-nested models, larger values of z indicate the
may be limited, thus implying relatively short interpurchase preferred model.
times, ceteris paribus. Alternatively, given limited income, Table 3 presents the likelihood function values for the 16
they may purchase smaller total amounts, thus implying estimated models. The combination of gamma heterogen-
longer interpurchase time. The variables SUBURB and eity and time-varying covariates are presented for each
RURAL are included to account for di erences in purchase cheese. Given the nested nature of the models with and
patterns for households that live in suburban and rural without covariates, likelihood ratio tests of the null hypo-
areas when compared to central cities. These unique pur- thesis of no impact of covariates is presented in the column
chase patterns may be the results of di erences in oppor- (4). For example, the x 2 -statistic of 23 764.2 is the result of
tunity costs of shopping trips, real incomes, etc. testing the null hypothesis of the Erlang-2 model for All
Table 3. L ikelihood function values for Erlang-2 interpurchase time models with and without covariates and
gamma heterogeneity
Model description
No No - 73789.4 23 764.2*
- 73 792.6
All Cheese No Yes - 61907.3 - 61 959.2
Yes No - 70737.0 - 70 743.5
25 475.4*
Yes Yes - 57999.3 - 58 054.5
No No - 47968.3 13 953.6*
- 47 971.4
Processed No Yes - 40991.5 - 41041.6
Yes No - 45787.5 17 091.2*
- 45 793.8
Yes Yes - 37241.9 - 37295.2
No No - 26213.6 - 26 216.5
7287.6*
Cheddar No Yes - 22569.8 - 22616.0
Yes No - 25874.3 10 326.0*
- 25 880.1
Yes Yes - 20711.3 - 20760.4
No No - 48880.4 8734.2*
- 48 883.5
Cottage No Yes - 44513.3 - 44563.5
Yes No - 46612.9 12 634.6*
- 46 619.2
Yes Yes - 40295.6 - 40348.9
*Indicates signi® cance at the 0.001 level.
452 B. W . Gould
cheese without gamma heterogeneity or time varying covari- chase time although their importance varied by character-
ates (5) versus a similar model with time-varying covariates istic and cheese. Of particular importance for this analysis is
(11). In order to compare non-nested models the z -statistic is the hypothesis test concerning coupon redemption and in-
presented in the last column of Table 3. The results of both terpurchase time. From the parameter estimates presented
the likelihood ratio tests and relative values of the z -statistics in Table 4, the COUPÐ V AL UE coe cients are of the ex-
indicate, that similar to Gupta (1991), the preferred model for pected sign for all cheese types. Likelihood ratios were used
each cheese is the version where gamma heterogeneity and to test the null hypothesis that coupon redemption does not
time-varying covariates are included. Table 4 presents the accelerate cheese purchases. This hypothesis is clearly rejec-
parameter estimates for this model for each cheese. ted for all cheese types given the large x 2 -statistics (Table 5).
The impact of changes in shelf-price is as hypothesized given
negative POSÐ CHANGE and positive NEGÐ CHANGE
Impact of purchase characteristics
coe cients for all cheeses. Interpurchas e time was found
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
Purchase characteristics coupon use, price change, lagged to increase with shelf price increases while a price de-
inventories and seasonality were found to impact interpur- crease increases the hazard rate and therefore decreased
Table 4. Parameter estimates of Erlang-2 interpurchase time models with time varying covariates and
gamma heterogeneity
Purchase characteristics
PURCHÐ L AG
- 0.004 - 0.031* - 0.106* 0.025*
(0.003) (0.005) (0.009) (0.005)
0.354* 0.538* 0.682* 0.648*
COUPÐ V AL UE
(0.002) (0.003) (0.009) (0.019)
POS Ð CHANGE
- 0.092* - 0.140* - 0.352* - 0.076*
(0.003) (0.006) (0.018) (0.007)
0.315* 0.438* 0.595* 0.452*
NEGÐ CHANGE
(0.017) (0.002) (0.037) (0.028)
SUMMER
- 0.029* 0.004 - 0.074* 0.051*
(0.009) (0.011) (0.018) (0.010)
HOL IDAY
- 0.002 - 0.008 0.009 - 0.120*
(0.012) (0.015) (0.022) (0.016)
Household characteristics
INV Ð HHSIZE - 0.450* - 0.438* - 0.402* - 0.191*
(0.035) (0.056) (0.067) (0.042)
0.340* 0.303* 0.688* 0.119
PER < 5
(0.072) (0.075) (0.075) (0.089)
- 0.047 0.066 - 0.217* - 0.005
PER6 Ð 13 (0.059) (0.075) (0.079) (0.078)
PER14 Ð 18
0.172* - 0.168* - 0.066 0.159*
(0.061) (0.058) (0.073) (0.067)
BL ACK
- 0.069 - 0.126 0.302* - 0.210
(0.096) (0.065) (0.084) (0.225)
HISPANIC
- 0.075 - 0.055 0.170* 0.310*
(0.043) (0.027) (0.059) (0.070)
POV Ð RA T IO
0.105 - 0.159* - 0.208* 0.138*
(0.052) (0.066) (0.072) (0.038)
SUBURB
- 0.025 - 0.000 0.061 - 0.022
(0.039) (0.063) (0.075) (0.049)
RURAL
- 0.049 0.135 - 0.008 0.020
(0.074) (0.076) (0.123) (0.061)
Null hypothesis All Cheese Processed Cheese Cheddar Cheese Cottage Cheese
interpurchase time. We tested the null hypothesis that the Impact of household characteristics
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
7
The inverse is used so as to avoid potential scaling problems when maximizing the various likelihood functions.
454 B. W . Gould
Table 6. Hazard rate price and coupon-value elasticities
models we simulate hazard rates pro® les for the above four cheese, after 4 weeks, the hazard rate for the simulated
cheeses. These hazard rate pro® les show the instantaneous single-person household is slightly more than 0.48 com-
probability of a household purchasing given the number of pared to more than 0.67 for Hispanic households. Single-
weeks since last purchase. Figure 2 shows hazard rates for person households have the lowest hazard rates except for
non-minority, Black, Hispanic, and single-person house- cottage cheese where Black households exhibit the lowest
holds.8 From this ® gure we see that as interpurchase time pro® le.
increases, Hispanic households exhibit the largest instan- In contrast to the above hazard rate pro® les, Fig. 3
taneous probability (hazard rate) of purchase for all cheese, shows `survival’ probability pro® les based on (3).9 These
while single person households show the lowest. For all pro® les show the probability of a household not purchasing
8
For each elasticity, they are calculated at mean values of the exogenous variables except for the exogenous variable that is changing. For
this exogenous variable conditional means are used.
9
The pro® les in Figs 2 and 3 are generated by using mean values of the demographic variables for the various subgroups. For example,
Hispanic hazard rate pro® les for all cheese, are calculated based on mean values of the demographic variables for Hispanic households
only. Non-minority households for these ® gures are assumed to be households that are classi® ed as being neither Black nor Hispanic.
Consumer promotion and purchase timing 455
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014
a particular cheese as interpurchase time increases. Given and cottage cheese. We incorporate within these duration
(3) the relative position of each household type will be the models household demographic and purchase character-
opposite of that observed in Fig. 2. For all cheese, after istics that allow for the distribution of interpurchase times
approximately 6 weeks, all households have less than a 10% to vary across households. The results of likelihood ratio
probability of not purchasing by this time. At 3 weeks, there tests indicate that these characteristics are statistically sig-
is a 41% probability that single-person households will not ni® cant factors impacting the distribution of interpurchase
have purchased any cheese. This compares with 36% for time.
Black, 28% for non-minority, and 28% for Hispanic house- A likelihood ratio test of the null hypothesis that coupon
holds. A similar pattern was observed for processed cheese use has no impact on the timing of cheese purchases is
with 3-week `survival’ probabilities of 58%, 54%, 47% and clearly rejected. As hypothesized, the use of coupons result
46% for Single, Black, non-minority, and Hispanic house- in reduced interpurchase times for all cheeses. This impact,
holds, respectively. however varies across cheese type, especially when consider-
ing the type of household doing the purchasing. Elasticity
estimates indicate inelastic hazard rate responses to price
Summary and areas of future research
changes and to the use of coupons. Only in the use of
The use of coupon-based incentive programmes continue to coupons in the purchase of cheddar cheese, is the hazard
be an important marketing tool. The present analysis inves- rate elasticity greater than 1.
tigates one facet of coupon usage, namely its impact on the The present analysis has been concerned with one facet of
timing of purchases for a frequently purchased non-durabl e determining the e ect of coupon use on overall commodity
commodity, cheese. We estimate a series of duration models demand. Previous analyses have shown coupon use has
for four cheese classi® cations: all, processed, natural cheddar, direct impacts on purchase time and quantity purchased
456 B. W . Gould
and indirect impacts on quantity purchased that may Heien, D. and Wessells, C. R. (1990) Demand systems estimation
counteract the direct impacts (Gupta, 1988; Neslin et al. with microdata: a censored regression approach, Journal of
1995). An area of future research is one of developing Business and Economic Statistics, 8, 365Ð 71.
Helsen, K. and Schmittlein, D. C. (1992) How does a product
a model which takes into account the simultaneous de- market’s typical price-promotion pattern a ect the timing of
cisions of coupon use, interpurchase time and quantity household’s purchases? An empirical study using UPC scan-
purchased. Previous analyses have not recognized the si- ner data, Journal of Retailing, 68, 316Ð 38.
multaneous nature of these consumer decisions. The present Helsen, K. and Schmittlein, D. C. (1993) Analyzing duration times
analysis provides the foundation for this future research in marketing: evidence for the e ectiveness of hazard rate
models, Marketing Science, 11, 395Ð 414.
activity necessary to determine the net e ect of coupon
Herniter, J. (1971) A probabilistic market model of purchase tim-
promotion on commodity demand and whether the costs of ing and brand selection, Management Science, 18, 102Ð 13.
such promotion are justi® ed. Jain, D. C., and Vilcassim, N. J. (1991) Investigating household
purchase timing decisions: a conditional hazard function ap-
Downloaded by [University of Colorado - Health Science Library] at 15:45 25 December 2014