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.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
JACQUELYN S. THOMAS, ROBERT C. BLATTBERG, and EDWARD J. FOX*
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
32 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 33
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
34 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 35
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
36 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
(1) Z horizon, in which case it is also right censored. Note that the
censoring value is known; it is a necessary condition to esti-
mate this model (Amemiya 1985, p. 363). We specify the
We model the latent dependent variable zi* with the linear latent duration of subspell si of customer i as
model
(4) ln(yrsi) = Xisi i3 + cisc
(2) z
where x;sii3 is the deterministic component, Eisi is the sto-
where wiNy is the deterministic component, is the sto-
chastic component, xisi is the customer's vector of predictors
chastic component, wi is the customer's vector of predictors,
during subspell si, and 13 is the associated parameter vector.
and y is the associated parameter vector. Probit specifica-
The dependent variable is log-transformed to approximate
tions have been used previously to model customer acquisi-
more closely the normality assumption about the residuals.4
tion (Hansotia and Wang 1997; Thomas 2001).
Given the likely relationship between the customer's
Modeling of the second tenure is somewhat complicated
acquisition and retention behavior (Thomas 2001), it is
by the firm's propensity to change the offer price during the
important that these components be linked. We explicitly
relationship. Although this occurs relatively infrequently
model this linkage, along with customer heterogeneity, by
(see the data description in the following section), we must
specifying variance components. We specify errors of the
nevertheless allow for price changes to affect the customer's
reacquisition and duration components, respectively, as
probability of terminating the relationship. We note that follows:
Amemiya (1985, pp. 433-35) shows that a split hazard
model of the form that we specify is simply a generalization (5) rh
of a single spell of a continuous Markov model. Recogniz-
ing this relationship, we adopt the continuous Markov
(6) Eisi
process assumption that the probability of the customer ter-
minating the relationship at any point in time is independent
where N(04) and N(0,0).
of the current duration (i.e., it is stationary). Using this For the reacquisition and duration components, respec-
assumption of stationarity, we partition the second tenuretively,
so ti and ai represent customer-specific preferences, and
that each period during which a given price is offered to the
vi and Eisi are random errors. As the subscripts oft and a
customer is a "subspell." Thus, each customer's second indicate, customer-specific preferences for reacquisition and
tenure consists of one or more subspells that differ only duration
in are not fixed; they are distributed across house-
the offer price. Moreover, the stationarity assumption holds. In this way, we allow for heterogeneity across cus-
implies that the duration of a subspell does not depend on tomers. Moreover, we allow the distributions of customer-
the length of prior subspells.2 To illustrate, consider a hypo-
specific preferences to be correlated so that BVN(0,X0),
thetical customer who is reacquired by the firm at time t as
where
a result of a reacquisition price offer, is offered a second
price at time t + 81, and then terminates the second tenure at
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 37
In this way, we allowoffers from the firm at any time before ourprefer
a customer's observation
relationship with theperiod. firm
Thus, our application
to does be not include
correlatsequential
offers to
tomer's preference for lapsed customers,
the duration and we conjecture
of aboutthat
their
ship. Allowing for correlation
response to such a sequence of between
offers. cus
for the discrete and Observations
continuous for the duration component
component differ from
reacquisition observations
ard specification is similar in in two ways. First,
spirit tosome acus- se
tomers have
which single stochastic multiple subspells.
error terms More specifically,
for of the
th
416 customers that werethe
are correlated. In summary, reacquired,error
140 received multiple
vari
2 and 4 are price offers during the observation period (an average of
2.19 price offers for customers who received multiple
(7) Var(rii prices), for a total of 582 subspells. Recall that each subspell
represents a different price offer from the firm during a
given customer's second tenure. Thus:
(8) Var(Eis)
We follow Ainslie and Rossi (1998) in estimating our
variance components specification in a Bayesian framework
by using Markov chain Monte Carlo methods. The
(9) Si
The second difference
simulation-based methods of estimating is thatdistribu-
posterior the price in ea
characteristic
tions of parameters in censored of that data
and missing subspell and so may
problems
have only recently become reacquisition price. Thus,
available (Casella andalthough most
George
1992; Gelfand and Smith 1990; tomers
forpaid a single price
application to(the reacquisition
censored
out their relationship
regressions, see Chib 1993). (A complete discussion of with the firm,
ourmore
received multiple price offers.
estimation procedure is available as a technical report on
request.) All the customers examined in this analy
newspaper seven days a week. When cust
DATA AND MODEL VARIABLES
receive the newspaper, they commit to a w
Data remains fixed for a given period (roughly
were unable to determine from the data wheth
The data we used to estimate the model come from a
precommits to buy for several consecutive p
newspaper subscription database and comprise 566 lapsed
less, the firm does not engage in price disco
customers targeted for reacquisition (i.e., C = 566).5 Of the
chase commitments of more than one perio
customers, 416 were successfully reacquired. The reacquisi-
unlike some other contractual selling agreem
tion component of the data includes a single observation for
tions are nonbinding; therefore, customers
each lapsed customer that consists of the reacquisition price
continue or terminate the subscription at an
offered, the result of that offer (i.e., successful or failed reac-
period. The average length of the second
quisition), the number of periods elapsed since the cus-
days (see Table 2). Of the customers who re
tomer's most recent purchase (each period as defined by the
tionships with the firm, 66.8% terminated
company is roughly one month), the price of the last pur-
during the two-year observation horizon. T
chase, and the length of the first tenure. Note that the firm
customer durations are right censored. Desc
made only one reacquisition offer per customer during our
are provided in Table 2.
observation period (i.e., customers who did not respond did
Model Variables
not receive multiple offers). Moreover, each customer in the
data set lapsed only once, so none had received reacquisition The first and most obvious variable included in the esti-
mation is price, that is, the current price offered/paid in the
reinitiated
5We removed customers who lapsed or defected as a result of relocation relationship (recall that customers are offered a
or vacation from the data.
single price for reacquisition but may be offered multiple
Table 2
DESCRIPTIVE STATISTICS
aBecause the price is time varying, we present the averages over the duration of the reinitiated relationship.
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
38 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
Table Tenure
prices over the duration of the second tenure). The 3 1
variable measures the total duration of the customer's rela- COVARIATE IMPACT
between the last price in the prior relationship and the cur- (.91)
rent price paid or offered (i.e., price difference = current Impact on Relationship Duration
Price 1.194 1.1936
price - reference price). A positive value for price difference
(1)
means that the current offer price is higher than the last price
Lapse duration -.047 -.0473
observed by the customer. The second approach to modeling (.78)
price differences enables us to assess asymmetric response Tenure 1 .454 .4539
to gains and losses by defining price decrease (i.e., a gain) (1)
Price decrease 1.556 .1802
and price increase (i.e., a loss) variables. Specifically, we
(1)
define price decrease as the dollar amount of price decrease Price increase -.312 -.0523
relative to the last price the customer was offered by the firm (.76)
before the relationship lapsed; price increase is the dollar
al3 is less than or greater than zero.
amount of the price increase relative to the last price the firm
offered the customer. In interpreting the results, note that
both variables are coded as positive values. The data show
for the reacquisition model as well as posterior probabilities
that 26.6% of the prices paid or offered in the restart rela-
(in parentheses) that the parameter is less than or greater
tionship were decreases relative to the last price paid in the
than zero, depending on the sign of the posterior mean.
prior relationship, and 34.7% of the prices paid or offered
In terms of relative impact, the elasticities indicate that
were increases relative to the last price paid in the prior
the offer price has the largest effect on reacquisition likeli-
relationship.
hood. Tenure 1 and lapse duration also have a material influ-
RESULTS ence on the reacquisition outcome. Notably, price difference
has a much smaller effect. Thus, the absolute effect of price
Covariate Effects on the Reacquisition Probability
is much more important than the effect of price relative to
Consistent with Hia and H2, the reacquisition model the last price paid in the prior relationship. This is an impor-
shows that the probability of a firm reacquiring a customer
tant insight for managers because it suggests that reacquisi-
is higher if the lapse duration is shorter and/or if the first
tion strategies that emphasize decreasing price relative to the
tenure is longer. Consistent with economic theory and prior
H4, relationship are not likely to be effective. A more fruit-
the results also show that customers are more likely toful beapproach to winning back lapsed customers is simply to
reacquired if the reacquisition price is lower. offer a low price, regardless of the price that the customer
As we noted previously, we estimated alternative specifi-
was accustomed to paying before the lapse. This also
cations, one with the price difference variable only and implies
one that customers who previously paid low prices
which separated gains and losses. Comparing the corrected should not be enticed with significantly lower prices.
Akaike information criterion (CAIC) and the Bayesian
information criterion (BIC) for the two specifications, Covariate
we Effects on Length of the Second Tenure
find that the model specification with price difference wasTable 3 also reports posterior means for parameters (and
preferred to the specification with separate price increase
associated posterior probabilities that parameters are less
and price decrease variables.6 Thus, we report parameterthan or greater than zero) for the duration equation. As in the
estimates from the former specification. The results support
reacquisition decision, the duration elasticity of price has a
H6: The likelihood of a customer being reacquired decreases
higher magnitude than other predictors. However, there are
with the difference between the reacquisition price and several
the notable differences between the factors that affect
last price offered in the prior relationship. This result implies
reacquisition and the length of the second tenure. An impor-
that pricing in the prior relationship anchors response totantthedifference is the sign of the price effect. Consistent with
reacquisition offer. More generally, this supports the asser-
Bolton and Lemon (1999) and contrary to economic theory,
tion that customers make decisions about reinitiating the we find that higher retention prices lead to longer relation-
relationship based on comparisons with the lapsed relation-
ship durations. Comparing this to the other relevant CRM
ship. Table 3 reports posterior means of parameter estimates
literature, we find that it is consistent with the assertion that
loyal customers are willing to pay higher prices (Reichheld
6For the model specification with price decrease and price increase vari-
1996) but contradicts Reinartz and Kumar's (2000)
finding
ables, CAIC and BIC are 1288 and 1281, respectively. For the specification
that long-life customers pay lower average prices
than do
with only price difference, CAIC and BIC are 1285 and 1279, respectively.
Thus, the second specification represents a better balance of fitshort-life
and customers (in a catalog retailing context). How-
parsimony. ever, Reinartz and Kumar (2000, p. 28) state, "we expect
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 39
sion processing can explain the effect of price comparisons, The asymmetric impact of price increases and decreases in
it does not explain our results about the effect of lapse dura- conjunction with the current price effect suggests that pric-
tion on the second tenure. Specifically, the results do not ing decisions for restart customers is complex. Specifically,
support H3, suggesting that there is no relationship between the parameter estimates suggest that the last price in the prior
the length of the lapse and the customer's second tenure. It relationship affects a customer's price sensitivity and behav-
is possible that this null result is due to the relationship ior in both reacquisition and retention. In this section, we
between lapse duration and the incorporation of customers' explore how the prior price affects the profitability of restart
preference heterogeneity. If unmodeled individual charac- customers and the optimal pricing strategy of the firm.
teristics, such as a propensity toward variety seeking or iner- Ultimately, a firm's targeting decision and offer decision
tial behavior (Bawa 1990) in newspaper subscription, were should be based on the expected profitability of a reacquired
correlated with the lengths of both the lapse and the second customer. We used estimates from the reacquisition and
tenure, any relationship between the lapse and the second duration models to predict the likelihood of a firm reacquir-
tenure would be obscured by the specification of preference ing and retaining a customer with certain characteristics. By
heterogeneity.? In summary, although firms are less likely to assuming specific marketing costs, we determined the
reacquire customers who have had longer lapses, when they expected SLTV of a potentially reacquired customer.9
have been reacquired, the length of the second tenure
appears to be unaffected.
8The price difference parameter for the reacquisition equation, which
Link Between Reacquisition and Length of the Second
Tenure was significant at a = .10 in the full data set, was significant at a = .05
(actual p-value = .041) in the second reduced data set (i.e., eliminating all
Another issue in our investigation is the link between second tenures of less than ten days).
91t is important to note that we computed expected SLTV and not the
reacquisition and duration of the second tenure. Using a NPV of the profits generated after a customer is reacquired. The difference
is that the expected SLTV takes into account the reacquisition profit or loss
7We thank an anonymous reviewer for pointing out the possibility that and discounts the post-reacquisition LTV by the probability of reacquiring
our heterogeneity specification might mask this relationship. the customer.
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
40 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 41
1
.869
91
13
$.65 $.68
Reacquird
5 6
.74
$.9 $.05
178
29
Reacquird
1 .968
7
$.68
Covarites
ValueofA Observal
90thPercnil
2
.847 5
192 15
$2. $1.75 $1.0 $6.7 $2.43 $.68 $1.0 $45.09 $38.19 $4.86
Covarites
ValueofA Observal
75thPercnil
6 .53 4
17 127
$2. $2.0 $1.0 $4.35 $2.43 $.23 $1.0 $36.9 $20.43 $24.79
Covarites
ValueofA Observal
50thPercnil
Table4
.146
3
17
50
103
$2. $2.75 $1.0 $1.5 $2.43 $(.32) $1.0 $29.7 $4.38 $5.3
Covarites
ValueofA Observal
TARGEINCUSOMPFL 25thPercnil
.01
1
5
35 38
Covarites
ValueofA Observal
10thPercnil
Reacquistonpr Lastpricenolh Lapsedurtion() Tenur1(days) Probailtyfecqusn Reacquiston Expectdraquisonmg Averagtniopcd Pricehang PredictTnu2(ays) Periodtncs Expectdrniomagvqus ExpectdRaquisonVl
PredictTnu2(pos) Expectdrniomag Notes:Fracquindpfl,mb.Whxv
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
42 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
$3.15 51.62 48.2 38. 37.49 36.17 34.87 3.60 32.98 32.6 31.6 29. 28.6 27. 26.71 25.70 24.7 23.78 2.87
$51.48
$3.05 49.51 46.52 37.9 36.1 34.8 3.60 32.8 31.79 31.20 30.4 28.93 27.84 26.80 25.78 24.81 23.87 2.97 2.10
$49.35
$2.95 47.38 4.59 35.92 34.70 3.49 32.1 31.5 30.58 30.2 28.91 27.85 26.81 25.81 24.8 23.91 23.01 2.14 21.3
$47.19
$2.85 45.2 42.6 34. 3.28 32.1 31.0 29.0 29.36 28. 27. 26.75 25.76 24.81 23.89 23.0 2.14 21.3 20.51
$4.9
19.7
$2.75 43.0 40.6 32.94 31.84 30.75 29.68 28.63 28.1 27.61 26.1 25.6 24.71 23.80 2.9 2.07 21.5 20.47
$42.78
19.62 18.90
$2.65 40.82 38.67 31.42 30.9 29.36 28.35 27.36 26.87 26.39 25.4 24.53 23.64 2.78 21.95 21.4 20.37
$40.5
$38.0
$36.04
RetnioPrc
$3.79
Averag
Table5
RegionB
EXPCTDSLV
$31.5
$30.41
$31.5
$32.56
$3.84
RegionA
$34.9
$35.81
.50
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 43
Table 6
EXAMPLE OF SLTV CALCULATION
Reacquisition
Reacquisition price $ .30 $ .30 $ .30 $ .30 $ .30 $ .30 $ .30
Last price in prior relationship $ 2.20 $ 2.20 $ 2.20 $ 2.20 $ 2.20 $ 2.20 $ 2.20
Lapse duration (periods) 6 6 6 6 6 6 6
Tenure 1 (days) 117 117 117 117 117 117 117
Probability of reacquisition .998 .998 .998 .998 .998 .998 .998
Reacquisition costs $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00
Expected reacquisition margin $ .20 $ .20 $ .20 $ .20 $ .20 $ .20 $ .20
Retention
Average retention price per period $ 1.75 $ 2.00 $2.25 $ 2.50 $ 2.75 $ 3.00 $ 3.25
Absolute price change $ (.45) $ (.20) $ .05 $ .30 $ .55 $ .80 $ 1.05
Predicted Tenure 2 (days) 178 142 118 123 128 131 134
Retention costs $ 1.00 $ 1.00 $1.00 $ 1.00 $ 1.00 $ 1.00 $ 1.00
Expected retention margin given reacquisition $35.68 $33.09 $31.39 $37.05 $42.66 $48.16 $53.48
Expected retention margin $35.61 $33.02 $31.33 $36.97 $42.58 $48.07 $53.37
Expected Reacquisition Value $35.81 $33.22 $31.53 $37.17 $42.78 $48.27 $53.57
There are two key factors that increase the expected Second, to hedge against the risk and potential costs of a
SLTV with the low-price strategy. The first is the signifi- restart customer lapsing again, firms may opt to increase
cantly higher reacquisition rate, relative to a price of $2.20. prices above their prior levels. Our analysis shows that the
Table 7 shows how the reacquisition rates vary with the firm can moderately reduce the reacquisition rate (see Table
reacquisition prices. The second factor is the impact of the 7) and still increase expected profits.11 To increase profits by
price decrease on length of the second tenure. Specifically, using this strategy, the firm must increase both the reacqui-
the length of Tenure 2 increases when the reacquisition price sition and the retention prices above the last price before
is less than $2.20. The increased duration compensates for lapse. Region B of Table 5 shows some of the price combi-
the lower retention margin that is received when the reten- nations for which expected SLTV exceeds the SLTV obtain-
tion price is less than $2.20. able when reacquisition and retention prices are both $2.20.
The key to a firm's successful implementation of a price
increase strategy is managing the trade-off between a lower
Table 7 acquisition rate (due to higher reacquisition prices) and
CUSTOMER REACQUISITION RATES increased margins (due to higher retention prices). Because
of this trade-off, there are limits on how high the firm can set
the reacquisition price relative to the retention price and
Reacquisition Price Reacquisition Probability
generate profits beyond those of a $2.20 fixed price. A reac-
$1.75 .677
1.80 .663
quisition price that is greater than $2.20 and significantly
1.85 .649 higher than the retention price will result in an expected
1.90 .635 SLTV below the expected SLTV attainable at a fixed price
1.95 .622 of $2.20.
2.00 .608
2.05 .595
Summary of Numerical Simulation
2.10 .582
2.15 .570 From these results, it might be concluded that the key to
2.20 .557
managing customer winback most profitably is successful
2.25 .545
2.30 .533
customer reacquisition. However, our targeting discussion
2.35 .521 has revealed that the firm may not want to price so as to
2.40 .510 reacquire all lapsed customers and instead may want to
2.45 .498
focus on customers with attractive profiles. Note that we
2.50 .487
2.55 .476
performed our simulation analysis by assuming a moder-
2.60 .466 ately attractive profile (i.e., at the median value of the prior
2.65 .455 relationship characteristics). When attractive customers
2.70 .445 have been recaptured, their behavior is such that the firm can
2.75 .435
recoup the losses from reacquisition by charging higher
2.80 .425
2.85 .416
prices. This approach maximizes margins but not long-term
2.90 .407
2.95 .398
3.00 .389
11 It is notable that when we computed expected profits for price increase
Notes: We calculated all probabilities at the median values of the prior strategies, the effect of price increase relative to prior price on relationship
tenure, lapse duration, and last prior price variables. duration was small (the effect of price decrease was irrelevant).
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
44 JOURNAL OF MARKETING RESEARCH, FEBRUARY 2004
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms
Recapturing Lost Customers 45
This content downloaded from 193.175.6.21 on Thu, 23 Nov 2017 16:02:55 UTC
All use subject to http://about.jstor.org/terms