Learning About Computers: An Analysis of

Learning About Computers:
An Analysis of Information Search and
Technology Choice
Tülin Erdem
University of California, Berkeley
Michael Keane
Yale University
Sabri Öncü
Stanford University
Judi Strebel
San Francisco State University
IO workshop, Duke University
October 26, 2004
A Quote
“How should we allocate resources between
channels? We really don’t know exactly
how consumers learn about technology”
Marketing manager for a major computer manufacturer
2
Motivation
• Issues for a major computer manufacturer in
1995/96:
1) What is the process consumers use to search
for information? What factors affect this
process?
• Information sources utilized in technology choice
• How do information sources interact
• How does accuracy/cost of information provided affect
search and technology process
2) How do consumers decide which technology
to choose and when to buy
• Role of expectations
3
Objective
• Develop and estimate a model of consumer information
search and choice behavior in high-tech durable goods
markets characterized by three key features:
1) there are two or more technological alternatives,
2) consumers have uncertainty about the quality of
each alternative (and/or its suitability to their
particular needs), and
3) there is a rapid pace of technological
improvement, reflected in a rapid rate of price
decline for any given level of quality.
 Our application is to the computer market.
4
Contribution
• Develop and estimate a structural dynamic model of
INFORMATION SEARCH (ACTIVE LEARNING) and
TECHNOLOGY CHOICE
• To understand consumer information search and
technology behavior in high-tech durable good markets
• To investigate the managerial/ consumer welfare
implications of changes in cost and variability of
information, pricing policy, consumer expectation
formation processes
• Use data on price expectations to estimate
consumer future price expectations (Manski 2003),
as well as use data on quality expectations, which
facilitate identification (McFadden 1989)
• Individual level panel data on “real” consumers
5
Relevant Previous Work
• Information search:
• Antecedents (Srinivasan and Ratchford 91)
• Patterns (Meyer 82, Hagerty and Aaker 84)
• High-Tech Durables
• Diffusion of innovation (Bass 69), prelaunch models
(Urban, Hauser, Roberts 90), market share models
(Bridges et. al. 95), expectations (Sultan and Winer
93)
• Consumer Choice
• Learning models (Roberts and Urban 88, Erdem 98,
Anand and Shachar 00, Ackerberg 02, Ching 02,
Crawford and Shum 02, Erdem and Keane 96)
• Forward-Looking expectations and consumer tradeoffs: Gönül and Srinivasan 96,
6 Erdem, Imai and Keane
03, Melinkov 2000, Song and Chintagunta 2003)
A Dynamic Structural Model of
Information Search and
Technology Choice
• Each period the consumer decides
• whether or not to obtain information from
several sources
• whether and what to buy
• Consumers are uncertain about quality
levels of alternatives, and changes in prices
• over time, they learn about quality levels,
• they form expectations about prices
7
Dynamics arise due to...
• The trade-off btw buying today versus delaying the
purchase
 Computer prices tend to drop over time for a given
configuration and uncertainty about quality levels
decreases over time  reasons for waiting
 Opportunity costs that arises from not having a new
computer during the period of delay  incentive to buy
now
8
Data
• Computer Purchase Panel: n = 345;
Six waves of surveys
• By t=6, 98 consumers bought, 102 consumers
did not buy, attrition: 145 (42%)
• Sample Characteristics
•
•
•
•
•
•
Expertise:
34% Novice, 52% Intermediate, 14% Expert
Past Purchase: 45% First time buyer
Gender:
62% Male
Income:
50% over 50K
Age:
Average 39
Education
58 % college or graduate school degree
9
Panel Data on...
 Search and Choice Information
• Information Channels visited each period
–
–
–
–
–
Retail Stores
Computer Sources
General Sources:
Advertisements
WOM
• Whether s/he had bought a PC (if so, its
description/cost)
• Ratings about perceived quality of each technology
• The individual’s perceived price of the type of
configuration considered a) at the present time, b) six
months earlier, c) price forecasts
10
Perceived Quality Construct
• Will meet my needs for a long time to come
• User friendly
• Powerful
• A large number of software titles
• All components operate together without any
problems (Hardware, software, peripherals)
11
A few words on the Data
• Search
• Individuals search the most intensively in the first period and
then search activity declines for all channels; no large
differences across channels
• Demographics do not affect the search patterns over time
but they affect slightly the amount of search in each period
(“Intermediate” expertise individuals search the most in each
period than experts and novices)
• Price Expectations
• Similar expectations for Apple versus IBM/Compatible
technology
• Demographics seem to affect expectations very little (experts
expect bigger declines than the rest)
12
13
14
The Model: The Preliminaries
• Let Uijrt denote the utility to person i from purchase of
technology j, where j=Apple (platform), IBM/compatible
(Windows Platform), in dollar amount Pr at time t=1,T. Let Pr
for r=1,R be a set of discrete dollar amounts that the
consumer may choose to spend on a computer.
• Assume that consumers have a utility function defined over
the efficiency units of computer capabilities they possess, G,
and consumption of an outside good, C.
• If a consumer sends Pr dollars on a computer, then his/her
consumption of the outside good is Cir = Ii – Pr, where Ii is the
consumer’s income.
15
Preliminaries continued
Gijrt = jtPrQj
• jt is an index of the efficiency units of
computer capabilities that one can purchase
by spending one dollar on technology j at
time t  can be thought as inverse price
indices.
• Qj is the per dollar quality of technology j
• jt Qj is the efficiency units of computer
capabilities that one can purchase by
spending one dollar on technology j at time t
Gijrt : the efficiency units of computer
capabilities that one can purchase by
spending Pr dollars on technology j at
time t
16
Utility Specification
U ijrt   i (1  exp{Gijrt})  Cijrt   ijrt
U ijrt   i {1  exp(  jt Pr Qij )} i Pr  ijrt
i indexes individual; j indexes technology; t indexes time.
Uijrt: the utility to person i from purchase of technology j in dollar amount Pr at time t.
 : risk aversion parameter
i : individual specific utility weight
 : price coefficient
jt: an index of the efficiency units of computer capabilities that one can purchase by
spending one dollar on technology j at time t
Qj : per dollar quality level of technology j.
Pr for r=1,..R is a set of discrete dollar amounts the consumer may choose to spend on a
computer
17
Forecasting Future Prices
ln  ij,t 1  E[ln ( j ,t 1 /  jt ) | I it ]  ij ,t 1
 ij ,t 1 ~ N (0,   )
2
ij,t+1: the inverse of the consumer’s report of his/her expectation of the
price decline from t to t+1.
E denotes the Expectation operator.
18
Price Expectations Process
Et [ln ( j ,t 1 /  jt )| I it ] 
 0 1 ln(  jt /  j ,t 1 ) 2 ln(  jt /  j ,t 2 )
If 0=2=0 and 1=1,
then consumers simply extrapolate the most recent
one period (inverse) price change into the future.
If 0=0, 1 > 2|2|, 2 < 0,
then consumers expect that whatever acceleration
or deceleration that occurred from t-2 to t will
continue in the future.
If 00, 0 < 1 < 1, 0 < 2 < 1, then the consumer expects the rate of price change
to revert to some “natural” rate.
19
Distribution of Future Prices
Point estimates of expected future prices are not sufficient
to solve the consumer’s dynamic choice problem  We
need to specify the expected distribution of future prices.
Assume that agents expect that the distribution of future
prices be:
ln  j ,t 1~N ( E[ln  j ,t 1 | I it ],   )
2
Iit : information set that individual i has at time t
20
Learning about Quality
• Consumers lean about quality of each technology through
information channel (source) signals in a Bayesian fashion;
for each signal we estimate the variability of the
information source
• Over time, their mean expectations will evolve and the
variance of their quality beliefs will shrink (provided that
the signals are not very noisy)
• We estimate mean quality perceptions for each technology
for two segments (discrete mass approach to capture
unobserved heterogeneity) but we also use consumer selfreports of quality expectations
21
Learning about
Quality of Technology
Consumer Priors about Quality Levels
Q j ~ N (Qoj , )
2
oj
Qj : quality of technology j.
Qoj : consumer’s prior expectation of the quality of computers type j
2oj: consumer’s prior uncertainty about about computers of type j
22
Information Channel Signals
S jkt ~ N (Q j ,  )
2
k
k indexes the information source,
Sjk : Signal from source k about technology type j.
k2 : Variance of the information provided by the k-th signal (inverse of
the precision of information provided by k)
[accuracy/precision of information]
23
Expected Quality of Technology
Q j ~ N ( E[Q j |I it ], ijt2 )
zijt ~ N (0, ijt2 )
E[Q j |I it ]  Q j  zijt ,
.,
  E{(Qj  E[Q j |I it ]) | I it }
2
ijt
2
z is the expectation error

2
ijt
is the variance of consumer beliefs of quality of
technology
I is the information set
24
Information obtained from
Information Channels
xijkt ~ N (0, )
Sijkt  Q j  xijkt
2
k
S denotes the signal received about the quality of technology
x is the random term (noise) associated with the information received
25
Expectations of Technology Quality
& Technology Quality Ratings
(8)
.
,
E[Q j |I it ]  Q j  zijt ,
qijt  L
qijt  M
qijt  H
if
zijt ~ N (0, ijt2 )
E[Q j | I it ]   jL
if  jL  E[Q j | I it ]   jH
if
E[Q j | I it ]   jH
26
Ordered Probit to Estimate the Cut-off
Points for the Signal Levels of L, M and
H for Quality of Technology
Pr ob(qijt  L)   ijt ( jL  Q j )
,
Pr ob(qijt  M )   ijt ( jH  Q j )   ijt ( jL  Q j )
Pr ob(qijt  H )  1   ijt ( jH  Q j )
•
 ijt
is the cumulative normal distribution function with
mean 0 and variance
2
 ijt
27
Consumer’s Dynamic Choice Problem
5
Vimt ( I it )   J km ck  E max{VitP ( I it ,m),VitN ( I it ,m)}  imt
k 1
V  max E[U ijrt |I it , m]
P
it
{ j ,r}
VitN U i 0   E max Vim,t 1[I i ,t 1 ]  eit
m
m indexes the combination of information sources (the m=1,..,32 search options, where
32=25); V denotes “Value function”; P denotes “Purchase”; NP denotes
“No=Purchase”
Jkm: an indicator for whether source k is included in combination m.
ck: the costs of obtaining information from each source k
imt: an i.i.d. stochastic shock to the cost of search option m at time t.
VP captures the maximum over all possible technology and price choices {j,r} of the
expected utilities associated with choices
Ui0 : per period utility from the current computer (if applicable)
28
Choice Probability for
Search Options
Pr( M imt  1 |  , zi,t 1 , ) 
exp(Vimt ( I it ))
32
 exp(Vilt ( I it ))
l 1
Pr(Mimt): Probability of search option m for person i at time t
Θ: the set of parameters
z: expectation errors
τ: latent class type
29
Choice Probabilities for Purchase
and No-Purchase
Pr( Dijrt  1 |  , zit , ) 
exp( E[U ijrt | I it ])
2
R
exp(VitN )   exp( E[U ilqt | I it ])
l 1 q 1
Pr( Di 0t  1 |  , z it , ) 
exp(VitN )
2 R
N
exp(Vit )    exp( E[U ilqt | I it ])
l 1 q 1
• D stands for probability, θ for the set of
parameters, z for expectation errors, and τ for
30
latent class type
Likelihood Function
Lit ( , zit , zi, t 1 , )  L1it ( , zi, t 1 , ) L2it ( , zit , ) L3it ( , zit , ) L4it ( )
•
L1 corresponds to information choices
•
L2 corresponds to purchase decisions
•
L3 corresponds to reported quality ratings
•
L4 corresponds to reported price expectations
31
Likelihood Function II
2
Li ( )     Li ( , zi , ) f ( zi )dzi
 1
zi
32
Parameter Estimates
Parameter Estimates
Price Process Parameters
Utility Parameters (continued)
Price Process Coefficents
0
1
2

(0.012)
3.917
(1.258)
1.239
(0.163)
Ownership of Computer
0.149
(0.061)
-0.958
(0.194)
Experience
0.009
(0.049)
0.088
(0.013)
Age
Education
Gender
0.644
-0.019
0.543
(0.278)
(0.089)
(0.245)
0.076
(0.043)
Income
-0.294
(0.124)
Measurement Error Std. Dev.

No Purchase Utility Coefficients
Constant
0.041
Utility Parameters
Information Search Process Parameters
Price Coefficent
0.478
Risk Aversion Coefficient
1.000
-
Prior Std.Dev. of Quality Perception
IBM
Apple
Discount Factor
0.995
-
Variability of Information
Utility Weight
Constant
Experience
Age
Education
Gender
Income
Quality Coefficients
Mean Quality( IBM)a
1.000
0.479
0.207
-0.734
0.639
-0.670
(0.238)
(0.103)
(0.257)
(0.283)
(0.399)
0.891
(0.397)
Mean Quality( Apple)a
-0.891
Mean Quality( IBM)b
-0.275
Mean Quality( Apple)b
0.275
Latent Class Probabilities
1st Latent Class
2nd Latent Class
0.879
0.121
a
1st Latent Class
b
2nd Latent Class
(0.193)
Store Visit
Ads and Articles
Word of Mouth
Information Costs
Store Visits
General Articles
Computer Articles
Advertising
Word of Mouth
(0.125)
-
0.927
2.071
(0.160)
(0.350)
1.977
4.741
3.128
(0.659)
(1.976)
(1.325)
1.415
0.642
1.137
0.379
0.318
(0.123)
(0.088)
(0.175)
(0.052)
(0.057)
Quality Perception Interval Coefficients
IBM-Left
-1.741
(0.267)
IBM-Right
Apple-Left
Apple-Right
2.867
-3.217
2.433
(0.419)
(0.569)
(0.388)
(0.155)
-
Note. Values enclosed in parentheses represent standard errors.
33
Parameter Estimates:
Price Process Parameters
Price process parameters
0=-0.041
, 1= 1.239
, 2= - 0.958,
= 0.088
 2 is negative and larger in absolute value than 1+ 2
consumers expect mean reversion in price declines
 If price stayed constant from t-2 to t then consumers would
expect a 4.1% price decline from t to t+1.
 A steady state expected rate of price decline of roughly
2.5% per two-month period.
 Substantial measurement error in consumer’s reports of their
own expectations.
34
Parameter Estimates II
•
Prior uncertainty is substantial. It is higher for Apple.
•
Two segments of consumers, first segment’s mean expected quality
is higher for IBM/Compatible than Apple; the opposite holds for the
second segment. The first segment constitutes 88% of the
individuals.
•
No-purchase utility is higher for individuals
• who 1) already own a computer, 2) older people, 3) women and 4)
lower income people. Education and experience do not have a
statistically significant effect on No-Purchase Utility.
•
Utility weight is higher
• 1) the more experienced they are with computers, 2) the older they
are, 3) the less educated they are, and 4) if they are male.
35
Parameter Estimates III:
Variability and Cost of the information Sources
• Computer magazines, and general sources and
advertising provide the noisiest information,
whereas store visits provide the most precise
information.
• Reading computer articles, followed by store
visits, seem to be the most costly sources,
whereas obtaining word of mouth information
seems to be the least costly one, followed by
reading advertisements.
36
Model Fit:
Comparing sample frequencies
of search and purchase behavior with
baseline simulations based on parameter estimates
• The model fits the data quite well.
It slightly overstates the No-Purchase rate in wave 6,
and overpredicts purchases in wave 1.
It slightly under predicts the extent of search.
It accurately predicts the relative utilization of each
source, as well as the time-path of utilization.
Percentage of people using each information source
declines over time both in the data and according to
the model. Evidence for duration dependence.
37
Model Fit: Table 4
TABLE 4
Sample Frequencies vs. Model Simulation: Purchase Behavior
Sample frequencies
Number of Purchases
No Purchase IBM Apple
281
0
0
261
19
1
225
18
2
179
22
1
145
17
0
102
17
1
Percentage of Purchases
No Purchase IBM Apple
100.0
0.0
0.0
92.9
6.8
0.4
91.8
7.3
0.8
88.6
10.9 0.5
89.5
10.5 0.0
85.0
14.2 0.8
Attrition Adjusted Sample Frequencies
Number of Purchases
Wave
Sample Size No Purchase IBM Apple
1
281
281
0
0
2
281
261
19
1
3
261
240
19
2
4
240
213
26
1
5
213
191
22
0
6
191
162
27
2
Total
281
162
113
6
Percentage of Purchases
No Purchase IBM Apple
100.0
0.0
0.0
92.9
6.8
0.4
91.8
7.3
0.8
88.6
10.9 0.5
89.5
10.5 0.0
85.0
14.2 0.8
57.7
40.2 2.1
Wave
1
2
3
4
5
6
Sample Size
281
281
245
202
162
120
38
Table 4 (con’t)
Model Simulation - Baseline
Wave
Sample Size
1
281.0
2
265.1
3
246.6
4
226.1
5
204.5
6
183.5
Out of Sample
7
163.6
8
142.4
9
120.7
10
100.9
11
81.1
12
64.8
First 6
Total
281
281
Number of Purchases
No Purchase IBM Apple
265.1
13.3 2.6
246.6
17.3 1.2
226.1
19.8 0.6
204.5
21.2 0.4
183.5
20.9 0.2
163.6
19.7 0.2
Percentage of Purchases
No Purchase IBM Apple
94.3
4.7
0.9
93.0
6.5
0.5
91.7
8.0
0.3
90.5
9.4
0.2
89.7
10.2 0.1
89.2
10.7 0.1
142.4
120.7
100.9
81.1
64.8
51.8
19.5
20.3
18.7
19.0
15.6
12.6
1.8
1.4
1.1
0.9
0.6
0.5
87.0
84.8
83.6
80.3
80.0
79.9
11.9
14.2
15.5
18.8
19.3
19.4
1.1
1.0
0.9
0.8
0.8
0.8
163.6
51.8
112.1 5.2
217.8 11.5
58.2
18.4
39.9
77.5
1.9
4.1
39
Model Fit II: Table 5
Sample Frequencies vs. Model Simulation: Search Behavior
Sample Frequencies
Number of Consumers Collecting Information
Wave
1
2
3
4
5
6
Sample Size
281
281
245
202
162
120
Store
Visits
181
115
82
77
46
35
General Computer
Word of
Articles Articles Advertising Mouth
164
148
188
247
116
98
163
204
101
87
132
136
87
83
121
116
64
53
77
94
50
41
59
76
Percentage of Consumers Collecting Information
Store
Visits
64.4
40.9
33.5
38.1
28.4
29.2
General Computer
Word of
Articles Articles Advertising Mouth
58.4
52.7
66.9
87.9
41.3
34.9
58.0
72.6
41.2
35.5
53.9
55.5
43.1
41.1
59.9
57.4
39.5
32.7
47.5
58.0
41.7
34.2
49.2
63.3
Attrition Adjusted Sample Frequencies
Number of Consumers Collecting Information
Wave
1
2
3
4
5
6
Sample Size
281
281
261
240
213
191
Store
Visits
181
115
87
91
60
55
Total
1467
589
General Computer
Word of
Articles Articles Advertising Mouth
164
148
188
247
116
98
163
204
107
92
140
144
103
98
143
137
84
69
101
123
79
65
93
120
653
570
828
975
40
Percentage of Consumers Collecting Information
Store
Visits
64.4
40.9
33.5
38.1
28.4
29.2
40.1
General Computer
Word of
Articles Articles Advertising Mouth
58.4
52.7
66.9
87.9
41.3
34.9
58.0
72.6
41.2
35.5
53.9
55.5
43.1
41.1
59.9
57.4
39.5
32.7
47.5
58.0
41.7
34.2
49.2
63.3
44.5
38.9
56.4
66.5
Table 5 (con’t)
Model Simulation - Baseline
Number of Consumers Collecting Information
Wave Sample Size
1
281
2
265
3
247
4
226
5
205
6
184
Out of Sample
7
164
8
143
9
122
10
102
11
82
12
65
Store
Visits
170
88
71
83
51
47
General Computer
Word of
Articles Articles Advertising Mouth
164
139
189
231
103
81
155
179
92
83
129
126
84
78
125
123
72
57
89
106
65
55
81
109
Percentage of Consumers Collecting Information
Store
Visits
60.5
33.2
28.7
36.7
24.9
25.5
General Computer
Word of
Articles Articles Advertising Mouth
58.4
49.5
67.3
82.2
38.9
30.6
58.5
67.5
37.2
33.6
52.2
51.0
37.2
34.5
55.3
54.4
35.1
27.8
43.4
51.7
35.3
29.9
44.0
59.2
42
35
29
24
19
16
59
47
39
36
27
21
46
40
34
27
23
19
70
58
46
41
31
26
98
76
62
50
40
32
25.6
24.5
23.8
23.5
23.2
24.6
36.0
32.9
32.0
35.3
32.9
32.3
28.0
28.0
27.9
26.5
28.0
29.2
42.7
40.6
37.7
40.2
37.8
40.0
59.8
53.1
50.8
49.0
48.8
49.2
First 6
1408
510
580
493
768
874
36.2
41.2
35.0
54.5
62.1
Total
2086
675
809
682
1040
1232
32.4
38.8
32.7
49.9
59.1
41
Policy Experiments I: Effects of
Expected Price Declines and Learning
on Purchase Delay
• No expectations of price decline
• Relatively small effect of expected price
declines
– Acceleration of purchases
– However, total purchases are almost unchanged at
T=12
• Substantial increase in search costs
• Lower computer sales overall
• Decreases positive duration dependence
42
Policy Experiments II:
Effects of Price Expectations on
Price Elasticities of Demand
• Temporary %20 price cut
 Expectations adjust versus
 They are fixed
• Elasticity of demand wrt transitory price cur is almost
50% greater if we allow price expectations to adjust
rather than holding them fixed
• Acceleration of purchases and total purchases
 Incremental sales due to the price cut only 4.6% at T=12
but 80% at T=6.
 Price elasticity higher for Apple
43
Policy Experiments III
• EFFECT OF DECERASE IN INFORMATION COSTS
ON PUCRHASE BEHAVIOR
Decreasing information costs leads to
acceleration of purchases
 Decreasing information costs affects Apple sales
more than IBM/Compatible sales
Advertising has the least effects
44
Policy Experiments IV
• EFFECTS OF INCREASING THE PRECISION OF INFORMATION
ON SEARCH AND PURCHASE BEHAVIOR
 Decreasing the variability of information sources
(increasing the precision) encourages consumers to search
slightly more (own variability effects are positive) and
accelerates purchases
 Most cross-effects are positive (except for negative effect
of store visits precision on advertising and general
sources)
 Increased precision helps Apple relatively more
45
Conclusions & Future Research
• Consumers are forward-looking
• Both learning and price expectations affect purchase timing and
purchase decision
• Ignoring price expectations severely bias price elasticities
• Cost and accuracy of information sources affect consumer search
behavior
• Availability of survey data about expectations facilitate
identification
• Future research:
 Information search: Learning to learn?
 Brand choice, consideration set formation
46
Appendix
47
Simulated Effects of Changing Price Expectations and Raising Search Costs
Table 6
A. Consumers Expect No Downward Trend in Prices
Number of Purchases
Wave
Sample Size
No Purchase IBM Apple
1
281.0
263.7
14.4
2.8
2
263.7
241.5
20.1
2.2
3
241.5
217.2
22.9
1.3
4
217.2
193.6
22.8
0.8
5
193.6
171.2
21.9
0.5
6
171.2
150.9
20.0
0.3
Out of Sample
7
150.9
131.4
18.1
1.4
8
131.4
111.2
19.0
1.1
9
111.2
92.2
17.7
1.3
10
92.2
75.8
15.8
0.7
11
75.8
62.1
13.1
0.5
12
62.1
50.9
10.9
0.3
First 6
Total
281
281
150.9
50.9
122.2
216.7
7.9
13.4
Percentage of Purchases
No Purchase IBM Apple
93.8
5.1
1.0
91.6
7.6
0.8
90.0
9.5
0.5
89.1
10.5
0.4
88.4
11.3
0.3
88.1
11.7
0.2
87.1
84.6
82.9
82.2
82.0
81.9
12.0
14.5
15.9
17.1
17.3
17.5
1.0
0.9
1.2
0.7
0.7
0.6
53.7
18.1
43.5
77.1
2.8
4.8
B. Increase Search Cost 60%
Number of Purchases
No Purchase IBM Apple
266.3
13.3
1.4
248.0
17.3
1.0
228.2
19.1
0.7
209.3
18.5
0.4
190.8
18.3
0.2
172.8
18.0
0.1
Wave
1
2
3
4
5
6
Out of Sample
7
8
9
10
11
12
Sample Size
281.0
266.3
248.0
228.2
209.3
190.8
172.8
154.4
135.7
117.6
100.8
86.4
154.4
135.7
117.6
100.8
86.4
74.0
First 6
Total
281
281
172.8
74.0
48
17.5
17.9
17.5
16.3
14.0
12.2
104.5
199.8
Percentage of Purchases
No Purchase IBM Apple
94.8
4.7
0.5
93.1
6.5
0.4
92.0
7.7
0.3
91.7
8.1
0.2
91.1
8.7
0.1
90.5
9.4
0.0
0.9
0.8
0.7
0.5
0.4
0.3
89.4
87.9
86.6
85.7
85.7
85.6
10.1
11.6
12.9
13.8
13.9
14.1
0.5
0.5
0.5
0.5
0.4
0.3
3.7
7.3
61.5
26.3
37.2
71.1
1.3
2.6
Price Cut Simulations
Table 7
A. 20% Price Decline in Wave 2 - Total Effect
Number of Purchases
Wave
Sample Size No Purchase IBM Apple
1
281.0
265.1
13.3
2.6
2
265.1
234.2
28.4
2.5
3
234.2
214.4
18.8
1.0
4
214.4
192.8
21.1
0.6
5
192.8
172.6
19.8
0.4
6
172.6
153.7
18.7
0.2
Out of Sample
7
153.7
133.4
18.5
1.8
8
133.4
112.8
19.3
1.4
9
112.8
93.9
17.8
1.1
10
93.9
75.0
18.1
0.8
11
75.0
59.5
14.9
0.6
12
59.5
47.1
11.9
0.5
First 6
Total
281
281
153.7
47.1
120.0
220.4
7.3
13.5
Percentage of Purchases
No Purchase IBM Apple
94.3
4.7
0.9
88.3
10.7
1.0
91.6
8.0
0.4
89.9
9.8
0.3
89.5
10.3
0.2
89.1
10.8
0.1
86.8
84.5
83.2
79.9
79.4
79.1
12.0
14.4
15.8
19.2
19.8
20.1
1.2
1.1
1.0
0.9
0.8
0.8
54.7
16.8
42.7
78.4
2.6
4.8
B. 20% Price Decline in Wave 2 - Expected Future Price Changes Held Fixed
Number of Purchases
Percentage of Purchases
Wave
Sample Size No Purchase IBM Apple
No Purchase IBM Apple
1
281.0
265.1
13.3
2.6
94.3
4.7
0.9
2
265.1
237.9
25.1
2.0
89.8
9.5
0.8
3
237.9
218.2
19.1
0.6
91.7
8.0
0.3
4
218.2
197.0
20.7
0.6
90.3
9.5
0.3
5
197.0
176.7
20.1
0.2
89.7
10.2
0.1
6
176.7
157.5
19.0
0.2
89.2
10.7
0.1
Out of Sample
7
157.5
137.1
18.8
1.7
87.0
11.9
1.1
8
137.1
116.2
19.6
1.3
84.8
14.3
1.0
9
116.2
97.1
18.1
1.0
83.5
15.6
0.9
10
97.1
77.9
18.3
0.8
80.3
18.9
0.9
11
77.9
62.2
15.1
0.6
79.8
19.4
0.8
12
62.2
49.6
12.1
0.5
79.7
19.5
0.8
First 6
Total
281
281
157.5
49.6
117.3
219.2
49
6.2
12.2
56.1
17.6
41.7
78.0
2.2
4.3
Table 8
Simulated Effects of Decreasing Search Costs
A. 20% Decrease in the Cost of a Store Visit
Wave
Sample Size
1
281.0
2
258.1
3
233.9
4
209.6
5
187.4
6
165.6
Out of Sample
7
145.8
8
125.1
9
104.8
10
86.1
11
68.8
12
54.6
Number of Purchases
No Purchase IBM Apple
258.1
18.2
4.7
233.9
21.4
2.7
209.6
22.7
1.6
187.4
21.3
1.0
165.6
21.2
0.6
145.8
19.3
0.5
Percentage of Purchases
No Purchase IBM Apple
91.8
6.5
1.7
90.6
8.3
1.0
89.6
9.7
0.7
89.4
10.1
0.5
88.4
11.3
0.3
88.1
11.7
0.3
125.1
104.8
86.1
68.8
54.6
42.4
18.9
18.8
17.5
16.4
13.6
11.7
1.9
1.5
1.2
0.9
0.7
0.4
85.8
83.8
82.1
79.9
79.3
77.7
12.9
15.0
16.7
19.0
19.7
21.5
1.3
1.2
1.1
1.1
1.0
0.8
First 6
281
145.8
124.1
11.0
51.9
44.2
3.9
Total
281
42.4
221.0
17.6
15.1
78.7
6.2
50
Table 8 (con’t)
B. 20% Decrease in the Cost of Gathering Information from Advertising
Wave
Sample Size
1
281.0
2
261.4
3
239.5
4
216.2
5
194.5
6
172.2
Out of Sample
7
151.8
8
130.8
9
110.0
10
91.5
11
73.5
12
58.9
First 6
Total
281
281
Number of Purchases
No Purchase IBM Apple
261.4
15.6
3.9
239.5
20.1
1.8
216.2
22.4
0.9
194.5
21.1
0.5
172.2
22.0
0.3
151.8
20.2
0.2
Percentage of Purchases
No Purchase IBM Apple
93.0
5.6
1.4
91.6
7.7
0.7
90.3
9.4
0.4
90.0
9.8
0.2
88.5
11.3
0.2
88.1
11.7
0.1
130.8
110.0
91.5
73.5
58.9
47.3
19.4
19.5
17.6
17.2
14.0
11.2
1.6
1.3
1.0
0.8
0.6
0.4
86.1
84.1
83.1
80.3
80.2
80.3
12.8
14.9
16.0
18.8
19.0
19.0
1.1
1.0
0.9
0.8
0.8
0.7
151.8
47.3
121.5
220.4
7.7
13.3
54.0
16.8
43.2
78.4
2.7
4.7
51
Table 9
Simulated Effects of Making Information Sources More Accurate
A. Store Visit - 20% Increase in Precision
Number of Purchases
No Purchase
IBM
Apple
256.7
18.7
5.6
231.5
21.7
3.5
206.9
22.9
1.7
183.3
22.6
1.0
161.4
21.2
0.7
141.7
19.2
0.5
Percentage
No Purchase
91.4
90.2
89.4
88.6
88.1
87.8
of Purchases
IBM
Apple
6.6
2.0
8.5
1.4
9.9
0.7
10.9
0.5
11.5
0.4
11.9
0.3
Wave
1
2
3
4
5
6
Out of Sample
7
8
9
10
11
12
Sample Size
281.0
256.7
231.5
206.9
183.3
161.4
141.7
121.1
101.0
83.3
66.2
52.4
121.1
101.0
83.3
66.2
52.4
41.4
18.7
18.6
16.5
16.1
13.2
10.6
1.9
1.5
1.2
0.9
0.7
0.4
85.4
83.4
82.5
79.5
79.1
79.0
13.2
15.3
16.4
19.4
19.9
20.2
1.3
1.2
1.2
1.1
1.0
0.8
First 6
281
141.7
126.2
13.1
50.4
44.9
4.7
Total
281
41.4
219.9
19.7
14.7
78.3
7.0
B. Advertising, Articles in General and Computer Publications - 20% Increase in Precision
Number of Purchases
No Purchase
IBM
Apple
256.3
18.0
6.7
230.2
22.5
3.6
204.3
24.2
1.7
179.4
23.9
1.0
156.7
22.0
0.6
136.5
19.8
0.4
Percentage
No Purchase
91.2
89.8
88.7
87.8
87.4
87.1
of Purchases
IBM
Apple
6.4
2.4
8.8
1.4
10.5
0.7
11.7
0.5
12.3
0.3
12.6
0.3
Wave
1
2
3
4
5
6
Out of Sample
7
8
9
10
11
12
Sample Size
281.0
256.3
230.2
204.3
179.4
156.7
136.5
116.3
96.8
78.6
62.3
49.2
116.3
96.8
78.6
62.3
49.2
38.2
18.5
18.2
17.2
15.6
12.5
10.5
1.7
1.3
1.0
0.8
0.6
0.5
85.2
83.3
81.2
79.2
79.0
77.6
13.6
15.6
17.7
19.8
20.1
21.4
1.2
1.1
1.1
1.0
0.9
1.1
First 6
281
136.5
130.5
14.0
48.6
46.4
5.0
Total
281
38.2
222.9
19.9
13.6
79.3
7.1
52
Table 9 (con’t)
C. Word-of-Mouth - 20% Increase in Precision
Wave
Sample Size
1
281.0
2
261.3
3
235.4
4
209.9
5
185.8
6
163.1
Out of Sample
7
142.3
8
121.5
9
101.5
10
84.0
11
68.1
12
55.0
First 6
Total
281
281
Number of Purchases
No Purchase IBM
Apple
261.3
15.2
4.5
235.4
23.5
2.5
209.9
24.4
1.1
185.8
23.5
0.6
163.1
22.3
0.4
142.3
20.6
0.2
Percentage of Purchases
No Purchase IBM
Apple
93.0
5.4
1.6
90.1
9.0
1.0
89.2
10.4
0.5
88.5
11.2
0.3
87.8
12.0
0.2
87.2
12.6
0.1
121.5
101.5
84.0
68.1
55.0
44.6
19.3
18.9
16.6
15.3
12.5
10.0
1.5
1.1
0.9
0.7
0.5
0.4
85.4
83.5
82.7
81.0
80.8
81.1
13.6
15.5
16.4
18.2
18.4
18.2
1.0
0.9
0.9
0.8
0.7
0.7
142.3
44.6
129.5
222.1
9.2
14.3
50.6
15.9
46.1
79.0
3.3
5.1
53