SIMS Hal Varian

November 1999

Pravin Varaiya, Hal VarianRichard Edell, Karyen Chu, Walter Beckert

EECS, SIMS, Economics at UC Berkeley

INDEX: a status report

INDEX Hal VarianPage 2

Internet use currently has zero marginal cost for useResult: congestion (see next slide)

Internet currently uses one quality of service (QoS)Differential quality of service requires differential pricing

Research on QoSHow to engineer QoS: a lot of workWhat QoS costs: a little workHow users value QoS: nothing

Goal of INDEX project: measure how users value QoS

Motivation

INDEX Hal VarianPage 3

Modem use by day and user

0

20

40

60

80

100

120

140

160

180

200

220

240

260

Time of day

Mod

ems

02 04 06 08 10 12 14 16 18 20 22 00 02

256 14.4 Kbps

112 2.4 Kbps

60 1.2 Kbps

32 19.2 Kbps (15min limit)

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

80

90

100

% Users

% T

ime

51.2%

69.8%

INDEX Hal VarianPage 4

INDEX = Internet DEmand eXperiment

Web page: http://www.index.berkeley.edu

Support: Cisco Systems, Pacific Bell, Hewlett Packard,and National Science Foundation.

Information about INDEX

INDEX Hal VarianPage 5

Goal: Measure how much users are willing to pay forInternet Quality of Service (QoS).

Dimensions of QoSbandwidth (symmetric and asymmetric)volumecongestiondelay to connectetc.

What is our goal?

INDEX Hal VarianPage 6

•Give 150 Berkeley people "free" ISDN servicehardware and setup costs are freewe cover PacBell’s monthly charges

•Offer them different QoS for different prices•Simulate the QoS

same QoS choices for several weeksdifferent prices each week

•Note: real money, simulated service

How do we do it?

INDEX Hal VarianPage 7

Architecture

CampusNetwork

&Internet

Cisco 7200(DCNS maintains)

BGW(Intel PCs)

OfflineAnalysis

Supervisor

(also WWW& DBMS)

Cat 5000SwitchedEthernet NMAS

Cisco7513

PSTN

Cisco762s

Cisco7507

ISDN PRIs

ISDN BRI

1 2 3 4

IP IP

PPPUDP

IP

PPP

IP

4

4

1

4

4

2

2

2

3

3

3

INDEX Hal VarianPage 8

Interface for INDEX: settings

INDEX Hal VarianPage 9

Interface: experiment

INDEX Hal VarianPage 10

Interface: pricing

INDEX Hal VarianPage 11

Interface: choices

INDEX Hal VarianPage 12

Interface: asymmetric BW

INDEX Hal VarianPage 13

Interface: volume pricing

INDEX Hal VarianPage 14

Interface: billing

INDEX Hal VarianPage 15

Started providing service April 1, 1998have about 70 subjectssubjects proceed through experiments asynchronously

Have runsymmetric bandwidthasymmetric bandwidthvolume pricingbandwidth + volume mixture (predetermined +self-selected)pay upfront for flat pricing

Where we are

INDEX Hal VarianPage 16

Usage overall:

Symmetric bandwidth

8 kbs

16 kbs

32 kbs

64 kbs

96 kbs

128 kbsNote: 8 kbs is free

INDEX Hal VarianPage 17

Scatter plots

1000 2000 3000 4000u16

0.5

1

1.5

2

2.5

price

5000 10000 15000 20000 25000u32

1

2

3

4

price

16 Kbps 32 Kbs

INDEX Hal VarianPage 18

Scatter plots, continued

2500 5000 7500 10000 12500 15000 17500u64

2

4

6

8

10

price

2000 4000 6000 8000u96

2

4

6

8

10

12

14

price

2500 5000 7500 10000 12500 15000 17500u128

2.5

5

7.5

10

12.5

15

17.5

price

64 Kbs96 Kbs

128 Kbs

INDEX Hal VarianPage 19

Reduced form: estimate demand as a function of priceamount consumed = function of prices and individualcharacteristics

Structural: estimate parameters of utility functionchoice of bandwidth depends on value of time, urgency,etc. as well as price and characteristics

First look at reduced form, then examine some options forstructural estimates

Demand estimates

INDEX Hal VarianPage 20

u128= -2.0p128 +. 80p96 +.25p64 -.02p32 -.16p16

u96 =+1.7p128 -3.1p96 +.43p64 +.19p32 +.18p16

u64 = +.77p128 +1.8p96 -2.9p64 +.59p32 +.21p16

u32 = +.81p128 -1.0p96 +1.0p64 -1.4p32 +.15p16

u16 = +0.2p128 -.29p96 +.04p64 +1.2p32 -1.3p16

Red = negative own price effectLight = not statistically significant

Lesson: large negative own price effect, positive cross priceeffect

Log Regressions

INDEX Hal VarianPage 21

R2

ISE no ISEu128 .95 .11u96 .93 .25u64 .92 .18u32 .95 .14u16 .90 .17

ISE = "individual specific effect"; interpretationConclusion: very good fits!

How good is the fit?

INDEX Hal VarianPage 22

Notationb = bandwidth chosenx = bits transferredt = time at bandwidth b = x/b

p(b) = price of bandwidth bUtility

u(x) - [c+p(b)]t

c is value of time; varies with circumstancesf(c) is probability distribution of c

Structural estimates

INDEX Hal VarianPage 23

Let (x,b*) be choice. Optimization requires:

Rearrange to find

Note that this gives upper and lower bounds on cost of time.

Choice model

minb∗<b

p(b∗)b − p(b)b∗b∗ − b ≥ c ≥ max

b∗>b

p(b∗)b− p(b)b∗b∗ − b

u(x)− [c+ p(b∗)]x

b∗≥ u(x)− [c+ p(b)]

x

bfor all b.

INDEX Hal VarianPage 24

Plot K(c) = [c + p(bi)] (1/bi) for i = 1,...,n

Observe (cU,cL,f ) for i= 1,...,n

Picture of estimation method

c

totalcost

b1

b2

b3

cL cU

i i i

INDEX Hal VarianPage 25

Histograms for one user

2 4 6 8 10 12 14

0.2

0.3

0.4

0.5

2 4 6 8 10 12 14

0.1

0.2

0.3

0.4

0.5

2 4 6 8 10 12 14

0.2

0.4

0.6

0.8

2 4 6 8 10 12 14

0.2

0.4

0.6

0.8

INDEX Hal VarianPage 26

Can calculate average time cost for each individual usingthese frequency distributions.Table shows number of subject with average estimated timecost that falls in indicated range:

Distribution of time cost

Range 1 2 3 4 5 6 7 8 9 10 11 12

Upper bound 39 8 3 4 1 2 2 1 2 0 3 0Lower bound 63 3 0 0 1 0 0 0 0 0 0 0

Average 47 7 2 3 3 3 1 3 1 1 0 0

INDEX Hal VarianPage 27

Who has high WTP?

Prof Tech Admin StudentOccupation

2

4

6

8

Est. Time Cost

INDEX Hal VarianPage 28

Very low WTP for time (= 1/2 cent per minute)• very small charges induce large changes in behavior• lot of low-value traffic out there• admin/technical have highest WTP

WTP for population as a wholecLowMean = 0.02cUpMean = 0.40cLowStd = 0.07cUpStd = 1.11

Average cost of time

INDEX Hal VarianPage 29

Suppose frequencies drawn from prob distribution p(c,β)

Find distribution that is close to all the frequencies

Closeness = Kullback-Leibler entropy measure, Σ f log p

Parametric fit

maxβ

n∑i=1

fi log∫ ciU

ciL

p(x, β) dx

INDEX Hal VarianPage 30

µ=.487, σ=.32

mean = .52

Truncated Normal

0.5 1 1.5 2 2.5 3 3.5

0.2

0.4

0.6

0.8

1

INDEX Hal VarianPage 31

µ=.48, σ=.32 µ=.42. σ=.25

max entropy min SSR

Picture of CDF

0.5 1 1.5 2 2.5 3

0.2

0.4

0.6

0.8

1

0.5 1 1.5 2 2.5 3

0.2

0.4

0.6

0.8

1

INDEX Hal VarianPage 32

What about population as a whole?cLowMean = 0.02, cUpMean = 0.40cLowStd = 0.07, cUpStd = 1.11

Estimated parameters: µ=-.16, σ=.64, mean time cost = .18

Entire Population

0.5 1 1.5 2 2.5 3 3.5 4

0.2

0.4

0.6

0.8

1

INDEX Hal VarianPage 33

Possible explanations1. Our population is not representative

can look at demographics, e.g., income effects2. We only get lower bound on WTP for 128 kbs

but this is only about 1/6 of use3. We can’t control QoS outside of Berkeley network

high bandwidth isn’t worth much if congestion iselsewhere

4. We can only measure value of existing applications

Why such a low WTP?

INDEX Hal VarianPage 34

Two bandwidths8 Kbs for free128 Kbs for money

Price volume downloadedprices range from 1 to 20 cents per megabytemonthly expenditures similar in bandwidth and volumeexperiments

Volume experiments

INDEX Hal VarianPage 35

Scatterplot

INDEX Hal VarianPage 36

no ISE: u128 = -0.29 p128 R2 = 0.02

ISE: u128 = -0.24 p128 R2= 1.00

Interesting finding:amount transmitted is not sensitive to priceentire effect is individual-specific effect

Why? Probably choice offered is too extreme...

Log Regressions

INDEX Hal VarianPage 37

Current experiment: pay flat fee on Sunday to opt out ofmetered pricingMove to ADSL or cable modem?What are you interested in?

What’s next?