Estimating Consumer Preferences for Internet Access Service
IV. Data and Econometric Method
A. Survey
Data for estimating consumer preferences for bundled Internet access attributes are obtained from a nationwide mail survey during September and October 2002. The survey questionnaire comprised of three sections: cognitive buildup; choice task; and demographics. Cognitive buildup asks respondent’s 15 questions about their access and use of IT and the Internet, and provides them with information to form preferences about Internet access service. Respondents are then required to evaluate eight choice questions, and provide answers to nine demographic questions. Prior to implementation, the survey questionnaire was pre-tested on ten respondents during May and July 2002, and refined accordingly.
PA Consulting Group administered the survey. Advance postcards describing the survey to residents were mailed on September 3, and an initial survey package (with the questionnaire and incentive) was mailed on September 6. Thank you/reminder postcards were sent on September 13, and follow up surveys on October 3. When the survey was closed on October 25, 378 completed questionnaires were obtained for a response rate of 32 percent, which is about average for surveys of similar length and complexity (Louviere et al 2000). The mean completion time for each questionnaire was 26 minutes (n = 322).
B. Econometric method
Conjoint analysis is the primary tool used to estimate the demand for Internet access service. Survey respondents answered a series of eight choice questions. Each choice occasion presented a pair of Internet access options, A and B, that differed by five attributes. Respondents indicated their preferred choice. In addition, respondents indicated whether they would switch to the service they had selected if they were already online, or if they would adopt the service selected if they were not (See
Figure B1 in Appendix B for a choice question example). The parameters of the representative individual’s utility function (the marginal utilities of the five attributes) are estimated from observed choices.
Internet access service is described by five attributes.51 Access is always on when no dialup is required for Internet connection, and respondents can use the Internet and place telephone calls at the same time. Cost is the fixed monthly price for unlimited usage, ranging from $10 to $85. Speed describes the time it takes to receive and send information to and from the home computer. Speed is either very fast for uploads and downloads (very fast), or fast for downloads but relatively slower for uploads (fast), or same as dial-up (slow). Installation of Internet access service can be immediate, within one week, and within several weeks. Finally, very reliable Internet access is never disrupted (i.e., there are no service outages); however, with less reliable Internet access users may occasionally experience outages that require customer support. Table 1 summarizes the levels of the five attributes.
Table 1. Internet access service attributes Attribute Levels
Always on (AO) 1 Always on 2 Not always on Cost per month (COST) $10 to $85
Access speed (SPEED) 1 Very fast (download is 20 × dial-up; upload is 20 × dial-up) 2 Fast (download is 10 × dial-up; upload is 5 × dial-up) 3 Slow (same as dial-up)
Installation (INSTALL) 1 Immediately 2 Within one week 3 Within several weeks Reliability (RELIABLE) 1 Very reliable
2 Less reliable
Theory indicates that respondents maximize their (household’s) utility of the service option conditional on all other consumption and time allocation decisions. A linear approximation to the household conditional utility function is:
51 Theory, received evidence, and industry discussion provided feedback on service descriptions, the definition of attributes, and their levels.
U* = β1AO + β2SPEED + β3COST + β4INSTALL + β5RELIABLE + ε (4)
where the β’s are parameters to be estimated, and ε is a random disturbance. Note that the alternative attributes have been coded for econometric estimation so that the expected signs for β1 through β5 are negative. For instance, utility is expected to be less when cost increases so β3 < 0, but we also expect β4 < 0, as higher values for INSTALL imply less desirable outcomes. The hypothetical utility of each service option, U*, is of course not revealed. Instead, what is known is which option has the highest utility. For example, when a respondent chooses A over B and then the status quo (SQ) over A, it is assumed that U*A > U*B and U*SQ > U*A. Therefore, for this kind of dichotomous choice data, the method of estimation is not linear regression, but rather a form of maximum likelihood analysis called bivariate probit. Essentially, the probability of the outcome for each respondent-choice occasion is written as a function of the data and the parameters (β’s). The probability of the entire set of outcomes (all individuals, all choice occasions) is called the likelihood, and this is maximized for choice of the parameters (See Appendix C for further description of the estimation method).
Interpretation of the parameters as marginal utilities is the same as a partial derivative: the increase in utility for a one unit increase in the variable. For example, when a less reliable service (RELIABLE = 2) can be made more reliable (RELIABLE =1), utility would increase by β5 units. Since utility does not have an understandable metric, it is convenient to put this change in dollar terms. This is done by employing the economic construct called willingness to pay. The WTP for a one unit decrease in RELIABLE (the discrete improvement from less to very reliable) can be interpreted as how much more the service would have to be priced to make a consumer just indifferent between the old (cheaper but less reliable) service and new (very reliable) service. The required change in cost to offset an increase of β5 in utility is, from equation (4), β5/β3. This is true for any attribute. The WTP for a one unit improvement in that attribute is the ratio of its marginal utility to the marginal utility of COST.
Individuals may not have identical preferences. An individual’s preference toward speed, for example, may differ because of observable demographic characteristics, or may be idiosyncratic. This issue can be examined by estimating (4) on sub-samples of the data. This has the effect of allowing all parameters to be different for individuals in different socioeconomic groups. It is also possible to observe differences in the marginal utility of specific service attributes by interacting those characteristics with demographic variables. For instance, suppose individuals with different levels of education value speed differently. A model that captures this difference is:
where η is an additional parameter to be estimated, and EDUC is education.52 Here, the WTP for a one-unit improvement in speed is β2/β3 when education is not important. When education is important, the WTP for a one-unit improvement in speed is now: 3 2
)
(
β
η
β
+
EDUC
(6)and is evaluated at different levels of education.
Finally, an individual's preference toward speed may differ because of unobservable characteristics. One parameterization of this is the random parameters model. For a random speed parameter, for example, the model is:
U* = β1AO + (β2 + ν)SPEED + β3COST + β4INSTALL + β5RELIABLE + ε (7)
where ν is a zero-mean, white noise disturbance uncorrelated with the attributes or ε. The additional parameter now estimated is the variance of ν. It is possible to estimate this variance with survey data because there are multiple observations (i.e., choice occasions) for each individual. Mean WTP for an improvement in speed is β2/β3, as in the fixed parameter case.