Neighborhood effects and trial on the internet: Evidence from online grocery retailing

For example, Progressive Grocer (April 1995) reports that location explains up to 70% of the variance in consumer choice of supermarket retailers. Moreover, the attractiveness of an outlet to a shopper declines exponentially the further the individual is from the store (e.g., Fotheringham 1988; Huff 1964).

For reasons of confidentiality we have excluded the dollar values from panel (a), however all 48 contiguous states generate revenue.

The former term is preferred by sociologists and the latter by economists, but have complementary interpretations: Neighborhood effects emanate from the influence of well-defined exogenous groups on a focal group, whereas social influence refers to the broader behavioral process. We focus on an empirically grounded neighborhood effect without speculation as to the exact mechanism.

Specifically, a discrete time hazard model estimated on time-dependent trial data is consistent with random utility maximization over binary outcomes.

This conceptualization can be traced back to early work by Howard and Sheth (1969).

For a complete treatment of this issue see Anselin (1988) and Manski (1993).

Please see Appendix. Parameters of discrete time models are usually not invariant to the length of time intervals chosen (Heckman and Singer 1984a; Ryu 1995); the discrete time model with complementary log-log link function is the exception (see also Allison (2001 p. 216-219.). Ter Hofstede and Wedel (1999) document aggregation biases in discrete time models. Ryu (1995) shows that even for a standard discrete time model a time interval to average event time ratio of 1/16 is generally sufficient to mitigate bias.

Conceptually, this says that first individual to try in a region was a “local innovator” and not influenced by others in the region. This is behaviorally plausible because none of the other same-region individuals had in fact tried. At the same time, the model will however allow for the first trier in region z to be influenced by prior triers in region j, if region j is a neighbor of z. Details follow shortly.

We also estimate \(V_{z}(t) = \alpha_z + \beta X_{z}(t) + \ln(n_z), \ \ \alpha_z = \alpha_0 + \eta_z \ \ \eta_z \sim N (0, \sigma^2)\). Results are discussed in the next section.

As noted earlier, we do not distinguish between the two. Passive observation is facilitated by individuals observing deliveries (each box is clearly marked with “netgrocer.com”). Unfortunately, we cannot address Internet-based communication directly as we have no way to track it.

In order to check that results are robust to alternative formulations, we estimate a wide variety of alternative models that use different standard definitions for these two constructs. The results are reported in a subsequent section. We thank an anonymous reviewer for suggesting that the number of individuals (or a function thereof) as the measure for π _z (t − 1) gives the most behaviorally meaningful interpretation. Our final specification relies on ln [w _z Y _z (t − 1) + 1], and a previous working paper reports (qualitatively similar) results from a model that uses equal influence weights and binary indicators of lagged neighbor behavior.

Manski (1993) distinguishes endogenous from contextual and correlated effects. In the context of this study a true endogenous neighborhood effect exists if, all else equal, the probability of trial for a focal region varies with a measure of the average probability of trial for the reference group. A contextual effect exists if probability of trial varies according to the socio-economic characteristics of the reference group. A correlated effect is present if individuals in the same location behave similarly as a result of selection or exposure to similar environmental stimuli (retail stores, etc.).

We thank an anonymous reviewer for explicitly suggesting this approach.

Netgrocer periodically ran specials to entice new customers, or encourage existing customers to buy larger orders. We are unable to separately account for orders that might have resulted from such promotions. Subsequent to the data collection period management introduced a non-linear shipping fee schedule. Fees are differentiated by order size and region of the country (larger orders to western states are more expensive).

Based on 161,778 shopping trips taken by the 1,042 consumers in the Stanford Market Basket Database. Those data cover June 1991–June 1993. The inflation-adjusted average order value for the period of the Netgrocer data is approximately $30.

Netgrocer shipped 25,132 orders to 1,508 non-residential zip codes which were predominantly Army Post Office (APO) addresses. The average dollar value of these orders is $69.12 and we have excluded them from our analysis. Detailed information is available upon request.

This visual pattern is representative of other months and regions. In the interests of brevity other figures are not shown but are available from the authors upon request.

We control for over time variation in zip code level access to the Internet; one could also use such data to model changes in the “at risk” population. While only relatively crude controls for the size of the risk set are available, the substantive results from a variety of formulations produce qualitatively identical results.

We are indebted to an anonymous reviewer for suggesting this check.

We thank an anonymous reviewer for prompting this analysis. Our choice of 5 month intervals is somewhat arbitrary, however popular press articles and some leading market research companies suggest that new product related “buzz” can last for up to 20 weeks. See Bzzagent.com.

We are grateful to anonymous reviewers for providing references that led to this Appendix.