Estimating level effects in diffusion of a new technology: barcode scanning at the checkout counter

Abstract

Cross-country or cross-industry studies of technology diffusion typically estimate how independent factors affect diffusion speed or timing, often based on a two-stage approach. In many applications, however, countries (industries) differ most in the saturation level of diffusion. In a single-stage econometric approach to a standard diffusion model, we therefore estimate how the saturation level covaries with independent factors. In our application to diffusion of an important retail information technology, we focus on the competitive effect of hypermarkets (superstores). We also find standard scale, income and labour substitution effects.

Acknowledgements

Thanks to two anonymous referees, Joe Clougherty, Lapo Filistrucchi, Oz Shy, Irina Suleymanova and participants at the EEA conference in Vienna, the EARIE conference in Porto, the 5th ZEW Conference on the Economics of ICT in Mannheim, the BDPEMS workshop, the InterVal yearly meeting and the WZB seminar for comments and helpful discussions. Anna Kälberer and Kemal Azun provided able research assistance. Financial support from the German Federal Ministry of Education and Research, project InterVal – Internet and value chains (01AK702A), is gratefully acknowledged. The authors are responsible for all remaining errors.

Notes

¹ A notable example are Liikanen et al. (2004), who study intergenerational effects in the diffusion of mobile phones. Whereas Comin et al. (2006) look at direct data on a large number of technologies and countries, Caselli and Coleman (2001) study the diffusion of computers using imports of computing equipment as an indirect measure.

² Regarding the retail sector, we are only aware of studies based on firm-level data, for example Foster et al. (2002) and Levin et al. (1987, 1992). Although rich in various aspects, firm-level data typically lack variation in the regulatory environment and hence provide little opportunity to examine policy issues.

³ With micro-level data, discrete choice and hazard rate models are commonly used; for example, see Karshenas and Stoneman (1993), Åstebro (2004) and the references therein. For a review see Hall and Khan (2003).

⁴ Of course, the usual econometric suspects such as autocorrelation may have to be dealt with (Section ‘Robustness’).

⁵ Most studies, however, use another version of Equation 1, where S_t = γN_t /[1 + exp(−α − βt)]. Whereas the advantage of that version is that it lends itself more easily to log-linearization, its disadvantage is that α is erroneously interpreted as a timing indicator. Instead, α = −βτ and hence ‘timing’ estimates for α resulting from the traditional version are strongly correlated with respective speed estimates for β.

⁶ The intensive margin corresponding to our data would be the share of retail sales that go through scanner checkouts.

⁷ An alternative approach is a linear cross-country panel regression analysis, in which a potentially nonlinear diffusion pattern is partly accounted for by time dummies (Caselli and Coleman, 2001; Comin and Hobijn, 2004). The respective coefficients are typically assumed to be constant across countries, such that the added independent variables capture cross-country differences in both timing and saturation level of technology diffusion.

⁸ The earliest EAN report available (at www.ean-int.org) is the 1983 report, which also gives figures for 1981 and 1982 for most countries (or indicates that there were no scanning stores before 1983 in a particular country).

⁹ In contrast to the Italian case, we are rather surprized by the low estimated saturation level for Ireland, since Ireland's retail structure is more comparable to that of the UK (see Table A1 in the Appendix). As Ireland has developed strongly throughout the 1990s, we presume that our data cover only the very beginning of a corresponding diffusion process, which complicates estimation (Debecker and Modis, 1994). We return to this point in the following text.

¹⁰ Measurement differences do not appear to be substantial, however, we obtain similar cross-country differences if we relate the number of barcode scanning stores to population instead of the total number of outlets (Section ‘Robustness’).

¹¹ For Italy and Spain, this indicator does not cover the whole sample period. For these two countries, we therefore constructed a comparable indicator based on Euromonitor and GGDC data (see Beck et al., 2005, for more details).

¹² Clearly, barcode scanning also facilitates other potentially productivity-enhancing practices, e.g. sophisticated logistics systems (‘efficient consumer response’, ‘category management’); but these systems did not develop before the mid-1990s and still seem to represent ‘untapped potential’ (Haberman, 2001).

¹³ Levin et al. (1987, 1992) study the adoption of barcode scanning in the US retailing. They analyse firm-specific data relating to the early years of the technology (1974–1985).

¹⁴ Two countries in our sample – Germany and Denmark – apply a slightly broader hypermarket definition which includes large supermarkets with a floor space between 1500 and 2500 square meters. In our pooled estimation below, we therefore allow for a different hypermarket effect for these two countries.

¹⁵ Carrefour, one of the world's largest retailers, claims to have invented the concept. It opened its first hypermarket in 1963 near Paris, ‘with a floor space of 2500 square meters, 12 checkouts and 400 parking spaces’ (www.carrefour.com/english/groupecarrefour/annees60.jsp).

¹⁶ In Section ‘Robustness’, we argue that reverse causality or endogeneity are not affecting the observed relationship between hypermarkets and IT diffusion.

¹⁷ When we include Ireland in calculating these correlation coefficients, only the coefficient for GDP changes qualitatively, resulting from Ireland's combination of strong GDP growth with a low γ-estimate.

¹⁸ We use the estimates from the country-wise regressions as initial values for country-specific effects. For the independent variables’ coefficients, we set initial values equal to 0.

¹⁹ The results are available upon request.

²⁰ Comparable OECD data for the retail volume indicator VOL indicate that, between 1990 and 2000, the US retail volume increased by about 67%, whereas it increased by about 30% in the UK and by about 7% in France. In Germany, retail volume decreased by about 1% between 1990 and 2000.