Methodology

---

The methodology consisted of three phases: (1) belief elicitation, (2) survey of online shopping habit and satisfaction, and (3) repurchase tracking.

To identify specific satisfaction factors, we relied on belief elicitation to develop a formative measurement model for each of the satisfaction dimensions; i.e., product satisfaction, sales process satisfaction and after-sale service satisfaction. For the remaining constructs, i.e., overall satisfaction and online shopping habit, we developed reflective measurement models. The dependent variable, repurchase, was operationalized as the number of repurchases made within a specific period of time (one month) since the first purchase.

Belief Elicitation

To develop formative items for the constructs, "product satisfaction", "sales process satisfaction" and "after-sale service satisfaction", we examined the literature and conducted a belief elicitation procedure. Sixty online shoppers were selected randomly from the customer base of a North American supermarket. The selected online consumers were invited to participate in focused group discussion. They were divided into six groups of 10 individuals each. In the focused group discussion, the participants were asked to identify important satisfaction factors and to categorize them under product satisfaction, sales process satisfaction and after-sale service satisfaction. Based on the literature review and the results of the belief elicitation process, we ended up with a list of formative satisfaction items represented in Table 1.

Table 1: Weights and Loadings for Formative and Reflective Measures.

Factors	Variables	Weights	Loadings	Std. Error	T- statistics
Overall satisfaction	Satisfaction 1		0.94	0.01	90.49
	Satisfaction 2		0.93	0.02	60.10
	Satisfaction 3		0.96	0.00	199.19
Online shopping habit	Online shopping habit 1		0.81	0.02	32.57
	Online shopping habit 2		0.85	0.03	30.96
	Online shopping habit 3		0.81	0.03	28.80
Repurchase	No. of repurchase in the last month		1.00	0.00	0.00
Product satisfaction	Quality of products	0.32		0.15	2.09
	Prices of products	0.60		0.15	4.07
	Packaging of products	0.06		0.14	0.45ns
	Product choices	0.29		0.14	2.10
	Product description	0.21		0.11	1.90
Process satisfaction	Transaction efficiency	0.40		0.09	4.51
	Privacy measures	0.13		0.08	1.60 ns
	Navigation efficiency	0.20		0.09	2.19
	Comparative shopping	0.36		0.08	4.52
	Convenience of shopping	0.08		0.10	0.75 ns
	Site accessibility	0.28		0.09	3.18
	Web page loading speed	0.27		0.14	1.98
	Security measures	0.19		0.11	1.72
	User-friendliness	0.19		0.07	2.58
After-sale service satisfaction	Delivery time	0.79		0.06	13.05
	Handling returns	0.29		0.09	3.08
	Customer service	0.02		0.08	0.21 ns
	Delivery care	0.05		0.06	0.71 ns

Survey of Online Shopping Habit and Satisfaction

A survey instrument was constructed based on reflective items for the "overall satisfaction" and "online shopping habit" constructs and the formative items developed in the belief elicitation stage for the constructs "product satisfaction", "sales process satisfaction" and "after-sale service satisfaction". The reflective items were validated using the card sorting procedure (Moore & Benbasat, 1991). The resulting survey was then administered to first-time online buyers of a major grocery retailer in North America that also operates an Internet store. For a period of six months, every new online shopper was invited to answer an online survey after the delivery of the grocery (within 24 hours from the online order). Once the respondent has completed the questionnaire, the responses were automatically sent to a database. Pitkow and Recker (1995) present all the advantages of this surveying method. A total of 391 new online shoppers completed the survey. The demographics of these respondents are shown in Table 2.

Table 2: Demographics of Respondents.

Demographics		%
Age	Less than 20	18
	20–30	37
	30–40	28
	40–50	16
	Greater than 50	1
Gender	Male	28
	Female	72
Household family income	Less than US$ 20,000	16
	US$20,000–35,000	28
	US$35,000–50,000	31
	Greater than US$50,000	25

Repurchase Tracking

The number of repurchases that every respondent of the survey made over a period of 30 days was automatically recorded. We assumed that one month would be sufficient for customers to repurchase from the store given the perishable nature of grocery products. This assumption is supported by the findings of the 10^th survey conducted by the Graphics, Visualization and Usability (GVU) Center at Georgia Tech in October 1998. It showed that the frequency of online shopping ranges from once a month to few times a week (GVU, 1998). The entire data collection process lasted for six months.

Data Analysis

The analysis of the data was done in a holistic manner using Partial Least Squares (PLS). The PLS procedure (Wold, 1989) has been gaining interest and use among researchers in recent years because of its ability to model latent constructs under conditions of non-normality and small to medium sample sizes (Chin, 1998; Compeau & Higgins, 1995; Chin & Gopal, 1995). It allows the researcher to both specify the relationships among the conceptual factors of interest and the measures underlying each construct. The result of such a procedure is a simultaneous analysis of (1), how well the measures relate to each construct and (2), whether the hypothesized relationships at the theoretical level are empirically confirmed. This ability to include multiple measures for each construct also provides more accurate estimates of the paths among constructs which are typically biased downward by measurement error when using techniques such as multiple regression. Furthermore, due to the formative nature of some of the measures and non-normality of the data, LISREL analysis was not appropriate (Chin & Gopal, 1995). Thus, PLS-Graph version 2.91.02 (Chin, 1994) was used to perform the analysis. Tests of significance for all paths were conducted using the bootstrap resampling procedure (Cotterman & Senn, 1992). For reflective measures, all items are viewed as parallel measures capturing the same construct of interests. Thus, the standard approach for evaluation, where all path loadings from construct to measures are expected to be strong (i.e., 0.70 or higher), is used. In the case of formative measures, all item measures can be independent of one another since they are viewed as items that create the "emergent factor". Under this situation, Chin (1998) suggested that the weights of each item to be used to assess how much it contributes to the overall factor. For the reflective measures, rather than using Cronbach's alpha, which represents a lower bound estimate of internal consistency due to its assumption of equal weightings of items, a better estimate can be gained using the composite reliability formula (Chin, 1998).

In formulating and testing the moderating effect of "online shopping habit" on the relationship between "overall satisfaction" and "repurchase" with PLS, we followed a hierarchical process similar to multiple regression, where we compared the results of two models (i.e., one with and one without the interaction construct: online shopping habit x satisfaction). We applied the procedure described by Chin et al. (1996). The standardized path estimate from the product construct (online shopping habit x overall satisfaction) to repurchase indicates how a change in the level of the moderator construct (online shopping habit) would change the influence of satisfaction on the dependent construct (repurchase). Thus, if satisfaction has an estimated beta effect of B on repurchase, a beta of M for the interaction path can be interpreted as a beta change of B+M for the estimated path from satisfaction to repurchase when habit increases by one standard deviation from the baseline of zero.

By comparing the R-square for the interaction model with the R-square for the main effects model (which excludes the interaction construct), we can assess the strength of the moderating effect. The difference in R-squares was used to estimate the overall effect size (f²) for the interaction where .02, 0.15, and 0.35 suggest small, moderate, and large effects, respectively (Cohen, 1988) ^[1]. It is important to understand that a small f² does not necessarily imply an unimportant effect. If there is a likelihood of occurrence for the extreme moderating conditions and the resulting beta changes are meaningful, then it is important to take these situations into account.

^[1]f² = [R-square(interaction model)-R-square(main effects model)]/[1-R-square(main effects model)]

---