Description of application 1
In our first empirical application, we investigated consumers’ purchase of organic products. We expected only high degrees of environmental concern to be strongly positively related to the purchase of organic products. For respondents with less extreme attitudes, the benefits of organic products would presumably not outweigh the costs, such as higher prices or higher transaction costs for travelling to a specialized store (Grunert & Juhl, 1995). That is, apart from attitudes, contextual factors are influential on pro-environmental behaviour (Stern, 1999). This should result in a zero or weak attitude-behaviour relationship in the range of low to moderate attitudes. The relationship thus resembles the upwards-shaped curve in Fig. 1.
We collected data by means of a questionnaire, in which respondents were inquired about their purchase of several categories of organic products, like meat, bread, and vegetables. We measured environmental concern by applying scales developed in literature (see appendix and Dunlap & Van Liere, 1978; Noe & Snow, 1990), and embedded these in a longer questionnaire. This way, respondents could not discern the goal of our research and were not inclined to pretend to having a higher attitude-behaviour consistency than they actually did. We mailed the questionnaire to randomly selected 2,000 Dutch households and obtained a response of 309 consumers (response rate of 15.5%). After elimination of the cases that were unusable due to missing values, 266 responses remained.
The five items for environmental concern had a good internal consistency with a Cronbach’s Alpha of 0.71, thus exceeding the critical threshold of 0.7 (Nunnally & Bernstein, 1994). We use the average of the items measuring environmental concern; where zero denotes a low environmental concern and six denotes a high environmental concern. The distribution of the responses is presented in Table 1. We assume that the data are of equal interval and apply ordinary regression analysis. An alternative would be to use ordinal regression analysis. However, this would require a large number of additional parameters. Furthermore, the dependent variable in our analysis is the average of multiple items instead of the value of a single rating scale. Finally, there is some empirical evidence that the assumption of equal intervals with data measure on a rating scale is often reasonable (Gregoire & Driver, 1987).
Table 1 Distribution of the responses
For the spline model, we define the spline adjustment variables Z
j
at j = 1.5, 2, 2.5, 3, 3.5, 4, 4.5, 5 and 5.5. We use the MKSPLINE procedure provided by STATA to accomplish this. To estimate our model, we use the stepwise regression procedure in STATA. To account for consumer heterogeneity, we include socio-demographic variables (gender, age, income, household situation, education, house ownership) as dummy variables.
As dependent variable, we use a variable denoting the self-reported number of organic product categories bought by a consumer on a regular basis. We include all the socio-demographics in the model and apply the stepwise selection technique to the attitudinal spline adjustment variables Z
j
.
Estimation results of application 1
Table 2 provides the estimation results of the linear regression model and the three alternative non-linear models.
Table 2 Model results for the effect of environmental concerns on the purchase of organic products
It can be seen in Table 2 that the effects of the socio-demographic variables are limited, but stable across all model specifications. In all models, female consumers and consumers between 55 and 64 years of age are more likely to purchase organic products.
In the linear model specification, environmental concern has a significant positive effect on the number of product categories in which a consumer purchases organic products. However, by allowing a non-linear model specification, the link between environmental concern and the purchase of organic products can be much better explained, which can seen from the considerable increase in R
2 for the non-linear models. According to the R
2, the spline regression model performs best. However, the R
2 does not account for the additional variables added to the model due to the non-linear specifications. Also, classical testing procedures for model selection cannot be used because the models are non-nested (e.g., Greene, 2003). Therefore, we take into account the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC), both of which penalize model expansion (Greene, 2003). When comparing non-nested models, the use of these criteria is most practical to balance model fit and model parsimony. With both AIC and BIC being lowest for the spline model, these criteria favour the spline model as well.
In the spline model, the stepwise procedure selects a change in effect parameter at the threshold value 5, while all other spline knots are not significant. Concerning the other two non-linear models, the exponential model is also better suited than the linear model to capture the stronger behavioural consequences of extreme attitudes. The quadratic model outperforms the linear model on only two out of three fit-statistics, which makes it less preferable than the spline and exponential specification.Footnote 1 The parameter pertaining to the non-linear term of the environmental concern effect is significant in both the quadratic and exponential model.
To further understand the non-linear effect of environmental concern on the purchase of organic products, we graphically show the relationship according to the four estimated models in Fig. 2. To obtain this figure, we varied environmental concern from zero to six and inserted average values for all other explanatory variables. The interpretation of the linear model is straightforward: the average respondent with minimum environmental concerns purchases organic products in 0.10 categories, and the expected number of product categories increases by 0.345 with any 1 point-increase in environmental concern. In the spline model, consumers with very low environmental concerns are predicted to purchase organic products in 0.98 categories, which is substantial higher than in the linear model. The effect of the “basis” environmental concern-variable is relatively small (0.083) and not significant. However, with a threshold value of five on the scale, environmental concern has a very strong impact (0.083 + 1.485). Hence, the relation between environmental concern and the purchase of organic products is negligible for environmental concern below five, but for extremely high levels of environmental concern the relation is much stronger than in the linear model. The curve of the exponential model shows the strongest resemblance to the spline model: Actually, there is not much of an effect before environmental concern exceeding a value of about five. The functional form of the quadratic model deviates from the other models and least resembles the relationship between attitude and behaviour we presume in Fig. 1.
Hence, we conclude that a linear model does not adequately capture the relationship between environmental concern and the purchase behaviour with respect to organic products. The true relations are in fact masked in the linear model, as the weak (or non-existing) relationship in the low to moderate regions of the attitudinal scale and the strong correlation in the endpoints of the scale are not captured. The spline model is best suited to describe the true attitude-behaviour relationship, although the exponential model also performs quite well.