The importance of non-linear relationships between attitude and behaviour in policy research

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.

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

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.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.

Open image in new window

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.

Description of application 2

In our second empirical application, we examine the number of loyalty cards consumers possess depending on their privacy concerns. Loyalty programmes generate a number of important benefits to consumers, such as presents and other rewards, (customized) promotions, sponsored magazines, and the thrill of collecting points (Dowling & Uncles, 1997). However, loyalty programmes may also have a negative side for consumers (O’Malley & Prothero, 2004). In particular, loyalty cards imply purchase data being tracked and observed by others (Dowling & Uncles, 1997). The attitude considered here (privacy concerns) is expected to be negatively related to the possession of loyalty cards. As loyalty cards are generally associated with benefits for customers, however, we expect only consumers with severe privacy concerns to display a significant attitude-behaviour relationship (Milne & Gordon, 1993). Thus, the relationship between privacy concerns and the possession of loyalty cards is expected to resemble the downwards shaped curve in Fig. 1.

A large-scale survey has been conducted in the Netherlands, as part of the so-called Relationship Builder project in collaboration with the Carlson Marketing Group. The project studies customer relationship management in general and loyalty programmes in particular. Six companies from six different industries, namely financial services, food industry, oil industry, telecom, utilities, and retailing, participated in the project. Each company provided a sample of 10,000 customers. In the spring of 2004, these customers were contacted by e-mail with the request to fill in an internet-based questionnaire. The questionnaire consisted of a general part on customer relationship management and loyalty programmes and a confidential company-specific part.

To indicate their possession of loyalty cards, respondents had to denote their membership of 24 widespread loyalty programmes in the Netherlands. Respondents could also specify the possession of other loyalty cards in addition to the ones included in our list. In a second task, respondents were asked to indicate how many loyalty cards they owned in total. We then used the maximum of the number of cards indicated in the first task and the number of owned cards reported in the second task as our behavioural measure. The questionnaire contained two items measuring privacy concerns we developed for this study, since existing scales on privacy concerns do not address the context of loyalty programmes. The scale ranged from one (weak privacy concerns) to five (severe privacy concerns) (see appendix). The items on privacy concerns showed sufficient internal consistency with a correlation of 0.53. Hence, we averaged the items to obtain our measurement of privacy concerns. The frequency distribution of the scale is presented in Table 1.

We eliminated cases with missing data and obtained a sample of 3657 respondents. We estimated regression models comparable to our first empirical application with a consumer’s level of privacy concerns as attitudinal variable and socio-demographics accounting for customer heterogeneity. Again we use ordinary regression analysis to estimate a linear model and three alternative non-linear models. For the spline models, we specified spline adjustment variables Z _j at j = 1.5, 2, 2.5, 3, 3.5, 4, and 4.5.

Estimation results of Application 2

It can be seen in Table 3 that the explanatory power of the model increases when allowing for a non-linear relation between privacy concerns and membership of loyalty programmes, albeit to a lesser extent than in our first empirical study. The R ² is highest for the spline regression model. According to the AIC-statistic, the spline model also performs best, but the BIC-statistic is slightly lower for the linear model. This is due to the fact that the spline model uses one parameter more than the linear model, which is stronger penalized in the BIC-statistic than in the AIC-statistic. According to two out of three fit statistics, the spline regression model is preferable to the other models. The exponential and quadratic models outperform the linear model according to two out of three fit statistics as well, but provide a worse fit than the spline model for each of the statistics.² Yet, we have to conclude from Table 3 that the non-linearities are weaker for this empirical application and that a linear model approximates the attitude-behaviour link quite well.

Figure 3 displays the functional forms of all four models. Compared to Fig. 2, the functional forms deviate less from each other, which reinforces the observation that the non-linearities in this example are weaker than in our first empirical study. Still, the parameter estimates of the linear and non-linear models reveal some differences. In the linear model, the possession of loyalty cards is predicted to decrease with 0.136 if privacy concerns increase with one point on the scale. In the spline model, only one spline knot is identified as being significant, namely, at a privacy concern score of 2.5. The basis variable “privacy concerns” shows a very small positive effect, which is not significantly different from zero. This means that up to scores of 2.5, privacy concerns have hardly any impact on the number of possessed customer cards. For privacy scores of 2.5 and higher, the average number of customer cards decreases more rapidly than in the linear model, the slope of the regression line with 0.237 being steeper. The exponential and quadratic models show a similar curve compared to the spline model and also suggest that privacy concerns matter only for scores exceeding three.

Open image in new window

With respect to the numerical interpretation for the linear model, the average individual having low privacy concerns holds 4.08 customer cards, with a reduction of 0.136 cards per scale point of privacy concerns. In the spline model, the average number of customer cards for a respondent with low privacy concerns is slightly lower (3.86). At the lower parts of the scale, the number of loyalty cards slightly increases (0.077) with a one point increase in privacy concerns, but the decrease rate for privacy concerns exceeding 2.5 is much higher than in the linear model (0.077–0.314 = −0.237). In the exponential (quadratic) model, the average individual with low privacy concerns possesses 3.95 (3.88) customer cards, with the rate of decrease with every point increase in privacy concerns being 0.005 multiplied with the exponential of privacy concerns (0.067 multiplied with the square of privacy concerns).

Public policy implications

This research has some implication for public policy makers. First, bearing in mind possible non-linearities in the attitude-behaviour link can be quite useful when designing marketing or public policy measures that aim at altering behaviour by changing consumers’ attitudes. As behavioural responses initiated by attitude changes can be different than a linear model specification would suggest, segmenting consumers according to their attitude extremity before designing policy measures can be worthwhile. In many cases, it makes sense to target only those consumers close to or at the extreme points of the attitudinal scale. In the field of marketing, this is in fact already practiced by some companies by focusing on the so-called “Bottom 2–Box” and “Top 2-Box,” respectively, for instance in the field of customer satisfaction. This implies that customers with a very low satisfaction with certain product or service features are identified and efforts are made to improve their satisfaction. The rationale behind this is that very low satisfaction evaluations have a greater propensity to drive behaviour, which in this case is churning behaviour (Anderson & Mittal, 2000). Likewise, following the notion of customer delight, only very high satisfaction evaluations in the “Top 2-Box” are likely to drive customer retention, which means that efforts should be made to shift customers just beneath the “Top 2-Box” above the threshold.

Second, in the field of public policy, it can also be advisable to focus on consumers with more extreme attitudes, in particular because attitudes such as, say, environmental concern should be harder to alter than satisfaction evaluations. Applied to our first empirical application, if one aims at stimulating the sales of organic products, it is most promising to target those consumers with already quite high environmental concerns. If one succeeds in changing their attitude, consumers still below the extremity threshold will cross it and consequently purchase more organic products. Consumers already above the extremity threshold will intensify their attitude-related behaviour in compliance with their attitude change. Consumers, whose environmental concern changes from “low” to “moderate” will, according to our results, hardly change their behaviour. Targeting these consumers is of not much use. Hence, segmenting consumers according to their attitude extremity prior to designing public policy measures might be worthwhile in terms of campaign efficiency.

Third, we demonstrate that for some attitudes, only extreme evaluations exert an influence on consumer behaviour, thus assuming linear and straightforward relations might not be adequate. Policy measures designed to change consumers’ behaviour through a change of attitudes, can thus have quite different behavioural consequences for consumers with moderate and those with more extreme attitudes.