Statistics

62 Attitudes

Attitudes were measured via a semantic differential, a technique often used to assess attitudes.

Participants are presented opposite adjectives as the two endpoints of a bipolar scale, and are asked to assess reduced consumption with regard to these adjectives. Examples of such bipolar scales used are ‘unimportant’ and ‘important’ or ‘worthless’ and ‘valuable’. Participants then give their rating towards either of the opposite adjectives, with the middle category being neutral. Eight bipolar pairs were used in Study 1, and the four most consistent with the attitude factor were taken over for Study 2 and 3.

Perceived behaviour control

Perceived behaviour control was assessed with 4 (Study 1) and 3 (Study 2 and 3) items asking for participants agreement, e.g. to ‘If I want to, I will be able to reduce my personal clothing consumption in the next three months’ or ‘It is mostly up to me whether or not to reduce my personal clothing consumption in the next three months’.

3.3.1. Ensuring data quality

Data quality across all three studies was ensured with multiple measures for identifying and screening out careless responses (DeSimone, Harms, & DeSimone, 2015; Meade & Craig, 2011). Exemplarily, in Study 1, we applied instructed items (e.g. ‘Please select strongly agree’) and bogus items (e.g. ‘I always sleep less than one hour per night’) as attention filters.

Participants failing such instructed items were filtered out automatically. Moreover, we assessed the quality of the received data through multiple quality checks, e.g. self-reported data on answer quality (e.g. ‘In your honest opinion, should we use your data in our analysis of this study’) or measures for answer patterns like straight-lining. Participants failing multiple quality checks were reported to Qualtrics and replaced.

63

framework in this thesis (see chapter 2, section 2.1.). Additionally, measurement variance tests were realised as basis for the cross-cultural comparison in Paper 2. For Study 3, mixed panel regression analysis are the method of choice to analyse developments between groups and across time. In the following, each method is described shortly.

3.4.1. Structural equation modelling

Structural equation modelling (SEM) is a multivariate data analysis strategy. It combines elements from confirmatory factor analysis and path analysis. It does so, firstly, by modelling constructs of interest, e.g. a person’s attitudes or perceived personal norm towards reduced clothing consumption, as latent variables. Secondly, it examines relationships between these latent variables through simultaneous multiple regressions, allowing for multiple dependent and independent variables in the same analysis (Wolf & Brown, 2013). As such, it makes it easy to estimate direct and indirect paths between all model variables at once.

There are two types of variables in SEM, latent and manifest or observed variables (Kline, 2011). A latent variable is a hypothetical construct, represented through multiple, highly correlated manifest or observed variables, also called indicators. Indicators are observed variables and assumed to measure the same latent, i.e. not directly observable, construct. As an example, an individual’s personal norms cannot be observed directly, but we can measure manifestations of his or her personal norm e.g. in the form of answers to a survey item like

‘Reducing my personal clothing consumption is the right thing to do’. Multiple such items are combined to represent the latent factor ‘personal norm’.

Using latent variables leads to improvements in construct validity of the variables of interest, as latent variables use information from multiple indicators and therefore are closer representations of the construct in question than single items (Wolf & Brown, 2013). In a SEM analysis, the relationships between such latent variables are analysed via multiple regression analysis.

Simultaneously, the analysis removes unique variance from each indicator, i.e. variance that is not explained by the latent variables the indicator represents, and models it as indicator error.

The SEM approach therewith includes an explicit representation of measurement error of observed variables, which yields more accurate estimation of relationships between different variables included in the model. This is a clear advantages compared to multiple regression

64

(Kline, 2011). It is important to note that within SEM we assume a certain directionality of the relationship between two variables, i.e. in the path modelling we draw an arrow for example from personal norms to intentions. However, SEM is only an analysis strategy and does not allow drawing any conclusion about causality between variables when applied to cross-sectional data. When interpreting SEM results we should therefore not forget that the proposed direction of relationships are theoretical assumptions that cannot be proven, and we cannot eliminate the possibility that relationships in reality might be reversed or alternative models with different associations between the model variables might fit the data, too (Raykov & Marcoulides, 2006).

3.4.2. Measurement invariance

In Study 1, all items, i.e. the indicators for latent variables, were developed in English and afterwards translated for the different countries. We assume that these indicators measure the same construct in the same way across all countries (Cieciuch, Davidov, Algesheimer, &

Schmidt, 2017). This, however, is not necessarily the case. Across different cultural groups, differences in e.g. familiarity with a particular item wording or in the prevalence of social desirability can lead to differences in the precision with which a construct is measured. Instead of assuming equivalence, we therefore need to test for so-called measurement invariance, i.e.

test that the internal structure of measurement instruments is equivalent between the countries (Fisher & Fontaine, 2011). This is only the case if there is no systematic bias in the response behaviour e.g., no bias due to translation problems or other cultural, unobservable differences (Steenkamp and Baumgartner, 1998). In our analysis we tested for measurement invariance in the framework of multiple-group confirmatory factor analysis (Cieciuch et al., 2017). There are three levels of measurement invariance which are commonly assessed. Configural invariance means that latent constructs can be conceptualized in the same way across all five countries, i.e.

that the same items measure them across all countries. The next level, metric invariance requires that factor loadings are equal across countries, i.e. ‘that each item contributes to the latent construct to a similar degree across groups’ (Putnick & Bornstein, 2016, p. 5). Lastly, scalar invariance assumes invariance of both factor loadings and item intercepts across countries. This implies that ‘mean differences in the latent construct capture all mean differences in the shared variance of the items’ (Putnick & Bornstein, 2016, p. 5). A comparison of latent factor means across groups is only possible when all three levels of measurement invariance are established.

65

Putnick & Bornstein (2016) provide a detailed procedure how to test for each level of measurement invariance, and we followed their procedure in our analysis for Paper II.

3.4.3. Repeated measurement analysis

In order to account for the panel structure of the data in Study 3 we applied repeated measurement analysis strategies. The data of Study 3 across the time points is nested in individuals, who are at the next level nested in the intervention groups. To account for both the variance between the groups and within the individual participants over time we applied repeated measures mixed regression models with repeated data over participants and estimated the influence of both time and group on the number of items purchased and the change in intentions and personal norms. To analyse the effects a change in different psychological determinants has on a change in behaviour, intentions and personal norms (e.g. the effect of a change in awareness of need on the change in personal norms) we employed multiple linear regression models with fixed effects and clustered standard errors across individuals. For behaviour, intentions and personal norms, respectively, we fitted models seperately for each group. The fixed effects approach is modelling the within variation, i.e. the difference in values across time for individuals. It ignores differences between the groups, hence we calculate the model seperately for each group.

66