*4.1. Analysis*

The TPB was tested by fitting a structural equation model (SEM) to the latent variables. First, each of the constructs was examined for reliability (i.e., if similar results are likely to be obtained with a retest) and unidimensionality (whether the items in a construct work together to measure one latent trait). The Guttman's lambda 6 (λ6) [62], as well as the criticized [63], but still widely used Cronbach alpha ( α) [64] and McDonald's omega ( ω) report reliability, and Revelle's beta (β) report internal consistency [65].

Thereafter, the mean score of the items within each construct was calculated in order to obtain a summary score for each construct (measuring instrument). The next step was to test the effect(s) of the constructs on recycling behavior. The objective of the statistical analysis was to test whether the theoretical TPB model could be supported with a statistical significant model based on the collected data. Structural equation modeling (SEM), using partial least squares path modeling (PLSPM) was used. As a final check, classification and regression trees and random forests (results not shown) were also used to confirm the predictive ability of the various constructs on recycling behavior. The package R was used for the statistical analysis [66,67].

MS Excel was used for descriptive statistics (graphs, frequency tables, etc.), to describe averages and for determining measures for variability and relationships between variables (correlation and regression analyses).

#### 4.1.1. Treatment of Inappropriate Answers

A small number of respondents gave "not applicable" or "do not know" answers which do not form part of the 7-point scales. Of the 33 respondents that indicated they do not know how often their households recycle, 29 respondents selected the "nothing" option in the statements that tested the

quantity households recycle and their "do not know" option was subsequently changed to "never". The remaining four respondents' "do not know" option for frequency of recycling was also scored to match their recycling quantity scores. The 33 respondents represent 1.65% of the total sample group. Given the large sample size (*n* = 2004), these adjustments should have no significant effect, even if the modified scores are incorrect by more than 2 units. The occurrence of inappropriate answers to other items was low and was thus treated as missing. The statistical methods make provision for such instance by dropping those observations with missing values in the specific analyses affected by the missing information, but retaining information of non-missing values in other analyses.
