In order to enhance the usability, transparency and translational nature of the instrument, all analyses were conducted using single items rather than multi-item constructs or factors. Yet, given that blocks 3-6 were also overarching constructs comprised of more than three scale-items (see Table 6.1), Cronbach’s alpha tests were conducted to explore correlations between items and test the reliability of these constructs (Rust and Golombok, 1999; Eisinga, Grotenhuis et al., 2013).
Fisher's Exact test was used to evaluate differences between the final sample and the general population (p < 0.05 was considered significant). Descriptive statistics, Pearson’s Chi-square (categorical variables) and t-tests (interval or continuous variables) were computed to gauge the relationships between each of the assessed variables and vaccination behaviour (p < 0.05 was considered significant). The outcome measure was having received an influenza vaccine in the last 6 months (latest influenza season).
Since the outcome measure was categorical, logistic regression analysis was conducted to uncover the predictors of influenza vaccination. Logistic regression does not make many of the key
assumptions of linear regression. Some assumptions, however, still apply. These are: 1) the
dependent variable is dichotomous; 2) the categories are mutually exclusive and exhaustive, that is, a case can only be in one group and every case must be a member of one of the groups; 3) the model should be fitted correctly, thus overfitting or underfitting should be avoided; 4) the model should have little or no multicollinearity; and 5) sample sizes are large; a minimum of 10 cases per predictor (Hosmer Jr and Lemeshow, 2004; Field, 2013).
Although stepwise regression is widely used in logistic regression, in recent years, the purposeful selection of variables has been favoured over deterministic model-building methods. This is because the latter rely on automatic selection of variables based upon mathematical criteria, which can lead to overfitting or underfitting the model (Hosmer Jr and Lemeshow, 2004; Field, 2013). Therefore, to achieve a parsimonious model, that is, one that offers the simplest explanation to vaccination behaviour, I used a forced entry, hierarchical approach. The specific procedure employed is described below.
111 When predictors are correlated, as it is often the case, the order of variable entry can have an effect on the parameters calculated. Variables were, therefore, entered in “blocks” using a hierarchical approach based on previous evidence (Field, 2013) and my aim of assessing whether socio-
psychological factors are better predictors of influenza vaccination than population characteristics and practical barriers. Blocks of variables which had predicted vaccination uptake in the past were entered first and those which had seldom or not been explored before were entered last
(Kohlhammer, 2007; Ward and Draper, 2008; Nagata, Hernandez-Ramos et al., 2013; Wheelock, Thomson et al., 2013). Priority was given to demographic, socio-economic and health-related variables, and practical vaccination barriers to evaluate the extent to which they contributed to explain the variance in vaccination behaviour before socio-psychological variables were
incorporated. Consequently, eight blocks of explanatory variables were entered in the following order: 1) demographic, socio-economic and health-related variables; 2) practical barriers to
influenza vaccination; 3) social influence; 4) influenza perceptions; 5) influenza vaccine perceptions; 6) trust and disposition toward vaccination and vaccination stakeholders; 7) general perceptions and constructs; 8) previous vaccine and health-related experiences.
Two goodness-of-fit tests were used to assess the overall model and each of the 8 models (blocks): chi-square and Nagelkerke R². If the chi-square of any given model was not significant (p>0.05), that is, the contribution of such model over the previous one was statistically negligible, its variables were systematically excluded one at a time, starting from those which were not significant, and regressions rerun until significance was attained (Field, 2013). Non-significant variables from models with a significant chi-square (p<0.05) were also excluded one at a time and regressions were rerun. If the predictability of the overall model decreased by more than 1%, the excluded variable was added back. This process continued until all of the important variables appeared to be included in the model. I then added back all the excluded variables to identify those which may not have been significant by themselves, but could be important in the presence of other variables (Hosmer Jr and Lemeshow, 2004).
An overall model with a Nagelkerke R² a value close to 1, which indicates optimal predictability, was sought. Odds ratios for the hierarchical logistic (logit) model, together with the standard errors, significance levels and confidence intervals, as well as the predictidibility of the model, are presented (p < 0.1 was considered significant). Thorough checks to ensure the robustness of the model were conducted, including variance inflation factor (VIF) to assess collinearity, standardised residuals to detect and evaluate outliers and Cook’s distance to identify influential cases. All analyses
112 were conducted by myself using IBM SPSS Statistics 22. Regression analyses were also carried out by expert statisticians (Kerry O'Neill, solutions-2) who validated the results reported in this chapter.
6.2.4.1 Variable coding
For bivariate and logistic regression analyses the variables age, eligible health condition (including pregnant or breastfeeding women), marital status, level of education and ethnicity were
dichotomised, both to aid interpretation and ensure there was a fair number of subjects in each cell for validity (a minimum of five per cell) (Greenland, 1989).
For logistic regression analysis, 11 variables with “I do not know” responses were also recoded as dummy variables in order to maximise the number of observations, as follows: values expressing agreement with a given statement (6-10) were coded as 1 = “yes” and the rest (0-5 and “I do not know”) were coded as 0 = “other than yes” (see recoded variables in Table 6.3).