A multi-dimensional model of the origins of attitude certainty: teachers’ attitudes toward attention-deficit/hyperactivity disorder

Abstract

This research tested and extended a laboratory-derived model of the origins of attitude certainty using a real attitude object: Teaching children with attention-deficit/hyperactivity disorder (ADHD). In Study 1, an experiment manipulated the amount of information, thought, and consistency of information delivered to pre-service teachers ($n = 224$) with no prior experience with ADHD. Structural equation modelling (SEM) of Study 1 did not support the original model, in which relationships between attitude certainty and objective knowledge, thought, and consistency were mediated by perceived knowledge, thought, and ambivalence. Instead, objective amount of information, thought, and consistency interacted in their effect on attitude certainty. Study 2 ($n = 368$) used a survey to test whether experiences with ADHD (personal, direct and indirect) among in-service and pre-service teachers were antecedents of attitude certainty, and to test perceived accessibility as a mediator. SEM supported both these hypotheses. Perceived accessibility and perceived knowledge mediated the relationship between attitude certainty and prior experiences with ADHD, and between attitude certainty and objective knowledge. Together, the results suggest that psychological processes underlying strong attitude certainty differ according to the familiarity and personal relevance of the attitude object, and the context and stage of attitude formation. The results have practical utility for teacher training at pre-service and in-service levels.

Appendix: Assumptions

1.1 Study 1

Following MacCallum et al. (1996) power tables for 72 degrees of freedom, $N = 168$ was needed for power of .80 and a close fit. Close fit refers to testing the null hypothesis that the model fits the population at $p<.05$ (MacCallum et al. 1996). The achieved sample size of 224 pre-service teachers was satisfactory. Multivariate normality, multicollinearity, linearity, homoscedasticity and multivariate outlier assumptions were met. Due to concerns about significant positive skewness at $p< .001$ for five variables, PK2 ($z_\mathrm{skew} =3.65$), C1 ($z_\mathrm{skew} = 3.44$), C2 ($z_\mathrm{skew} = 3.42$), PT1 ($z_\mathrm{skew} = 3.78$), and PT3 ($z_\mathrm{skew} = 4.06$), bias-corrected bootstrapped standard errors were used in the structural equation analyses (Byrne 2001). The bootstrapped procedure is free of assumptions of normality and provides more accurate estimates of parameter values than parametric methods when there are non-normal variables (Byrne 2001). All variables were left untransformed to maintain interpretability and to allow comparison to Smith et al.’s (2008) results.

1.2 Study 2

1.2.1 Missing data: combined sample

Data were collected from 327 pre-service teachers and 127 in-service teachers. Contrasts of missing cases (five cases on total knowledge, 63 cases on objective elaboration and objective ambivalence) versus complete cases on all the perceived attitude strength variables and total knowledge were non-significant ($p> .001$). Thus, data was Missing Completely at Random (MCAR) (Byrne 2001; Kline 2005). Considering the large sample size and MCAR evidence, the 63 cases with missing data, which included 5 cases with missing scores on total knowledge, were excluded from the analysis, leaving complete cases for 111 in-service teachers and 280 pre-service teachers.

1.2.2 Model 1 assumptions

Assumptions for univariate outliers, homoscedasticity and linearity were met. Exclusion of 11 multivariate outliers using Mahalanobis distances for 15 predictors, $p< .001$ (Tabachnick and Fidell 2007), left a sample of 269 pre-service teachers, which had good statistical power (MacCallum et al. 1996). Due to concerns about normality (Mardia criterion for multivariate kurtosis, $z = 5.98$, PAM3, $z_\mathrm{skew}= - 8.61$) bias corrected bootstrapped standard errors were used in the analyses (Byrne 2001). The variables were left untransformed to aid interpretability. A multicollinearity issue was detected. The items PK2 and PK3 were strongly correlated, $r (267) = .94, p< .001$, and loaded onto the smallest eigenvalue at .73 and .76, respectively. This was resolved by using the average of the two variables (abbreviated PK2a, see Table 3) in the SEM analyses (Tabachnick and Fidell 2007). The final Durbin-Watson statistic was 2.06, which supported independence of errors among the endogenous variables.

1.2.3 Model 2 assumptions

Exclusion of 14 multivariate outliers detected using Mahalanobis distances for 19 predictors at $p< .001$, left a sample size of 366 (267 education students and 99 in-service teachers). Linearity assumptions were met, however heteroscedasticity was detected. Due to concerns about normality (direct experience, $z_{skew }= 17.92, z_{kurtosis} = 23.19$; personal experience, $z_\mathrm{skew} = 6.98$; indirect experience, $z_\mathrm{skew} = 5.10$; Mardia criterion for multivariate kurtosis, $z = 8.82)$, bootstrapped bias corrected standard errors were used (Byrne 2001). The variables were left untransformed to aid interpretability. The items PK2 and PK3 were strongly correlated, $r (364) = .93, p< .001$, and loaded onto the smallest eigenvalue at .77 and .75, respectively. The average of these two variables was used in the SEM (PK2a).