Table 3.

Direct and Indirect Effects in Weighted Cross-sectional and Longitudinal Mediation Analyses of the Scanning – Norm – Behavior Pathway

Direct Pathways	B	OR	95% CI
H1. Media scanning → E-cigarette use
T1 → T1		1.23***	1.14, 1.33
T1 → T2		1.26*	1.05, 1.52
H2. Media scanning → Norm perceptions
T1 → T1	0.12***		0.09, 0.14
T1 → T2	0.04*		0.00, 0.08
H3. Norm perceptions → E-cigarette use
T1 → T1		2.38***	2.16, 2.64
T1 → T2		2.00***	1.52, 2.63
Indirect Pathways	Indirect Effects		Total Effects
	Effect Size	BC CIs	Effect Size
H4. Media scanning → Norm perceptions → E-cigarette use
T1 → T1 → T1	.010	.008 – .012	.022
T1 → T1 → T2	.009	.004 – .015	.025
T1 → T2 → T2	.004	.001 – .007	.018

Note: Sampling weights applied. T1 = variable measured at first interview. T2 = variable measured at the re-contact interview. B = unstandardized regression coefficient. OR = adjusted odds ratio. CI = confidence interval.

^*p < .05,

^**p < .01,

^***p < .001.

For the mediation results, BC CIs = Bias-corrected bootstrap confidence intervals.

The bootstrapping procedures were conducted with 500 simulations given that this size is considered sufficient for general standard bootstrapping method in most cases (Efron & Tibshirani, 1993). Simulation (Pattengale, Alipour, Bininda-Emonds, Moret, & Stamatakis, 2010) and empirical (Deng, Allison, Fang, Ash, & Ware, 2013) evidence also confirmed that 500 resamples were more computationally practical, yielded robust estimates, and had little impact on either the bootstrapped standard errors or confidence intervals compared to larger sample sizes.

Indirect and total effect sizes are standardized. Nonzero indirect effects are bolded. These analyses report the effects of the compound path from the independent variable to the dependent variable through the mediator, adjusting for demographic variables and potential confounders at T1 as listed in regression result tables found in Tables 5 and 6. Because of the non-linear nature of logistic regression (i.e., the b path – norm- behavior – in the mediation model), the Hayes macro applied different standardization procedures for the a and b coefficients (Hayes, 2009; Hayes et al., 2011; Mackinnon & Dwyer, 1993; MacKinnon, Lockwood, Brown, Wang, & Hoffman, 2007). Specifically, standardized a* = a (SD_X/SD_M), where a refers to the unstandardized coefficient of the first path, and SD_X and SD_M refer to the standard deviations of X (scanning) and M (norm). For the second path, because Y (e-cigarette use) is binary, standard b* = b (SD_M/SD_Y), with SD_Y calculated differently as ${\text{SD}}_{\text{Y}}=\sqrt{b^{2}S{D}_M{}^2+{\pi }^2/3}$. The constant π²/3 is an estimate of the binomial distribution variance. The product of the standardized paths, a*b*, is then used as the effect size estimate for the indirect effect.