Hester van Eeren

Comparative effectiveness of MST and FFT | 5 101 | PS score divided by one subtracted by the PS score (Harder et al., 2010). The PS scores that showed no overlap in the treatment groups were removed. The treatment effect was then estimated in the sample containing overlap with the estimated PSs. Though this restricts the generalizab7ility of the results to cases for which overlap is present, removing cases without overlap allows for more precisely balancing the treatment arms (Harder et al., 2010). Missing indicator approach The baseline covariates in the dataset of 697 adolescents who completed either FFT or MST had missing values. To manage these missing values, a missing indicator approach was used while estimating the PS (Cham & West, 2016; D’Agostino, Lang, Walkup, Morgan, & Karter, 2001; Harder et al., 2010; Rosenbaum & Rubin, 1984; West et al., 2014). This method can be theoretically justified and works well to balance observed and missing value patterns across treatment groups without removing cases from the analysis (Cham &West, 2016; Harder et al., 2010; Rosenbaum & Rubin, 1984). In applying this method, the covariate and a missing indicator for this covariate were included in the PS estimation, coded 1 if there was a missing value for the covariate and 0 if not (D’Agostino et al., 2001; Haviland, Nagin, & Rosenbaum, 2007; Rosenbaum, 2010). The missing values of the covariates included in the PS were replaced with an arbitrary value in the range of the values of the covariate itself (Haviland et al., 2007; Rosenbaum, 2010). Using a missing indicator and the covariate with substitution of missing values in the PS estimation enables the use of all cases and balances observed values in the covariates, as well as the missing patterns of these covariates. After PS estimation, balance was assessed for the missing indicators and covariates without missing value substitution. In estimating treatment effects, the missing value substitution was also removed. Thus, this substitution does not affect the evaluated balance, nor does it affect the treatment effect estimation. Furthermore, missing indicators were not taken into account in estimating treatment effects (Haviland et al., 2007; Rosenbaum, 2010). Balance assessment An important step in applying the PS is to assess the balance of the observed covariates between the two treatment arms instead of assessing the parameter estimates of the PS model itself (Stuart, 2010). Balance was evaluated for the covariate without missing value substitution and for the missing indicators of the covariates (Harder et al., 2010; Haviland et al., 2007). Balance is achieved when the distribution of the baseline covariates is similar for the two interventions. Balance was assessed with the standardized bias which is independent of the sample size of the study. It was calculated by dividing the difference of the means of the covariates between the treated (MST) and comparison (FFT) group by the standard deviation of the treated group (Ali et al., 2015; Austin, 2009; Austin, 2011; Harder et al., 2010; Stuart, 2010; West et al., 2014). As such, the difference in means was divided by the standard deviation of the MST group (Harder et al., 2010; McCaffrey et al., 2013). For the categorical covariates, the standardized bias was estimated per level. For instance, if the covariate had three levels, the standardized bias was calculated for all three (Harder et al., 2010).