Table 4 Summary of statistical approach of included studies
Study ID | Authors and year of publication | Was variance in SRI (or any injury) due to neighborhood-level differences assessed or reported for unconditional or null model(s) to justify the use of multilevel model? | Was variance in individual slopes (random effects) evaluated or reported and was cross-level interaction tested for to account for the variance where it existed? | Was the unexplained variance at neighborhood-level or the proportion of variance explained by the neighborhood-level variables assessed or reported for the final multilevel model(s)? | Model-building process employed to develop final model(s) | Multilevel modeling type and software used |
|---|---|---|---|---|---|---|
1 | Haynes, Reading, Gale. 2003 | Yes. Unexplained variance for the null or unconditional models were reported and tested for significance; however, ICCs were not reported | No | Yes. The unexplained variances at both the enumeration district and social area levels were reported | A single multilevel model was developed; however, several variations of multilevel models (including null model, single level, and different combinations of two- and three-level models) were developed to assess the unexplained variance in models | Multilevel logistic regression using MLwiN software package |
2 | Sellström, Guldbrandsson, Bremberg, Hjern & Arnoldsson. 2003 | No | No, variance in individual slopes was not evaluated or reported. But cross-level interactions were tested for. Interaction terms were, however, excluded from final models because they were not statistically significant | Yes | Three multilevel models of increasing complexities were developed to adjust for potential confounders | Multilevel logistic regression using SAS makro Glimmix software package |
3 | Kendrick, Mulvaney, Burton, & Watson. 2005 | No | Yes. Two-way interactions were examined where it appeared they might exist; however, it is not stated if the interactions included cross-level interactions | Yes | A forward selection approach was used | Multilevel Poisson regression using MLwiN software package |
4 | Simpson, Janssen, Craig, & Pickett. 2005 | No | Yes, it was reported that there was no significant variation in the slopes of the relationship between each socioeconomic variables and medically-treated injury across neighborhoods. For other injury outcomes, the variation in slopes were therefore assumed to be non-significant | No | A two-step process was employed which include: (1) fitting bivariate models, and (2) fitting multivariable multilevel models from significant socioeconomic exposure variables in the bivariate models. Age and sex were included in all multilevel models | Multilevel logistic regression using MLwiN software package |
5 | Pattussi, Hardy, & Sheiham. 2006 | Not clear. It appears that a null model may have been fitted based on descriptive results of between-neighborhood variation in dental injuries; however, this was not clearly reported | No, variance in individual slopes was not reported. However, results of the sex-stratified multilevel models mean that cross-level interactions were tested for between sex and neighborhood-level variables | Yes, the variance explained by neighborhood-level variable was assessed based on the report that most of the between-neighborhood variation in dental injuries was explained by social capital; however, no numerical value of the proportion of variance explained was reported | Four series of pre-determined multilevel models were developed including final models containing both individual and neighborhood-level predictor variables | Multilevel logistic regression using MLwiN software package |
6 | Mecredy, Janssen, & Pickett. 2012 | No | No | No | Three series of pre-determined multilevel models were developed including a final model for total street injuries that contained both individual- and neighborhood-level predictors fitted using a backward elimination multilevel regression method. Predictor variables in the final reduced multilevel model were also then used to develop four physical activity-specific injury multilevel models | Multilevel logistic regression using SAS Glimmix procedure |
7 | Mutto, Lawoko, Ovuga, & Svanstrom. 2012 | Variance due to neighborhood level was reported for the null model; however, statistical significance of the variance, and/or the variance partition coefficient (VPC) for the model was not reported to justify the use of a multilevel model | No | Yes, the neighborhood level variance and the proportional change in variance (PCV) explained by the final model was reported | Four series of pre-determined multilevel models were developed including a null model, and final model that included both individual and neighborhood-level predictor variables | Multilevel logistic regression using STATA |
8 | Gropp, Janssen & Pickett. 2012 | Yes. Intraclass correlation coefficient (ICC) was reported | No | No | A backward elimination approach was used to select statistically significant individual and area-level predictor variables | Multilevel logistic regression using SAS software package |
9 | Byrnes, King, Hawe, Peters, Pickett & Davison. 2015 | No | No | No | A backward elimination approach was used but sex, grade, and relative family affluence were retained in the final model based on a priori decision | Multilevel, multivariable, log binomial regression model was used to analyze risk of injury among northern youths only. Modeling was carried out with SAS software package |