The variable that led to the largest decrease in deviance was added at each step.
It is expected that as models increase in complexity by adding predictor variables, a significant decrease in Deviance is expected to occur, and the significance of the decrease is tested with a chi-square test [ 35].
The significance of each of the factors was obtained by calculating the decrease in deviance when the factor was included in the model (given the null model including age and sex only) and comparing this to the appropriate chi-square distribution.
In backward elimination, variables are eliminated from the full model based on a certain criteria mostly based on a decrease in Ror deviance.
If adding a new covariate to a given multivariate model does not lead to a reduction in deviance, its inclusion has little benefit because it does not decrease the discrepancy between the model and the data.
Values were retained if on their removal there was a significant increase of the residual deviance of the model with likelihood ratio statistics (LRS) of p>0.05 and they were removed from the model if they caused an insignificant increase or decrease in the residual deviance with LRS of p<0.05.
Variables which when added to the model didn't result in a significant decrease in the residual deviance or when included gave a higher Akaike Information Criteria were rejected from the model.
Other variables, such as smoking, physical activity, alcohol consumption, education, and heart rate, were included as variables in early analyses, but because they only marginally decreased the deviance in the multivariate analyses they were excluded in the final analysis.
Results clearly show that the developmental pattern found with the DBVS is similar to the one found in previous research, except that girl's involvement in deviance after 16 years old kept more or less stable instead of decreasing.
If I may be thoroughly unorthodox for one sentence, and quote Anton La Vey, "There is less room for deviance in deviance, than in any other human endeavor".
Although variations were not statistically significant, the cumulative arsenic exposure effect decreased with time since last arsenic exposure by factors of 1.0, 0.8, and 0.2 for < 5, 5 14, and ≥15 years since last exposure, respectively (Table 3; change in deviance = 2.6 and p = 0.27 comparing models B0-R and T1-R; or change in deviance = 1.2 and p = 0.55 comparing models T2-R and T3-R).
