This analysis was based on a large individual-level data set
collected under service conditions across all regions of England.
It allowed us to estimate with adequate statistical power how
outcomes of the MEND intervention varied by participant,
family, neighbourhood and programme characteristics, estimates
that would be underpowered in most research studies.
However, in the absence of controls, such associations cannot
be equated with effectiveness. We were able to compare the
size of change in BMI in the service data with that observed in
the RCT, and they were similar, but we had limited power to
assess differences between them (there were only 47 children in
the intervention arm of the RCT12). The data we analysed were
collected for service provision and not for research. We used a
range of techniques for improving data quality, including
algorithms for data cleaning, and multiple imputation for
missing data. These techniques were aimed at maximising the
value of the observed data and minimising bias. We fitted
models using imputation and complete case approaches and
found that, other than the differences in statistical significance
that are to be expected given the greater power of imputation
models, findings were similar. Nevertheless, some bias may still
be present.