Data were imputed for ethnicity (33% missing), family structure (36%),
housing tenure (35%), employment status (63%) and percentage of
sessions attended (42%). A multilevel (participants nested in programmes)
multiple imputation model (N = 13 998) was used to adjust for betweenprogramme
variation in missingness in MEND 7–13 programmes. The
model assumed that data were missing at random—that missingness on
variables was associated with other variables included in the multiple
imputation model. Ten imputed data sets were produced and analysis
results were combined using Rubin’s rules.21 We followed the guidelines of
Sterne et al.22 for the analysis and reporting of missing data and multiple
imputation (available on request). To test whether our findings were
influenced by using imputed data, we also conducted sensitivity analyses,
including analysis using complete case data with and without the variable
describing parental employment status, as missingness was relatively high
for this variable (data provided in Supplementary Information).
We also used unpublished data from participants in the intervention arm
(N = 47) of the RCT of MEND 7–1312 to compare change in BMI under trial
and service conditions. Height and weight were measured in the first and
penultimate sessions of the trial as in the service data. Age, sex, baseline
BMI, ethnicity and housing tenure were also measured.
Following the Sterne guidelines,22 analysis outcomes were included in
the multiple imputation model where they were missing to ensure that
covariates were imputed correctly. However, analysis data sets excluded
cases where outcomes were not completely observed at both baseline and
follow-up. Sample sizes of the four data sets for analysis of change in BMI,
zBMI, self-esteem and SDQ, respectively, are given in Figure 1.
Four sets of two-stage analyses were conducted, one for each outcome.
In the first stage, relationships between the outcome and each covariate
were tested in multilevel models adjusted for the outcome measured at
baseline (‘baseline-adjusted’ models). If the relationship between the
covariate and the outcome was statistically significant, the covariate was
carried forward to a multilevel multivariable model. The intercept of the
multivariable model describes the mean change in the outcome for a
given ‘reference group’, which for categorical variables was the largest
group, whereas for continuous variables were grand mean centred
(allowing the intercept of the model to be interpreted as the mean
change). Coefficients in the model describe the amount and direction of
change per unit change in the covariates, relative to the reference group
Data were imputed for ethnicity (33% missing), family structure (36%),
housing tenure (35%), employment status (63%) and percentage of
sessions attended (42%). A multilevel (participants nested in programmes)
multiple imputation model (N = 13 998) was used to adjust for betweenprogramme
variation in missingness in MEND 7–13 programmes. The
model assumed that data were missing at random—that missingness on
variables was associated with other variables included in the multiple
imputation model. Ten imputed data sets were produced and analysis
results were combined using Rubin’s rules.21 We followed the guidelines of
Sterne et al.22 for the analysis and reporting of missing data and multiple
imputation (available on request). To test whether our findings were
influenced by using imputed data, we also conducted sensitivity analyses,
including analysis using complete case data with and without the variable
describing parental employment status, as missingness was relatively high
for this variable (data provided in Supplementary Information).
We also used unpublished data from participants in the intervention arm
(N = 47) of the RCT of MEND 7–1312 to compare change in BMI under trial
and service conditions. Height and weight were measured in the first and
penultimate sessions of the trial as in the service data. Age, sex, baseline
BMI, ethnicity and housing tenure were also measured.
Following the Sterne guidelines,22 analysis outcomes were included in
the multiple imputation model where they were missing to ensure that
covariates were imputed correctly. However, analysis data sets excluded
cases where outcomes were not completely observed at both baseline and
follow-up. Sample sizes of the four data sets for analysis of change in BMI,
zBMI, self-esteem and SDQ, respectively, are given in Figure 1.
Four sets of two-stage analyses were conducted, one for each outcome.
In the first stage, relationships between the outcome and each covariate
were tested in multilevel models adjusted for the outcome measured at
baseline (‘baseline-adjusted’ models). If the relationship between the
covariate and the outcome was statistically significant, the covariate was
carried forward to a multilevel multivariable model. The intercept of the
multivariable model describes the mean change in the outcome for a
given ‘reference group’, which for categorical variables was the largest
group, whereas for continuous variables were grand mean centred
(allowing the intercept of the model to be interpreted as the mean
change). Coefficients in the model describe the amount and direction of
change per unit change in the covariates, relative to the reference group
การแปล กรุณารอสักครู่..
