| Measures of Treatment Effect
If participants, interventions and outcome measures are sufficiently similar, meta‐analyses will be carried out. Data from the extraction sheets will be entered into an Excel spreadsheet and copied into Review Manager version 4.2.8 (RevMan). Data entry will be checked by entering the same data from a different author's extraction sheets into RevMan using the double data entry facility. Dichotomous data will be pooled using odds ratios. Ordinal data from rating scales will be treated as continuous data. Where the same rating scale has been used for all studies, data will be pooled using weighted mean differences; where different rating scales have been used to measure the same outcome, standardised mean differences will be used.
Dichotomous Outcomes
Response to medication (secondary outcome) will be defined as a 25% reduction in scores on a validated ADHD rating scale. Reported response rates will be pooled where they are sufficiently similar, however if this information is not reported, the raw data will be requested from study authors and response rates calculated.
Cross‐over Trials
Data from cross‐over trials will be pooled according to the methods described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2006b) and by Elbourne et al. (Elbourne 2002). Mean within‐participants difference and standard error of the mean difference will be entered into RevMan using the generic inverse outcome type. Where the standard error of the mean difference is not reported, the original data will be requested from study authors or the value will be imputed. Correlation coefficients will be calculated from studies where sufficient data is available and if these values are consistent they will be used to calculate the missing standard errors for other studies.
Dealing with Missing Data
If studies do not report intention‐to‐treat analyses, attempts will be made to obtain missing data by contacting the study authors. If dichotomous outcome data are not available, values will be imputed by assuming that all participants for whom data is missing did not experience a favourable outcome. If missing continuous data is not forthcoming, an available case analysis will be performed.
Where studies do not report response rates, values will be imputed using the method described by Furukawa et al. (Furukawa 2005). Outcome data will be assumed to be normally distributed. Number of responders (n) will be calculated using the formula in Figure 01, where N is the number of participants at endpoint, x is 75% of the baseline mean score, mu is the endpoint mean score, sigma is the standard deviation of the endpoint mean and phi is the cumulative distribution function. The validity of this method will be checked by applying it to studies which do report response rates and then calculating the correlation coefficient between reported and imputed response rates.
Assessment of Heterogeneity
The chi‐squared and I‐squared tests will be performed to assess for heterogeneity between studies. Graphical representations will also be inspected. Where significant heterogeneity is suspected (p > 0.1, I‐squared > 50% or on visual inspection), a random effects meta‐analysis will be used. Fixed effects meta‐analyses will be used where significant heterogeneity is not suspected.
Subgroup Analysis and Investigation of Heterogeneity
Where heterogeneity is identified, it will be investigated by performing the following subgroup analyses:
1) Subgroups by severity of intellectual disability (mild, moderate, severe or profound);
2) Subgroups by different doses of drug (fixed doses rather than mg/kg as this reflects clinical practice)
3) Subgroups by whether participants have comorbid pervasive developmental disorder or not;
4) Subgroups by age ‐ whether participants are adults (18 years or over) or children (under 18 years old);
5) Subgroups by whether treatment is with dexamfetamine or mixed amfetamine salts, and with sustained release or immediate release preparations.
Sensitivity Analysis
A sensitivity analysis will be performed as follows to explore whether the results of the review are robust.
1) Studies will be grouped qualitatively into low, moderate or high quality and meta‐analyses performed by group.
2) Comparisons will be made between studies which use a crossover design and those with a separate control arm.
3) If missing data have been imputed for intention‐to‐treat analysis with dichotomous outcomes, a "best case/worst case" analysis will be performed. All participants for whom data is missing will be assumed to have had a favourable outcome (response to medication) and the results compared to the original, more conservative, analysis.
These analyses will be compared with the original meta‐analysis and any effect on the results noted.
Assessment of Reporting Bias
1) Publication bias will be assessed by constructing funnel plots.
2) If unpublished data is included in the review, a subgroup analysis will be performed to compare published and unpublished data. |
