Types of study

We included randomised controlled trials (RCTs) for evidence of effectiveness and economic evaluations for evidence of cost-effectiveness (including cost-effectiveness, cost–utility and cost–benefit analyses).

Types of population

We included adolescents or adults (mean sample age ≥ 13 years) who have participated in a study designed to evaluate the effectiveness of e-learning to promote dietary behavioural change. We included all clinical conditions for which dietary advice plays a major part in case management.

Types of intervention

Interventions were included if there were interactive computer software programs that tailored output according to user input (i.e. second- and third-generation interventions). These include interventions for which users enter personal data, or make choices about information, that alter pathways within programs to produce tailored material and feedback that is personally relevant. Users may interact with the programs as members of a small group, as well as individually.

Programs should be available directly to users and allow independent access without the need for any expert facilitation.

Interventions were excluded if they were:

• first-generation tailored ‘information only’ (e.g. providing a leaflet or PDF file)
• simple information packages with no interactive elements
• non-interactive mass media interventions (such as TV advertisements)
• interventions designed to be used with others' help (e.g. teacher or health professional)
• interventions targeted at health professionals or teachers
• computer-mediated delivery of individual health-care advice (e.g. online physicians)
• electronic history-taking or risk assessment with no health promotion or interactive elements.

Outcome measures

We anticipated that most interventions would be aimed at dietary behaviours and were unlikely to have followed participants long enough to obtain changes in clinical measures. However, as measures of dietary behaviour tend to be based on self-report, they are prone to error (e.g. recall bias). Biological outcomes [e.g. body mass index (BMI)] tend to be measured more objectively (e.g. using measures of weight and height) and are also the necessary inputs to economic models of cost-effectiveness. We therefore specified dietary behaviour as our primary outcome, and also obtained data that allowed us to model the relationship between behaviours and clinical changes.

Search process

Our search comprised the following:

• a search of electronic bibliographic databases for published work
• a search of trial registers for ongoing and recently completed trials
• inspection of the reference lists of all included studies and previously published reviews
• contact with authors of included studies and e-health research groups to check for more trials.

There were no restrictions by language. The search strategy comprised two concepts: ‘computer-/ internet-based interventions’ and ‘dietary behaviour’ (see Appendix 2 for full electronic search strategies).

Eleven electronic databases, two trials databases and two theses databases were searched using the search strategy (Table 1). Searches covered the period January 1990 to November 2009 (we assumed that any studies of e-learning conducted in the 1980s would be identified through inspection of the reference lists of all included studies).

TABLE 1

Databases searched for relevant references.

Selection process

All studies identified through the search process were exported to a bibliographic database (EndNote version X3; Thomson Reuters, CA, USA) for de-duplication and screening. Two review authors independently examined the titles, abstracts and keywords of electronic records according to the eligibility criteria above. Results of this initial screening were cross-referenced between the two review authors, and full-text records obtained for all potentially relevant reports of trials. These potentially eligible trials went through a secondary screening by each reviewer using a screening form based on the eligibility criteria (see Appendix 3) for final inclusion in the systematic review, with disagreements resolved by discussion with a third author.

Data extraction

Two review authors extracted relevant data into a Microsoft Access 2007 (Microsoft Corporation, Redmond, WA, USA) database specifically designed for the review (available from the authors on request). Corresponding authors of included studies were contacted directly by e-mail when required information or data were not reported in the published report, using a pre-notification e-mail followed by up to two contact attempts.

Descriptive analysis

We described all studies that met the inclusion criteria, including (where reported):

• study design:

  –

  study objectives (i.e. target outcomes)

  –

  trial design and quality

  –

  data collection methods, modes and techniques; validity of tools

• participants:

  –

  socioeconomic and demographic characteristics

  –

  health status: diagnosed disease versus no diagnosed disease

• intervention:

  –

  components of the intervention, including delivery and content

  –

  frequency, intensity and duration of the intervention

  –

  behaviour change theories employed in intervention design, and postulated mediators

• outcomes:

  –

  primary and secondary outcomes measured

  –

  information on process (ease of use) and usage (compliance).

Information on the sociodemographic characteristics of participants was used to address concerns over the ‘digital divide’. Where sufficient data were provided by the primary studies, we planned to undertake subgroup analyses of intervention effects in low-income and low-educational-status users.

Analysis of effectiveness

When studies reported the same outcome (e.g. servings of fruit and vegetables eaten per day, percentage of energy from fat), we pooled the results using a random effects model, with weighted mean differences (WMDs), and calculated 95% confidence intervals (CIs) and two-sided p-values for each outcome. When outcomes were assessed more than once during follow-up, the final assessment was used in analysis. In studies in which the effects of clustering were not taken into account, we adjusted the standard deviations (SDs) by the design effect, using intraclass coefficients if given in papers or using external estimates obtained from similar studies.⁴¹

We assessed evidence for selection bias using Egger's test for small study effects. Heterogeneity among the trial results was assessed using both a chi-squared test and the I² statistic, the percentage of among-study variability that is due to true differences between studies (heterogeneity) rather than to sampling error. We considered an I² value > 50% to reflect ‘substantial heterogeneity’. We conducted sensitivity analyses in order to investigate possible sources of heterogeneity including study quality (adequate vs inadequate allocation concealment, low vs high attrition) and sociodemographic factors that could act as effect modifiers [e.g. gender and socioeconomic status (SES)].

When studies reported more than one measure of a single outcome, the measure used was that for which greatest validity had been demonstrated [e.g. if a validated Food Frequency Questionnaire (FFQ) was used to measure the number of portions of fruit and vegetables eaten daily, as well as an unvalidated single item ‘how many portions of fruit and vegetables do you eat each day?’, we chose the former for inclusion in any subsequent meta-analysis].

Causes of heterogeneity and subgroup effects were assessed using random effects meta-analysis. This was implemented in Stata (StataCorp LP, College Station, TX, USA) using the ‘metareg’ command, and included trial characteristics as covariates. All statistical analysis was conducted using Stata statistical software version 11.