Study design

The steps of the proposed systematic review will include outlining the search strategy, the inclusion and exclusion criteria, finding studies, selecting those studies that address the review question and meet the inclusion criteria, and extracting and synthesising data. This systematic review protocol is guided by the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) Protocol checklist (online supplemental file 1).37

Search strategy

We will perform an electronic database search in JSTOR, PsycARTICLES, PsycINFO, PubMed, ScienceDirect and Web of Science to identify an initial set of records. A two-part search strategy will be used to identify studies that are eligible for inclusion. First, we will search electronic bibliographic databases for published studies using a string of preselected key words and phrases. Second, we will search the reference lists of records that meet the criteria for inclusion in the review and the reference lists of relevant, previously published review articles for other records that may be eligible for inclusion.

Inclusion and exclusion criteria

The study population will include PLHIV and who are on ART treatment. There will be no limits on study participants in terms of age, gender and ethnicity. The selection of studies will be restricted to RCTs and observational studies involving cross-sectional, case–control and longitudinal designs. Studies that use either validated and non-validated measures of one or more dimensions of body image and ART adherence will be included, regardless of the measurement approach (eg, self-report, report cards). We will exclude studies that examine lipodystrophy as an independent variable, because it is a factor that may contribute to a negative body image rather than a dimension or component of body image itself.32 38 However, if there are sufficient studies that both meet the inclusion criteria and report on lipodystrophy, we will include lipodystrophy as a covariate and report the percentage of participants who have lipodystrophy. Eligible records will consider body image as either an independent variable or a mediator, while ART adherence will be assessed as a criterion or outcome variable. Due to the ambiguity of parsing out effects of specific treatment components when interventions are multifaceted, only RCTs that exclusively and explicitly target body image or examine body image as a mechanism (ie, mediator) by which the treatment is associated with ART adherence will be included. We will include studies that have been peer-reviewed and published. We will exclude papers that are not available in English. Moreover, qualitative studies, mixed-method studies, commentaries, study protocols, literature reviews, conceptual papers and conference abstracts will also be excluded. We will manually check the reference lists of included articles and relevant review papers to identify additional studies that were not revealed during the formal database search. The Rayyan QCRI (https://rayyan.qcri.org) cloud-based platform will be used to manage the review process from initial screening of records through to selection of studies that are eligible for inclusion.39

Time period

All eligible studies published from 1 January 2000 to 31 March 2021 will be included. The period from early 2000 to the present marks an era of large-scale ART coverage, classifiable as (1) ART introduction 2000–2007, (2) expanded ART (2008–2010) and (3) scaled-up ART (2011–present). The progression in ART availability and usage during this time frame offers a suitable context for documenting the possible implications of long-term use of ART for PLHIV, such as the decline in healthy body image and its impact on ART adherence.

Study screening and selection

The first review author will use a set of keywords to search databases to identify studies that potentially meet the inclusion criteria outlined above. The search words for body image were informed by current literature and by Thompson et al40 who identified a range of terms that either referred to body image or a subdimension of body image. The search terms will be combined using the Boolean operators ‘AND’ and ‘OR’. The preliminary search strategy that was developed is detailed in table 1. The search strategy will be adjusted to fit the syntax specifications of each electronic database. Given the iterative nature of systematic reviews and the potential for additional search terms to be discovered during the search process, final search queries in each database will be published alongside the systematic review as online supplemental file 2.

Data extraction

When feasible, we will restrict the screening process to the title and abstract. However, if all the keywords necessary to include the record are not present in the abstract or title, the first review author will screen the entire article to decide whether or not it should be included. References of records that meet the criteria for inclusion will be searched manually to identify further eligible studies. The first and second review authors will extract the relevant data from eligible records and independently assess each for eligibility using a data extraction form that will be developed to standardise the data extraction process. The third review author will arbitrate any disagreements between the two review authors over the eligibility of particular records. A PRISMA flow diagram will be prepared to summarise the record screening and assessment process.

Data items

The authors will extract the following information from each of the included studies: (1) general characteristics of the publication: title, year of publication, first author last name and purpose of the study; (2) characteristics of the participants: age, sex, weight gain, mode of HIV transmission (ie, vertical vs horizontal), ART regimen, concurrent medication use and sexual orientation of participants; (3) research strategy: overall research design, number of participants, and country where participants were sampled from; (4) instruments: type of instruments used to measure body image and ART adherence, dimensionality of measures and method of data collection (eg, self-administered, interviewer administered); instrument reliability; (5) analysis and results: primary statistical analysis, variables included in the analyses and main outcomes (including potential confounding factors that were taken into consideration). If any included studies report results for multiple groups independently, we will (A) specify the groups used to subset the results, (B) report the overall effect size and (C) report descriptive statistics for relevant groups. We will contact the authors of included studies publications for missing outcome data or unclear information.

Data management

We will implement the search strategies and import all references identified into EndNote. We will record and report details on the number of full text papers obtained and the number of included and excluded articles. The search results from the different electronic databases will be combined in a single EndNote library and duplicate records will be removed.

Risk of bias in individual studies

The quality of the identified studies will be assessed independently by two reviewers using the Quality Assessment of Controlled Intervention Studies, the Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies, and the Quality Assessment of Case-Control Studies tool provided by the National Institute of Health (2015).41 Studies will be rated as good, fair and poor quality, where high risk of bias is translated to a rating of poor quality (‘−’) and low risk of bias is translated to a rating of good quality (‘+’). A traffic light plot will be generated to illustrate the proportion of studies classified as good, fair and poor, both for each category and overall.

Data synthesis

Effects from records included in the proposed review and meta-analysis will be coded to reflect the relation between positive body image and ART adherence. To aggregate the results across the studies identified, we will first convert all study results to a common effect size metric (ie, correlation coefficient). Fisher transformations will be used to convert correlations to approximate z-scores, which helps to stabilise the variance.42 The estimated sampling error of each correlation is approximated using the SE resulting from the Fisher transformation method. Given a sample correlation in study i with study sample size,the second order approximations43 to the corresponding are

(1)

(2)

(3)

Some studies may conduct logistic regression using ART adherence as a dichotomous criterion or outcome variable. In such cases, we will use the resulting beta-coefficients and the corresponding log OR to compute the between positive body image and ART adherence. For a given odds ratio (), we will use the Ulrich and Wirtz approximation to transformation ORs to correlations,44 that is

where with and being the number of subjects in each adherence group. We will then use this in (1) and (2) to transform the converted correlation to a z-score. Lastly, some studies may report the association between negative body image and ART non-adherence. In such cases, we will reverse the sign so that scores correspond to the association between positive body image and ART adherence. If additional unforeseen transformations are needed, we will make note and describe how these additional transformations were conducted.

The effect size estimates will be aggregated using a random effects meta-analysis approach. We will conduct a linear mixed-effects model with normally distributed errors. The within-study variation will be captured by the approximate sampling variance given by Equation (2). An unconditional model and a full conditional model will be considered for analysis. The unconditional model will be a random effects only model. For this, we mean that only the average effect size and the heterogeneity in effect sizes will be estimated. For the full conditional model, we intend to explain potential sources for why studies may differ in effect sizes. The size of the full model will depend on the final number of studies eligible for analysis (see inclusion criteria). If between three and five studies are included in the analysis, we will restrict our model to only approximate the average effect size, the heterogeneity of effect sizes, and one covariate (eg, research design, gender). If more than five studies are extracted, we will expand the full model to include additional characteristics. For every three additional studies an additional design factor will be included. For example, with six studies, two study characteristics will be included; with nine studies, three characteristics; and with 12 studies, four characteristics. The specific characteristics included in the full model will be determined by what characteristics are available, substantive importance and magnitude of effect. The effects of each design factor will be estimated as fixed across studies. The full random effects model can be expressed as

Where,

is the normally distributed sample z-transformed correlation where has known variance, that is . Each study may have its own subpopulation but the overall distribution of across studies is the main focus of this meta-analysis. That is, each comes from a population with mean θ,

θis the population effect-size,

is the study characteristic (ie, covariate) included in the study where is the estimated differences in the population effect size relative to changes in the covariate. All ’s will be mean centred such that the θ does not change meaning,

oFor example, if represents the proportion of females in the study, we would mean centre each estimate relative the average proportion of females across all studies,

is the deviation score associated with each study. The deviations are assumed to be normally distributed, that is . An important result is the variance of . That is, represents the between-study heterogeneity.

This relationship results in a linear hierarchical model where each effect size is drawn from a population, which we can restate the model without the deviation scores by using . The more elaborate approach makes it easier to see the deviation between the overall estimate of effect size (θ) and the study estimates.

Models will be estimated using a Bayesian framework.45 Estimation will be carried out with Stan,46 using the rstan package47 in R. The prior specification depends on the number of studies included in the meta-analysis. Generally, the prior choice for the fixed effects make little difference in the estimation of a random effects model. The priors are the variance components and tend to be highly influential, especially when the number of studies is below five.45 48 This is the statistical reason why only one fixed effect is to be included in the full model when the number of studies is five or fewer. We will use multiple sets of priors for each model to investigate the sensitivity of the meta-analysis results to selection of prior.

For the unconditional model, the full prior specifications are shown in figure 1, online supplemental file 3. For the relatively uninformative prior structure detailed in figure 1 (Panel A), all priors were selected to be flat (or essentially flat) over a logical range of possible values for θ and τ. The upper bound of two was chosen for the uniform prior on τ because each was mapped to a standard normal distribution with SD of 1. This means that the SD of the estimated effect sizes is most likely less than 1, but using a uniform prior with a relatively high upper bound of 2 allows for all plausible values with equal likelihood. However, the uniform prior for the SD can lead to potential bias in the estimates, as the use of the uniform prior has been shown to overestimate effects if the upper bound is too high.48 The relatively informative prior structure is shown in figure 1 (Panel B). The prior θ was chosen to be more informative of our belief that the association between positive body image and ART adherence is positive. This means that previous research has shown that the association is positive.31 The centre of the normal prior was chosen to be the Fisher transformation of the average of reported correlations () in two previous studies.21 30 To provide a relatively informative comparison to the prior used in (Panel A), the SD of this prior was chosen to be one. The half-t prior for τ was chosen based on the Gelman’s48 recommendations for priors when the number of studies is below five. The variance of the half-t prior was selected to 0.25, which means that we suspect the between-study variability will be small relative the standard normal distribution. The df for the half-t was selected to be one.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Prior specifications for unconditional random effects model. (A) The prior structure is relatively non-informative with diffuse priors for the population mean effect size (θ) and between-study heterogeneity (τ) parameters; (B) prior structure is relatively informative with lower variance priors for the population mean effect size and between-study heterogeneity parameters. Similar in both models is the transformation of the observed correlations () and study sample sizes () into the effect sizes () and study variance estimates ().

For the full conditional model, the prior structure will be nearly identical but the priors for the regression weights () parameters were added. In figure 2, online supplemental file 4, the priors for the regression weights are relatively diffuse to provide a relatively non-informative prior for the effect of each study characteristic. The relatively non-informative priors for each were chosen to be the same as the prior for (ie, normally distributed with large variance). In figure 2B, the priors for the regression weights are specified to be relatively informative to match the relatively informative prior structure specified for the other model parameters. The relatively informative priors for covariate effects () will be based on content expertise from the first author and relevant a priori literature. For example, under a potential gender effect for studies with different proportions of females, we selected the regression weight to be positive (). This specification indicates that studies with more females are likely to exhibit a stronger association between positive body image and ART adherence.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Prior specifications for full conditional random effects model. (A) The prior structure is relatively non-informative with diffuse priors for the population mean effect size,(θ between-study heterogeneity, and regression weight of study design characteristic () parameters; (B) Prior structure is relatively informative with lower variance priors for the population mean effect size, between-study heterogeneity and regression weight of study design characteristic parameters. Similar in both models is the transformation of the observed correlations () and study sample sizes () into the effect sizes () and study variance estimates ().

A prior-sensitively analysis will be conducted for each prior structure. The sensitivity analysis will test how dependent the posterior of the variance τ is to the chosen prior. For the relatively non-informative uniform prior, this will be done by changing the upper bound to smaller and larger values. For the half-t prior, the variance will be decreased and increased. We will indicate whether the initial prior structures were influential and how we selected priors to be less influential.

After we have completed necessary assessments of the posterior distributions, we will test for evidence of a positive association between positive body image and ART adherence by computing the probability that the pooled effect size is positive. This will be done using the posterior distribution of θ from the above models. If , then we have evidence that the correlation is positive. To facilitate interpretation of results, we will back transform the results to the correlation scale. We will note any potential differences in conclusions about the association between body image and ART adherence across the four models. We will use forest plots to display the results of the estimated study effect sizes and the pooled effect size estimate. If quantitative synthesis is not appropriate, the extracted studies will be described narratively to summarise and explain the characteristics and findings of the included studies.

Meta-bias(es)

To evaluate publication bias, we will follow the recommendations of Sterne et al49 by graphically representing the relationship between the observed effect sizes and estimation precision using funnel plots. Ideally, the plot appears symmetrical and points on the plot form an inverted funnel. Studies with smaller sizes tend to be scattered widely at the bottom of the plot whereas studies with larger sample sizes typically have greater precision in the estimates of effect sizes and are centrally located at the top of the funnel. Asymmetrical plots or blank spots within plots are evidence of potential bias. The asymmetry of the plot will be tested using Eggar’s test.50 The statistical test of asymmetry will only be conducted if a minimum number of studies (10) are included in the meta-analysis.51 Additionally, the trim-and-fill method will be applied to assess the impact of potential publication bias.52

Another potential source of bias in our results is a lack of available information about confounding variables. One such source is that different antiretroviral drugs/classes of drugs are related to weight gain, which is linked to body image. Knowledge of which classes of drugs were used in a particular study may then be linked to the observed effect size of the relationship between ART adherence and body image by the impact that particular classes of drugs have on weight gain. Although weight gain and class of drug may potentially be reported in each study that meet inclusion criteria, our preliminary searches indicate that such information is not widely reported. We expect a large amount of missing data for some of these potentially confounding variables. The potential bias resulting from lack of information about confounders will be investigated using sensitivity analyses.53 Similar to how we will be conducting prior-sensitivity analyses, additional sensitivity analyses will be performed to investigate how strong the effect of an omitted confounder would need to be in order to change the substantive conclusions of the meta-analysis.

Confidence in cumulative evidence

The Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach will be used to report on the overall quality of evidence for the outcome of interest.54 In brief, the two review authors will assess the quality of evidence across the domains of risk of bias, consistency, directness, precision and publication bias. Additionally, we will assess the magnitude of effect. A quality of evidence assessment presented alongside the effect estimates provides an indication of how much confidence can be placed in the findings. As a result of applying GRADE, two authors will independently rate the quality of the reviewed records and assign an overall quality rating of very low, low, moderate or high.55 These four categories reflect the level of confidence the reviewers have in the quality of the evidence. The high category indicates that further research is very unlikely to change the level of confidence in the estimated effect, whereas a very low category indicates uncertainty about the estimated effect.