Introduction

Health promotion and the prevention of ill-health via population and individual level interventions are key recommendations of the World Health Organization for the management of communicable and non-communicable diseases [1, 2]. Specific health education interventions are considered integral to system-wide public health strategies [3, 4]. Such educational interventions commonly aim to promote understanding about how behaviours impact health, and require individuals to have the capacity to acquire, understand and operationalize the content of health education in order to improve their health status [4, 5]. These capacities are influenced by the social and economic circumstances of individuals’ lives [6, 7].

Social and economic circumstances also importantly contribute to inequalities in health. This is depicted by the ‘social gradient’ in health, [8] whereby the lower a person’s socio-economic position, the poorer their health status. ‘Unhealthy’ behaviours associated with the development of chronic disease, such as smoking, poor diet, too little physical activity, and low engagement with preventative (e.g. screening) healthcare, are more prevalent among individuals who are socially or economically disadvantaged [9, 10]. Public health interventions to promote healthy behaviours may therefore be of most importance for these populations.

Socio-economically determined disparities in health outcomes can sometimes be further increased by behavioural health promotion initiatives, particularly those that are delivered across a large population. Benefit seems to be related to individuals’ access to social and economic resources and improvement is lowest in disadvantaged groups [10, 11]. For example, peoples abilities to respond to health promotion messages by changing health behaviours (such as improving diet and exercising regularly) vary widely – but changes are less likely to be adopted amonst low-income groups [10]. Similarly, technological interventions to improve health outcomes “work better for those who are already better off”(p. 1080), for reasons that stem from discrepancies in accessibility, adoption, and adherence [12]. Intensive, small-scale interventions targeted to high risk populations may be more likely to generate benefits, but economic and practical issues commonly limit broad implementation. Even the best-intentioned interventions frequently fail to reach, and to impact, those whose health needs are greatest.

Although specific educational interventions to improve health literacy and health-related behaviours are considered integral to public health interventions, little is known about the extent to which educational interventions that target disadvantaged populations are effective, nor about the intervention characteristics that are associated with success. Our principal objective was to identify and synthesize evidence of the effectiveness of health-related educational interventions in adult disadvantaged populations. Our primary outcome was health related behaviour, and our secondary outcome was a biomarker related to the health intervention. Our secondary objective was to summarise the characteristics of effective interventions.

Methods

We registered our full protocol a priori on Open Science Framework (https://osf.io/ek5yg/). Our study is reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement, [13, 14] the Checklist of Items for Reporting Equity-Focused Systematic Reviews (PRISMA-E 20,212 Checklist), [15] and the Synthesis Without Meta-analysis (SWiM) [16] reporting guidelines. We deviated from the registered protocol by reconsidering our approach to addressing the secondary objective of this study and undertaking an additional vote-count analysis.

Search strategy and selection criteria

We developed a comprehensive search strategy with the assistance of a health librarian and systematically searched five electronic databases (MEDLINE, EMBASE, EMCARE, and the Cochrane Central Register of Controlled Trials (CENTRAL)) since inception to 20th May 2020 to identify eligible studies. We updated these searches on 5th April 2022. Studies were limited to those involving human participants and available in English. Details of the search strategies are provided in Appendix 1.

We searched for studies that assessed the effectiveness of any health-related educational intervention delivered to socio-economically disadvantaged adults in any country. We defined health according to the World Health Organization definition, as: “a state of complete physical, mental and social well-being and not merely the absence of disease or infirmity” [17]. We defined socio-economically disadvantaged adults as belonging to a socio-economically disadvantaged population, classified as: “an area, neighbourhood or community with residents clearly defined as disadvantaged, relative to the wider national population” [18] (p. 372). Socio-economic disadvantage could be defined by factors including (but not limited to) income, educational level, living standards, and minority grouping. To be eligible for inclusion, at least 75% of participants in the included studies were required to meet this definition of belonging to a socio-economically disadvantaged population and be aged 18 years or over.

Published, peer-reviewed experimental studies investigating the effectiveness of an educational intervention on health-related outcomes were considered for inclusion. Eligible designs included (but were not limited to): randomised controlled trials, quasi-randomised and cluster-randomised trials. We excluded studies that were not published in English, pilot studies, reviews, commentaries, and case study reports, studies that did not describe the study population sufficiently to enable classification as ‘socio-economically disadvantaged’, and studies that did not report at least one outcome of interest.

Interventions and outcomes

Studies included in this review must have evaluated the effectiveness of an educational intervention. Interventions were considered to be ‘educational’ if the authors described the intervention as having intent to ‘educate’ or ‘inform’. Studies evaluating an educational intervention as their main objective or as a component of a comprehensive intervention were eligible for inclusion. Individual, group, community or population-based health education interventions, delivered through any medium (e.g. face-to-face, telephone, text, online, mass media) were considered. Included studies needed to have compared the educational intervention to any type of intervention, placebo, or no-treatment control. The primary outcome was health-related behaviour, or actions that individuals take that affect their health [19]. All behavioural outcomes that were considered to be health related and related to the study intervention were regarded as relevant. The secondary outcome was any biomarker related to the health condition the intervention was targeting (e.g. body mass index (BMI) as a biomarker of weight loss; or Haemoglobin A1C as a biomarker of diabetes control).

Screening and data extraction

Identified studies were retrieved and exported into Endnote citation management software (Clarivate Analytics, Philadelphia), and then imported into Covidence systematic review management system (Veritas Health Innovation Limited, Australia). Duplicates were removed. Pairs of reviewers independently screened all titles and abstracts for relevance according to the inclusion and exclusion criteria (AG, CP, TA, LW and RS). The full texts of potentially eligible studies were obtained, the article further screened for eligibility and reasons for exclusion recorded. Any discrepancies or disagreements between the two reviewers were discussed. If agreement was not met, a third reviewer (EK) was consulted to provide opinion and a majority decision was made.

Pairs of reviewers independently extracted the relevant data from each study using a standardised and pilot-tested spreadsheet. The results were compared, discrepancies discussed, and a third reviewer was consulted to resolve disagreements if required. The data extraction template included the fields: study design, health ‘condition’, population characteristics (including reason for classification as socio-economically disadvantaged), participant characteristics, sample size, details of study intervention(s) and comparator, assessment time points, outcomes, and results.

Risk of bias assessment

Pairs of authors independently evaluated the risk of bias (ROB) for each study using the Cochrane Collaboration’s tool for assessing ROB in randomised trials [20]. Six domains were evaluated: selection bias, performance bias, detection bias, attrition bias, reporting bias, and ‘other’ bias. We used the guideline provided by the Cochrane Handbook to assess each item as high, low or unclear ROB. A third reviewer was consulted to resolve any disagreements between the independent evaluations if required. Overall ROB was also assigned according to the Cochrane Handbook. Low overall ROB was assigned for studies where all key domains were low risk; unclear overall ROB was assigned when key domains were either low or unclear; and high overall ROB was assigned when one or more of the key domains were assigned a high ROB.

Data analysis

To address our primary aim – to identify and synthesize evidence of the effectiveness of health-related educational interventions in disadvantaged populations – we extracted effect sizes and precision estimates from the included studies where available. If an effect size was not reported we extracted the number of participants in each condition, the means and standard deviations of the observations (at the longest follow-up timepoint). We examined the clinical and methodological heterogeneity between the included studies to determine the appropriateness of combining the effect sizes to estimate an overall effect for our primary and secondary outcomes. To determine the appropriateness of data pooling we primarily considered homogeneity of outcomes, follow-up durations and comparison groups. In cases where studies were considered to be sufficiently (clinically and methodologically) homogenous for pooling, but data were missing, we contacted study authors to request the missing data. Authors were emailed, with a follow-up email sent two weeks later. In the case of no reply a further email was sent after another week, and if there was still no reply the data were not included. Random effects meta-analysis (DerSimonian and Laird model [21]) was conducted using Comprehensive Meta-Analysis software (version 3We evaluated the quality of the evidence of the included studies and rated the certainty of recommendations using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework [22]. Publication bias was assessed by visual inspection of a funnel plot; Egger’s test was applied if there were 10 or more studies in the meta-analysis [23].

Since meta-analysis could only be performed on a proportion of the studies, we summarised the overall effectiveness of interventions for our primary and secondary outcomes using a vote-counting approach [20]. When studies specified a single primary outcome, we determined intervention benefit from that outcome. We classified ‘intervention benefit’ using a standardised binary metric assigned according to the observed direction of effect. This classification was based on the point estimate of effect, without consideration of statistical significance or the size of the effect. Studies with a point estimate of effect in favour of the intervention were counted as [1]; studies with a point estimate of effect in favour of the control were not counted. When studies had two or more outcomes, we applied a decision rule to identify a single outcome from which to classify intervention benefit (Appendix 2). We calculated the number of effects showing benefit as a proportion of the total number of studies and determined a confidence interval using the Agresti-Coull interval method recommended for large sample sizes [24]. We undertook a subsequent calculation in which we determined the proportion of effective interventions by classifying benefit (for the outcome of interest) according to the conclusions of the individual studies, rather than using the point estimate to indicate effect. This approach minimised the risk of an inflated vote-count result.

To address our secondary objective – to summarise the characteristics of effective interventions – we tabulated details of the intervention (setting, type, dose, description) in a format to facilitate reader interpretation. Classification of intervention dose [25] (as low, moderate, or high) considered intervention duration (in months), frequency (number of contacts), and amount (in hours) (see Appendix 3 for details). We aimed to provide a summary of the features of the effective interventions.

Role of the funding source

The funder of this study played no role in the study design, data collection, data analysis, data interpretation, writing of the report or decision to submit the paper for publication.

Results

Our searches identified 8618 records; 200 full text articles were screened for eligibility; 96 studies were included (Fig. 1). Key characteristics of the included studies are provided in Tables 1 and 2. Eighty studies (83%) were undertaken in high-income countries; four studies (4%) were undertaken in upper-middle income countries; ten studies (10%) were undertaken in lower-middle income countries; and 3 studies (3%) were undertaken in low-income countries (see Tables 1 and 2). Seventy-seven (80%) of the included studies were randomised controlled trials (RCTs); 12 were cluster RCTs (13%); 7 were quasi-experimental studies (7%). The educational interventions addressed a wide range of health issues. The most common education topics were parenting skills, pregnancy and newborn health, (14 studies each) cancer screening, multi-factorial healthy lifestyle interventions (11 studies each), diet (9 studies), smoking cessation (8 studies) and sexual health (5 studies). The total number of adult participants exceeded 57,000, residing in 22 different countries.

Fig. 1
figure 1

PRISMA flow chart

Full size image
Table 1 Characteristics of studies included in Meta-analyses (n = 16)
Full size table
Table 2 Characteristics of studies not included in Meta-analyses
Full size table

Risk of bias

All included studies had either high or unclear overall ROB. The ‘other’ ROB domain of ‘intention to treat analysis’ was most frequently assessed as high. High ROB ratings were also common for ‘number lost to follow up’ and participant blinding (Fig. 2; see Appendix 4 for full details). Visual inspection and interpretation of the funnel plots for each main meta-analysis (to evaluate publication bias) identified no major asymmetries in the distribution of effects for any of the outcomes (Appendix 5), suggesting a low risk of publication bias. Egger’s tests were not conducted because there were < 10 studies in each analysis [23].

Fig. 2
figure 2

Risk of bias summary

Full size image

Certainty in evidence

Our evaluation of certainty in the evidence for each main meta-analysis was conducted using GRADE. Our results are summarised in relation to each meta-analysis (below); detailed results are provided in Appendix 6.

Data synthesis

High clinical and methodological heterogeneity amongst the included studies precluded overall meta-analysis of effect sizes for the primary and secondary outcomes of this review. Instead, we considered outcomes that were evaluated in three or more of the included studies for meta-analysis. Pre-planned subgroup analyses (specified in the protocol) were explored for intervention complexity, the level of intervention and intervention dose.These were undertaken if there were two or more studies in a subgroup. Results of the main meta-analyses of behaviour outcomes are detailed below; results of subgroup analyses and the meta-analyses of biomarker outcomes are detailed in Appendices 7–9.

Meta-analyses: Behavioural outcomes

Fifteen studies had physical activity or exercise outcomes; nine had dietary outcomes; eight had smoking cessation outcomes; seven had cancer screening outcomes; and five had vaccination and breast-feeding outcomes. Meta-analysis was not conducted for studies involving dietary, smoking cessation, vaccination, and breast-feeding outcomes because of varied study designs, outcome measures, follow-up durations and comparison groups.

Moderate intensity physical activity

We evaluated the 15 studies with physical activity or exercise outcomes for clinical heterogeneity. Six of these studies (total n = 1330) used ‘moderate intensity physical activity’ as a primary or secondary outcome; the intervention group was compared with a minimal intervention, standard care or control group; and effectiveness was evaluated at ‘long term’ follow up [26,27,28,29,30,31]. We downgraded certainty in the evidence by one level due to high risk of bias. There is moderate certainty that the pooled effect of educational interventions, when compared to standard care, minimal intervention or control, is 0.05 (95% CI = -0.09–0.19; Tau2 = 0.01%) (Fig. 3). There was moderate heterogeneity (I2 = 31%), which we explored by removing one study that used a differing outcome measure (i.e. the percentage of participants who improved their physical activity in contrast to post-intervention physical activity measures) from the analysis (2011) [31]. This reduced I2 to 0.0% and the pooled effect increased to 0.11 (95% CI = -0.01–0.22). Subgroup analysis of studies with complex or ‘non-complex’ interventions were possible; the results are reported in Appendix 7.

Fig. 3
figure 3

The effectiveness of educational interventions at improving moderate intensity physical activity outcomes in socio-economically disadvantaged populations: random effects meta-analysis

Full size image

Cancer screening

We evaluated for clinical heterogeneity the ten studies that had cancer screening outcomes. Five of these studies (n = 2388) used rates of cancer screening as their primary or secondary outcome; the intervention group was compared with a minimal intervention, standard care or control group; and effectiveness was evaluated at ‘long term’ follow up [32,33,34,35,36]. We downgraded certainty in the evidence by four levels due to risk of bias, inconsistency (two levels), and imprecision in trial results. There is very low certainty that the pooled effect of educational interventions, when compared to standard care or minimal intervention is 0.29 (95% CI = 0.05–0.52; Tau2 = 0.24) (Fig. 4). The I2 value of 83% indicates a considerable degree of heterogeneity across trial results. We explored this heterogeneity by removing individual studies from the analysis, which had only a minor impact. Removal of one study [32] reduced statistical heterogeneity to a small degree (I2 = 75%). Subgroup analysis of studies with moderate or low-dose interventions were possible; the results are reported in Appendix 8.

Fig. 4
figure 4

The effectiveness of educational interventions at improving cancer screening outcomes in socio-economically disadvantaged populations: random effects meta-analysis

Full size image

Overall synthesis: Vote-counting

We performed separate vote-counting syntheses for the behavioural outcomes and biomarker outcomes. Vote counting based on direction of effect found that 67 of the 81 studies with behavioural outcomes had point estimates that favoured the intervention (83% (95% CI 73%-90%), p < 0.001); ten studies favoured the control, and four studies demonstrated equal effects for intervention and control conditions. Twenty-one of 28 studies with biomarker outcomes had point estimates that favoured the intervention (75% (95% CI 56%-88%), p = 0.002); four studies favoured the control. Calculation of votes based on ‘effectiveness’ being determined by individual studies found 47% of interventions were effective on behavioural outcomes, and 27% were effective on biomarker outcomes. The votes assigned to each study by both vote-count methods are presented alongside the available data and/or effect estimates in Table 3.

Table 3 Intervention characteristics and effectiveness
Full size table

Secondary objective: Characteristics of effective interventions

Narrative synthesis of the features of ‘effective’ versus ‘ineffective’ interventions was precluded by the high clinical and statistical heterogeneity of the included studies. We have organised the studies according to the health focus of the intervention in Table 3. This table provides descriptions of the main characteristics of the interventions alongside indications of effectiveness in order to facilitate reader interpretations.

Discussion

We aimed to (i) identify and synthesize evidence of the effectiveness of health-related educational interventions in adult disadvantaged populations, and (ii) summarise the characteristics of effective interventions. When studies were sufficiently homogenous to allow data pooling, meta-analyses revealed that health education interventions targeting socially disadvantaged populations produced positive behavioural effects that were small or negligible in magnitude. The certainty of evidence was low (at best). Our vote-count syntheses found a marked discrepancy in the proportion of effective interventions depending on the method applied to classify benefit (i.e., 85% versus 43% for behavioural outcomes and 83% versus 31% for biomarker outcomes). The evidence included in this review did not demonstrate consistent, positive impacts of educational interventions on health behaviours or biomarkers in socio-economically disadvantaged populations. We were unable to draw conclusions related to the common features of ‘effective’ interventions due to the high clinical and statistical heterogeneity of the included studies.

Meta-analysis of the six sufficiently homogenous studies aiming to increase physical activity showed no effect, but the four studies that were not included in the meta-analysis due to heterogenous outcomes reported significant improvements in the physical activity outcome compared to control interventions [37,38,39,40]. Of these four positive studies however, two had fewer than 50 participants [37, 38] and two had drop-out rates exceeding 48% [37, 39]. Thus, evidence suggests it is unlikely that educational interventions had changed physical activity in disadvantaged populations.

Educational interventions were shown to have a small, pooled effect (Hedges g = 0.3) on cancer screening rates, however certainty for this evidence was rated as ‘very low’. Five studies investigating cancer screening uptake were not included in this meta-analysis – two used varied outcomes (self-reported breast self-examination), [41, 42] two were low-dose, [42, 43] and two had comparison groups that were active interventions [44, 45]. From these studies, the interventions were effective for the two studies with breast self-examination outcomes, one of which analysed only 21 participants at follow-up. Based on the findings of the studies not included in the meta-analysis, the lack of evidence of benefit combined with the low quality of evidence reinforces that educational interventions to boost cancer-screening had, at best, small effects on cancer screening.

This review of evidence concerning the effectiveness of health-related educational interventions that target socio-economically disadvantaged populations is less encouraging than reviews of other health interventions in socio-economically disadvantaged groups. One review of mixed interventions for diabetes care [46] including novel providers’ roles, education and resources, found positive outcomes in 11 of the 17 included studies. The authors suggested that cultural tailoring, individualised components, multiple contacts (> 10), providing feedback, and involving community educators or lay people in delivery, were associated with better outcomes.

Our findings also show a stark contrast to the positive effect observed from health education interventions in non-disadvantaged socio-economic groups. Educational interventions designed to improve health-related behaviours such as oral health practices (15 studies), [47] foot self-care practices amongst diabetics (14 studies), [48] and cervical cancer screening rates (17 studies), [49] seem to provide mostly meaningful benefit. Education to promote self-management of hypertension demonstrated benefits on blood pressure outcomes in a systematic review of education programs that also targeted self-efficacy (14 studies) [50]. This contrast seems critically important because it raises the distinct possibility that educational interventions that are widely endorsed on the basis of their apparent effects, are often failing to meet the needs of the very people most likely to need them [51].

There are strengths and limitations of this work. We applied contemporary standards of transparency [52] and rigour, and reporting was in line with the PRISMA and PRISMA-E templates, and SWiM guidelines. We were unable to perform meta-analysis on a large majority of included studies due to heterogeneity. We synthesised data from these studies using two vote-counting methods: 1) studies were categorised as positive or negative based on direction of effect, regardless of effect size, and 2) studies were categorised as positive if the authors concluded the intervention was effective. The former method is recommended by Cochrane and does not consider statistical or clinical significance. Critically, neither approach provides estimates of the size of effects which is needed for policy or clinical decisions. The two synthesis methods provided very different results. Method 1 resulted in 83% of positive studies for behavioural outcomes and 75% for biomarkers, Method 2 resulted in 47% and 27% respectively. This inconsistency casts significant doubt over the usefulness of vote-counting approaches and means that we have very low certainty in our conclusions.

There may have been studies eligible for inclusion that were not identified by our database searches. For example, searching for specific conditions (e.g. diabetes) may have identified relevant studies not identified in our more general search for ‘health-related’ interventions; and studies that involved education as components of an intervention without explicit mention of this may have been missed. Citation chaining may also have identified further eligible studies. While not searching grey literature can contribute to an over-estimation of effectiveness (since null findings are less likely to be published in peer reviewed journals), this is unlikely to impact the findings of our review since most of the included studies concluded a lack of effect. Our evaluation of publication bias also suggests that this is not likely to be of major concern. Finally, it is important to acknowledge that we applied a very broad definition of socio-economic disadvantage when selecting studies for inclusion. While included studies most commonly involved participants with low income, types of disadvantage were also widely disparate (e.g., low educational attainment, living in rural areas, ethnic minority groups). Subgroup analyses of these factors was precluded due to study heterogeneity, such that it remains undetermined whether these varied types of disadvantage differentially impacted involvement in clinical trials or responsiveness to interventions. The impact of contextual factors associated with the economic classifications of the countries in which the study was conducted (e.g., lower middle income vs high income) is also unknown.

Conclusions

This review highlights that health-related educational interventions tested to date have not consistently demonstrated positive impacts on health behaviours or biomarkers in socio-economically disadvantaged populations. Based on this conclusion – along with the low certainty of findings and the high ROB of the majority of included studies – we suggest that targeted approaches must continue to be pursued, concurrent with efforts to gain a greater understanding of factors associated with their successful implementation and evaluation. This investment is likely to be important to reduce inequalities in health.