Chronic or long-term conditions such as diabetes have a significant impact on the individual’s quality of life and are the major cause of disability and premature death worldwide [1,2]. In high-income countries, chronic conditions are estimated to account for 87% of all deaths [3]. Due to their nature, they cause illness over long periods and their management is complex and costly [2]. In countries where health care is universally provided (such as the National Health Service [NHS] in the United Kingdom), services are struggling with the increasing demand, partly driven by the increasing number of people with chronic conditions [4-7]. Even in high-income countries, people living in constrained conditions and with a lower socioeconomic status (SES) experience chronic illness more commonly and severity is greater than average [2,8]. For example, 52% of those from unskilled occupations suffer from chronic conditions in comparison with 33% of those in professional occupations [9]. These inequalities in health have been attributed to social determinants of health (SDH) and inequity in access to health care [10]. SDH are the complex interacting elements in the physical and social environment that contribute toward disparities in health status. Inequalities in the distribution of good quality health care mean that people do not have equal access to treatments that can improve health outcomes. Taken together, SDH mean disadvantaged groups suffer more illness and more severe illness, but are least likely to receive effective treatment, which together result in disparities in health outcomes [11].

Both self-care and web-based interventions have been proposed as methods for increasing access to health care while relieving pressure on health care services. The underlying assumption with self-care interventions is that they provide health care where there was none, by encouraging people to be their own health resource [12]. Web-based interventions have the potential to increase access to good quality health care by providing support to an almost unlimited number of people from the same digital platform at the same time, and the interventions can be tailored to individual needs [13-16]. More recently, combinations of the 2 approaches, in the form of web-based self-care interventions, have become more prevalent for a range of health conditions, particularly for chronic conditions [17,18]. These digital self-care interventions are viewed as playing a vital role in the prevention and treatment of long-term illnesses such as chronic lung disease, cancer, cardiovascular disease, and diabetes [18,19].

Evidence of the impact of these approaches on health inequalities has been mixed. Nondigital self-care interventions in the form of community-based training courses have been found to improve health status, health behaviors, and the quality of life of patients with chronic conditions [20-24]. However, there is evidence that disadvantaged groups face barriers to accessing these interventions, such as the high levels of health literacy that are often required to understand the training materials as well as language barriers where the training is only conducted in English [8]. Web-based interventions designed specifically for those from underserved and disadvantaged groups have been found to benefit these populations [14,25,26]. However, there is also evidence that access and usability for disadvantaged groups remain to be barriers [27-30]. People from lower socioeconomic groups and older adults are less likely to seek health information on the web and have problems using the web-based information available [31-38]. There is an absence of systematic review evidence investigating whether web-based self-care interventions designed for people with chronic conditions are equally effective for people with different social characteristics.

This study aimed to synthesize evidence investigating whether web-based self-care interventions designed for people with chronic conditions are equally effective for people with different social characteristics.

There were no language restrictions. The publication dates were limited from January 1, 2006, to February 20, 2019, to ensure that the review included interventions with recent technology.

One author (ST) extracted the data from the included studies. Where more than one outcome was reported in each category, the primary outcome was included. Quality of life (QoL) was categorized as a health or psychosocial outcome depending on whether there was a greater balance of health or disability or psychosocial questions in the QoL tool.

The PROGRESS Plus (PP) framework was used to identify the SDH that could contribute toward health inequalities in the included health conditions and in the context of web-based interventions [51]. The analyses that had explored the modification of intervention effectiveness by social characteristics in the PP framework were extracted.

Methodological quality was assessed independently by ST and either PL, CC, or AH. Disagreement was resolved by discussion, and consensus was reached for all of the risk of bias (RoB) domains. The Cochrane collaboration RoB tool 1.0 was used to assess the quality of RCT studies [52]. The newly updated version 2.0 was not used as it does not allow for assessment of other bias and would therefore not allow us to capture issues with selective recruitment. The Risk Of Bias In Non-randomized Studies-of Interventions (ROBINS-I) tool was used to assess the quality of non-RCT studies [53]. These tools were used to produce an overall RoB rating for each study.

Much of the Cochrane RoB assessment is focused on ensuring that there is balance in the samples in the two arms of the study. However, the potential for selective recruitment is also important to ensure that the sample is representative of the population with the condition. Here, selection bias was assessed under the other category. Inclusion and exclusion criteria were examined to determine whether they potentially excluded people who experienced greater health inequity (eg, no access to the internet, not having the skills to use it, language barriers) and whether there was a discussion of the study population being representative of those with the condition.

Descriptive tables were populated using the data from the included papers, accompanied by a narrative synthesis. A meta-analysis was not possible due to the heterogeneity of outcomes. A novel summary figure was developed for this systematic review, which was based on an adapted version of the Harvest plot, referred to here as the Adapted Harvest plot [54]. The Adapted Harvest plot allows for a direct comparison of the sample size of the studies where the effect was found (or not found) across outcomes, and an impression of the quality of the study through RoB. This gives an indication of the strength and validity of the findings in relation to each outcome. A key explanation of the features and representation of the Adapted Harvest plot is shown in Multimedia Appendix 2, and information is provided in each plot. To be inclusive, all reported trends (P<.10) were included as evidence regardless of whether they fell under the standard P<.05 probability cut off and were reported in the text.

We intended to conduct subgroup analysis of differences in the application of behavioral change techniques and theory in intervention development and differences in use as potential mechanisms for modification of intervention effect by the PP group. However, these were inconsistently reported, making it challenging to analyze and draw any conclusions. Only 7 of the studies reported modification of use by PP characteristics, and 11 described the use of theory in intervention development. The selection and application of behavioral change techniques have largely not been described. Therefore, the exploration of mechanisms for modification of the intervention effect did not progress to a full analysis.

After the removal of duplicates, 7346 records were obtained (Figure 1). A total of 18 studies (reported across 19 papers) were eligible for inclusion in the review (Figure 1).

Of the 18 included studies, 1 focused on asthma [55], 2 on COPD [56,57], 13 on diabetes (5 on type 1 [58-62], 6 on type 2 [63-68], and 2 on both [69,70]), and 2 studies included participants with osteoarthritis [71,72]. The study characteristics are described in Multimedia Appendix 2. In total, 9 of the studies were conducted in the United States, 3 in the Netherlands, 2 in Canada, and 1 in the United Kingdom, Australia, France, and Israel; 13 of the studies were RCTs and 5 were non-RCTs. In total, 6003 participants were included, and the study size ranged from 38 to 1799. The reported characteristics of the study participants are presented in Multimedia Appendix 3. Studies of both children and adults were included, with an age range of 12 to 75 years across the studies. Half of the participants were female (50.1%).

Only 1 study reported that the patients enrolled in the study were representative of the population with diabetes [69]. In total, 4 studies purposely recruited a highly diverse sample in terms of ethnicity [62,67,69,73]. Follow-up times varied from 2 weeks to 12 months, and the follow-up time point was not clear for 1 study [69]. Study attrition ranged from 0% [67,70] to 31% [65], and attrition was not clearly stated in 2 studies [69,71].

In total, 4 of the studies explored the effectiveness of smartphone apps and the remaining 14 explored interventions delivered through websites. Descriptions of the intervention content are provided in Multimedia Appendix 4 and summarized below. In addition to the website or app, 3 interventions provided remote support by phone or video call [55,63,65,72], 1 provided a workbook [67], 1 provided a blood glucose monitor with wireless transfer [69], and 1 provided motivational interviewing before use of the website [61]. In 3 studies, there were 2 versions of the intervention, 1 with and 1 without external support (eg, email or phone support) [60,64,67].

A total of 15 of the 18 studies that explored the modification of intervention effectiveness by PP categories reported effect modifiers. There was evidence that people were more likely to benefit from the intervention if they were from a minority ethnic group, were employed, had a higher level of education, and had divorced parents (study of adolescents). The findings for age, gender, and health literacy were mixed. There was no evidence of an interaction income, numeracy, or the number of people living in the household. The full key to the Adapted Harvest plot is shown in Table 1. Further details including outcomes and estimates (where provided) where interactions were found are presented in Multimedia Appendix 5 and a matrix table containing an overview of interactions across the conditions and PP groups is provided in in the Multimedia Appendix 6.

This review examined the modifying effects of participant characteristics on the effectiveness of web-based interventions for chronic health conditions. We found evidence that intervention effectiveness was modified by participants’ social characteristics. In the diabetes literature, there was evidence that people from minority ethnic groups gained greater benefits from interventions than majority ethnic groups. There was evidence from single studies with variable quality that those with higher education, divorced parents (adolescent), and who were employed benefited more from interventions. A small high-quality diabetes study indicated that those with a higher level of education benefited more from the intervention. A small low-quality diabetes study found that a higher proportion of adolescents with divorced parents (adolescents) were in the high-effect intervention group compared with the low-effect group. A high-quality osteoarthritis study indicated that employed participants were more likely to benefit from the intervention than unemployed participants. The findings for modification of the effect by participant age were mixed. Older people were found to benefit less from the intervention in high-quality COPD studies, and they were found to benefit more from the intervention in 2 low-quality diabetes studies and a low-quality asthma study. Gender and health literacy were only considered in studies on diabetes and showed mixed effects. A total of 2 small low-quality studies indicated that male participants benefited more from diabetes interventions, and 1 small high-quality study indicated that female participants benefited more. Two studies had contradicting findings for health literacy: a small study with an unclear RoB indicated that those with higher health literacy benefited more, while a small low-quality study found that those with lower health literacy benefited more. There was no evidence of a modification to intervention effectiveness by income, numeracy, or the number of people living in the household.

The strength of evidence across the studies could not be assessed through a meta-analysis as not all studies provided estimates, and the outcomes were heterogeneous. Therefore, the strength of the evidence was explored through study size, RoB, and estimates, where possible. The majority of the evidence was drawn from studies with high and uncertain RoBs.

To our knowledge, this is the only systematic review that has investigated whether there are equal benefits of web-based self-care interventions for people with different characteristics. The breadth of the studies reviewed here is both a strength and a weakness. The inclusion of 4 different physical health conditions meant that it was not possible to combine the evidence in cases where the outcomes targeted were heterogeneous, which limits what can be inferred from the results. However, we summarized the available evidence, providing the first systematic exploration of how PP characteristics modify intervention effectiveness. Simple vote counting was not judged appropriate for the synthesis of findings that could not be meta-analyzed. Instead, we created a novel summary figure based on the Harvest plot referred to here as the Adapted Harvest plot. The Adapted Harvest plot provided an indication of the strength of the evidence for narrative synthesis by including study size and RoB.

The majority of the screening was conducted by one person, and only 10% of the abstracts and titles were double screened. This is in line with previous practice where a large number of studies were located [46,74-77]. Every effort was made to locate all relevant literature; however, it is possible that some relevant studies may have been missed.

There were several limitations in the methodology and data reported by the studies included in the review. It was not possible to conduct a meta-analysis because the outcomes were heterogeneous, and essential data were not reported. A high proportion of the included studies found that the intervention effect was modified by at least one participant characteristic on at least one outcome. This suggests that teams may be more likely to publish these analyses when they find evidence of a difference in effect [78].

There was a high risk of selection bias across the included studies, which may have excluded people from lower SES groups. The few studies that did comment on the representativeness of the study population indicated that the samples tended to be more white people, with higher levels of education. This not only limits the generalizability of the study findings but also potentially masks differences in effectiveness that may have been present between more and less advantaged groups because the sample is underpowered to detect differences in these subgroups.

The complex relationship between social characteristics and potential effectiveness or engagement with online interventions was not considered carefully in the included studies. The range of different PP characteristics explored and the comparisons within the PP characteristic appeared to be decided post hoc and did not appear to be guided by theory. Therefore, the evidence on the influence of PP on effectiveness was limited.

The Cochrane RoB 1.0 tool was limited in evaluating bias in the data that were important for this systematic review. Although the tool is very effective at identifying bias that can arise from the arms of a study with an unbalanced sample, it does not provide a category that evaluates the risk of the sample not being representative of the general population with the condition. RoB 2.0 was not used in this study because it does not allow for other RoBs and would therefore not allow us to capture issues with selective recruitment.

There is no previous systematic evidence comparing the health equity effects of web-based self-care interventions for people with different social characteristics. This review agrees with previous evidence from single studies that have found that web-based self-care interventions can benefit underserved and disadvantaged groups when the intervention has been designed specifically for such groups [14,25,26]. Some of the included studies that found evidence that underserved groups benefited more from the intervention had modified their interventions to be more accessible, useable, or engaging for these groups. Two studies designed the intervention, so it was in an accessible format for those with different educational abilities. A study found that those with lower health literacy benefited more and designed interventions to maximize usability for people with lower literacy [65]. A study found that those with divorced parents and lower baseline knowledge benefited more from the intervention and opted for a serious-game intervention that was designed to be appealing to a range of ages (11-18 years) and therefore a range of baseline knowledge in boys and girls [59].

Of the 4 studies that found that ethnic minority groups benefited more, 2 studies had adapted the intervention to be appealing to the study population targeted. One study targeting Latinos with T2D provided the intervention in Spanish and English [64,73]. Another study that found that ethnic minority youth with T1D benefited more from using the intervention than majority ethnic groups used an intervention with a graphic novel format and a cast of ethnically diverse characters with T1D [62]. However, there are no specific design features that suggest why ethnic minority groups may benefit more than majority ethnic groups in these studies. Indeed, the other 2 studies that found ethnic minority groups benefited more did not adapt the interventions to make it more accessible to ethnic minority groups. It is possible that in these studies, ethnic minority groups may have benefited more from the intervention because they had less exposure to health care support before using the digital intervention. There has been previous evidence that those from ethnic minority groups and those with lower SES face greater challenges accessing health care services and support [79-81]. Therefore, interventions that reduce barriers to access and use may be more effective for populations currently underserved by health care services.

Common to the 4 studies that found minority ethnic groups benefited more from the intervention was their sampling strategy aimed at maximizing the recruitment of minority ethnic groups [62,67,69,73]. Subsequently, all 4 studies had a high representation of people from ethnic minority groups in the sample, resulting in the sample being powered to detect differences in effectiveness by ethnicity. Webb et al [82] similarly found that recruitment sampling was an important predictor of the effectiveness of an intervention. They found that when theory or predictors were used to select recipients for the intervention, the intervention had the greatest improvements in behaviors [82]. The 4 studies in this review cited the potential for digital interventions to increase access to health care in these minority groups, as motivation for their study design, and target the recruitment of minority groups [62,67,69,73]. Therefore, consideration of the sample where the intervention was evaluated appears to be important in addition to considering the needs of the target population. 

Interventions designed without considering the needs of the users can exclude social groups, and this is the likely cause of the difference in effectiveness found by education and employment. Van Dijk’s theory of the digital divide proposed that if the content of the technology only fulfils the needs of the dominant group (eg, high education, employed) or is challenging to use, those users from the less dominant group will benefit less from the use of the technology [83]. This supposition has been supported by findings that the design of web-based health information can limit the usability of digital interventions for people of lower SES [36,37,84,85].

The mixed findings for age, gender, and health literacy may be associated with whether the interventions were designed considering the needs of people with those characteristics. This was illustrated by the findings for health literacy in this review. Davis et al [65] designed their intervention with the needs of low literacy individuals in mind and found that those with lower literacy levels benefited more from the intervention. Huang et al [58] found that those with higher literacy benefited more and acknowledged that they would need to provide additional support for users with lower health literacy levels at baseline. Alternatively, mixed findings may be related to the participant’s other social characteristics. For example, in studies where older people were found to benefit more from the intervention, they may have had a higher level of education and consequently higher digital skills relative to their younger counterparts. There is growing evidence that individual social characteristics do not work in isolation but interact in complex ways that influence health outcomes [86]. As such, conducting an analysis involving the comparison of individual groups may not be sufficient to establish how digital self-care interventions may impact health inequities.

There was evidence that web-based self-care interventions for chronic conditions can benefit some (minority ethnic groups, divorced parents) and disadvantage other (low education, unemployed) social groups who have historically experienced health inequity. However, these findings should be treated with caution as most of the evidence came from a small number of low-quality studies. The findings for gender and health literacy were mixed across diabetes studies, and the findings for age were mixed across asthma, COPD, and diabetes studies. There was no evidence that income, numeracy, or the number of people living in the household modified intervention effectiveness. We conclude that there appear to be interaction effects that warrant exploration in future research, and a priori consideration of predicted interaction effects is recommended.