Cancer accounts for 1 in 6 deaths globally. The number of cancer cases worldwide may increase by 60%, especially in low-income and middle-income countries where the increase may be as high as 81%, as estimated by the latest World Cancer Report [1]. This trend will undoubtedly lead to socioeconomic pressure and a shortage of medical resources. The current clinical research on cancer medicine mainly focuses on finding the cause of cancer and preventing the spread of cancer cells in the early stage [2,3]. However, these findings can be helpful only if people are sufficiently aware of adopting healthy behavior (ie, prevention and detection). This is a unique form of a health crisis that requires intensive communicative efforts. Therefore, messages need to be carefully designed in cancer communication to achieve positive health outcomes.

Message framing is an effective technique to change health behavior [4-10]. Health messaging attempts to change people's attitudes, intentions, or behaviors toward a specific health topic by emphasizing the expected benefits of undergoing specific health behaviors (ie, gain-framed messaging) or the possible loss if specific health behaviors are not done (ie, loss-framed messaging), to persuade people to follow healthy guidelines. According to the theory of reasoned action, attitude (or intention) is an essential direct predictor of behavior [11-13]. Moreover, attitudes, intentions, and behaviors are the standard measures of the effectiveness of health messages [4,14].

Health behaviors include disease prevention and detection behavior [4,6]. The former aims to avoid illness or deterioration, while the latter aims to reflect the presence or absence of risk. Gain-framed messaging fulfills the promise of a safer and more certain disease prevention measure in terms of disease prevention, so it is more effective than loss-framed messaging [4,15]. O’Keefe and Jensen [16] also showed the advantage of the gain-framed messaging in disease prevention, but they indicate that it does not apply to skin cancer [17,18]. On the other hand, in terms of disease detection, it was thought that the loss-framed messaging showed a higher persuasive effect [4,15]. However, the meta-analysis results show no significant difference between gain and loss framing [6,16]. Thus, apart from demonstrating the contribution of loss-framed messaging to breast cancer detection [17], there has been no definitive conclusion regarding cancer and message framing.

The concept of message framing is an essential strategy in health promotion, but research on how to optimally frame cancer prevention and detection messaging is scarce. Furthermore, while the effectiveness of message framing in cancer prevention and detection remains unclear, there is still an increasing number of related studies that combine message framing with other variables (ie, color) to verify the strengthening or weakening of the framing effect. Therefore, a thorough examination of the effectiveness of the message framing in the context of cancer is needed.

This review mainly aims to systematically summarize the characteristics of the relevant intervention studies and then pool the effect sizes from the relevant studies to quantify the effects of those interventions on attitude, intention, and behavior change. The findings of this review can provide recommendations for researchers and clinicians to design effective messages in cancer communication.

This review was conducted and is reported according to the PRISMA (preferred reporting items for systematic reviews and meta-analyses) guidelines. As mentioned in the initial registration (registration ID CRD42021252658), a series of structured electronic searches were performed in three English databases, including Web of Science, Scopus, and PubMed, focusing on the effectiveness of message framing related to a cancer topic and dated from January 1, 2000, to December 31, 2020. The procedures guiding article inclusion are presented in the flow chart in Figure 1. Example search terms are as follows: goal fram*, loss fram*, message fram*, goal fram*, health message, health infor*AND attitude, intention, behavio*, behavio*intention, behavio* chang* AND cancer, screen*, prevention, detection, AND specific validated database filters for randomized controlled trials

All articles identified in the search strategy were exported into reference management software (version 2.48.0; Mendeley) for duplicate checking and further screening. The reference lists of eligible articles were further reviewed to identify other relevant studies. Relevant reviews that emerged from the search strategy were checked for any additional studies. Grey literature (ie, working papers, unpublished studies, conference proceedings or abstracts, and dissertations) was not considered eligible. The preliminary retrieval process was taken as the total literature volume combining the reviews independently completed by two authors. The authors selected a sample of eligible studies after achieving good agreement (κ=0.8356), with the remainder being selected by one author.

Studies were included if they met the following criteria: (1) participants (not cancer patient or survivors) of both genders and any age range exposed to either the loss-framed or the gain-framed cancer prevention or detection (not treatment) message in experimental research; (2) interventions that have been delivered via either emphasizing the expected benefits of taking specific health behaviors (gain-framed message) or the possible loss of not taking specific health behaviors (loss-framed message); (3) the gain-framed and loss-framed message must come in pairs as a comparison and related to a specific cancer topic, not cancer in general; (4) the effectiveness of cancer-related messages was measured in terms of change in attitude (by scale), intention (by scale), and behavior (by yes or no), and the analysis of primary data was reported, and sufficient quantitative data was provided for estimating the total effect size. In addition, only randomized controlled trials (RCTs) were included. Finally, published studies that repeatedly used the same data were excluded, leaving only one study that reported the most complete results.

Data from a sample of eligible studies was extracted by two authors, and after they achieved good agreement (κ=0.9804), the remainder was extracted by one author. The main framework of the extraction criteria was drawn by the authors altogether. The following information was extracted: (1) basic study characteristics including the first author, publication date, country, and funding of the study; (2) participant characteristics including sample size, age, ethnic, and gender ratio (female); (3) intervention characteristics including cancer type, research setting (lab or not), message contents, message resources, message delivery channel, intervention duration, and underpinning theories; (4) outcome measures including measurement of the outcomes; and (5) main results including intervention completion ratio and converted effect size (standardized mean difference [SMD] and odds ratio [OR]).

The risk of bias for RCTs was independently assessed by two authors (κ=0.8137) using the Cochrane Collaboration tool for assessing bias [19]. Differences of opinion were discussed and agreed upon by the two authors.

All meta-analyses were performed with Review Manager 5.3 (version 5.3; The Cochrane Collaboration) [20]. This study performed meta-analyses on the three message effectiveness indicators (attitude, intention, and behavior). Among them, for the two continuous variables (attitude and intention), the SMD of the gain-frame (as the intervention group) and the loss-frame (as the control group) are calculated to represent an effect size. Moreover, two subgroups of cancer prevention and cancer detection were designed separately and analyzed under each indicator. For the two-category variable (behavior), the OR of the gain-frame (as the intervention group) and the loss-frame (as the control group) are calculated to represent an effect size. In addition, two subgroups of cancer detection within 6 months (<6 months) or more than 6 months (>6 months) after the experiment were designed for comparison and analysis of the impact of the timing of the framing effect on cancer detection behavior. All meta-analyses used random-effects models. I² statistics are used to determine heterogeneity and measure the degree of inconsistency. I² values are 25%, 50%, and 75%, corresponding to low, medium, and high levels of heterogeneity, respectively. For high heterogeneity, the source of heterogeneity was investigated by conducting a subgroup analysis to explore potential moderators and demonstrate why heterogeneity existed. We used the funnel plot to observe whether there is publication bias and used the Egger regression line to confirm possible publication bias further.

The literature search identified a total of 319 studies extracted by two authors. After deduplication, a total of 168 studies’ titles and abstracts were screened, with 53 remaining for full-text screening. Among those, 24 articles met the inclusion criteria, and 4 studies met the inclusion criteria but did not meet the exclusion criteria. The authors attempted to obtain the research data of 8 studies that met the first two inclusion criteria but failed to provide usable data by contacting their corresponding authors via email and received 2 replies [21,22]. Due to the oversize issue, the main content of each study is summarized in Multimedia Appendix 1. Of the 24 studies used in the final analysis, 13 were conducted by the United States [21,23-34], 2 by the United Kingdom [35,36], South Korea [22,37], and China [38,39], and 1 by Australia [40], Ireland [41], Italy [42], Singapore [43], and South Africa [44]. Among the studies used, 5 were funded by the government or institute grant [22,23,25,26,38], 2 by the institute and the center [24,29], 2 by the institute and universities [30,34], 2 by universities [28,37], and 1 by the health charity and behavioral insights project [35]. The remaining 12 studies did not report receiving any funding [21,27,31-33,36,39-44].

The total sample size was 11,637, ranging from n=85 (intervention group) to n=752 (control group). The participants’ average age ranged from 12 to 91 years. Because some cancers are associated with a particular sex, 2 studies notably recruited all-male participants [32,42], while another 10 studies recruited all-female participants [21,23,24,26,30,34,35,37,39,44]. Also, there was a range of different cancer topics targeted during the study interventions, including 9 studies for skin cancer [22,25,27,31,33,36,40,41,45], 2 for colorectal cancer [28,38], and 1 for lung cancer [29].

Most of the experiments were conducted online, through email [27-29,40,43,44], a phone call [23] or text message [35], and other methods like webpage URLs [31,37]. There are also a small number of experiments that were done in a laboratory setting using traditional printed materials [21,22,25,32,33,41,42,45]. Messages were delivered in a variety of ways, including text only [27,28,36,44], text plus images or graphics [26,29,38,40], and video [24,30,34,39,43]. The majority of studies were based on prospect theory [21,22,24,27,29,32,34,37,41,43-45], and some of them used theory of planned (or reasoned) behavior [28,31,38,39], self-affirmation theory [36], and the health belief model [30,42]. There are also some studies that did not use theories or models [23,25,35]. The majority of studies focused on attitude [28,31,32,37-40,42,43] and intention [21,22,25-29,33,36,37,39-43,45], whereas 6 studies used behavior to examine the message framing effect directly [23,24,30,34,35,44].

As shown in Figures 2 and 3, the risk of bias assessment indicated that the included articles were relatively high in quality. For the publication bias examination, the funnel plot was used for preliminary identification and the Egger regression line to confirm the possible publication bias further.

In the following sections, the synthesized results regarding attitude-related, intention-related, and behavior-related outcomes are introduced individually. The main results are visualized in the forest plot.

This study meta-analyzed 24 empirical studies on the persuasive effect of message framing in cancer prevention and detection. By subdividing the convincing effect into three dimensions, attitude, intention, and behavior, the results show that gain-framed and loss-framed messages have no significant differences in people's attitudes and intentions to change their cancer prevention or detection behavior. However, for cancer detection behavior, especially breast and cervical cancer detection, loss-framed messaging is more persuasive than gain-framed messaging, and this effect gradually weakens over time.

This meta-analysis did not identify any advantages of gain-framed messaging. Previous studies have proposed that gain-framed messaging is more persuasive than loss-framed in disease prevention, especially preventative behavior [6,16]. Since no relevant data on prevention behavior in the cancer-related literature are included in this meta-analysis, it cannot further verify whether the predecessors' benefits in disease prevention behavior can be equally applicable to the cancer topic. The results also negated the comparative advantage of loss framing in disease detection. There are several possible reasons for this contradictory conclusion. First, the previous meta-analysis did not distinguish between the three measurement indicators, so the results were affected by a large number of measurement indicators [16]. Second, the amount of cancer-related literature included in the previous meta-analysis is small, so the conclusions drawn do not apply to all cancer topics [6]. Therefore, it is necessary to verify whether these existing conclusions can be applied to all health topics based on previous studies. Moreover, because of different health topics, the focus also varies. Some emphasize daily prevention (such as using sunscreen), while some require regular detection (such as a colonoscopy).

This meta-analysis also supports some of the results of previous studies. Neither gain-framed nor loss-framed messaging can exert the framing effect of changing people's attitudes or intentions [6]. Our meta-analysis includes multiple cancer themes, and it also expands the positive impact of the loss framing found in breast cancer detection [46].

One interesting finding of this meta-analysis is that the message framing effect is time-limited. Others have questioned the duration of the framing effect in explaining the declining return rate [34]. On the other hand, this meta-analysis evaluates a larger scale of data and finds that the framing effect diminishes over time. Therefore, future research could further investigate how often health information stimulation can exert the framing effect and determine the effect of repeating the same framing.

Another notable result of this study is the disconnect between people's attitudes, intentions, and behavior. When reorganizing data related to attitudes and intention for breast and cervical cancer prevention and detection, these results also show that attitudes and intentions cannot effectively predict behavior, indicating that there may indeed be a disconnect between people's attitudes, intentions, and behavior. This conclusion has also been confirmed by previous studies [6,47]. The ultimate purpose of any framed message, but especially a health message, is to promote a specific behavior. In the research process, paying too much attention to the influence of other variables besides behavior may outweigh the gains. Further verification is worthwhile to measure whether subjective attitudes and intentions to target behaviors can replace or predict the subsequent behavior after reading the framed message. The exploration in this area cannot stop at testing latent variables such as attitude and intention, and it should return its focus to “behavior” itself [48].

As with all meta-analyses, this study was limited by the included literature measurement indicators and provided data. For example, the measurement data of attitude and intention included in the literature far exceed behavior-related data. Cancer detection behavior data were obtained only due to the lack of actual behavioral data related to cancer prevention. Also, the data are limited to two types of cancer (breast cancer and cervical cancer). Its generalizability also needs to be further verified based on more relevant research. Due to the low email response rate (2 out of 8), this meta-analysis excluded some studies meeting the inclusion criteria.

This meta-analysis shows that gain framing and loss framing have no significant difference in attitudes and intentions about cancer prevention and detection. On the other hand, loss framing can promote breast cancer and cervical cancer detection better than gain framing. However, the effect of the loss framing gradually weakens over time. Therefore, when constructing a message to promote cancer detection, more considerations can be given to loss framing’s promised short-term effects. People can be screened for related cancers quickly by emphasizing the possible risk of not performing cancer detection.