Effect of interventions including provision of personalised cancer risk information on accuracy of risk perception and psychological responses: A systematic review and meta-analysis

Highlights • Conceptualisation of risk is a complex cognitive process.• Individuals tend to overestimate their risk of cancer at baseline.• Immediately after risk information over 80% of people are able to recall the number.• However, less than half believe that to be their risk, thinking their risk is higher.• Risk information has either no effect or reduces worry, anxiety and depression.


Introduction
An increasing number of risk models are now available that enable estimation of an individual's future risk of cancer. Although providing individuals with a personalised risk estimate in isolation is unlikely to lead to behaviour change [1,2], personalised risk communication may complement educational interventions and increase motivation and health-related behaviour change over and above risk factor awareness education and lifestyle advice alone [3]. There is also increasing interest in the potential benefits of incorporating risk stratification into cancer screening programmes to enable the screening frequency, modality, and/or eligible age range to be adjusted to potentially optimise the benefit-harm ratio [4].
However, the general population does not easily understand the concept of risk [5,6], with lay perceptions of risk often being resistant to change and differing substantially from those of experts [7]. These discrepancies are potentially consequential. Risk perception, particularly when assessed using high quality measures, has been shown to predict behaviour [8], and cancer risk perception specifically is associated with health-related quality of life, depression, anxiety and cancer worry [9][10][11]. Understanding the impact of providing personalised cancer risk information on perceptions of risk and psychological responses is, therefore, important.
Previous reviews have shown that provision of cancer-based risk information in genetic counselling centres can increase accuracy of risk perception while leading to either no change in psychological outcomes or psychological benefits [12][13][14]. Individuals attending genetic counselling centres, however, are typically referred by healthcare professionals due to a family or personal history of cancer. These individuals are, therefore, already aware that they are potentially at high risk and their responses to risk information may differ from those at population level risk. To inform future population-based communication of cancer risk, we aimed to synthesise the effects of interventions incorporating nongenetic personalised cancer risk information on accuracy of risk perception and psychological responses in individuals not already identified as at high risk on the basis of a personal or family history of cancer or following referral to specialist cancer risk services.

Methods
We performed a systematic literature review following an a priori established study protocol (available on request). Reporting followed the PRISMA statement [15].

Search strategy
We used the same search strategy as for a previous review of the effect of interventions incorporating personalised cancer risk information that focused on intentions and behaviour [16]. This included an electronic literature search of Medline, EMBASE, CINAHL and PsycINFO from 1 st January 2000 until 1 st July 2017 with no language limits, using a combination of subject headings and free text incorporating 'cancer', 'risk/risk factor/risk assessment' and 'prediction/model/score/tool' (see Appendix File A.1 for the complete search strategies). We manually screened the reference lists of all included papers to identify additional papers. As the outcomes of interest for this review are not collected routinely within healthcare and both CINAHL and PsycINFO include citations to books, reports, dissertations and theses, we did not specifically search for additional grey literature.

Study selection
We included studies if they met the following criteria: 1) were published as a primary research paper in a peer-reviewed journal; 2) included adults with no previous history of cancer; 3) included provision to individuals of a personal estimate of future cancer risk based on two or more non-genetic variables, either alone or as part of a larger intervention; and 4) included data on either accuracy of risk recall or risk perception at the level of the individual or psychological measures (including cancer worry, anxiety, depression, affect and quality of life). As in our previous review [16], in order to focus on the provision of personalised cancer risk information to the general population, we excluded studies that had recruited participants on the basis of a personal or family history of cancer or following referral to specialist cancer risk services. We also excluded vignette studies, qualitative studies, conference abstracts, editorials, commentaries and letters.
Two reviewers (JUS and BS) each screened half of the titles and abstracts to exclude papers that were clearly not relevant. A third reviewer (SG) independently assessed a random selection of 5% of the papers screened by each of the first reviewers. The full text was examined by two reviewers (MB and MF) independently if a definite decision to exclude could not be made based on title and abstract alone. A third reviewer (JUS) then assessed all those for which it was unclear at full text level whether or not the inclusion criteria were met.

Data extraction
At least two researchers (JUS/BS/MB/MF) independently extracted data from studies included in the review directly into data tables. This included data on: (1) study characteristics (cancer type, study design, study setting, duration of follow-up); (2) selection of participants (inclusion criteria, method of recruitment/randomisation); (3) participant characteristics (age, level of cancer risk, sample size); (4) the intervention (risk tool used, method and format of risk communication, additional information or follow-up provided), and (4) measured outcome(s). Reviewers were not blinded to publication details. If numerical data were not included in the published articles, we wrote to the authors requesting additional information.    Comparative risk was expressed in terms of one of seven categories: "very much below average'', "much below average,'' "below average,'' "average'', "above average,'' "much above average,'' and "very much above average'' alongside an oval window with the risk marked on a horizontal hairline CRCcolorectal cancer.

Quality assessment
Quality assessment was performed by two reviewers (MB and MF) using a checklist based on the Critical Appraisal Skills Programme (CASP) guidelines [17]. This includes eight questions concerning whether the study addressed a clearly focused issue, the method of recruitment and randomisation, whether blinding was used, the measurement of the exposure and outcome, the comparability of the study groups and the follow-up. Each study was then classified as high, medium or low quality. We did not exclude any studies based on quality alone.

Data synthesis and statistical analysis
As data on psychological outcomes (worry, anxiety, fear, depression and quality of life) used different measurement scales and variably reported change from baseline to follow-up and mean values at follow-up, it was only possible to pool results for accuracy of risk perception. For the comparison between risk information and no information we used random effects meta-analysis [18] and the 'metan' package in Stata and present intervention effects as relative risk (RR) rather than odds ratios (OR) to avoid overestimating the risk [19]. If there were zero participants in any group, we added 0.5 to each of the cells of the 2 Â 2 table in both the control and intervention group [20]. For the study by Timmermans et al. [21] in which data were reported for accuracy in the same participants for colon cancer and lung cancer separately, we included only the results for colon cancer in the meta-analysis to avoid including the same participants twice in the same analysis. The results were similar when lung cancer was included instead (data not shown). To pool the percentage who were able to recall the risk information provided to them accurately and those whose risk perception accurately matched the risk estimate that they had been provided we also used the 'metan' package in Stata with a random effects model. In both cases we quantified the heterogeneity between studies using the I 2 statistic. All analyses were conducted using statistical software package Stata/SE version 14.

Results
As reported previously [16], we identified 35,802 unique papers from the electronic search. Of these, 35,604 were excluded at title and abstract level. After screening by the first reviewer (JUS/BS), no additional papers met the inclusion criteria in the random 5% screened by the second reviewer (SG). A further 180 were excluded after full-text assessment against the inclusion and exclusion criteria specific for this review question. The most common reasons for exclusion at this stage were that the papers did not include provision of a personal risk estimate (n = 69), did not include any data on predefined outcomes (n = 32), were conference abstracts (n = 20), or were not primary research (n = 16) (Fig. 1). We identified four additional papers through citation searching, leaving us with 22 papers describing 23 studies in the analysis. Table 1 summarises the design, setting and key outcomes of the 23 included studies and Table 2 provides additional details about the tools used to estimate the personalised risk and the format in which the risk information was provided. The majority (n = 15) focused on provision of breast cancer risk derived from the Gail model [22], four provided risk information about colorectal cancer, one lung cancer, one cervical cancer, one colorectal and lung cancer, and one colorectal, breast and ovarian cancer. All but two studies [23,24] were conducted in the USA. Twelve were assessed as high or medium/high quality, seven as medium quality and four as medium/low based on the CASP guidelines (Appendix File A.2) [25][26][27][28][29][30].

Recall of risk information
Three studies reported recall of absolute risk [31][32][33]. Immediately after being provided with risk information, 87% (95% CI 84%-91%, I 2 = 0%) of those given absolute risk information were able to recall their numerical absolute risk estimate accurately (defined as exact agreement) (Fig. 2), with no difference between those presented risk of breast cancer as either a point estimate on a 0-100% scale, as a range, or as a point estimate plus a range [33]. Comparative risk, where individuals were provided with estimates of their risk in comparison with others, was reported in only one study where 64% were accurate [31].
Two of these studies additionally compared recall of risk information and risk perception. In the study by Lipkus et al. only 17% (n = 19/102) of those who were able to recall their risk estimate perceived that to be their risk within 0.5%, with 71% (n = 72/102) believing their risk to be higher and 12% (n = 12/102) their risk to be lower [32]. Similarly, in Weinstein et al., those who had received absolute risk information gave the same answer for their own beliefs as their recollection of what they had been told only 45% of the time, giving a higher value for their own beliefs 47% of the time and a lower value 8% of the time [31]. Corresponding percentages for comparative information were 39%, 46% and 15% respectively. A further study did not compare recall with perceived risk but instead asked women at a follow-up telephone interview how they would compare their actual risk with the estimate provided in the study. 53% thought that their actual risk was 'just the same,' while 38% thought that their risk was greater than what they had been told [34].  $"

Seitz 2016
The degree to which participants overestimated their risk 0 Consistent improvement across six intervention groups when risk was measured as a percentage but not when risk was measured as a frequency out of 1000. For women with an estimated risk <1.5%, this effect was moderated by numeracy, with women with high numeracy having greater increases in accuracy than women with low numeracy. No significant moderation effects were seen for women with an estimated risk !1.5%.

"
Indirect assessment of accuracy in populations who all overestimated risk at baseline Dillard 2006a -0 The mean estimate of absolute risk among 72 undergraduate women decreased from 56.4% to 28.4% two weeks after absolute and comparative risk information. These, however, remained significantly higher than the estimated risk (mean 11.2% difference) p < 0.01. No significant differences were seen among those who were asked to provide a pre-intervention risk estimate, those who were led to believe that all the factors they considered possibly responsible for their own breast cancer risk were used to compute their risk, or those who completed a self-affirmation task. 2) for those in a risk only and risk plus social comparison groups respectively). Their estimates remained higher than the estimated risks they had been given (mean 16.9% vs 10.9%, p < 0.001).

"
McCaul 2003 -1 week Women who received absolute risk reported both lower absolute risk perceptions (mean 34.9% compared with mean 52.1%, p < 0.01) and lower comparative risk perceptions (mean 4.10 compared with mean 4.43, p = 0.05) immediately and at one week follow-up than women who did not. The effect for comparative risk information was not quite significant (p = 0.07) but women who received comparative risk estimates did report lower risk (mean 4.11) than those who did not (mean 4.43) " * Computed by first subtracting the participants' personalised risk estimate from the risk estimate of the average same-aged woman with no risk factors, then subtracting participants' estimates of their own and the average woman's absolute numerical risk, and then comparing the two differences and categorising participants as accurate if the differences were within 5%.

Accuracy of risk perception
Thirteen studies reported data on accuracy of risk perception. Eight of these reported accuracy as the agreement between the perceived risk estimates participants gave and the estimated personalised risks they had been presented with. The other five studies reported accuracy indirectly, either as the extent of overestimation or the change in risk perception in groups known to all either over-estimate or under-estimate their risk at baseline.
Definitions of what constituted "accurate" and the time interval between provision of risk information and follow-up varied widely between studies (Table 3). This made pooling many of the results inappropriate. It was possible, however, to pool data from three studies that measured accuracy of absolute or comparative risk perception immediately after provision of risk information about colon cancer compared with no information [21,31,35]. Those who received both absolute and comparative risk estimates were more likely to have accurate absolute risk perceptions immediately post risk information (pooled RR 2.59 (1.40 to 4.81) I 2 = 81.2%) (Fig. 3), with no difference between those provided with absolute risk alone or absolute plus comparative risk (data not shown). There was no significant effect on comparative risk accuracy (pooled RR 1.11 (0.74 to 1.66) I 2 = 82.9%) (Fig. 3).
The findings from these and the other studies that could not be pooled are summarised in Table 3. Overall, eight showed improvements in accuracy, two no effect and three mixed results. One study directly compared the effect of alternative formats on risk accuracy. In that study, Emmons et al. showed that those who were randomised to have the opportunity to see how adopting or changing any of the risk factors would impact on their total risk profile had greater improvement in accuracy immediately post information for both comparative and absolute risk accuracy compared to those who did not [35]. A further study assessed the role of numeracy and found that among women with an estimated risk <1.5%, the degree to which participants overestimated their risk was moderated by numeracy, with women with high numeracy having greater increases in accuracy than women with low numeracy [36]. No significant moderation effects were seen for women with an estimated risk !1.5%.
Having the opportunity to see how changing any of the risk factors would influence their risk, as well as inclusion of social comparison information [37], appeared to be associated with greater improvements in accuracy of perceived risk. By comparison no differences were seen for providing pre-intervention risk estimates, self-affirmation, providing data so that individuals believed that all factors they considered possibly responsible for their own risk were used to compute their risk [37], or with race or education level [38,39].

Cancer specific worry, anxiety or fear
Thirteen randomised controlled trials (RCTs) reported cancer specific worry, anxiety or fear. As the studies used different scales and variably reported change from baseline to follow-up and mean values at follow-up, it was not possible to pool the studies. Instead, the findings are summarised in Table 4. Ten reported no significant change and three a reduction.

General anxiety, depression, affect and health-related quality of life
Three studies reported general anxiety using versions of the Spielberger State Anxiety Inventory (STAI) [40]. Two RCTs showed non-statistically significant differences between women randomised to receive personalised estimates of the risk of cervical cancer during cervical screening appointments or routine care (-1.6 (95% CI: -3.5 to 0.2), p = 0.084) [23] and among 314 participants randomised to complete a self-administered decision aid for colorectal cancer (CRC) screening that included personalised Fig. 3. Forest plot showing the relative risk of having an accurate perception of absolute or comparative risk immediately after receiving it compared to controls who did not receive risk information.
information on risk of developing CRC or to receive a booklet about the Australian CRC screening guidelines [41]. The third study by van Erkelens et al. [42] measured anxiety using a Dutch version of the STAI alongside the Hospital Anxiety Depression Scale before and two weeks after 287 women had completed an online self-test that identified those at increased Familial Breast Cancer risk based on the Dutch breast cancer guidelines. It was the only study to report results separately for women at population risk and those at moderate (relative risk !2-3) or high risk (relative risk >4) of breast cancer. In women at population risk of breast cancer (n = 272), state-anxiety significantly decreased immediately after taking the test (mean change from baseline -2 (95% CI -2 to -1), p < 0.001) and both state anxiety and trait anxiety significantly decreased at two weeks (mean change from baseline -3 (95% CI -5 to -2) and -1 (95% CI -2 to -1) respectively, p for both 0.002). There was no change in distress among those participants at two weeks and no significant changes in any outcomes in the 15 women at increased familial breast cancer risk.
Affect was measured in one RCT using the Positive and Negative Affect Scale (PANAS) [43] in which female undergraduates received absolute risk feedback with or without comparative information [37]. No significant between-group differences in affect were observed. Health-related quality of life was additionally measured in two RCTs [44,45] using the SF-36 [46]. Both reported a significant increase in score on the SF-36 at follow-up in the intervention group compared with the control group.

Discussion
This study is, to our knowledge, the first comprehensive review of the impact of interventions incorporating provision of personalised cancer risk information based on non-genetic risk factors on accuracy of risk perception and psychological responses among individuals at population level risk. A particularly novel aspect is that in the synthesis we have been able to distinguish between recall of risk information and risk perception and have shown that, while immediately after provision of risk information 87% of individuals were able to recall the absolute risk estimate, less than half believed that to be their risk, with up to 71% believing their risk to be higher than the estimate. These findings in particular highlight the conceptual problems in understanding risk information and the tendency for people to resist information that is communicated to them by experts that have previously been reported across both cancer and other diseases [5]. Among these, qualitative studies have shown that risk perception is not as simple as recalling a number and that the processing of risk information is not purely 'rational' or 'objective' [47]. Instead, an individual's perception of risk is based on a complex integration of cognitive and social biases arising from cultural, personal or lay theories of disease and risk, and past experiences, expectations and beliefs [32,34,[47][48][49][50][51][52]. The studies included in this review support the view that, rather than simply replacing their prior beliefs concerning their risk of developing cancer with new information, individuals appear instead to be using the new risk information to update their prior beliefs, analogous to Bayesian inference. The extent to which individuals over-or under-estimate their risk at baseline decreases after provision of risk information (reflected by an increase in accuracy) but many individuals continue to, in most cases, overestimate their risk.
The complex cognitive processes involved in this conceptualisation of risk may in part also explain our finding that risk-based inventions improve accuracy of absolute risk perception but not comparative risk. By its very nature comparative risk is a more emotive and less abstract construct [8]. It may therefore be more prone to cultural, cognitive and social biases and in turn more resistant to change. For the same reasons, however, comparative risk may sometimes play a more important role in influencing decisions concerning health behaviours.
The observed discrepancy between the risk estimate and perceived risks may also reflect varying levels of numeracy and the difficulties people often have understanding risk information [53,54]. This is supported by the finding in this review that among women with an estimated risk <1.5%, those with high numeracy had greater increases in accuracy than those with low numeracy [36]. Numerical misunderstanding was also given as a reason for feeling that their risk was higher or lower by women who recalled their risk estimate correctly but gave a different response when asked about their perceived risk in the study by Lipkus et al. [32].
The finding that individuals tend to overestimate their risk prior to receiving risk information and that provision of risk information has no effect or reduces cancer worry, anxiety and depression has also been reported for other diseases, including diabetes [55] and  Single question -'How worried are you about developing breast cancer?' No significant differences were found between women who were asked to provide a pre-intervention risk estimate, those who were led to believe that all the factors they considered possibly responsible for their own breast cancer risk were used to compute their risk, or those who completed a self-affirmation task, or between those provided with their risk alone and those provided with their risk plus social comparison $

Timmermans 2012
Percentage who agreed or disagreed with the statement 'I am more worried now about my risk of cancer than before I did my cancer risk test' After receiving a combination of information on average population risk, personal risk and the relative risk reduction after changing lifestyle, 55.4% of participants disagreed with the statement for colon cancer and 61.4% for lung cancer and 12.1% and 11.18% agreed for colon cancer and lung cancer respectively, indicating that worry had stayed the same or reduced in most individuals

Lipkus 2005
Combined responses to three questions about how worried, fearful and anxious they were about developing breast cancer No difference between participants provided with either no risk information or absolute or absolute plus comparative risk information and no effect of age, race or education $

Lipkus 2006
Combined responses to three questions about how worried, fearful and anxious they were about developing breast cancer No difference between participants provided with either no risk information or absolute or absolute plus comparative risk information but those told that they "did not have more than the average number of risk factors" had lower combined worry, anxiety and fear at follow-up than those told that they had more than the average number (mean at follow-up adjusted for baseline 5 [56], and following communication of genetic risk [13,57]. Cancer specific worry has been reported to predict engagement in prevention initiatives [58]. This observed reduction in cancer worry, anxiety and fear may in part, therefore, explain the lack of association between provision of risk information and behaviour change [16]. These findings must however be interpreted within the limitations of this review. We performed it following accepted best practice with independent screening of full-text articles for inclusion and double data extraction and quality assessment [15]. Nevertheless, there are a number of limitations. Firstly, while we screened over 35,000 articles from four electronic databases and the reference lists of included articles, we did not specifically search for additional grey literature and were unable to assess publication bias formally. It is therefore possible that there are additional studies of relevance to this review question that we did not include. Given the number of articles screened and the high proportion of those with negative findings, however, we think it unlikely that these would change the overall findings. Secondly, the design of the included studies, definitions of accuracy of risk perception, and the range of ways in which psychological outcome measures were collected and reported varied substantially. For example, the 23 included studies incorporated 12 different measures of risk accuracy and eight of worry. This range of measures has been reported previously [12,14] and made summarising and pooling the findings difficult and meant we were only able to include a small number of the studies in the metaanalysis, limiting the strength of those results. This was further limited by many of the included studies also only presenting data for outcomes where significant changes had been observed, including only a statement of no change for other outcomes. Thirdly, risk was communicated to individuals in different formats and many of the interventions included written or verbal information alongside risk estimates. Isolating the effect of the risk information or any differences between formats was not possible. This is likely to have less of an impact on the measures of risk perception but may have influenced the psychological outcomes. Fourteen of the 23 studies also looked at breast cancer, all but two were in the US and all were at risk of potential recruitment bias. Together these limit the generalisability of the findings. Particularly for accuracy of risk perception, most of the studies only reported outcomes either immediately or a few weeks after provision of risk information. The findings therefore largely reflect the short term impact of provision of risk information.

Conclusion
This review shows that immediately after provision of risk information 87% of individuals were able to recall the absolute risk estimate that they had been given. However, less than half believed that to be their risk, with up to 71% believing their risk to be higher than the given estimate. Provision of risk information increased accuracy of perceived risk immediately after risk information and reduced mean perceived risk among groups who overestimated their risk at baseline. However over half of individuals remained inaccurate, with most perceiving their risk as higher than the risk estimate that they had been provided with. By comparison, there was no significant effect on comparative risk accuracy, either immediately or in the short/medium term and either no effect or a reduction in worry, anxiety or depression, with no evidence of differences with age, race, level of education or presentation of risk.
The review itself also highlights a number of important messages for researchers. These include the need for: consistent measures of risk accuracy and psychological responses to facilitate comparison across studies; sub-group analyses, particularly for psychological responses, in individuals who over-estimate or under-estimate their risk at baseline; studies including other cancer types, outside the US, and among men and people of diverse socioeconomic and cultural groups to improve the generalisability of the results; and better reporting of negative results. Attempting to measure risk perception with a single number is also unlikely to capture the complex cognitive processes involved in the conceptualisation of risk. Researchers should therefore consider using broader risk perception instruments, such as the Tripartite model of risk perception which includes assessment of susceptibility to disease (deliberative risk perception) alongside measures of the affective and experiential components of risk perception, including cancer-specific worry, anxiety and fear [59]. Not only is this model more likely to capture the range of cognitive processes, but it has been shown to predict intention to change health-related behaviour more accurately than unidimensional models of risk perception. Risk conviction, the subjective sense of certainty that one knows what one's perceived risk is and the confidence that this risk perception is accurate [60] may also be a more sensitive measure of the impact of provision of risk information.

Practice implications
Perhaps the most important message from this review for clinical practice is the recognition that individuals who appear to understand and be able to recall risk information provided to them most likely do not believe that the risk information reflects their own risk. As described above, the reasons for this are complex and, as a result, are unlikely to be specific to cancer or overcome within a single consultation or by a single intervention. However, by being aware of the limits of provision of information and cognisant of the context in which each person is using the information to construct an individual perception of risk, clinicians will be better able to tailor the explanations of risk to their patients and support their understanding and shared-decision making. has received salary support in respect of SJG from the NHS in the East of England through the Clinical Academic Reserve. The views expressed are those of the authors and not necessarily those of the NHS, Department of Health or the US NIH. All researchers were independent of the funding body and the funder had no role in data collection, analysis and interpretation of data; in the writing of the report; or decision to submit the article for publication.

Author statement file
The contributions of authors to the manuscript are as follows: a) study concept and design: Usher-Smith, Griffin, Silarova; b) acquisition, analysis or interpretation of data: Usher-Smith, Bayne, Fairey, Silarova, Griffin, Sharp, Klein, Sutton; c) drafting of the manuscript: Usher-Smith, Bayne, Fairey; d) critical revision of the manuscript for important intellectual content: Griffin, Klein, Silarova, Sharp, Klein, Sutton; e) obtained funding: Usher-Smith, Griffin.

Data sharing
All data are available from the reports or authors of the primary research. No additional data is available.

Declaration of Competing Interest
None.