- Research article
- Open Access
- Open Peer Review
Words or numbers? Communicating risk of adverse effects in written consumer health information: a systematic review and meta-analysis
© Büchter et al.; licensee BioMed Central Ltd. 2014
- Received: 17 September 2013
- Accepted: 20 August 2014
- Published: 26 August 2014
Various types of framing can influence risk perceptions, which may have an impact on treatment decisions and adherence. One way of framing is the use of verbal terms in communicating the probabilities of treatment effects. We systematically reviewed the comparative effects of words versus numbers in communicating the probability of adverse effects to consumers in written health information.
Nine electronic databases were searched up to November 2012. Teams of two reviewers independently assessed studies. Inclusion criteria: randomised controlled trials; verbal versus numerical presentation; context: written consumer health information.
Ten trials were included. Participants perceived probabilities presented in verbal terms as higher than in numeric terms: commonly used verbal descriptors systematically led to an overestimation of the absolute risk of adverse effects (Range of means: 3% - 54%). Numbers also led to an overestimation of probabilities, but the overestimation was smaller (2% – 20%). The difference in means ranged from 3.8% to 45.9%, with all but one comparison showing significant results. Use of numbers increased satisfaction with the information (MD: 0.48 [CI: 0.32 to 0.63], p < 0.00001, I2 = 0%) and likelihood of medication use (MD for very common side effects: 1.45 [CI: 0.78 to 2.11], p = 0.0001, I2 = 68%; MD for common side effects: 0.90 [CI: 0.61 to 1.19], p < 0.00001, I2 = 1%; MD for rare side effects: 0.39 [0.02 to 0.76], p = 0.04, I2 = not applicable). Outcomes were measured on a 6-point Likert scale, suggesting small to moderate effects.
Verbal descriptors including “common”, “uncommon” and “rare” lead to an overestimation of the probability of adverse effects compared to numerical information, if used as previously suggested by the European Commission. Numbers result in more accurate estimates and increase satisfaction and likelihood of medication use. Our review suggests that providers of consumer health information should quantify treatment effects numerically. Future research should focus on the impact of personal and contextual factors, use representative samples or be conducted in real life settings, measure behavioral outcomes and address whether benefit information can be described verbally.
- Consumer health information
- Health communication
Ideally, patient decisions for and against medical treatments are made in the presence of knowledge of the best available evidence for the benefits and harms of these treatments. Personal preferences and values can influence treatment decisions and may – legitimately – lead people to make choices which are not necessarily in line with the evidence. There are, however, some cognitive biases that may interfere with treatment. In particular, various types of data framing can influence risk perceptions .
Poorly framed information on the risk of adverse effects of drugs or other medical interventions may cause misinterpretation of the risks of harms. This may have an impact on treatment decisions and might also affect medication adherence. The 1995 contraceptive pill scare in the UK highlights the importance of helping doctors and patients understand risk information: media reports and “Dear Doctor” letters reported that the third-generation contraceptive pills increased the (relative) risk of blood clots by 100%, which caused many women to stop taking the pill and led to many unwanted pregnancies and abortions – although the absolute risk increase was as small as 0.014% .
European commission nomenclature for communicating frequency of adverse effects of drugs
(≥1/100 to <1/10)
(≥1/1000 to <1/100)
(≥1/10000 to <1/1000)
cannot be estimated from the available data
Several studies have compared the use of verbal terms versus numbers for communicating the frequency of adverse drug effects. However, to our knowledge no systematic review on the comparative effects of verbal versus numerical presentations of the frequency of adverse effects has been conducted. Risk communication has become a vast field which is difficult to keep up with. Thus, current recommendations on risk communication are often based on expert consensus or a selective review of the literature. For example, both the International Patient Decision Aid Standards (IPDAS) and the FDA’s user’s guide on communicating risks and benefits currently do not cite many of the studies we identified in our preliminary searches. The aim of this systematic review is to improve the evidence base of risk communication strategies by gathering and synthesizing the results from studies that examined different terms, scenarios and probabilities.
We included studies examining the effects of words versus numbers in communicating harms of treatments to consumers in written health information. Our inclusion criteria were: (1) study design: randomized controlled trials (RCTs); (2) outcomes: interpretation of probability, comprehension, recall, satisfaction, impact on decision, likelihood of treatment utilization, adherence and psychological outcomes (e.g. anxiety); (3) context: treatment effects were communicated through written health information only and (4) language: studies published in English or German.
Data sources and search methods
We searched MEDLINE, Embase, PsycINFO, CINAHL, ERIC, DARE, the CDSR, CENTRAL and the Campbell Library. Searches were developed and conducted by an information specialist using a combination of MeSH-terms, free text and validated search filters for specific study designs, where available. See Additional file 1 for the search strategy used to identify relevant studies in MEDLINE. This was adapted as required to other databases. Searches were conducted up to the 9th of November 2012. Titles and abstract of search results were assessed for eligibility independently by three reviewers in pairs. Full texts of potentially relevant studies were retrieved and assessed for eligibility independently by two reviewers. Reference lists of articles eligible for inclusion were screened for further potentially relevant studies.
Data extraction and risk of bias assessment
Data were extracted into standardized extraction sheets and double checked in pairs by three reviewers. These included data on study design, risk of bias items, population characteristics, study setting, study intervention and results for the relevant outcomes (means and standard deviations). In studies that only reported p-values, t-values or confidence intervals, we derived standard deviations from these statistics using the methods described in Chapter 7 of the Cochrane Handbook for Systematic Reviews .
Risk of bias was assessed for RCTs by random sequence generation, allocation concealment, completeness of follow-up and selective reporting bias. Judgements were made in accordance with the guidelines for the Cochrane risk of bias tool .
Data synthesis and analysis
Data were entered into RevMan 5 and pooled. Mean differences (MD) and their corresponding 95% confidence intervals (CI) were calculated for outcomes that were measured on scales of considerable similarity. Otherwise standardised mean differences were calculated. Meta-analyses were conducted using random-effects models as the underlying rationale of random-effects models may be more appropriate when pooling heterogeneous data, while fixed and random-effects models produce the same result if data are homogenous. A downside of random-effects models is that more weight is given to small studies which may have a higher risk of bias (small study bias), but this was not an issue in our review. Heterogeneity was measured using Chi2-tests and the I2 statistic. If heterogeneity was detected, subgroup analyses were conducted to explore reasons for heterogeneity. Subgroups were planned a priori for age, gender, socioeconomic status, type of illness (mild or severe), size of absolute effect and severity of side effects. Where statistical heterogeneity remained, but there was strong contextual homogeneity, we opted in favour of pooling the data into meta-analyses, because of their additional informational value and the problems associated with narrative or pseudo-quantitative interpretation of results . However, in these cases we did not pool results across subgroups.
Some studies had three comparison groups: two studies compared a verbal, percentage and natural frequency presentation; one study compared a verbal, numerical and combined verbal/numerical presentation . In this case we used data from both comparisons in our analyses and divided the number of participants in the verbal group by two in order not to artificially inflate the statistical power of these studies in the meta-analyses. In two studies participants received two scenarios with different adverse effects. In cases where both scenarios were relevant to the same meta-analysis, we averaged the results across the two scenarios. The standard deviations for these comparisons were recalculated to account for statistical dependence assuming a correlation of 0.5 (sensitivity analyses with correlations of 0.1 and 0.9 produced similar results).
Description of studies
All studies were randomized controlled trials, many of which used a factorial design. Some studies were reported in more than one publication. All studies randomized participants to short information leaflets on drugs for a particular condition, which only differed in whether the information on the frequency of the adverse effects of the drug were presented verbally or numerically. One study examined a combination of a verbal and numerical description, as it is currently included in the 2009 European Commission Guideline on the readability of package leaflets . The interventions and outcomes of the studies were very similar and mainly differed with respect to the conditions and drugs that were used in the scenarios as well as the frequency and the severity of the side effects. The studies included five outcomes of interest to our review: estimation of probabilities (in percentages), likelihood of occurrence, satisfaction, intention to take or continue to take the medicine and the impact of the information on the decision. The last four outcomes were all measured as one item on a 6-point Likert scale. All outcomes were measured shortly after distribution of the information leaflets, and none of the studies had a follow-up. In many cases the participants received information on more than one adverse effect, resulting in a higher number of comparisons than studies for the outcome estimation of probability.
In all but one study participants were recruited from the general population or via a cancer website and confronted with a hypothetical scenario. The studies were all conducted by two groups of authors from the UK, who were interested in evaluating the effects of the nomenclature used in drug package inserts in the European Union. Thus, the verbal descriptors that were studied in the trials were: very common, common, uncommon, rare and very rare. See Additional file 2 for detailed characteristics of the included studies with additional results from individual studies regarding effect modifiers.
Risk of bias
Risk of bias of included studies
Random sequence generation
Incomplete outcome data
Berry 2002 Study 1 
Berry 2002 Study 2 
Berry 2003 Study 1 
Berry 2003 Study 2 
Berry 2004 
Berry 2006 
Knapp 2004 
Knapp 2009a 
Knapp 2009b Study 1 
Knapp 2009 Study 2 
One study used an unconcealed allocation. However, the authors of the study argued that this was unlikely to bias the results, because it seems unlikely that the researcher could be able to anticipate the participants’ response to verbal or numerical information. Furthermore, excluding this study did not alter the results. There were no signs of selective reporting.
Effects of interventions
Estimation of probabilities
Interestingly, even participants who received a probability estimate of the frequency of the adverse effects often overestimated these values. Only between 9% and 50% of the participants in the numerical groups gave a correct probability for the adverse effects (see Additional file 3). However, this was not always reported. Furthermore, the variability in responses between participants was large, which is indicated by large standard deviations and wide ranges.
See Additional file 3 for a detailed table of the results of the comparisons from each study by verbal descriptor and type and frequency of adverse effect together with the results of the significance tests as they were reported in the primary studies.
Likelihood of occurrence
One trial compared a numerical presentation with a combined format. Splitting this trial from the others we conducted a second, exploratory subgroup analysis on this outcome. This suggested that the verbal presentation may dilute the effects of a numerical presentation on this outcome (test for subgroup difference, p = 0.003, analysis not shown) .
Likelihood of taking the medicine
Impact of information on decision
This systematic review provides evidence that compared to numerical information verbal descriptors commonly used to communicate the frequencies of adverse effects in written health information including “common”, “uncommon” and “rare” lead to an overestimation of the probability of adverse effects, when they are used as previously suggested in the Guidelines of the European Commission.
It could be argued that other verbal terms are needed to describe frequencies. We are not aware of any studies comparing verbal terms other than those suggested in the 1998 European Commission’s guidelines though. Some studies have asked patients to assign probability values to a range of different verbal frequency terms . According to these studies, other words do not appear to be better suited to describe frequencies than those previously suggested by the European Commission. For example, in one study in a general practice setting, the terms “almost never” and “rarely” were associated with the lowest frequencies . The probabilities assigned to these terms were still very high with 9.9% and 7.5%, respectively. Furthermore, the standard deviations in these studies were large, which is in line with our results and suggests a large variance in the frequencies assigned to different terms. This indicates that risk expressions should be tested for understanding before being routinely used. Furthermore, it suggests that there may be no verbal labels that are suited to convey frequencies, particularly of rare adverse effects.
Even participants who received numerical information overestimated the risk of adverse effects. This is in line with other findings showing that people are generally poor at estimating risks . Low numeracy in some of the patients may also explain this finding. In the UK, for example, one study suggested that one third of adults above the age of 50 had limited functional health literacy . Another possible explanation for this finding is that patients may perceive their personal risk of experiencing adverse effects to be larger than average.
People seem to be more satisfied with numerical presentations and that they would be more likely to take the drugs or continue taking them. Participants also stated that they would be less affected in their decisions by numerical presentations. These outcomes were measured on a 6-point Likert scale. Converting difference into percentages on the scale suggests changes between 7% and 24%, which can be considered to be in the small to moderate, but important range. Most effects were also in the small to moderate range based on Cohen’s interpretation, when converting effects into standardised mean differences. Some of the effects may be considered relatively large, since there is a tendency for people to avoid extreme answers on scales where extreme values are labelled in absolute terms, as it was the case in the studies included in this article .
Subgroup analyses suggested that combined verbal and numerical formats may dilute the effects of the numerical presentation on two outcomes, namely likelihood of occurrence (as measure on a Likert scale) and impact on treatment decision. However, these results are based on a post-hoc analysis and comparison with the combined format was restricted to a single trial with 100 participants.
Challenges for providers of patient information
Providers of patient information often have a broad audience and face the problem that people have different preferences regarding the need and use of risk estimates. The meaning that is ascribed to such information varies greatly. While some express a clear need for risk estimates, others are confused by numbers and prefer to make decisions based on other types of information . Different preferences imply that using a combined verbal and numerical format may be the best compromise to suit various needs. This is also reflected in the current European Commission Guideline on readability from 2009 as well as the current EU template for patient leaflets [16, 23]. Providing different information for different groups according to their preferences would be an option, but it may be difficult to direct patients to the information that best suits their needs.
Unfortunately, data on adverse effects are often poorly reported in trials and systematic reviews, which complicates the issue [24, 25]. Furthermore, there might still be a role for verbal terms in written information, for example for people with difficulties in understanding numbers, or when large amounts of numbers make information too difficult to comprehend. It is difficult to draw a clear recommendation for providers of patient information as it is unlikely that there is a one-size-fits-all approach. This will depend on many other factors such as the context and the target group of the information.
Limitations of the review
Our review is based on a comprehensive search and used rigorous methods for assessing and synthesising the included studies. However, it has some limitations. We restricted our search to English and German studies. It is reported in accordance with the PRISMA statement (Additional file 4). This may have introduced a language bias. We do not consider this to be a major weakness though, since it is questionable whether results can be generalized from one language to another due to semantic differences.
Limitations of the included studies
Many studies were conducted with healthy volunteers and used fictional scenarios. There were some exceptions: one study included patients admitted to a cardiac rehabilitation centre and produced similar results. Three studies of users of a patient information website partially included women with experiences of breast cancer. While they also produce similar results, these trials had some limitations, too. Some of the women in these studies were already taking the medication which was used in the scenario, which questions the applicability to other populations. An important caveat of all studies was that they used convenience samples, which may lack representativeness. Lastly, all outcomes were measured as single items. This may be problematic for an outcome such as satisfaction, which represents a complex construct. However, information leaflets only differed in one sentence and the results for this outcome were very homogenous, adding strength to the findings.
Our review suggests that – whenever possible – adverse treatment effects should be quantified numerically, because they lead to better estimates of risks. Verbal risk expressions should be tested for understanding before being routinely used.
Further studies should focus on the impact of personal and contextual factors, including the setting, disease, numeracy and educational level. Furthermore they should use representative samples or be conducted in real life settings and measure potentially more relevant outcomes such as actual behavior (including decisions and medication adherence for example) and whether decisions are in line with personal values. After all, risk communication is not an end in itself, but a means to the end of making better decisions. On a more critical note, it is questionable whether a difference solely in how information on adverse effects is communicated could have a detectable effect on behavioral outcomes. A recent systematic review examined whether informing patients about benefits and harms of medicines compared to usual care has an impact on behavior at all . Overall, the results did not show a significant effect. This systematic review had some limitations including heterogeneous results and statistical imprecision and there is some difficulty in interpreting the results. However, it suggests that we may need to focus on more general questions regarding the effects of provision of information on behavioral outcomes.
A further unanswered question is how different formats for describing the frequency of adverse effects are interpreted when they are presented together with treatment benefits, since these are also often overestimated by patients . Qualitative research methods may be able to shed some light into how people come to assign probabilities to words. On a final note, further research should be conducted within the framework of a systematic review of the literature.
We thank Ulrich Grouven for his kind statistical advice and Stefan Lange for critically reviewing the manuscript.
- Edwards A, Elwyn G, Covey J, Matthews E, Pill R: Presenting risk information–a review of the effects of “framing” and other manipulations on patient outcomes. J Health Commun. 2001, 6: 61-82. 10.1080/10810730150501413.View ArticlePubMedGoogle Scholar
- Gigerenzer G, Gaissmaier W, Kurz-Milcke E, Schwartz LM, Woloshin S: Helping doctors and patients to make sense of health statistics. Psychol Sci Public Interes. 2007, 8: 53-96.Google Scholar
- European Commission (EC): A Guideline on the Readability of the Label and Package Leaflet of Medicinal Products for Human Use. [http://pharma.be/assets/files/854/854_128901376878944246.pdf]
- European Commission (EC): A Guideline on Summary of Product Characteristics (SmPC). [http://ec.europa.eu/health/files/eudralex/vol-2/c/smpc_guideline_rev2_en.pdf]
- Cochrane handbook for systematic reviews of interventions. Edited by: Higgins JPT, Green S. [http://handbook.cochrane.org/]
- Higgins JP, Altman DG, Gøtzsche PC, Jüni P, Moher D, Oxman AD, Savovic J, Schulz KF, Weeks L, Sterne JA, Cochrane Bias Methods Group; Cochrane Statistical Methods Group: The cochrane collaboration’s tool for assessing risk of bias in randomized trials. BMJ. 2011, 343: d5928-10.1136/bmj.d5928.View ArticlePubMedPubMed CentralGoogle Scholar
- Ioannidis JP, Patsopoulos NA, Rothstein HR: Reasons or excuses for avoiding meta-analysis in forest plots. BMJ. 2008, 336: 1413-1415. 10.1136/bmj.a117.View ArticlePubMedPubMed CentralGoogle Scholar
- Knapp P, Gardner PH, Carrigan N, Raynor DK, Woolf E: Perceived risk of medicine side effects in users of a patient information website: a study of the use of verbal descriptors, percentages and natural frequencies. Br J Health Psychol. 2009, 14: 579-594. 10.1348/135910708X375344.View ArticlePubMedGoogle Scholar
- Moher D, Liberati A, Tetzlaff J, Altman DG, PRISMA Group: Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. BMJ. 2009, 339: b2535-10.1136/bmj.b2535.View ArticlePubMedPubMed CentralGoogle Scholar
- Berry DC, Knapp PR, Raynor T: Is 15 per cent very common? Informing people about the risks of medication side effects. Int J Pharm Pract. 2002, 10: 145-151.View ArticleGoogle Scholar
- Berry DC, Raynor DK, Knapp P: Communicating risk of medication side effects: an empirical evaluation of EU recommended terminology. Psychol Health Med. 2003, 8: 251-263. 10.1080/1354850031000135704.View ArticleGoogle Scholar
- Berry D, Raynor T, Knapp P, Bersellini E: Over the counter medicines and the need for immediate action: a further evaluation of European commission recommended wordings for communicating risk. Patient Educ Couns. 2004, 53: 129-134. 10.1016/S0738-3991(03)00111-3.View ArticlePubMedGoogle Scholar
- Berry DC, Hochhauser M: Verbal labels can triple perceived risk in clinical trials. Drug Inform J. 2006, 40: 249-258. 10.1177/009286150604000302.View ArticleGoogle Scholar
- Knapp P, Raynor DK, Berry DC: Comparison of two methods of presenting risk information to patients about the side effects of medicines. Qual Saf Health Care. 2004, 13: 176-180. 10.1136/qshc.2003.009076.View ArticlePubMedPubMed CentralGoogle Scholar
- Knapp P, Raynor DK, Woolf E, Gardner PH, Carrigan N, McMillan B: Communicating the risk of side effects to patients: an evaluation of UK regulatory recommendations. Drug Saf. 2009, 32: 837-849. 10.2165/11316570-000000000-00000.View ArticlePubMedGoogle Scholar
- European Commission (EC): Guideline on the Readability of the Labelling and Package Leaflet of Medicinal Products for Human Use. [http://ec.europa.eu/health/files/eudralex/vol-2/c/2009_01_12_readability_guideline_final_en.pdf]
- Eiser JR: Communication and interpretation of risk. Br Med Bull. 1998, 54: 779-790. 10.1093/oxfordjournals.bmb.a011729.View ArticlePubMedGoogle Scholar
- Woloshin KK, Ruffin MT, Gorenflo DW: Patients’ interpretation of qualitative probability statements. Arch Fam Med. 1994, 3: 961-966. 10.1001/archfami.3.11.961.View ArticlePubMedGoogle Scholar
- Lichtenstein S, Slovic P, Fischhoff B, Layman M, Combs B: Judged frequency of lethal events. Exp Psychol Hum Learn Memory. 1978, 4: 551-578.View ArticleGoogle Scholar
- Bostock S, Steptoe A: Association between low functional health literacy and mortality in older adults: longitudinal cohort study. Brit Med J. 2012, 344: e1602-10.1136/bmj.e1602.View ArticlePubMedPubMed CentralGoogle Scholar
- Streiner DL, Norman GR: Health Measurement Scales: A Practical Guide for their Development and Use. 2008, New York: Oxford University PressView ArticleGoogle Scholar
- Fisseni G, Lewis DK, Abholz HH: Understanding the concept of medical risk reduction: a comparison between the UK and Germany. Eur J Gen Pract. 2008, 14: 109-116. 10.1080/13814780802580247.View ArticlePubMedGoogle Scholar
- European Medicines Agency (EMA): Quality Review of Documents Human Product-information Annotated Template (English) Version 9. [http://www.ema.europa.eu/ema/index.jsp?curl=pages/regulation/document_listing/document_listing_000134.jsp]
- Cornelius VR, Perrio MJ, Shakir SA, Smith LA: Systematic reviews of adverse effects of drug interventions: a survey of their conduct and reporting quality. Pharmacoepidemiol Drug Saf. 2009, 18: 1223-1231. 10.1002/pds.1844.View ArticlePubMedGoogle Scholar
- Ioannidis JP, Lau J: Completeness of safety reporting in randomized trials: an evaluation of 7 medical areas. J Amer Med Assoc. 2001, 285: 437-443. 10.1001/jama.285.4.437.View ArticleGoogle Scholar
- Crockett RA, Sutton S, Walter FM, Clinch M, Marteau TM, Benson J: Impact on decisions to start or continue medicines of providing information to patients about possible benefits and/or harms: a systematic review and meta-analysis. Med Decis Making. 2011, 31: 767-777. 10.1177/0272989X11400420.View ArticlePubMedGoogle Scholar
- Hamrosi K, Dickinson R, Knapp P, Raynor DK, Krass I, Sowter J, Aslani P: It’s for your benefit: exploring patients’ opinions about the inclusion of textual and numerical benefit information in medicine leaflets. Int J Pharm Pract. 2013, 21: 216-225. 10.1111/j.2042-7174.2012.00253.x.View ArticlePubMedGoogle Scholar
- The pre-publication history for this paper can be accessed here:http://www.biomedcentral.com/1472-6947/14/76/prepub
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.