Negative versus positive framing of health information
Abstract
This is a protocol for a Cochrane Review (Intervention). The objectives are as follows:
To evaluate the effects of negative versus positive framing of the same health information on persuasiveness, understanding, attitude and behavior of healthcare providers and consumers.
Background
Kahneman and Tversky originally used the term 'framing effect' to describe the general tendency of risk aversion with positively‐framed problems, and risk seeking with negatively‐framed problems (Kahneman 1979; Tversky 1981). In their classic 'Asian disease' scenario, Kahneman and Tversky studied the responses of people assigned to two problems which were framed differently (but had equivalent outcomes). In each problem, they asked participants to choose between two alternative programs to combat an outbreak of an unusual Asian disease which is expected to kill 600 people. In Problem 1, Program A would result in 200 people out of 600 being saved, while with Program B there is a 1/3 probability that 600 people will be saved and a 2/3 probability that no people will be saved. In Problem 2, Program C would result in 400 people dying, and in Program D there is a 1/3 probability that nobody will die, and a 2/3 probability that 600 people will die. Problem 1 and Problem 2 are effectively identical. In both Problems, the first option is certain (200 will live and 400 will die) while the second is risky. When the outcomes of the scenario were framed positively as in Problem 1 (e.g. people will be saved) participants were risk averse (i.e. selected the certain alternative (Program A) over the risky alternative (Program B)). When the outcomes of the scenario were framed negatively as in Problem 2 (e.g. people will die) participants were risk taking (i.e. selected the risky alternative (Program D) over the certain alternative (Program C)).
Schneider and colleagues proposed a taxonomy classifying the framing effect into three types (Schneider 1995). The first type is the 'risky choice framing' as described by Kahneman and Tversky. The second type is the 'attribute framing', and relates to positive or negative framing to encode or evaluate a specific attribute of a single item (e.g. procedure 90% safe versus procedure 10% risky). The third type is the 'action or goal framing' and relates to positive or negative framing of the consequences of performing or not performing the act (e.g. by doing a mammogram you increase your chance of survival versus by not doing a mammogram you decrease your chance of survival).
The same authors later published a systematic review on framing effect based on their proposed categorization but not restricted to the health setting (Levin 1998). They showed that within each type of framing effect, results show substantial consistency. In risky choice framing, a choice shift (but not necessarily a reversal) typically occurs such that positive frames generally enhance risk aversion relative to negative frames. In attribute framing, attributes are judged more favorably when labeled in positive terms rather than negative terms. And in a goal framing, a negatively‐framed message emphasizing losses tends to have a greater impact on a given behavior than a comparable positively‐framed message‐emphasizing gains.
More specifically to the health domain, Rothman and Salovey built on Kahneman and Tversky's work and hypothesized that gain frames would be more effective for disease prevention behaviors and recuperative (therapeutic) behaviors (considered as risk aversion behaviors) (Rothman 1997). They also hypothesized that loss frames would be more effective for disease detection behaviors (considered as risk seeking behaviors).
Many studies have evaluated the effects of message framing in the health domains, with conflicting results (Marteau 1989; Donovan 2000; van Assema 2001; Williams 2001). The objective of this systematic review is to compare the effects of negative versus positive framing of the same health information on healthcare providers and consumers. We are interested in attribute framing and goal framing but not in risky choice framing. While the first two are relevant for both clinical and public health practice and present exactly the same message using different frames, the latter does not.
We review separately the evidence for using different statistical formats to present health information (Akl 2007b).
Objectives
To evaluate the effects of negative versus positive framing of the same health information on persuasiveness, understanding, attitude and behavior of healthcare providers and consumers.
Methods
Criteria for considering studies for this review
Types of studies
Randomized controlled trials (RCT), quasi‐RCTs, controlled before and after studies (CBA).
Types of participants
Healthcare providers, policy makers, patients, and the general public.
Types of interventions
Interventions that compare a positively‐framed message to a negatively‐framed message on the same evidence about health. As mentioned above, we are interested in attribute framing and goal framing but not in risky choice framing.
We will exclude from this review interventions that compare different statistical presentations of the same health information (see Akl 2007b), different graphical or verbal presentations of the same health information, different orders of comparing risks or comparisons, different amount of information or different media to present the same information.
Types of outcome measures
Any outcome measures (including self‐reported) of persuasiveness, understanding, attitude and behavior (including choice made). Persuasiveness refers to the power to induce taking a course of action or the embracing of a point of view by means of argument or entreaty. We will consider these outcomes in the setting of both real and hypothetical decisions.
Search methods for identification of studies
The search for this review will be part of a larger search for studies assessing different presentations of the same health information. We will search MEDLINE (Ovid), EMBASE (Ovid), PsycINFO (Ovid) and the Cochrane Central Register of Controlled Trials (CENTRAL) using no language or date restriction.
Search strategies for MEDLINE, EMBASE and PsycINFO are presented in Appendices.
We will search the CENTRAL using FRAM* and PRESENT* as text words. In addition, we will search MEDLINE, EMBASE and PscyINFO using "framing" as title word (framing.ti).
We will use the 'Related Articles' feature of PubMed MEDLINE to find additional articles. We will search MEDLINE and PsycINFO databases for articles published by the first authors of included articles and of excluded but closely related articles. We will review the reference lists of related systematic reviews, included articles and excluded but closely related articles. Finally, we will contact experts in the field.
Data collection and analysis
Selection of trials
Two review authors will independently screen the title and abstract of identified articles for relevance. We will retrieve the full text of articles judged potentially relevant by at least one author. Two review authors will then independently screen the full text article for inclusion or exclusion. The review authors will resolve their disagreements by discussion or by consulting a third author.
Methodological quality
Two review authors will independently assess the methodological quality of each included study and resolve their disagreements by discussion or by consulting a third author. We will assess the following methodological data::
1. Randomization (according to the criteria set out in Ryan 2007)
2. Allocation concealment (according to the criteria set out in Ryan 2007)
3. Outcome type; we will use the following categorization:
-
Objective (example: actual use of mammography)
-
Self‐reported past/present (example: self‐reported use of mammography, personal attitudes towards mammography)
-
Self‐reported future (example: self‐reported intention to use mammography)
-
Hypothetical (example: preferences of non‐cancer patients for chemotherapy versus none)
4. Follow‐up; we will record the percent follow‐up for each study.
Assessment of methodological quality will be reported in an additional table.
Data extraction
We will develop a data extraction form. Two review authors will independently extract data from each included study and resolve their disagreements by discussion or by consulting a third author. We will extract data relating to study design, intervention type (goal versus attribute framing; and preventive versus screening versus recuperative behavior), number and type of participants, outcomes assessed and study results. We will contact the authors for incompletely reported data.
Analysis
We will analyze the results of included studies to compare positive and negative framing. Because outcomes in these studies are typically scaled responses to survey questions, we will standardize the effects using Hedges adjusted standardized mean difference (SMD). For comparisons where we would not be able to calculate the SMD directly, we will estimate t‐values for the study and the corresponding SMD using SMD=2t/sqrt(N) (Cooper 1994) and adjust it using the same adjustment factor; in all cases, we will calculate the adjusted standard error for the resulting SMD. We will pool multiple outcome measures for a single trial ‐ for example, three different questions about attitude, or responses to three different scenarios conducted in the same group of participants ‐ using fixed‐effect models into a single SMD for that comparison.
We will pool data from different studies when appropriate using random‐effects models with the inverse variance approach.
If a paper reports the results of two or more separate comparisons enrolling different participants, we will treat these as such. If a study uses a between‐subjects factorial design to compare an intervention of interest across another factor, we will treat these as separate comparisons. When practical, we will conduct subgroup analyses for different types of participants and types of behavior as defined by Rothman and Salovey (Rothman 1997). We will conduct sensitivity analyses by comparing the results of studies of lower methodological quality with those of higher methodological quality.
We will create inverted funnel plots of individual study results plotted against inverse of the variance in order to check for possible publication bias.
