|
Article Excerpt
ABSTRACT: Graphs are frequently used as decision aids. When properly designed, graphs facilitate decision making by highlighting important trends and relationships in the data. It is all too easy, however, to design graphs so that they do not accurately portray the underlying data (Tufte 1983). Indeed, there is considerable evidence that annual reports contain such improperly designed graphs (Beattie and Jones 1992a, 1992b; Courtis 1997; Johnson et al. 1980; Jones and Beattie 1997; Steinbart 1989). This paper reports the results of three experiments that investigated the effects of such improperly designed graphs on subjects' choices. The results of all three experiments show that improperly designed graphs can alter subjects' choices. These findings not only have important implications for the design and use of graphs as decision aids, but also raise potential policy issues.
Keywords: graph design; presentation format; decision making; choice.
Data Availability: Please contact the first author.
I. INTRODUCTION
For more than 70 years, beginning with Washburne (1927), researchers and systems designers have investigated the issue of how to best display quantitative data. During that time, a number of guidelines and rules for designing graphs have been proposed (Bertin 1983; Jarvenpaa and Dickson 1988; Kosslyn 1989, 1994; Stevens 1975; Tufte 1983, 1990, 1997; Wainer 1997). Nevertheless, violations of those guidelines frequently occur, even in academic papers (Editor's note in introduction to Gillan et al. [1998]). Part of the problem may be the consistent finding of empirical research that there appears to be no optimal way to design graphs, but rather the relative efficacy of different types of graphs depends upon the nature of the task and of the data being presented (Benbasat and Dexter 1986; DeSanctis and Jarvenpaa 1989; Dickson et al. 1986; Meyer 1997; Meyer et al. 1997; Meyer et al. 1999; Tan and Benbasat 1993; Wilson and Addo 1994). Consequently, the choice of how to present quantitative data in graphs depends o n "both the characteristics of the readers and of [the data]," (Gillan et al. 1998, 30).
The literature on the design of graphs, however, has focused almost exclusively on the use of graphs as decision aids to support and improve task performance. Although graphs are often designed to support decision making, they are also often prepared to persuade or convince the viewer (Meyer et al. 1997). In such cases, graphmakers may violate one or more of the suggested guidelines for graph design, in order to create a more persuasive presentation that directs the viewer's attention to some particular feature in the data set. For example, Tractinsky and Meyer (1999) found that people preferred and created bar graphs with more depth when the task involved creating a favorable impression, but switched to simpler two-dimensional bar graphs when the task was to provide information to improve decision accuracy.
There is also considerable evidence that published annual reports frequently contain graphs that are not designed in accordance with suggested guidelines (Beattie and Jones 1992a, 1992b; Johnson et al. 1980; Courtis 1997; Jones and Beattie 1997; Steinbart 1989). Tractinsky and Meyer (1999) note that participants in their experiment were more likely to violate the principles of graph design when the data itself reflected undesirably on the presenter. Similarly, Steinbart (1989, 69) reports that graphs violating one or more suggested design guidelines are more likely to be found in the annual reports for companies that experienced a decline in net income from the prior year than in the annual reports of companies that experienced an increase in net income.
These findings suggest that preparers of graphs believe that they can design graphs to manage the viewer's impressions. Evidence suggests that such a strategy may be successful. An experiment by Lawrence and O'Connor (1993) showed that seemingly minor formatting options, such as including or omitting horizontal grid lines, significantly affected users' predictions about future values of time-series data. Likewise, Carswell (1991) found that bar charts, embellished to show depth, reduced decision accuracy and induced more favorable attitudes toward the data.
An important question, however, is whether such attempts at impression management can actually alter choices. When used for decision making, graphs are likely to include quantitative labels. Thus, even if the graphs are designed to selectively portray a favorable picture of the data, decision makers will also have access to accurate numeric information describing each alternative. The study described in this paper was designed to investigate the effects of such a combination of accurate numeric data and impression-inducing graphs on decision makers' choices.
This paper reports the results of three experiments that investigate the effects of improperly designed graphs on decision making. Our results indicate that decision makers' choices are indeed influenced by improperly designed graphs, even though such graphs also present precise, accurate numeric data.
The next section of this paper reviews the relevant literature and develops the research hypothesis. The following sections describe our research method, and then our results. The paper concludes with a discussion of the implications of our findings.
II. BACKGROUND
Benefits of Graphs
There is considerable evidence that well-designed graphs facilitate task performance (Benbasat and Dexter 1986; Blocher et al. 1986; Davis 1989; DeSanctis and Jarvenpaa 1989; Dickson et al. 1986; Vessey 1991). Well-designed graphs make it easier to identify trends and relationships among variables (Tufte 1983, 1990, 1997; Vessey and Galletta 1991). Consequently, it is not surprising that graphs are particularly useful for tasks that involve forecasting (DeSanctis and Jarvenpaa 1989; Umanath and Vessey 1994).
However, a graph's potential benefits depend on the nature of the task. It is easier to identify relationships and trends among variables, but harder to extract precise values from graphs as opposed to tables of numbers (Coil 1992; Coll et al. 1994; Davis 1989; Umanath and Scamell 1988; Umanath et al. 1990; Vessey and Galletta 1991). Consequently, it is not surprising that performance on forecasting tasks that require identification of trends and estimations of future values improves when decision makers are provided graphs that also contain precise numeric data (Wright 1995).
There is also evidence that different graph formats are not equally useful (Benbasat and Dexter 1986; Lawrence and O'Connor 1993; Tan and Benbasat 1993; Wilson and Addo 1994). For example, decision makers can extract specific point values faster from bar charts than from line graphs (Tan and Benbasat 1993). More importantly, graph design can also affect users' perceptions and judgments. For example, seemingly minor formatting options, such as including or omitting horizontal grid lines, significantly affect users' predictions about future values of time-series data (Lawrence and O'Connor 1993). Such findings point out the importance of carefully designing graphs to support the tasks for which they are intended to be used.
Principles of Graph Design
There exists a large body of literature that prescribes how to design graphs properly (Bertin 1983; Ives 1982; Kosslyn 1989; Tufte 1983, 1990, 1997; Wainer 1997). Those principles can be briefly summarized as follows: graphs should be drawn in a manner that leads viewers to reach conclusions consistent with those that they would reach when analyzing the underlying numeric data upon which the graphs are based. Tufte (1983, 77) presents six rules for achieving this objective:
1. The magnitude of change depicted graphically should be directly proportional to the numerical change in the data. Graphs depict both the quantitative size and the direction of change. If sales have doubled over the past five years, then the height of the bar representing current year sales should be twice as tall as that of the bar representing sales five years ago. Showing the direction of change correctly, but distorting the magnitude, is misleading.
2. Graphs should be clearly labeled to avoid any ambiguity.
3. Variation in design should mirror numerical change in the data. Visual effects, such as bizarre shadings, should not distract the reader from focusing on changes in the data being graphed.
4. When graphing time-series data, deflated and standardized units of monetary measurement are usually better than nominal units.
5. The number of dimensions used to depict change should be limited to the number of dimensions in the data. Thus, three-dimensional graphs are inappropriate for depicting changes in a single variable because the magnitude of the change in volume does not correlate with the change in the underlying data, thereby violating Principle 1.
6. Proper context to accurately interpret the data should be provided. For example, at least three data points should be included when graphing trends in data.
Examples of Improperly Designed Graphs
Empirical evidence shows that annual reports in the United States, Canada, the United Kingdom, and Hong Kong frequently contain graphs that do not follow the aforementioned prescriptions (Beattie and Jones 1992a, 1992b; Courtis 1997; Johnson et al. 1980; Jones and Beattie 1997; Steinbart 1989). For example, Johnson et al. (1980, 52) examined 50 annual reports of Fortune 500 companies for the years 1977 and 1978, and found that 21 contained at least one incorrectly drawn graph. A decade later, Steinbart (1989,68) examined 319 annual reports of large U.S. companies, and found that 38 contained graphs that distorted the underlying financial information. Beattie and Jones (1992b, 298) examined 240 annual reports of U.K. companies in 1989, and found that 30 percent contained graphs with material measurement distortions. Jones and Beattie (1997) report finding graphs that distorted financial information in annual reports from six countries, including France and Australia. Finally, Courtis (1997) examined 114 Hong K ong companies' annual reports, and found that close to half of the graphs in those reports distorted the financial data being presented. Figure 1 presents examples of four common types of...
|
