U.S. Department of Transportation
Federal Highway Administration
1200 New Jersey Avenue, SE
Washington, DC 20590
Federal Highway Administration Research and Technology
Coordinating, Developing, and Delivering Highway Transportation Innovations
|This report is an archived publication and may contain dated technical, contact, and link information|
Publication Number: FHWA-RD-96-145
Date: February 1998
Development of Human Factors Guidelines for Advanced Traveler Information Systems and Commercial Vehicle Operations: The Effects of Inaccurate Traffic Information on Driver Behavior and Acceptance of an Advanced In-Vehicle Traveler Information System
CHAPTER 3. RESULTS
Three objective dependent variables are reported: penalty cost, convergence relative to the baseline route, and system query frequency. Figure 4 shows mean penalty costs were lower when traffic information was accurate, F(1,44) = 108, MS e= 2.89, p < 0.001. This result is consistent with previous research (Kantowitz, et al., 1994) and shows the greater benefit of using the simulated ATIS when it was accurate. Penalty costs were higher in the familiar Seattle network, F(1,44) = 15.5, MSe = 2.54, p < 0.001, suggesting that drivers were more likely to follow ATIS advice in an unfamiliar setting. When traffic information was 100 percent accurate, penalty costs were higher in the familiar setting ($10.34) than in the unfamiliar setting ($8.67), t(44) = 4.02, p < 0.01. The interaction shown in figure 4 between Accuracy and Familiarity while significant, F(1,44) = 6.81, MSe = 4.15, p < 0.02, is unimportant; the less rapid rise in cost for the familiar setting is probably a ceiling effect (Kantowitz, Roediger, & Elmes, 1994, p. 335) due to a maximum penalty of $13.00 on any one trial. Penalty costs were reduced on Trial 2 ($9.20) versus Trial 1 ($9.80), t(44) = 2.34, p < 0.02, showing that repetition allowed drivers to use the simulated ATIS more effectively.
The higher penalty costs for the familiar network cannot be explained by a lower frequency of queries. Query frequency did not differ between Seattle (32.3) and New City (29.8), F(1,44) = 3.95, MSe = 75.9, p > 0.05. Thus, although the same amount of traffic information was received in the familiar setting, it was not used effectively. Perhaps drivers rely more upon their internal mental representations of traffic density in a familiar setting and so tend to discount external information or possibly create a mental weighted average of old internal representation and new external information. However, more queries were made when information was Inaccurate (34.8) than when Accurate (27.3), F(1,44) = 34.7, MSe = 79.1, p < 0.001. This seems to be a reasonable strategy for drivers: when the world has greater uncertainty, drivers seek more information to resolve that uncertainty.
A convergence score of 100 percent indicates that a driver perfectly followed his or her preferred route marked on a paper map at the start of the experiment; a score of 0 percent indicates no common links between the paper map and the route chosen on the RGS. Convergence was higher when information was Accurate (64.6 percent) than when Inaccurate (47.5 percent), F(1,44) = 11.8, MSe = 1192, p < 0.001. This agrees with previous results (Kantowitz, et al., 1994), showing that drivers are less likely to diverge from preplanned routes when traffic information is accurate. However, a significant three–way interaction, Order X Accuracy X Familiarity, F(1,44) = 9.03, MSe = 811, p < 0.005, revealed the importance of the first city encountered on Trial 1. Drivers who started with New City had greater differences in Convergence between Accurate and Inaccurate conditions than did drivers who first encountered the Seattle map. These drivers in Order 1 had higher convergence scores for Accurate information, and lower scores for Inaccurate information, than did drivers first encountering New City. Convergence was lower on Trials 1 and 4 (but equal on Trials 2 and 3) for Seattle (48.2 percent) than for New City (66.7 percent). Since the topographies of Seattle and New City were identical, this interaction can be interpreted as consistent with the result that drivers exhibited greater self–confidence in familiar environs.
Five subjective dependent variables were analyzed: trust, self–confidence, trust minus self–confidence (T–SC), traffic expectations, and estimated link travel times. All F–ratios involving the Link Position independent variable used the Greenhouse–Geiser correction for repeated measures.
Figure 5 shows mean rated trust as a function of Link Position and Information Accuracy. Trust was higher when information was accurate, F(1,44) = 31.6, MSe = 1031, p < 0.001. When information was inaccurate, trust was higher for the 71 percent condition than for the 43 percent condition, F(1,44) = 5.02, MSe = 2539, p < 0.03. While there was a significant effect of Link Position, F(6, 264) = 11.9, MSe = 240, p < 0.001, this effect was due to the inaccurate information on Trials 3 and 4 rather than the accurate information on the first two trials, F(6,264) = 6.17, MSe = 139, p < 0.001. When inaccurate information was presented, trust recovered on subsequent links when accurate information was presented. Figure 6 shows mean rated trust as a function of network familiarity and information accuracy. While there was no main effect of familiarity, F(1,44) < 1.0, the interaction shown in figure 6 reveals that 43 percent––accurate information decreases trust more than 71 percent––accurate information in both settings, F(1,44) = 5.01, MSe = 630, p < 0.03.
Trust did not differ according to type of inaccurate information: Harmless (71.3) versus Harmful (68.8), t(1338) = 1.89, p > 0.05. While this differs from earlier results (Kantowitz, et al., 1994), we believe the present results are more definitive. In the previous experiment, type of inaccurate information depended upon the path taken and was based upon a small number of observations. This experiment controlled the type of inaccurate information (table 1).
Rated self–confidence was higher in Familiar (76.6) versus Unfamiliar (71.7) settings, F(1,44) = 6.92, MSe = 1196, p < 0.02.
Figure 7 shows T–SC as a function of Link Position and Information Accuracy. Results are similar to those in figure 5. T–SC was higher when information was accurate, F(1,44) = 26.6, MSe = 1135, p < 0.001. When information was inaccurate, T–SC was higher for the 71 percent condition than for the 43 percent condition, F(1,44) = 5.97, MSe = 5209, p < 0.02. While there was a significant effect of link position, F(6,264) = 11.2, MSe = 196, p < 0.001, this effect was due to inaccurate information on Trials 3 and 4, F(6,264) = 3.15, MSe = 170, p < 0.02. Note that T–SC became negative on the last two trials for the 43 percent accuracy condition. Figure 8 shows T–SC as a function of network familiarity and information accuracy. Unlike figure 6, there was a main effect of familiarity, F(1,44) = 6.10, MSe = 1229, p < 0.02, with T–SC being higher for the unfamiliar setting. Also unlike figure 6, there was no interaction between variables, F(1,44) < 1.0. Note that T–SC becomes negative only for the familiar setting with inaccurate information.
In general, T–SC scores are not as well behaved as either of their two components considered in isolation. This is a well–documented problem with difference scores from empirical data (Cronbach & Furby, 1970; Bittner, Carter, & Kennedy, 1986). For example, the divergence on Trial 2 of the two Information Accuracy groups was unexpected and inconsistent with results for the Trust dependent variable in isolation (figure 5). Similarly, Trust for the 100 percent/71 percent group does not show the continuing decline over Trials found for T–SC. While it might be tempting to interpret this decline in figure 7 as indicating that with enough trials T–SC will eventually become negative even for the 100 percent/71 percent group, the decline of the T–SC scores could also reflect methodological issues associated with difference scores. Since any in–vehicle ATIS will provide the driver with far more than four trials, it would be prudent to replicate this experiment over several days to investigate a variety of learning and practice effects.
Figure 8 shows rated traffic expectations as a function of information accuracy. Expectations were better met when information was accurate, F(1,44) = 14.9, MSe = 1037, p < 0.001. The interaction shown in figure 8 reveals that inaccurate information did not alter expectations when information was 71 percent accurate, but expectations were not met when information was only 43 percent accurate, F(1,44) = 11.5, MSe = 1037, p < 0.001. Mean estimated link travel time was greater for Familiar (2.50 minutes) versus Unfamiliar (2.20) settings, F(1,44) = 19.6, MSe = 1.7, p < 0.001. Accuracy of information did not influence estimated travel time, F(1,44) < 1.0.