An analysis of subjects' expectations obtained in 1998 regarding the reported percentage in each of the 21 statements in the first task revealed that most expectations agreed with those collected from the 1992 study. For both studies, 18 out of 21 statements (86%) were expected in the presumed direction. This supported our intended manipulation in order to invoke different expectations in subjects about the likelihood of events stated in 21 statements. This context effect of expectation was presented in Figures 1A through 11A for the 1998 data and in Figures 1B through 11B for the 1992 data. In each of these figures, the vertical axis represented the percentage of subjects who selected that particular probability expression to match the numerical probability embedded in each statement (the medians for the 7 probability intervals were indicated on the horizontal axis). The patterns of the figures for each probability expression were very similar between 1992 and 1998. For example, the patterns of Figures 1A and 1B were similar. Likewise, Figures 2A and 2B were similar, and so on.

Figures 1A and 1B show that the numerical meaning of "almost impossible" fell into the lowest probability interval (0% to 14.3%) and subjects' expectations had an impact on their numerical interpretation of this word. When subjects expected 7% for an event's occurrence to be higher than their perceived actual probability, they were more inclined to use "almost impossible." If subjects' expectation for 7% of an event's occurrence was either lower than, or equal to, their perceived actual probability, they were not as inclined to choose "almost impossible." These observations suggested that the exact meaning of "almost impossible" might be lower than 7%. Figures 2A and 2B show that the numerical meaning of "very unlikely" also fell into the lowest probability interval and the subjects' expectation influenced their choice of this word. When subjects expected 7% for an event's occurrence to be lower than the actual probability, they were more inclined to use "very unlikely" This suggested that the exact meaning of "very unlikely" might be higher than 7%. Figures 3A and 3B were consistently flat across the 7 intervals at a low response rate. This suggested that subjects seldom used "improbable" to express uncertain information. Figures 4A and 4B indicate that "unlikely" covered the lowest three intervals (0% to 42.9%). These figures further suggest that the exact meaning of "unlikely" might be higher than 22% yet lower than 36%. Figures 5A and 5B indicate that "an even chance" fell into the interval centering around 50%. When the expectation effect was taken into consideration, the exact meaning of "an even chance" might be higher than 50%. Figures 6A and 6B demonstrate that "possible" covered a broad range of probabilities, most of which were in the range of 14.4% to 42.9%. Figures 7A and 7B imply that "probable" conveyed meanings of probabilities spanned two intervals, i.e. 28.7%-42.9% and 57.3%-71.5%. Numbers in both intervals were equally likely to be equivalents for "probable." Figures 8A and 8B indicate that while the numerical meaning of "likely" was broad, its numerical equivalent was mostly centered around 64% and, to a less degree, around 36%, but not around 50%. This was a point of difference between "probable" and "likely." Figures 9A and 9B illustrate the numerical meaning of "very probable" to be in the range of 57.3% to 100%. "Very likely" covered the similar range but peaked around 79% (see Figures 10A and 10B). In addition, "very likely" was used more often by subjects than "very probable," as shown by its higher response rate. Lastly, subjects selected "almost certain" to represent extremely high probabilities. Its numerical meaning(s) were mostly found in the interval centered at 93% (see Figures 11A and 11B).

The second (paired-comparison) task was included in both studies not only to reduce the recall of information between task 1 (expectation) and task 3 (probability restatement), but also to investigate the relative magnitudes of the 11 probability expressions under the context-free condition. Two statistical methods were used to uncover the derived numeric likelihood of the 11 probability expressions. The first method was the least squares unidimensional scaling applied to incomplete data, assuming a constant variance of discriminal differences for all pairs of stimuli (Gulliksen, 1956; Torgerson, 1958). This assumption is identical to Thurstone's case V or Torgerson's condition C (Thurstone, 1927; Torgerson, 1958). In the 1998 data, "likely" was judged to imply greater probability than "very unlikely" by all subjects so the transformation of this observed proportion (100%) to the corresponding unit normal deviate could not be made. Thus, the solution for incomplete data was used. The second method was the logit modeling using 11 indicator variables to model a total of 55 (11*10/2) contrasts within 55 pairs of probability expressions (Agresti, 1990; Bradley & Terry, 1952; Flury, 1997). For each pair of expressions ( i ¹ j), suppose the expression i is chosen n_{i j} times (out of N_{i j}) over the other expression j. When the N_{i j} comparisons are independent, with the same probability ?_{i j} applied to each comparison, n_{i j} has a binomial (N_{i j} , ?_{i j}) distribution. If the comparisons for different pairs of expressions are also assumed to be independent, logit modeling can be applied to these pair comparison data.

Scale values estimated by these two methods for the 1992 data and the 1998 data were listed in Table 1. The scale values estimated by Reagan et al. (1989) are listed in ascending order in the first column as a reference. Table 1 demonstrates that our subjects from both 1992 and 1998 judged "possible" to imply a higher probability than "an even chance". This finding was contrary to those reported in the first column, which was derived from the lexical-to-numeric mapping without a context (Reagan et al., 1989). Rank orders of scaled values corresponding to high probability expressions (i.e. "probable," "likely," "very probable," "very likely," and "almost certain") were completely consistent with those found in the first column. In terms of low probability expressions (i.e. "very unlikely," "improbable," and "unlikely"), the least squares estimates of scale values (less than 0) were even lower than estimates (exactly 0) for "almost impossible". The relative ranks of these four expressions contradicted those reported by two previous studies which included both lexical-to-numeric and numeric-to-lexical mappings without a context (Reagan et al., 1989; Wallsten et al., 1986).

In Table 2, the paired-comparison scales estimated by the same method (either least-squares or logit) between years (1992 and 1998) agreed with each other closely; the correlation for least squares scales was 0.97 and the correlation for logit scales was 0.99. The scales derived by the two different methods for the same data were also consistent with each other–the correlation for the 1998 scales was 0.82; the correlation for the 1992 scales was 0.85.