Optimal number of response categories in rating scales: reliability, validity, discriminating power, and respondent preferences

Using a self-administered questionnaire, 149 respondents rated service elements associated with a recently visited store or restaurant on scales that diered only in the number of response categories (ranging from 2 to 11) and on a 101-point scale presented in a dierent format. On several indices of reliability, validity, and discriminating power, the two-point, three-point, and four-point scales performed relatively poorly, and indices were signi®cantly higher for scales with more response categories, up to about 7. Internal consistency did not dier signi®cantly between scales, but test±retest reliability tended to decrease for scales with more than 10 response categories. Respondent preferences were highest for the 10-point scale, closely followed by the seven-point and nine-point scales. Implications for research and practice are discussed. Ó 2000 Elsevier Science B.V. All rights reserved. PsycINFO classi®cation: 2220


Introduction
Rating scales are among the most widely used measuring instruments in psychology, and it is therefore not surprising that a great deal of research has been devoted to the eects of variations in rating scale format, including dierences in the number of response categories. In current practice, most rating scales, including Likert-type scales and other attitude and opinion measures, contain either ®ve or seven response categories (Bearden, Netmeyer, & Mobley, 1993;Peter, 1979;Shaw & Wright, 1967).
In spite of decades of research, the issue of the optimal number of response categories in rating scales is still unresolved. Some investigators have studied response patterns and information retrieval. Schutz and Rucker (1975) found in their study of response patterns that``the number of available response categories does not materially aect the cognitive structure derived from the results'' (p. 323), which seems to suggest that the number of response categories has little eect on the results obtained. This conclusion is not in line with the ®ndings of other studies, which have provided support for the use of scales with more than two or three response categories. For example, Garner (1960) suggested that maximum information is obtained by using more than 20 response categories. Green and Rao (1970), on the other hand, found that information retrieval is maximized by using six or seven response categories, with little extra information being gained by increasing the number of categories beyond seven. Symonds (1924) was the ®rst to suggest that reliability (in this case inter-rater reliability) of scores is optimized by the use of seven categories. This suggestion was contested by Champney and Marshall (1939), who advocated the use of ®ner scales, but the case for seven-point scales was strengthened by Miller (1956), who suggested in an in¯uential article that the human mind has a span of apprehension capable of distinguishing about seven dierent items (plus or minus two), which implies a limit of about seven on the number of categories that people are able to use in making judgments about the magnitudes of unidimensional stimuli. This has implications for rating scales: the limit on the human span of apprehension suggests that little if any additional information can be obtained by increasing the number of response categories beyond about seven. The reliability of scores derived from scales with different numbers of response categories was later investigated by Bendig (1953Bendig ( , 1954, who found relatively constant test±retest reliabilities over scales with two, three, ®ve, seven, and nine response categories, relatively constant inter-rater reliability over scales with three, ®ve, seven, and nine response categories, and a decrease in reliability for 11-point scales. A few subsequent researchers con®rmed Bendig's ®nding that reliability is largely independent of the number of response categories (e.g., Boote, 1981;Brown, Wilding, & Coulter, 1991;Komorita, 1963;Matell & Jacoby, 1971;Peabody, 1962;Remington, Tyrer, Newson-Smith, & Cicchetti, 1979).
Some researchers in this area have arrived at dierent conclusions regarding reliability. In a study based on Monte-Carlo simulation methods, Cicchetti, Showalter and Tyrer (1985) found evidence for an increase in inter-rater reliability from twopoint to seven-point scales; beyond this ± even up to 100 response categories ± no substantial increase in reliability was found. These researchers concluded that``the dierences in scale reliability between a 7-, 8-, 9-, or 10-category ordinal scale on one hand, and a 100-point or continuous scale on the other is trivial . . . 7 ordinal categories of response appear at least functionally interchangeable with as many as 100 such ordered categories'' (p. 35). Similar conclusions were drawn by Oaster (1989) with regard to test±retest reliability and inter-item consistency, and a number of other researchers have reported that reliability is maximized with seven-point scales (Finn, 1972;Nunnally, 1967;Ramsay, 1973). These studies provide support for the early ®ndings of Symonds (1924), mentioned above, and more generally for the continued popularity of seven-point scales in practice (see Cox, 1980, for a review). A few researchers have, however, reported higher reliabilities for ®ve-point scales (Jenkins & Taber, 1977;Lissitz & Green, 1975;McKelvie, 1978;Remmers & Ewart, 1941), and a recent study using the multitrait-multimethod approach found evidence for higher monotrait-monomethod (MTMM) reliability in four-point than six-point scales (Chang, 1994).
In a comparatively small number of studies, validity has been used as a criterion for judging the performance of scales with dierent numbers of response categories. Matell and Jacoby (1971) carried out a thorough empirical study comparing scales with varying numbers of response categories (from 2 to 19) and concluded that as few as two response categories may be adequate in practice. They suggested that both reliability and validity are independent of the number of response categories, and their results implied that collapsing data from longer scales into two-point or three-point scales would not diminish the reliability or validity of the resulting scores. Loken, Pirie, Virnig, Hinkle and Salmon (1987) examined the criterion validity of various scales through their ability to dierentiate between dierent population groups and found 11-point scales to be superior to three-point or four-point scales. Hancock and Klockars (1991) found that nine-point scale scores correlated better than ®ve-point scale scores with objective measures of the original stimuli. In a comparison using a MTMM covariance matrix, Chang (1994) found approximately similar criterion validity coecients for four-point and six-point scales but higher convergent validity coecients for the six-point scales. Taken together, these studies tend to suggest that validity increases with increasing numbers of response categories or scale points.
Respondent preferences have not been investigated in depth in previous studies of rating scales. However, Jones (1968) examined respondents' preferences for scales with two or seven response categories and a graphic rating scale and reported that the dichotomous scale was judged to be less``accurate'', less``reliable'', less``interesting'', and more``ambiguous'' than both the seven-point and the graphic rating scales, but the two-point and graphic rating scales were judged to be easier to use. Respondents expressed a clear preference for multiple-category over dichotomous scales.
The aim of the investigation reported below is to provide a thorough assessment, using multiple independent criteria, of the reliability, validity, and discriminating power of scores from rating scales varying widely in number of response categories. A secondary aim is to investigate the other important issue in rating scale design, namely respondent preferences. Our research is more detailed and thorough than most previous studies in the area, and its design allows us to examine not only several objective indices of reliability, validity, and discriminating power, but also subjective measures of respondents' opinions about the scales. For example, if a scale is too dicult to use, or too simple to allow respondents to express themselves, then respondents may become frustrated and demotivated, and the quality of their responses may decrease. In addition to considerations of reliability, validity, and discriminating power, a test designer may wish to take respondent preferences into account when constructing a rating scale.

Materials and methods
A questionnaire was administered to 149 respondents (45 males and 104 females, aged from 18 to over 60 with a mode of 20), the majority of whom (134) were undergraduate students at the University of Leicester. The respondents were recruited by a form of snowball sampling in which students who volunteered to participate as respondents also recruited additional respondents in return for course credits. The sample thus consisted of undergraduate students and their friends (some of whom were also undergraduate students) and relatives.
Each respondent ®lled in a questionnaire consisting of a number of rating scales with varying numbers of response categories relating to the quality of service provided by either a store or a restaurant of which he or she had personal experience. These two service categories were chosen on the assumption that all respondents would have visited a store or restaurant in the recent past, and this turned out to be the case. Respondents were ®rst required to provide a global rating of their chosen store or restaurant for overall quality from 0 (extremely bad) to 100 (extremely good); this global rating served as a criterion measure for later assessments of validity. The rest of the questionnaire consisted of 11 sets of ®ve questions each, in which respondents rated the quality of ®ve service elements for their chosen store or restaurant. To avoid order-of-presentation eects that might have arisen from using a single random order of presentation, the 11 sets were presented in a dierent randomized order to each respondent. For the restaurant, the ®ve service elements werè`c ompetence of sta'',``promptness of service'',``range of choice'',``cleanliness of premises'', and``individual attention''; for the store, the service elements werè`h elpfulness of sta'',``promptness of service'',``range of products'',``tidiness of premises'', and``individual attention''.
The 11 sets of questions to which the participants responded were identical apart from the number of response categories in the rating scales associated with each of the ®ve service elements in each set. The number of response categories of the rating scales were 2, 3, 4, 5, 6, 7, 8, 9, 10, and 11. Each scale was end-anchored with very poor on the left and very good on the right and was presented as a series of numbers spaced out across the page representing the response categories. Respondents were required to circle the appropriate number in each case depending on their opinions or judgments. A 101-point scale was also included, with respondents being asked to rate each service element from 0 (very poor) to 100 (very good), but in this case no numbers were displayed: respondents were required simply to write down appropriate numbers between 0 and 100 re¯ecting their ratings.
In order to measure respondents' preferences for the various scales, they were also asked to rate each scale on three dierent aspects of scale performance. Using the 101-point scale described above, from 0 (very poor) to 100 (very good), they rated each scale on the following three criteria:``ease of use'',``quick to use'', and``allowed you to express your feelings adequately''.
In order to provide data for an assessment of test±retest reliability, each respondent completed an identical retest questionnaire between one and three weeks after the original testing ± the second questionnaire was issued to the respondents one week after the ®rst and had to be returned within two weeks from that date. Of the 149 respondents who completed the ®rst questionnaire, 129 completed the second within the two-week time limit, and the retest results are based on their responses. This response rate of 86% is high for research of this type.

Preliminary checks
The psychometric criteria that are used to compare scales with dierent numbers of response categories make sense only if such scales measure the same construct. We therefore carried out three preliminary checks.
First, we used a procedure described by Green (1992) to determine whether the 12 correlation matrices representing the associations between the ®ve items in each scale, one correlation matrix per scale length, diered signi®cantly from one another. The weighted least squares chi-square value was 144.09, df 155Y p 0X72, suggesting no evidence of signi®cant dierences between the correlation matrices.
Second, the unnormed and normed ®t indices (Bentler & Bonnet, 1980) that were achieved were 1.00 and 0.96, respectively; according to Bentler and Bonnet (1980), this indicates that little additional increase in goodness of ®t would be likely to be achieved by allowing some of the correlation matrices to dier from others.
Third, we performed a maximum-likelihood factor analysis on each of the 11 correlation matrices and then tested whether the ordinal structure of the factor loadings was the same across matrices. The rank-order of the factor loadings within each scale length (matrix) turned out to be identical across all 11 matrices, yielding a Kendall coecient of concordance of 1.00.
All three of these preliminary checks con®rm that the scales with dierent numbers of response categories all measure the same underlying construct.

Reliability
Ratings derived from each scale were evaluated for test±retest reliability (stability) and also, using Cronbach's alpha, for consistency over the ®ve questions within each scale type (internal consistency reliability). Table 1 shows the reliability coecients for the test±retest reliability analysis and the alpha coecients for the internal consistency reliability analysis.
The reliability coecients are all relatively high (0.79 or above) and statistically signi®cant beyond p`0X05, and the eect sizes are all large according to Cohen's (1988Cohen's ( , 1992 criterion. Test±retest reliability coecients were lowest for two-point, three-point, and four-point scales and highest for scales with about 7 to 10 response categories; there was a slight decline in test±retest reliability coecients for scales with 11 and 101 response categories. Statistical tests for the signi®cance of dierences between correlations were carried out using Fisher's r to z transformation (Howell, 1992, p. 251). Statistically signi®cant at p`0X05 were the dierences between the two-point scale and the scales with 6, 7, 8, 9, and 10 response categories; between the three-point scale and the scales with 6, 7, 8, 9, 10, and 11 response categories; between the four-point scale and the eight-point and nine-point scales; and between the 101-point scale and the eight-point and nine-point scales. All other dierences between the test±retest reliability coecients were statistically nonsigni®cant.
Cronbach alpha coecients were lowest for two-point and three-point scales, and like the test±retest reliability coecients they increased with increasing numbers of response categories up to seven. Little further increase in reliability was found above seven response categories: alpha coecients for scales with 7, 8, 9, 10, 11, and 101 response categories were all very similar. Using Feldt's test for the signi®cance of dierences between more than two related alpha coecients (Woodru & Feldt, 1986), it was determined that none of the dierences between the alpha coecients were statistically signi®cant: using either of the recommended statistics HAN2 or UX1, v 2 10`18X31, p b 0X05.

Validity and discriminating power
Validity and discriminating power were assessed in several dierent ways. First, an index of the criterion validity of scores derived from each scale was created by computing an aggregate score for each respondent over the questions relating to the ®ve service elements and then correlating these aggregate scores with the respondents' scores on the criterion global rating of overall quality of service. The global rating of overall service quality was included in the same questionnaire as the ratings of particular service elements, but it was separated from these other more speci®c measures. There are many precedents for the use as validity criteria of scales that were included in the same questionnaire as the scales in question (Althauser, & Scott, 1971;Campbell & Fiske, 1959;Messick, 1993). The global rating that was used in this study was chosen because there was no genuinely external criterion available, and because it is reasonable to assume that ratings of particular service elements, if they measure what they purport to measure, ought to correlate with global ratings of overall service quality.
Next, an index of the intertertile discriminating power of scores from each scale was obtained by calculating a t statistic relating to the mean dierence in aggregate scores for the scale's ®ve questions between the tertile of respondents who provided the highest global ratings of overall service quality (over 80 on the scale from 0 to 100) and the tertile who provided the lowest global quality rating (0±25). To provide a second index of the discriminating power of each scale, item-whole correlations were calculated between the ratings of each of the ®ve service elements and the aggregate score from all scales relating to the corresponding service element. For each scale, the index of discriminating power was provided by the mean item-whole correlation over the ®ve questions in the scale. For these calculations, ratings from all scales apart from the 101-point scale were rescaled to make them comparable using the following formula: rating À 1anumber of response categories À 1 Â 100X Finally, convergent validity (Campbell & Fiske, 1959) was evaluated by examining the correlations of scores on each scale with scores on each of the others. Scores from a scale were assumed to show convergent validity to the extent to which they correlated with scores from other scales measuring the same underlying construct (see also Althauser et al., 1971;American Psychological Association, 1985, pp. 9±10;Chang, 1994;Messick, 1993).
Data concerning criterion validity, intertertile discriminating power, and itemwhole correlations are presented in Table 2, and data concerning convergent validity are presented in Table 3.
The results in Table 2 show that the pattern for validity and discriminating power is similar to the pattern that was found for reliability. Correlations between scale scores and the criterion variable, t values for intertertile discriminating power,  and item-whole correlations were statistically signi®cant for all scales, and the eect sizes are all large according to Cohen's (1988Cohen's ( , 1992 criteria. As regards criterion validity, scores from the scales with two, three, and four response categories produced the lowest correlations (below r 0X86) with the criterion. Correlations for scores from 5-point to 101-point scales were higher and very similar to each other (around r 0X88). The scales that yielded scores with the highest criterion validity coecients r 0X89 were those with 9 and 101 response categories. However, none of the criterion validity coecients shown in Table 2 diers signi®cantly at p`0X05 from any of the others according to Williams's t statistic (Howell, 1992, p. 254). Turning to intertertile discriminating power, scales with two, three, and four response categories again performed least well, with scores from the three-point scale showing the lowest intertertile discriminating power of all g`20). Scores from the other scales showed similar discriminating power to one another, with nine-point, 11-point, and 101-point scales performing best. The signi®cance of the dierences between the intertertile discriminating power t values was tested by transforming the t values to z scores (Howell, 1992, p. 251). The only statistically signi®cant dierences at p`0X05 were between the t values for the three-point scale on the one hand and the scales with nine and 101 response categories on the other.
Item-whole correlations were also lowest for the scales with two, three, or four response categories and increased with increasing numbers of response categories up to about six. There was no substantial change in item-whole correlation coecients for the scales with more than six response categories. Fisher's r to z transformation was used to evaluate the signi®cance of dierences between the item-whole correlations. The coecient for the two-point scale diered at p`0X05 from all of the others. Coecients for the three-point and four-point scales also diered at p`0X05 from all the others but did not dier statistically signi®cantly from each other. All other dierences were nonsigni®cant. Table 3 shows the intercorrelations between scores from scales with dierent numbers of response categories. Every scale correlated highly and statistically sig-ni®cantly with each of the others, and all of the correlation coecients represent large eect sizes according to Cohen's (1988Cohen's ( , 1992 criterion. These results, which provide evidence of convergent validity, also indicate that the scales with relatively more response categories (six or more) correlated best with one another, and that the two-point and three-point scales correlated less highly with the longer scales.

Respondent preferences
Using a 101-point scale described above, respondents rated each scale for its``ease of use'', whether it was``quick to use'', and whether it``allowed you to express your feelings adequately''. Analysis of variance was carried out on the ratings given in response to each of these questions. The mean score for each scale on each of the three questions designed to measure respondent preferences, and the associated standard deviations, are shown in Table 4, along with the corresponding F values. The eect sizes were all large according to Cohen's (1988Cohen's ( , 1992 criterion:``ease of use'', g 2 34;``quick to use'', g 2 0X63;``allowed you to express your feelings adequately'', g 2 0X94. Table 5 shows the scales ordered from the one that received the lowest mean score (that is, the one that was rated least favorably) to the one that received the highest mean score (that is, the one that was rated most favorably) on each of the three questions related to respondent preferences. Scales in the same row that share the same subscript are not signi®cantly dierent; all other dierences between scales in the same row are statistically signi®cant at p`0X05 according to the Tukey-HSD multiple comparisons.
These respondent preference scores show several statistically signi®cant dierences between the scales. For``ease of use'', the scales with ®ve, seven, and 10 response categories were the most preferred, and the scales with 11 and 101 response categories were rated as least easy to use. The scales that were rated as most``quick to use'' were the ones with the fewest response categories: the two-point, threepoint, and four-point scales were rated most favorably on this criterion, and once Table 5 Scales ranked in order of increasing respondent preference: statistically signi®cant dierences a Scales (No. of response categories)

Lowest Highest
Ease 101 11 2 9 a 6 ab 3 b 8 b 4 7 10 5 Quick 101 11 9 10 8 a 7 a 6 5 b 4 b 2 c 3 c Express feelings 2 3 4 6 a 5 a 8 7 9 b 11 b 10 101 a Note. Scales in the same row that share the same subscript do not dier signi®cantly; all other dierences between scales in the same row are statistically signi®cant at p`0X05 according to the Tukey-HSD comparison. Table 4 Respondents' preferences for scales: mean ratings ( again the scales with 11 and 101 response categories were least preferred. Ratings of the degree to which the scales``allowed you to express your feelings adequately'' showed the greatest dierentiation between the scales. The two-point and threepoint scales received extremely low ratings on this criterion (well below 50 on the 0±100 scale). In general, longer scales tended to receive more favorable ratings on this dimension: the scales with 9, 10, 11 and 101 response categories were rated highest.
Taking into account all three questions related to respondent preferences, the shortest scales (two, three and four response categories) generally received the lowest ratings, but the longest scales (11-point and 101-point scales) also received relatively low ratings. The scale that scores best overall according to respondent preferences was the 10-point scale, closely followed by the seven-point and ninepoint scales.

Discussion
The rating scales that yielded the least reliable scores turned out to be those with the fewest response categories. Test±retest reliability (stability) was lowest for twopoint, three-point, and four-point scales and was signi®cantly higher for scales with more response categories; the most reliable scores were derived from scales with 7, 8, 9, or 10 response categories. Internal consistency was lowest for scales with two or three response categories and highest for those with seven or more, although on this criterion of reliability the dierences between the reliability coecients were not statistically signi®cant. Our results also provide evidence of a decrease in test± retest reliability for scales with more than 10 response categories, although only the decrease from the eight-point and nine-point scales to the 101-point scale attained statistical signi®cance. Bendig (1954) found a similar decrease in inter-rater reliability, but Cicchetti et al. (1985), using a Monte-Carlo simulation rather than a fully empirical research methodology, found no decrease in reliability for long scales.
According to the indices of validity and discriminating power that we examined in this study, the scales with relatively few response categories performed worst. The criterion validity coecients were lowest for the scales with two, three, or four response categories and were generally higher ± though these dierences were not statistically signi®cant ± for scales with ®ve or more response categories. Discriminating power was lowest for the scales with two, three, or four response categories and statistically signi®cantly higher for scales with 9 or 101 response categories. Item-whole correlations told a similar story: scales with two, three, or four response categories performed worst, and scales with six or more response categories performed generally better ± the coecient for the two-point scale was signi®cantly lower than the coecients for all other scales in the investigation. Finally, the table of intercorrelations between scores on all of the scales suggested that the scales with two or three response categories yielded scores with lower overall convergent validity than the others. However, it should be borne in mind that a scale with relatively few response categories tends to generate scores with comparatively little variance, limiting the magnitude of correlations with other scales (e.g., Chang, 1994;Martin, 1973Martin, , 1978Nunnally, 1970). This restriction-of-range eect tends to depress the convergent validity of scores from scales with few response categories, but it is worth remembering that this arises ultimately from the inherent bluntness of such scales, which also limits their usefulness for many practical psychometric purposes.
These ®ndings provide no corroboration for Matell and Jacoby's (1971) suggestion that reliability and validity of scores are independent of the number of response categories and that nothing is gained by using scales with more than two or three response categories; but neither do they corroborate Garner's (1960) suggestion that scales with 20 or more response categories are necessarily best. As regards reliability, the results reported above tend to con®rm the ®ndings of Symonds (1924), Nunnally (1967, Green andRao (1970), Finn (1972), Ramsay (1973), Cicchetti et al. (1985), and Oaster (1989), which suggested that reliability of scores tends to increase from two-point to six-point or seven-point scales. As regards validity and discriminating power, our results provide a remarkably consistent picture across four independent indices, tending to con®rm evidence from a small number of studies using much more restricted criteria of validity (e.g., Chang, 1994;Hancock & Klockars, 1991;Loken et al., 1987) that, statistically, scales with small numbers of response categories yield scores that are generally less valid and less discriminating than those with six or more response categories.
Respondent preference ratings diered substantially between the scales, and the dierences were statistically signi®cant. Scales with 5, 7, and 10 response categories were rated as relatively easy to use. Shorter scales with two, three, or four response categories were rated as relatively quick to use, but they were rated extremely unfavorably on the extent to which they allowed the respondents to express their feelings adequately; according to this criterion, scales with 10, 11 and 101 response categories were much preferred. On the whole, taking all three respondent preference ratings into account, scales with two, three, or four response categories were least preferred, and scales with 10, 9, and 7 were most preferred.
From the multiple indices of reliability, validity, discriminating power, and respondent preferences used in this study, a remarkably consistent set of conclusions emerges. Our results provide no support for the suggestion of Schutz and Rucker (1975) that the number of response categories is largely immaterial. On the contrary, scales with two, three, or four response categories yielded scores that were clearly and unambiguously the least reliable, valid, and discriminating. The most reliable scores were those from scales with between 7 and 10 response categories, the most valid and discriminating were from those with six or more response categories or-in the case of intertertile discriminating power ± those with nine or more. The results regarding respondent preferences showed that scales with two, three, or four response categories once again generally performed worst and those with 10, 9, or 7 performed best. The superiority of scales with around seven response categories is in line with Miller's (1956) theoretical analysis of human information-processing capacity and short-term memory, subsequently re®ned by Simon (1974) in his characterization of information``chunks''. For several decades, the vast majority of rating scales and related psychometric instruments have used ®ve or seven response categories (Bearden et al., 1993;Peter, 1979;Shaw & Wright, 1967). In the light of our ®ndings, there is some support for seven-point scales, but the popularity of ®vepoint scales seems to be less justi®ed.
Taken together, the results reported above suggest that rating scales with 7, 9, or 10 response categories are generally to be preferred. In this study, participants responded to the scales consecutively, and this may have led to factors other than scale format aecting their ratings. Respondents may, for example, have been inconsistent through carelessness or boredom arising from having to respond to similar questions again and again. However, the satisfactory levels of scale reliability, ranging from 0.79 to 0.94 (see Table 1), and the relatively high correlations between scores from dierent scales, ranging from 0.84 to 0.98 (see Table 3), suggest that the respondents rated carefully and consistently across scales. Also, the fact that the scales were presented in a dierent random order to each respondent ensures that any inconsistencies in responding could not contribute to signi®cant dierences found between scales. Caution should, however, be exercised in generalizing the conclusions beyond the types of respondents and scales used in this study.
The research reported here examined responses to real-life experiences in recent visits to stores or restaurants. Some previous studies have focused on ratings of reallife experiences (e.g., Neuman & Neuman, 1981), whereas others have involved ratings of hypothetical experiences or abstract concepts (e.g., Matell & Jacoby, 1971;Oaster, 1989) or have used computer simulations (e.g., Cicchetti, Showalter, & Tyrer, 1985). These methodological dierences probably explain, in part at least, the discrepant ®ndings reported in these dierent studies. However, our aim was to investigate the eects of scale length on participant ratings using real-life experiences in order to produce ®ndings that would have practical relevance to ratings in everyday situations such as those studied in market research.
A careful reading of our results suggests that dierent scales may be best suited to dierent purposes. Circumstances may, for example, require respondents to use a rating scale under conditions of time pressure, and in such cases it may be necessary, in order to prevent the respondents from becoming frustrated and demotivated, to use ®ve-point or even three-point scales, because our ®ndings show that these scales are likely to be perceived by the respondents as relatively quick and easy to use. On the other hand, where considerations of face validity are regarded as paramount, it may be important for the respondents to perceive the scales as allowing them to express their feelings adequately, and in such cases 10-point scales may be most appropriate. Before deciding on the optimal number of response categories for a rating scale, researchers and practitioners may therefore need to perform a trade-o, in the light of the prevailing circumstances, between reliability, validity, discriminating power, and respondent preferences.
The ®ndings reported in this article relate to ratings of service quality in restaurants and stores; further research in a similar vein, using objective ratings of behavior by others, self-ratings of behavior, ratings of personality traits, ratings of the quality of products, and so on, would be needed to determine the extent to which the conclusions generalize to other domains.