The evidence - evaluative measures
Life evaluations are often assessed through a single question, which means the number of response options offered is a particularly important determinant of scale sensitivity (discriminating power). Cummins (2003) recommends that 3-point response scales should be eliminated from life satisfaction research because they are too crude to be useful in detecting variation in responses. As Bradburn, Sudman and Wansink (2004) point out, on a single-item scale with three response categories, anchored by extremes at either end (for example, “best ever”, “worst ever” and “somewhere in between”) most people will tend to select the middle category. Alwin (1997) meanwhile argues that, if one is interested in attitudes that have a direction, intensity and region of neutrality, then a minimum of five response categories is necessary (three-category response formats communicate neutrality and direction, but not intensity).
Increasing the number of response options available is unlikely to make a difference unless the options added represent meaningful categories that respondents consider relevant to them. Smith (1979) examined time-trends in US national data on overall happiness and reported that extending the scale from 3 to 4 response options through the addition of a not at all happy category captured only a small number of respondents (1.3%), and did not lower the mean of responses (the not very happy category simply appeared to splinter). Conversely, adding a fifth completely happy response category at the other end of the scale both captured 13.8% of respondents and drew respondents further up the scale, with more respondents shifting their answers from pretty happy to very happy.
For evaluative measures with numerical response scales, longer scales (up to around 11 scale points) often appear to perform better. Cummins (2003) argues that there is broad consensus that a 5-point scale is inferior to a 7-point scale. Alwin (1997) compared 7-point and 11-point scales on multi-item measures of life satisfaction. Using a multi-trait-multi-method design, Alwin found that across all 17 domains of life satisfaction measured, the 11-point scales had higher reliabilities than the 7-point scales. In 14 out of 17 cases, the 11-point scales also had higher validity coefficients; and in 12 of 17 cases, 11-point scales had lower invalidity coefficients, indicating they were affected less, rather than more, by method variance - i.e. systematic response biases or styles. This overall finding is supported by Saris et al. (1998) who used a similar multi-trait-multi-method analysis to compare 100-point, 4 or 5-point and 10-point satisfaction measures, and found that the 10-point scale demonstrated the best reliability. Similarly, in a German Socio-Economic Panel Study pre-test, Kroh (2006) also found evidence that 11-point satisfaction scales had higher validity estimates than 7-point and open-ended magnitude satisfaction scales.6
Where it is desirable to make direct comparisons between measures (e.g. for international analysis, or between differently-worded questions), it will be important that measures adopt the same number of response options. Although procedures exist whereby responses can, in theory, be mathematically re-scaled for the purposes of comparison, there is evidence to suggest that, when life evaluation data are re-scaled in this way, longer scales can produce higher overall scores, thus potentially biasing scores upwards. For example, among the same group of respondents measured at the same time point, Lim (2008) found that the mean average level on an 11-point scale of global happiness7 was significantly higher than recalibrated 4- and 7-point measures (although, curiously, not the 5-point measure). Lim attributed this to the negative skewness typically observed in the distribution of evaluative measures of happiness and life satisfaction. Cummins (2003) reports a similar finding, and further argues that this negative skewness means that life satisfaction measures are perhaps best scaled with at least 11 points, as most of the meaningful variance is located in the upper half of the distribution.