IMPUTATION OF DATA

Two types of survey responses require special discussion. The first occurs when respondents choose a “don't know” (DK) response. For frequency distributions, DK responses present no difficulty: we simply report the proportion of respondents who choose this option along with those choosing other responses, all as proportions of all valid responses (excluding missing data). For statistical analysis, however — such as calculating means or correlation coefficients — DKs can present special problems. On the questionnaire, these responses are often placed at the end of response scales and numbered accordingly, such as a 5 on a scale that runs from “0 = never” to “4 = every day.” However, in this case, computer statistical programs read the DK to signify a substantive value of 5. To prevent this, a standard procedure is to set all DKs as “missing” to make them invisible to statistical analysis. One drawback to this option is that it reduces the number of effective cases, thus limiting our confidence that the results apply across the entire sample, and the population that it represents (more on this below).

Moreover, “don't know” is often a legitimate, substantive answer that should be taken into account rather than discarded. Wherever possible, our preferred alternative is to recode DK to theoretically defensible spots on the response scale. For example, for many of the questions about political participation, we assumed that those who did not know whether they had ever taken a specific course of action had in fact never done so.