Data and Methods

The Kaiser Family Foundation/Harvard School of Public Health Survey, Health Care Agenda for the New Congress, was designed and analyzed by researchers at the Kaiser Family Foundation and Harvard School of Public Health. Fieldwork was conducted by telephone by ICR/International Communications Research between November 4 and November 28, 2004, among a nationally representative sample of 1,396 respondents ages eighteen and over. The survey included an oversample of respondents ages sixty-five and over (with a total of 407 respondents in this age group interviewed). Results for all groups were weighted to reflect the actual distribution in the nation based on 2004 U.S. Census Current Population Studies estimates. The questionnaire was offered in Spanish to those respondents wishing to be interviewed in Spanish.

For demographics and sample variables, the study utilized dichotomous measures for gender, married and never married, metropolitan status (rural or not), and race (African American, Hispanic, and “other race” all being dichotomous and therefore contrasted to whites). Continuous variables were also utilized, including age, number of adults in the household, and number of children in the household.

Party identification, voter registration, and voting behavior were also measured dichotomously. To measure potential impact within these measures more effectively, the raw scores were recoded using a contrast coding method. As defined by McKee McClendon, contrast coding, like the more ubiquitous dummy coding, is a process whereby g-1 variables are created to present g categories within a nominal variable. However, contrast coding differs from dummy coding in that contrast coding can define a single group or subset of groups and compare them to another group or subset of groups within the nominal variable. There are two mathematical restrictions to the procedure, namely, that the numerical designations given to each group within each created variable must sum to zero. The second restriction is that the sum of the products for each pair of contrast coded variables must equal zero. These restrictions ensure the creation of a set of uncorrelated variables.

For party identification, the original leaned party variable included responses for Democrat, Republican, Independent, other, and don’t know. This was recoded into a trichotomous variable coded as Democrat, Republican, and other. From this nominal variable, two contrast codes were created. “Republican versus Democrat” was created by coding republicans as -.5, Democrats as .5, and all others as 0. “Republican or Democrat versus neither” was created by coding Republicans and Democrats as -.25 and all others as .5.

Similarly, within voting behavior, two sets of variables were created. “Kerry versus Bush” was created by coding Kerry voters as -.5, Bush voters as .5, and all others as 0. “Bush or Kerry versus neither” was created by coding Kerry and Bush voters as -.25 and all others as .5. “Registered versus not registered” was created by coding registered as .25, and respondents who were not registered as -.5. “Voted versus registered but did not vote” was created by coding voters as .5, registered nonvoters as -.5, and unregistered respondents as 0.

By far the most important, and complicated, variables to construct were the dispositional variables. These include not only the number of call attempts required to reach a completion, but also the prior dispositions recorded by the CATI system during the field period. There are a vast array of dispositions recorded by most CATI software, including “resolved” dispositions such as nonworking, fax/data/modem, businesses, ineligible, not available for the duration of the survey, terminate (in the case of non-general-population surveys), failed refusal conversion, and, of course, completed interview.  “Active” dispositions include first refusal, no answer, uncalled, answering machine, busy, language barriers (for interviews where the contact interviewer was not able to speak the language of the respondent), privacy manager, and callback. Once a sample piece has reached a resolved disposition, there is no change of completion. However, completed interviews, at least those for which the completion occurred after the first call attempt, can be coded as any of the dispositions listed above as active.

However, to make for a more parsimonious analysis, a number of “like-minded” dispositions have been folded into more general overarching dispositions. Indeed, some CATI systems never break out busy, no answer, and answering machine into separate codes but rather fold them into one overarching “no answer” disposition. A similar collapsing of these codes was conducted for this analysis.

Most importantly, it is important to understand that every call attempt receives its own disposition, and the question then naturally follows as to which of these dispositions should be used for the analysis. We followed the most reasonable avenue of analysis, which is to say that a dichotomous variable was constructed for each disposition, and any given respondent was coded for that disposition as long as at any point in the respondent’s call history he or she was coded as that disposition. It is possible (but unlikely), after all, that a single respondent could have been coded as a no answer, a refusal, a language barrier, and a privacy manager, all before becoming a complete, and essentially in any plausible order. And, it stands to reason that once a household refuses to be interviewed, that household should be considered a refusal conversion if it later completes an interview, no matter how many subsequent callbacks it takes.

The same is true for privacy managers, language barriers, and callbacks. The exception to this is no answer/busy/answering machine dispositions. For these dispositions, it was determined that the variable would be coded as such only if that sample piece did not attain any of the other dispositions in prior call attempts.

Overall, then, the table below shows which variables were identified by disposition. The table also shows, for comparison purposes, what the last noncompletion disposition was for each disposition type.

Given the overlap between variables, such that any one respondent could have been scored a “one” on any or all of the dispositional dichotomous variables (not including “no answer” or “first call complete”), data analysis was conducted using a logistic regression analysis on each of seven dichotomous variables: “First call complete,” “fifth or more call complete,” “refusal conversion,” “callback,” “privacy manager,” and “no answer only.” The model for Spanish language barrier was not run, however, because of the low response within this variable and high multicolliniarity with other variables. In fact, only one of these respondents actually said he or she voted for one of the two major candidates. As such, the model was unstable and would not produce a stable iteration.