How to Deal with Missing Data from Modular Surveys?
At the recent 2013 Market Research in the Mobile World North America conference, Edward Paul Johnson, director of analytics at SSI, and Christi Walters, principal owner of Gongos Research gave a talk centered on the intricacies of conducting mobile surveys, including addressing the issue of missing data from a sequence of modular surveys.
Examining a study that was designed to test both mobilization and modularization and was jointly conducted by Gongos, the analytics firm SSI, and Coca-Cola, the two presenters gamely described how they managed to split up the results of long surveys into several parts and how they integrated the corresponding data sets.
Walters tackled how long surveys, such as those spanning twenty to thirty minutes of questioning, can unravel the psyche of respondents but may be difficult to modularize into smaller time increments. She said that fusion techniques can be used to deal with missing data resulting from modular surveys, raising intriguing points such as the following:can modular surveys mirror the results of a survey where a respondent decides to answer, for instance, all the questions in one sitting? Also, are the long survey results the same if they were modularized?
Within-respondent modularization necessitates one respondent filling all the survey modules. Between-respondent modularization, on the other hand, involves picking any respondent to fill out modules of his choosing, which results in missing data. To avoid missing data, respondents are given incremental incentives to motivate them to complete all the modules. They are also encouraged to complete the modules using their smartphones–which is a great step toward leveraging mobile research.
Meanwhile, Johnson described two assembly methods for dealing with missing data resulting from having respondents who fail to complete all survey modules. There’s respondent matching, wherein two people are treated as a single respondent. One person completes the first module, while the other person sets out to answer the second module. In respondent matching, there is no made-up data because the two respondents’ answers are simply correlated and treated as if they originated from a single respondent. The other method involves hot-deck imputation, where the missing information is assigned to a randomly chosen similar data that were previously recorded.
In addition, hook questions can be introduced to address the problem of missing data. The hook questions must be introduced in the first module. That way, respondents always see them. After ensuring that the hook questions are delivered, give incentives to goad respondents to proceed to subsequent modules.
You can view the full video presentation below: