Classification Scoring for Cleaning Inconsistent Survey Data
Thissen, M. (2017). Classification Scoring for Cleaning Inconsistent Survey Data. International Journal of Data Engineering, 7(1), 1-14.
Data engineers are often asked to detect and resolve inconsistencies within data sets. For some data sources with problems, there is no option to ask for corrections or updates, and the processing steps must do their best with the values in hand. Such circumstances arise in processing survey data, in constructing knowledge bases or data warehouses  and in using some public or open data sets.
The goal of data cleaning, sometimes called data editing or integrity checking, is to improve the accuracy of each data record and by extension the quality of the data set as a whole. Generally, this is accomplished through deterministic processes that recode specific data points according to static rules based entirely on data from within the individual record. This traditional method works well for many purposes. However, when high levels of inconsistency exist within an individual respondent's data, classification scoring may provide better results.
Classification scoring is a two-stage process that makes use of information from more than the individual data record. In the first stage, population data is used to define a model, and in the second stage the model is applied to the individual record. The author and colleagues turned to a classification scoring method to resolve inconsistencies in a key value from a recent health survey. Drawing records from a pool of about 11,000 survey respondents for use in training, we defined a model and used it to classify the vital status of the survey subject, since in the case of proxy surveys, the subject of the study may be a different person from the respondent. The scoring model was tested on the next several months' receipts and then applied on a flow basis during the remainder of data collection to the scanned and interpreted forms for a total of 18,841 unique survey subjects. Classification results were confirmed through external means to further validate the approach. This paper provides methodology and algorithmic details and suggests when this type of cleaning process may be useful.