Bayesian Solutions to Missing Data Problems in a Consumer Payments Survey

• Activity Number: 121
• Type: Topic Contributed
• Date/Time: Monday, August 5, 2013 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Christopher Henry*+ and Kyle Vincent
• Companies: Bank of Canada and Bank of Canada
• Keywords: survey ; Bayesian statistics ; missing data ; imputation ; MCMC ; payments

Abstract

In this paper we utilize Bayesian methods to solve a unique missing data problem in the Canadian Financial Monitor (CFM) survey. The CFM is a comprehensive annual survey of the financial attributes of Canadian households, which includes a module on the use of emerging payment methods for retail transactions, e.g. credit cards with a 'contactless' feature. From 2009 to 2010 there was a change in the survey design which made it impossible to distinguish between missing and negative responses to questions regarding the extensive margin of use for these payment methods. This resulted in item non-response rates of 40%-80% for these and related questions in 2010 and 2011. By specifying a probability distribution for the complete data, we first implement a data augmentation routine to impute missing values in the 2009 data. This approach allows us to formulate a prior distribution which we use to solve the more extensive missing data problems in the 2010 and 2011 surveys. We compare our method to other well know imputation procedures by intentionally deleting observed data at random, imputing based on the candidate routines, and comparing the results via the mean squared error statistic.