CENTER FOR EDUCATION STATISTICS
OFFICE OF EDUCATIONAL RESEARCH AND IMPROVEMENT 
U.S. DEPARTMENT OF EDUCATION


                       
SUBJECT:   STANDARD FOR TREATMENT OF ITEM NONRESPONSE        
EFFECTIVE DATE: 03/16/87   CES STANDARD 87-05-01
                       
PURPOSE: To ensure consistency of data over time and to ensure internal consistency of data before analysis.      
                       
  o Best efforts should be made to minimize the occurrence of item nonresponse. Normally, adequate pre- and pilot testing will identify likely item nonresponse problems and protect against them.   Geniunely  
    unanticipated problems are to be dealt with post facto. The respondents should be contacted to complete the questionnaire before resorting to use of weighting or imputation.
                       
  o If more than thirty percent of the key items in a record are missing, consideration should be given to deleting the record and making a weighting adjustment for survey nonresponse, rather than imputing items.
                       
  o If less than thirty percent of the key items in a record are missing, the missing items should be imputed. The type of imputation will depend on the type of survey and the type of data.
                       
  o Any imputation on a data tape should be clearly indicated by a separate field with a flag set to indicate the data values which have been imputed.  
                       
  o For continuing or longitudinal surveys where categorical data have been gathered in the past on the same items to be imputed, missing values should be imputed directly from the previous survey.        
                       
  o For continuing or longitudinal surveys where continuous data have been gathered in the past on the same items to be imputed, a previous observation for the same unit is available, and no strong trends are perceived in the data set, missing values should be imputed directly from the previous survey. A strong trend is defined as one where a linear or curvilinear regression has at least one coefficient significantly different from zero at the 95 percent confidence level.    
                       
  o For continuing or longitudinal surveys where two or more rounds of continuous data have been gathered in the past on the same items to be imputed, at least one previous observation for the same item for this same unit is available, and strong trends are perceived in the overall data set for this item, missing values should be imputed by estimating the trend using linear regression on the data collected in the previous time period and data in the current time period. The regression would be run only on cases where observations were present in both time periods, and a separate regression would be run for each variable to be imputed.  
                       
  o For continuing or longitudinal surveys where continuous data are not available for more than 2 time periods, but a previous observation for the same unit is available, a ratio factor can be computed from those cases where observations are available in both time periods. This can be used to inflate or deflate prior observations to provide an imputed value for missing data. Alternatively, if an inflation factor is known from outside sources (like CPI), it can also be used. Ratios computed from the data are preferable.    
                       
  o For longitudinal surveys where previous observations are not available, or for one-time surveys, missing items should be imputed using the best practical estimator available. Normally, this means using hot deck, typically nearest neighbor, or other techniques that, for each imputation case, assign a donor case that is as similar as possible in terms of the characteristics being used to impute.
                   
  o Zero-filling a record because of missing data, or zero filling certain items which are missing, should only be used as a last resort, and reasons for these actions should be documented with the data tape.  
                   
RELATED STANDARDS:              
                   
  o CES Standard for Planning Statistical Surveys, 87-01-01        
  o CES Standard for Testing Data Collection Instruments, 87-03-01 
o CES Standard for Minimizing Survey Nonresponse, 87-03-03
     
                   
RELATED CHECKLISTS AND FORMS:            
                   
  o Kalton, Graham, and Kasprzyk, "Imputing for Missing Survey Responses, 1982 Proceedings, American Statistical Association, p. 22-31.