Evaluated by C. Poirier, Statistics Canada, 2003
Last updated in January, 2010
SYSTEM INFORMATION
Full name:  BANFF  Generalized Edit and Imputation System 
Version:  2.3 
Year:  2008 
Developer:  Statistics Canada 
DESCRIPTION
The Banff functions are close to the GEIS ones (see the evaluation of GEIS), but the system works in a SAS environment rather than Oracle. Banff works on any platform where SAS is available, as opposed to GEIS which worked only in Unix or mainframe. An interesting improvement in Banff is its modular aspects where all the functions are independent from each another.
Banff is not an editing system as such but targets more the imputation process. It is usually used after preliminary editing associated with the collection and capture phases and respondent followup have been completed. Linear programming techniques are used to conduct the localization of fields to be imputed and search algorithms are used to perform automatic imputations. The processing is entirely driven by edit/imputation linear rules defined by means of numeric variables. More details are given in Statistics Canada (2009). Banff steps are:
Edit specification and analysis: This step serves to identify the relationships which characterize acceptable records. The relationships are expressed as a set of n linear edit rules in the form:
a_{11}x_{1}+ a_{12}x_{2}+ ... + a_{1m}x_{m} <= b_{1}
. . .
a_{n1}x_{1}+ a_{n2}x_{2}+ ... + a_{nm}x_{m} <= b_{n}
where the a_{ij}'s and b_{i}'s are userdefined constants, and the x_{j}'s represent the m survey variables. The rules are connected with logical 'and's, which means each rule must be satisfied for a record to pass the edits. The system checks for edit consistency, redundancy and hidden equalities. This step permits an iterative approach to the design of the best possible set of edits.
Outlier detection: This step aims at the detection of univariate outliers. It performs comparisons of selected variables across records and identifies outlying observations based on the median M, and the first and third quartiles Q_{1} and Q_{3} of the population. An observed value x will be identified as an outlier if it is outside the acceptance interval (MkQ_{1}, M+kQ_{3}), where k is set by the user. This method can be used to identify variables to be imputed or to be excluded from subsequent calculations.
Error localization: The error localization uses a linear programming approach to minimize the number of fields requiring imputation. This is an application of the rule of minimum change. The step identifies the fields that need to be imputed in order for the record to pass all the edit rules. The problem is expressed as a constrained linear program and solved using Chernikova's algorithm. The system also allows the use of weights for each variable when the user wishes to exert some influence on the identification of the fields to be imputed. Although the algorithm is costly to run, it constitutes one of the main features of Banff.
Automatic imputation: There are five imputation functions offered: Deterministic, Donor, Estimators, Mass Imputation and Prorating. The imputation function three imputation methods: Deterministic, Donor, and Estimators. Based on the edit rules, the deterministic imputation identifies cases in which there is only one possible solution that would allow the record to satisfy the rules. The donor imputation replaces the values to be imputed using data from the closest valid record, also referred to as the nearest neighbour. For a given record, a subset of the fields which do not need imputation are automatically used as matching fields, and the maximum standardized difference among these individual fields is used as the distance function. The user can specify postimputation edits to make sure the nearest neighbour is close enough to be used as a donor. The imputation by estimators provides a wide set of techniques using historical or current information. Builtin estimators are: Previous values, previous/current means, trends, and multiple regressions. If a nonstandard estimator is required, a userdefined estimator can also be specified. Mass imputation is a special case of donor imputation where the variables to impute are always the same for each record. The required conditions are such that records will pass the edits no matter what values are imputed. The prorating function will adjust components of a sum so that it matches the total of a linear edit with an equality sign.
Banff allows the use of different imputation techniques across questionnaire sections and subpopulations. The use of a sequence of techniques is also possible where, at each step, the user can include/exclude previously imputed data in the process. The system works with SAS as each of the nine Banff functions is a customized SAS procedure. It is also available in SAS Enterprise Guide as Banff tasks. Those tasks help the user in specifying the parameters and edit rules using the SAS Enterprise Guide interface. Banff was developed in C language. The functionality described above is quite adequate for economic surveys as Banff works only with numeric data. Newly initiated developments will allow the use of metadata. The users will define the parameters in a spreadsheet or XML and the system will automatically generate all SAS code to perform edit and imputation. Programming knowledge will no longer be a requirement for the users.
STRENGTHS
The strengths of Banff are its capacity to find minimum changes for any set of rules being expressed as a series of linear equations, its automated donor imputation function driven by the edit rules, and its flexibility within a SAS environment. This imputation function runs with almost no intervention from the user since it derives the matching fields by itself. It simply uses the response pattern, whatever it is, to look for a donor. The minimum change rule contributes to increase the chance of preserving a relatively good data integrity given the data in error. Banff can deal with both positive and negative values. The flexible estimator module of Banff, the diagnostic reports and the online tutorial, coupled with a continuous user support constitute the desirable aspects of the system.
WEAKNESSES
Banff only deals with numeric variables. Altough the imputation module can now process negative values, the editing process still cannot. This may cause problems to financial surveys. The system now requires SAS to be installed on the platform of interest.
FUNCTIONAL EVALUATION
LEGEND  
***  The implementation offers subfunctions or options being required by a wide range of survey applications.  
**  The implementation have a less complete set of options.  
*  The implementation offers a partial functionality. Options are too restrictive or not generalized enough.  
  No stars are assigned when the functionality is not offered at all. 

TYPE OF DATA 


Quantitative data  *** 

Qualitative data  * 

EDITING FUNCTIONS 


Data verification  * 

Online correction   

Error localization 


Minimum changes 


Userdefined changes   

Outlier detection 


IMPUTATION FUNCTIONS 


Deterministic imputation 


Donor imputation 


Imputation by estimators 


Multiple imputation   

GENERAL FEATURES 


Graphical user interface 


Userfriendliness  ** 

Online help   

Online tutorial   

Documentation 


Diagnostic reports 


Integration   

Reusable code 


Portability 


Flexibility 


User support 


Acquisition cost  30,000 USD 

REFERENCES
Statistics Canada (2009). "Functional Description of Banff  the Generalized Edit and Imputation System". Statistics Canada Technical Report.