Preventive Actions
The best way to create high quality data is to prevent poor quality data from entering an organization. Preventive actions stop known errors from occurring. Inspecting data after it is in production will not improve its quality. Approaches include:
- Establish Data Entry Controls: Create data entry rules that prevent invalid or inaccurate data from entering a system.
- Train Data Producers: Ensure staff in upstream systems understand the impact of their data on downstream users. Give incentives or base evaluations on data accuracy and completeness, rather than just speed.
- Define and Enforce Rules: Create a ‘Data Firewall’, which has a table with all the business data quality rules used to check if the quality of data is good, before being used in an application such a data warehouse. A data firewall can inspect the level of quality of data processed by an application, and if the level of quality is below acceptable levels, analysts can be informed about the problem.
- Demand High Quality Data from Data Suppliers: Examine an external data provider’s processes to check their structures, definitions, and data source(s) and data provenance. Doing so enables assessment of how well their data will integrate and helps prevent the use of non-authoritative data or data acquired without permission from the owner.
- Implement Data Governance and Stewardship: Ensure roles and responsibilities are defined that describe and enforce rules of engagement, decision rights, and accountabilities for effective management of data and information assets (McGilvray, 2008). Work with data stewards to revise the process of, and mechanisms for, generating, sending, and receiving data.
- Institute Formal Change Control: Ensure all changes to stored data are defined and tested before being implemented. Prevent changes directly to data outside of normal processing by establishing gating processes.
Corrective Actions
Corrective actions are implemented after a problem has occurred and been detected. Data quality issues should be addressed systemically and at their root causes to minimize the costs and risks of corrective actions. ‘Solve the problem where it happens’ is the best practice in Data Quality Management. This generally means that corrective actions should include preventing recurrence of the causes of the quality problems. Perform Data Correction in three General Ways:
- Automated Correction: Automated correction techniques include rule-based standardization, normalization, and correction. The modified values are obtained or generated and committed without manual intervention. An example is automated address correction, which submits delivery addresses to an address standardizer that conforms and corrects delivery addresses using rules, parsing, standardization, and reference tables. Automated correction requires an environment with well-defined standards, commonly accepted rules, and known error patterns. The amount of automated correction can be reduced over time if this environment is well-managed and corrected data is shared with upstream systems.
- Manually-Directed Correction: Use automated tools to remediate and correct data but require manual review before committing the corrections to persistent storage. Apply name and address remediation, identity resolution, and pattern-based corrections automatically, and use some scoring mechanism to propose a level of confidence in the correction. Corrections with scores above a particular level of confidence may be committed without review, but corrections with scores below the level of confidence are presented to the data steward for review and approval. Commit all approved corrections, and review those not approved to understand whether to adjust the applied underlying rules. Environments in which sensitive data sets require human oversight (e.g., MDM) are good examples of where manual-directed correction may be suited.
- Manual Correction: Sometimes manual correction is the only option in the absence of tools or automation or if it is determined that the change is better handled through human oversight. Manual corrections are best done through an interface with controls and edits, which provide an audit trail for changes. The alternative of making corrections and committing the updated records directly in production environments is extremely risky. Avoid using this method.
Statistical Process Control
Statistical Process Control (SPC) is a method to manage processes by analyzing measurements of variation in process inputs, outputs, or steps. The technique was developed in the manufacturing sector in the 1920s and has been applied in other industries, in improvement methodologies such as Six Sigma, and in Data Quality Management. Simply defined, a process is a series of steps executed to turn inputs into outputs. SPC is based on the assumption that when a process with consistent inputs is executed consistently, it will produce consistent outputs. It uses measures of central tendency (how values cluster around a central value, such as a mean, median, or mode) and of variability around a central value (e.g., range, variance, standard deviation), to establish tolerances for variation within a process.
SPC is used for control, detection, and improvement. The first step is to measure the process in order to identify and eliminate special causes. This activity establishes the control state of the process. Next is to put in place measurements to detect unexpected variation as soon as it is detectable. Early detection of problems simplifies investigation of their root causes. Measurements of the process can also be used to reduce the unwanted effects of common causes of variation, allowing for increased efficiency.
Root Cause Analysis
A root cause of a problem is a factor that, if eliminated, would remove the problem itself. Root cause analysis is a process of understanding factors that contribute to problems and the ways they contribute. Its purpose is to identify underlying conditions that, if eliminated, would mean problems would disappear.
Common techniques for root cause analysis include Pareto analysis (the 80/20 rule), fishbone diagram analysis, track and trace, process analysis, and the Five Whys (McGilvray, 2008).
