Validation Relaxation: A Quality Assurance Strategy for Electronic Data Collection.

BACKGROUND: The use of mobile devices for data collection in developing world settings is becoming increasingly common and may offer advantages in data collection quality and efficiency relative to paper-based methods. However, mobile data collection systems can hamper many standard quality assurance techniques due to the lack of a hardcopy backup of data. Consequently, mobile health data collection platforms have the potential to generate datasets that appear valid, but are susceptible to unidentified database design flaws, areas of miscomprehension by enumerators, and data recording errors.

OBJECTIVE: We describe the design and evaluation of a strategy for estimating data error rates and assessing enumerator performance during electronic data collection, which we term "validation relaxation." Validation relaxation involves the intentional omission of data validation features for select questions to allow for data recording errors to be committed, detected, and monitored.

METHODS: We analyzed data collected during a cluster sample population survey in rural Liberia using an electronic data collection system (Open Data Kit). We first developed a classification scheme for types of detectable errors and validation alterations required to detect them. We then implemented the following validation relaxation techniques to enable data error conduct and detection: intentional redundancy, removal of "required" constraint, and illogical response combinations. This allowed for up to 11 identifiable errors to be made per survey. The error rate was defined as the total number of errors committed divided by the number of potential errors. We summarized crude error rates and estimated changes in error rates over time for both individuals and the entire program using logistic regression.

RESULTS: The aggregate error rate was 1.60% (125/7817). Error rates did not differ significantly between enumerators (P=.51), but decreased for the cohort with increasing days of application use, from 2.3% at survey start (95% CI 1.8%-2.8%) to 0.6% at day 45 (95% CI 0.3%-0.9%; OR=0.969; P<.001). The highest error rate (84/618, 13.6%) occurred for an intentional redundancy question for a birthdate field, which was repeated in separate sections of the survey. We found low error rates (0.0% to 3.1%) for all other possible errors.

CONCLUSIONS: A strategy of removing validation rules on electronic data capture platforms can be used to create a set of detectable data errors, which can subsequently be used to assess group and individual enumerator error rates, their trends over time, and categories of data collection that require further training or additional quality control measures. This strategy may be particularly useful for identifying individual enumerators or systematic data errors that are responsive to enumerator training and is best applied to questions for which errors cannot be prevented through training or software design alone. Validation relaxation should be considered as a component of a holistic data quality assurance strategy.