The Vermont-Oxford Trials Network is a voluntary collaborative research group of neonatologists
that maintains a database for very low birthweight infants (501-1500 g). The database
(1) provides core data for randomized trials, (2) serves as a resource for outcomes
research in neonatology, and (3) generates quality management reports for participating
sites. To assess the reliability of this database and to determine the sources of
error, we reviewed 635 medical records chosen at random from among the 4341 eligible
infants born at 40 participating data generating sites during an 18-month period beginning
January 1, 1990. The estimated frequencies of disagreement between the medical record
and database for each of the 10 data items studied and the standard errors of the
estimates (in parentheses) were: date of birth 1.3% (0.4), date of admission 2.5%
(0.6), date of discharge 8.8% (1.0), birthweight (difference > 50 g) 2.9% (0.6), location
of birth (inborn or outborn) 2.1% (0.5), multiple birth 2.2% (0.5), cesarean section
2.5% (0.6), gender 2.1% (0.5), status 28 days after birth 3.4% (0.6), final status
2.9% (0.6). The overall proportions and mean values for items in the database were
close to the estimated values based on the random sample of records. There were a
total of 247 disagreements between the database and the medical records in the sample.
Twenty-three were due to data keying errors. Two hundred twenty-four were due to errors
in transcription or interpretation. The rate of data keying errors decreased from
over 50 errors per 10,000 fields to less than 15 errors per 10,000 fields when specific
quality control procedures, including visual inspection, were instituted. Data keying
errors accounted for 13.7% of all disagreements between the database and medical record
before improved data entry methods were introduced, and only 3.7% of all errors after
they were introduced. We concluded that the Vermont-Oxford Trials Network Database
is reliable. Data keying errors have been reduced by the introduction of additional
quality control measures. Further reductions in database errors will require measures
aimed at minimizing transcription or interpretation errors by individuals completing
the data forms.