What Is Information Bias In Epidemiology?
Statistical Design of Self-Reporting Instruments for Medical Research
Random sampling should not be confused with measurement error variability. Statistical methods can address the variability of the data, but they don't account for uncertainty due to measurement error. Careful planning is needed in each step of the research design to minimize or eliminate bias.
Rules and procedures should be followed when designing self- reporting instruments. Training interviewers is important to minimize bias. Measurement error can be difficult to eliminate since measuring devices and algorithms are often imperfect.
Before using a measuring instrument for data collection, it is recommended to revise the level of accuracy. Researchers should be aware of the sources of bias in their results and the effort is needed to minimize the effects of bias. The medical undergraduate level should be the first to learn about the possible drawbacks and pitfalls of decision making that can result in bias.
Effects of misclassification on statistical analyses
When the degree of misclassification of exposure status is different, it's called differential misclassification. Bias will always be caused by differential misclassification. If no datare available, a judgement of the direction of bias will be subjective, and based on knowledge of the potential degree and direction of bias.
J- or U-shaped associations are well-known in the literature on alcohol and cardiovascular disease 16 where non-drinkers have been shown to be at increased risk compared to those with low exposure. If the rationale for categorization is not explained, it is not uncommon for a variable to be categorized prior to performing statistical analyses. The age may be a variable that is not continuous.
Simulation studies suggest that if the cutting point is from the median of the variable, the greater the increase in the measure of association, which in turn causes confounders 21. All the above measurements are prone to being misclassified. Bleeding episodes in early pregnancy may be wrongly interpreted as menstruation, and often there areIrregularities in cycle length and timing of ovulation, which are not taken into account.
The measurement error seems to be constant when calculated as a percentage of the fetal dimensions or when reported as a Z-score 24, but it appears to be hampered by both the interobserver variability and the increase in the number of days that have passed. If the measures are performed in the second or third trimester, they may be misclassified because of decelerated or accelerated growth or a genetic predisposition to being small or large for a child. Simulation studies have shown that even moderate mismeasurement of a potential confounder may bias the estimates, regardless of whether the exposure, confounder or outcome are measured on continuous scales or categorized 27.
Mismeasurement of a confounder, which is a strong risk factor for the outcome and by definition associated with the exposure, will lead to underestimation of the true risk associated with the confounder. The bias 24 is affected by the degree of mis-measurement. The risk estimate may be altered by up to 50% 27 if the measurement error is less than 15%.
Nondifferential Misclassification in Medical Records
If there are multiple exposure categories, nondifferential misclassification may bias the estimate towards the null or away from it, depending on the categories into which subjects are misclassified. There are a number of ways in which nondifferential misclassification can occur. In a medical record, no healthcare workers remember to ask about tobacco use, so the record may be incomplete.
There may be errors in recording or interpreting information in records, or there may be errors in assigning codes to disease diagnoses by clerical workers who are unfamiliar with a patient's hospital course, diagnosis, and treatment. People completing questionnaires may have difficulty remembering past exposures. If the difficulty in remembering past exposures is the same in both groups, then there will be nondifferential misclassification, which will bias towards the null.