Back to Annual Meeting

Back to Annual Meeting

APHA Scientific Session and Event Listing

Recorded presentation

David Reichel, DC, MPH, Department of Health and Human Services, Health Statistics and Data Management Section, State of New Hampshire, 29 Hazen Drive, Concord, NH 03301, (603) 271-4652, dreichel@dhhs.state.nh.us

Statisticians are experts at sampling error, but rarely appreciate the unique problems public health epidemiologists face with administrative data. Since data for the entire population is available, there is apparently no error and no need for confidence intervals when calculating rates. Unfortunately, variability with small numbers and data quality issues makes interpreting this data more difficult than might be expected. Improper coding, missing data, errors in population estimates are just a few examples of data quality issues that produce error and uncertainty. Additionally, rare events in small populations tend to fluctuate from year to year making interpretation of rates and the underlying risk that theoretically produces them, problematic. Two events in a small rural county may result in the highest rate in the State for one year, and zero events the next year may result in the lowest State rate. This makes for interesting GIS plots of “hotspots”, but is otherwise useless and misleading. Confidence intervals help to quantify such variability. However, confidence intervals and statistical tests can sometimes cause more problems than they solve and these statistical procedures usually don't account for the additional uncertainty resulting from data quality issues. Once we have numbers we trust, what do we do with them? This may be the greatest source of uncertainty! If we appreciate uncertainty, even the uncertainty present with administrative data, it changes the way we think about data. We make better decisions about interpreting and using the data. We find ways to make data more valuable and effective.

Learning Objectives:

• Rates derived from non-sampled, complete, administrative data theoretically shouldn't require confidence intervals, but the way these data are frequently used to monitor trends and compare sub-populations or geographic areas according to theoretical underlying risk, makes confidence intervals a wise and proper choice.
• Confidence interval calculations are based on the random variability of the model and do not generally account for additional uncertainty caused by differential data quality errors.
• An appreciation of factors relating to uncertainty becomes especially important when dealing with rare events and small numbers.

Keywords: Data/Surveillance, Biostatistics

Presenting author's disclosure statement:

Any relevant financial relationships? No

Recorded presentation