Related Experiment Videos
Imputation of data values that are less than a detection limit
Paul A Succop1, Scott Clark, Mei Chen
1University of Cincinnati, Cincinnati, Ohio 45267-0056, USA. Paul.Succop@UC.EDU, psuccop@fuse.net
Journal of Occupational and Environmental Hygiene
|July 9, 2004
Summary
A new data imputation method accurately estimates lead dust loadings below detection limits. This method is superior to common practices that can significantly overestimate lead exposure levels.
Area of Science:
- Environmental Science
- Analytical Chemistry
- Occupational Health
Background:
- Analytical laboratories often report results below detection limits as "less than a specified value," creating left-censored data.
- Approximately 37% of household dust lead loadings in a large study were below the method detection limit, rendering them unusable for statistical analysis.
- Accurate estimation of these censored values through data imputation is crucial for reliable exposure assessment.
Purpose of the Study:
- To evaluate a novel data imputation procedure for estimating dust lead loadings below the method detection limit.
- To compare the performance of the new imputation method against commonly used techniques, including imputing the minimum detectable level divided by the square root of 2.
- To assess the bias and correlation of imputed values with actual reported dust lead loadings.
Main Methods:
- A new imputation procedure was developed, substituting values associated with the median percentile below each laboratory's method detection limit.
- The performance of this new method was compared to imputing the minimum detectable level by the square root of 2.
- Structural equation models were also used for an alternative imputation approach, with predictions centered to match censored data.
Main Results:
- The new imputation procedure yielded a correlation of r = 0.50 between predicted and reported dust lead loadings, with only 2.9% bias.
- Common imputation methods (e.g., minimum detection limit divided by the square root of 2) resulted in significant overestimation of dust lead loadings by 348%.
- An estimator combining the new procedure and structural equation models slightly improved the correlation to r = 0.51.
Conclusions:
- The newly formulated data imputation procedure is a more accurate and less biased method for estimating dust lead loadings below detection limits compared to common practices.
- Relying on common imputation methods can lead to substantial overestimation of environmental lead exposure.
- Analytical laboratories are encouraged to report all numerical results, flagging those below detection limits, as these may be more accurate than imputed values.