From: Evaluation of regression methods when immunological measurements are constrained by detection limits
Methods | Description | Software | Disadvantage |
---|---|---|---|
Deletion | Remove NDs. | Any statistical package | Bias |
DL/2 | Substitute each ND with half of the value of DL. | Any statistical package | Large RMSE for large proportion of NDs |
ROS | After computing a linear regression for data versus their normalized scores below-DL values are extrapolated under distributional assumption. | Underestimation of variance for large proportion of NDs | |
MI | Estimation of mean and standard deviation by MLE. Creating 10 complete samples. Pool the results from 10 individual analyses. | R (software available on request) | Bias for small proportion of NDs |
TOBIT | Parametric estimation method for incorporating NDs. | Sensitive to heteroscedastic errors | |
LOGIT | Create binary dependent variable of NDs (0s) and detects (1s). | Any statistical package | Loss of information & parameter estimates are less interpretable |