Computing values for viral load below test detection limit and comparing its effect on measures of central tendency and outcome (medication error)

Introduction: Researchers are often required to analyze data where levels of a substance are below the detection limit (DL) of the test. We analyzed the distribution of viral load (VL) using six methods and compared the effect on measures of central tendency and outcome using the original dataset and a multiply imputed (MI) dataset from a sample of 149 HIV+ patients admitted to UMass Memorial Medical Center. Method: We used deletion, substitution with 0, substitution with DL/2, random number generation, substitution with DL, and binary variable (for outcome only) methods to test the effect on measures of central tendency and outcome using the original dataset and MI dataset. We measured central tendency by calculating the mean, median, standard deviation, standard error, and skewness, and the outcome by calculating the intercept and standard error. Results: Of VL values, 85% were below DL. Deletion method affected measures of central tendency and deletion and binary variable methods affected the outcome the most. The remaining methods were similar in all measured aspects of central tendency except median. The median took on the substituted value in all methods except deletion and random number generation. There was no difference in measures when comparing the original dataset and the MI dataset. Conclusion: Except deletion, all other methods provide a suitable estimate of the data. In our dataset, 85% of VL values were below DL so there was minimal variability in the results. Similar procedures need to be studied in datasets with fewer values below DL.