OBJECTIVES: To illustrate a simple approach to adjusting for bias due to drop-outs (i.e., attrition bias) when evaluating the effect of a certain therapy in HIV clinical trials using the mean change in plasma viral load. To evaluate its validity and to compare its performance with that of another simple method for handling drop-outs: the last observation carried forward (LOCF) method. DESIGN: Data from a notional treated group of 100 patients followed up to 52 weeks were generated. Attrition bias was introduced by mimicking selective patient drop-out (i.e., more likely in patients doing badly). METHODS: The difference between the true mean change in HIV RNA levels at 52 weeks and the observed mean change because of drop-outs was calculated (attrition bias). The reduction in bias obtained by using the proposed approach was then calculated and compared with that obtained by using the LOCF method. To assess the performance of the methods over the entire follow-up, the mean areas under the curves were considered. RESULTS: Our method reduced the bias by a clinically relevant amount in a variety of different settings. In most of our simulations, bias was reduced by a larger amount than that obtainable from using the LOCF method. CONCLUSIONS: The current situation is that results from trials in HIV infection are invariably presented with no associated attempt to quantify the attrition bias present. Attrition-adjusted plots of mean change in HIV RNA should, we believe, be presented alongside usual plots as a form of sensitivity analysis.
OBJECTIVES: To illustrate a simple approach to adjusting for bias due to drop-outs (i.e., attrition bias) when evaluating the effect of a certain therapy in HIV clinical trials using the mean change in plasma viral load. To evaluate its validity and to compare its performance with that of another simple method for handling drop-outs: the last observation carried forward (LOCF) method. DESIGN: Data from a notional treated group of 100 patients followed up to 52 weeks were generated. Attrition bias was introduced by mimicking selective patient drop-out (i.e., more likely in patients doing badly). METHODS: The difference between the true mean change in HIV RNA levels at 52 weeks and the observed mean change because of drop-outs was calculated (attrition bias). The reduction in bias obtained by using the proposed approach was then calculated and compared with that obtained by using the LOCF method. To assess the performance of the methods over the entire follow-up, the mean areas under the curves were considered. RESULTS: Our method reduced the bias by a clinically relevant amount in a variety of different settings. In most of our simulations, bias was reduced by a larger amount than that obtainable from using the LOCF method. CONCLUSIONS: The current situation is that results from trials in HIV infection are invariably presented with no associated attempt to quantify the attrition bias present. Attrition-adjusted plots of mean change in HIV RNA should, we believe, be presented alongside usual plots as a form of sensitivity analysis.