The estimation of blood platelet survival. VI. Evaluation of the graphical method

An experiment was performed in which 13 scientists experienced in analyzing data on platelet survival were sent blind duplicate sets of ten platelet survival curves and requested to fit the "best" straight line through the data by eye and determine its point of intersection with the time axis. The duplicate sets labeled "control" and "aspirin" respectively were artificial data generated from evaluation for various times of a gamma curve of order five and adding random independent normal (Gaussian) errors having a standard deviation of 0.03. The results were analyzed to examine inter- and intra-observer variance, reproducibility and bias. The range of means of estimated survival was 8.68 days to 11.59 days and in the control data showed about 63 times as much variance among observers as that due to random error of measurement alone. The "aspirin" data showed similar results. Correlation coefficients within observers for the pairs of estimates varied from -0.307 to +0.920. Three observers showed strong systematic differences between the duplicate readings. These results provide clear evidence of the need for formal, rather than graphical, analysis of the data.