Table 2.
Bias in hypothetical study results due to errors in algorithms identifying and assigning dates to second breast cancer events
| Scenario | Description | Mean Estimated hazard ratio | Percent bias* |
|---|---|---|---|
| 1 | Exposure† is unrelated to errors in outcome classification and date assignment Date error in unexposed and exposed (years): mean=0, SD=1 Example: Exposure is a genetic factor, such as CYP2D6 genotype |
1.47 |
−2.2% |
| 2 | SBCE are detected later in exposed group Date error in unexposed (years): mean=−0.25, SD=1; exposed: mean=0.25, SD=1) Example: Exposure is annual vs. biennial follow-up visits, where all SBCEs are eventually detected, albeit at different times. |
1.51 |
0.4% |
| 3 | SBCE are detected earlier and with more precision in the exposed group Date error in unexposed (years): mean=0.25, SD=1.25; exposed: mean=−0.25, SD=0.75) Example: People who are exposed (e.g., medication users) visit their providers more frequently |
1.49 |
−0.9% |
| 4 | SBCE are detected later and with less precision in the exposed group Date error in unexposed (years): mean=−0.25, SD=0.75; exposed: mean=0.25, SD=1.25) Example: People who are exposed (e.g., high body mass index) visit their providers less frequently |
1.44 | −3.8% |
| 5 | SBCE sensitivity is higher and specificity is lower in the exposed group vs. unexposed, but no difference in date error Classification accuracy in unexposed: sensitivity = 0.86, specificity = 0.99; exposed: sensitivity = 0.91, specificity = 0.982 Date error in unexposed and exposed (years): mean=0, SD=1 Example: Exposure is provider type and certain providers tend to give more diagnosis codes than others |
1.60 | 6.5% |
| 6 | SBCE sensitivity is higher and specificity is lower in the exposed group vs. unexposed, and SBCE are detected earlier and with more precision in the exposed group Classification accuracy in unexposed: sensitivity = 0.86, specificity = 0.99; exposed: sensitivity = 0.91, specificity = 0.982 Date error in unexposed (years): mean=0.25, SD=1.25; exposed: mean=−0.25, SD=0.75 Example: People who are exposed (e.g., medication users) come in more frequently and get more procedure and diagnosis codes |
1.62 | 8.1% |
SCBE: second breast cancer events; SD: standard deviation
*
Relative to true hazard ratio of 1.5
†
In all scenarios, exposure was binary and non-time-varying