Protecting Privacy of Shared Epidemiologic Data without Compromising Analysis Potential
Table 2
Results of fitting a linear dose response using binary regression.
Dose masking scheme
ERRa (per Gray)
Standard error
Deviance
LR statistic ( value)
Relative bias (%)
MSE
None
0.5235
0.1548
826.36
9.27 (0.0023)
β
0.0240
Rounded to three decimal digits
0.5237
0.1548
826.35
9.28 (0.0023)
0.038
0.0240
Rounded to two decimal digits
0.5235
0.1547
826.35
9.28 (0.0023)
0
0.0239
Rounded to nearest centiGray
0.5235
0.1548
826.36
9.27 (0.0023)
0
0.0240
Rounded to nearest deciGray
0.5228
0.1547
826.36
9.27 (0.0023)
0.13
0.0239
Stratifiedb
0.5320
0.1553
826.11
9.52 (0.0020)
1.6
0.0242
Randomizedc
βΒ± 0.001
0.5235
0.1548
826.36
9.27
0.015
0.0240
ββ(min, max)
(0.5234, 0.5238)
(0.1548, 0.1549)
(0.0023)
(0, 0.057)
(0.02397, 0.02398)
βΒ± 0.01
0.5239
0.155
826.36
9.28
0.16
0.0240
ββ(min, max)
(0.5226, 0.5266)
(0.1548,0.1551)
(0.0023)
(0, 0.59)
(0.02396, 0.02407)
βΒ± 0.1
0.5271
0.155
826.33
9.31
1.6
0.0243
ββ(min, max)
(0.5159, 0.5573)
(0.1537,0.1584)
(0.0023)
(1.4, 6.5)
(0.02415, 0.02557)
aERR: excess relative risk (relative riskβ1). Precision is overrepresented for comparison.
bDoses were stratified according to the categories used in Life Span Study Report 13 [27]. The dose value assigned to each individual was the mean of all database AHS stomach dose values in that group.
cA random uniform deviate between the specified range was added to the dose; if this operation resulted in a negative value, the masked dose was set to zero. Results are the averages from 500 simulations.