Table 3.

Results of logistic regression analyses comparing AUCs for different rules (test data)

Independent variableCoefficient estimate [Average (interquartile range*)]Z statistic [No. runs with P < 0.05]
A: TPSA compared with logic rules
Test type
    TPSABaseline
    Logic0.21 (0.14, 0.28)3
Time from test to diagnosis (y)−0.11 (−0.14, −0.08)21
Age at time of test
    ≤60Baseline
    >60−0.13 (−0.27, 0.02)4
B: TPSA compared with CPSA
Test type
    TPSABaseline
    CPSA−0.11 (−0.14, −0.08)21
Time from test to diagnosis (y)−0.11 (−0.14, −0.08)20
Age at time of test
    ≤60Baseline
    >60−0.21 (−0.37, −0.05)8
Test type × age0.11 (0.07, 0.15)14
C: CPSA compared with logic rules
Test type
    CPSABaseline
    Logic0.27 (0.20, 0.35)10
Time from test to diagnosis (y)−0.11 (−0.15, −0.08)21
Age at time of test
    ≤60Baseline
    >60−0.07 (−0.21, 0.08)2
  • NOTE: Results from 35 test-train splits are presented. SDs for coefficient estimates are estimated based on 250 bootstrap samples. A positive coefficient estimate indicates that an increase in the independent variable is associated with an increase in the AUC (i.e., an improvement in diagnostic performance). The coefficient values for the test type variable are interpretable as follows: Exp(coefficient) gives the amount by which the AUC odds [AUC/(1 − AUC)] are increased for the given test type relative to the baseline test type (14). Interaction terms in A and C were rarely significant, so only models with main effects are presented. A positive (negative) coefficient implies that increasing values of the corresponding covariate are associated with better (poorer) diagnostic performance.

  • * The elements of the tuple are the 25th and 75th quintiles, respectively.