Study/test changes in relative size influenced accuracy only when targets were displayed in large test size and when target size changed between study and test; accuracy was higher for objects displayed in large test size when target size was the same (0.76) at study and test than when it was different (0.58).
If the prediction error does not increase significantly, the larger test set size is recommended for a less variable estimator of the prediction error.
In this simulation study, the variability of the error estimates changed considerably for test data set sizes up to n test = 7 and then changed only slightly for larger test set sizes.
It decayed quickly for larger test set sizes.
If the ascent is mild (or if there is even a plateau), larger test set sizes should be used to estimate the prediction error since variability often decreases dramatically for larger test data set sizes in the outer loop.
A follow-up ANOVA in the large target test size conditions with study/test relative size condition as a within-subject factor was significant, F 2,126) = 79.67, MSe = 0.02, η2 = .56.56
Thus, the variability of the error estimates in the outer loop (ave.vb (PE )) decreased with larger test data set sizes owing to less variable test data.
Large test data set sizes yielded highly fluctuating error estimates in the outer loop due to higher model uncertainty.
Lots of managers run sequential tests — e.g., testing size first (large versus small), then testing color (blue versus red), then testing typeface (Times versus Arial) — because they believe they shouldn't vary two or more factors at the same time.
In univariate analysis, using the log rank test, larger tumor size (p = 0.0004), nodal involvement (p = 0.0036), SBR grade (p = 0.0028), lack of estrogen and progesterone receptor expression (p = 0.0123 and 0.0132 respectively), CCR6 expression (p = 0.0316) and infiltration by CCL19/MIP3β-expressing DC (p = 0.0417) were associated with shorter relapse-free survival (RFS).
