By univariate analysis, MLN was significantly associated with melanoma-specific survival, whether modeled as a continuous variable (HR 1.82 for each micrometer increase;
P = 0.016) or a categorical one (HR 1.57 for each category increase, by tertile;
P = 0.007;
Table 3 ).
The second observer who graded a subset of 63 tumors drew conclusions qualitatively and quantitatively similar to those of the first observer, based on the same set (hazard ratio, 2.0 vs. 1.6; Wald χ2 5.29 vs. 4.80, P = 0.021 vs. 0.029, respectively). In addition, four of five of the bivariate associations reported for the entire data set of 126 patients (described later) were also identified by using data of either observer for the 63 patients.
The presence of ciliary body involvement; large LBD; presence of epithelioid cells, high grade of pigmentation; presence of microvascular loops and networks, as modeled by assuming networks to be an advanced stage of loops (HR 1.80 for each category increase;
P < 0.001); and high MVD (HR 1.02 for each unit increase in square-root–transformed count;
P < 0.001) were also significantly associated with melanoma-specific survival
(Table 3) .
MLN remained an independent predictor of prognosis, both as a continuous and categorical variable when adjusted in turn for the effect of ciliary body involvement, LBD, presence of epithelioid cells, and microvascular loops and networks
(Table 3) . When adjusted for MVD, it was of borderline significance as a continuous variable (
P = 0.11;
Table 3 ). Of the five bivariate models tested, those that combined MLN with cell type and MVD best predicted melanoma-specific survival, whether MLN was modeled as a continuous or categorical variable (e.g., model 3A vs. 4A, difference in −2 log likelihood 465.7 − 448.9 = 16.8, 1
df,
P < 0.001 χ
2 test with Bonferroni adjustment for nine possible comparisons between the five bivariate models in
Table 3 ). Of these two models, the one that included MVD was the better predictor (model 5A vs. 3A, 448.9 – 438.7 = 10.2, 1
df,
P = 0.013 with Bonferroni correction). The model that combined MLN with microvascular loops and networks was significantly better than those that combined MLN with ciliary body involvement and LBD (e.g., model 4A vs. 2A, 494.1 – 465.7 = 28.4, 1
df,
P < 0.001 with Bonferroni correction).
When the four best predictors were combined in trivariate models (see
Table 3 for models that included MVD), MLN lost statistical significance when modeled as a continuous variable (
P = 0.11–0.18) but retained significance as a categorical one (
P = 0.023–0.067). No model was notably better than the others. Combined in a single model, all four variables remained independent predictors of prognosis when MLN was modeled as a categorical variable (
P = 0.084, model 8B), but as a continuous one, MLN was not statistically significant (
P = 0.28, model 8A). Compared with the best trivariate models, the final models predicted survival significantly better (e.g., model 8B vs. 7B, 416.8 – 402.4 = 14.4, 1
df,
P < 0.001; Bonferroni adjustment for three possible comparisons).