A statistically significant relationship was found at the
block group level between risk for NHL and dioxin exposure,
in both a univariate analysis (p-value = 10-5), and a
multivariate analysis (p-value = 0.04) (Table 2). The best
goodness-of-fit was obtained with a model including
three confounding factors (departments, log(industry ×
years), and spline-smoothed log(population density)
with 1.963 effective degrees of freedom. In the first rounds
of backward stepwise iterations, the introduction of population
density heavily decreased the strength of the association
with dioxin exposure (results not shown).
Figure 3 displays log-transformed SIR for NHL vs. square
root transformed dioxin concentration, while adjusting
for departments, log(industry × years), and splinesmoothed
log(population density). A clear positive and
linear trend is noticeable.