Classification and regression trees for epidemiologic research: an air pollution example

Table 2 Quartile contrasts at each internal (Non-Terminal) node

Internal node no.^a	N	Quartile contrast^b	Wald P-value^c	Subset of pollutant quartiles to which contrast applies^d
				CO	NO2	O3	PM2.5
1	3879	PM2.5: 4 vs. 1-3	0.000	All	All	All	All
2	2878	NO2: 3–4 vs. 1-2	0.003	All	All	All	1-3
3	1001	NO2: 3–4 vs. 1-2	0.019	All	All	All	4
4	1560	O3: 4 vs. 1-3	0.096	All	1,2	All	1-3
5	1318	PM2.5: 2–3 vs. 1	0.123	All	3,4	All	1-3
7	685	O3: 4 vs. 1-3	0.128	All	3,4	All	4
8	1401	NO2: 2 vs. 1	0.086	All	1,2	1-3	1-3
9	159	CO: 3–4 vs. 1-2	0.043	All	1,2	4	1-3
14	244	NO2: 4 vs. 3	0.096	All	3,4	1-3	4
16	703	O3: 3 vs. 1-2	0.140	All	1	1-3	1-3
17	698	O3: 2–3 vs. 1	0.062	All	2	1-3	1-3
33	309	PM2.5: 3 vs. 1-2	0.033	All	1	3	1-3

^aThe node numbers correspond to the numbering in Figure 1 (where each node, n, produces two child nodes numbered 2n and 2n + 1).
^bBased on the indicator variable chosen for the best split.
^c P-value based on a Wald test that the beta coefficient for the quartile contrast indicator is zero.
^dEach subset of pollutant concentration levels represents an effect modifier of the quartile contrast and relates directly to the branching of the tree in Figure 1. Note that in the first split of the tree there is no effect modification by any of the pollutants because the entire dataset is used.

ISSN: 1476-069X