Parallel Computing for Data Science

280 CHAPTER 13. SUBSET METHODS

Figure 13.2: CA Performance, Forest Data

as seen in Figure 13.2. The speedup starts out linear, then becomes less

dramatic but still quite good, especially in light of the factors mentioned

earlier.

But what about the accuracy? The theory tells us that the CA estimator

is statistically equivalent to the full estimator, but this is based on asymp-

totics. (Though it should be noted that even the full estimator is based on

asymptotics, since glm() itself is so.) Let’s see how well it worked here.

Table 13.1 shows the values of the estimated coefficient for the first pre-

dictor variable, for the different numbers of cores (1