Early investments
Important features of these data are immediately visible in Figure 1. First, the citation frequency rises rapidly in the first few years after the cited patent, then peaks and declines slowly over time. A significant number of citations are still being received many years after initial grant. Second, a U.S.-invented patent is much more likely to be cited by a U.S.-invented patent than it is by a foreign-invented patent. Finally, even putting aside the domestic/foreign distinction, there are noticeable differences in the citation frequencies across citing countries. For example, the likelihood of a random U.K.-invented patent citing a U.S.-invented patent is about 40% higher than the likelihood of a random German-invented patent citing a U.S.-invented patent.

Although all of these qualitative features illustrated in Figure 1 are in some sense “real,” raw citation frequencies are afflicted by numerous theoretical and actual biases that make their interpretation dangerous. First, the observation of citations is always subject to truncation bias. Since we can observe only the citations already granted, we can see citations at long lags only for citations from very recent cohorts to very old cohorts. The significance of the truncation problem is greatly exacerbated by the fact that the number of citations made per patent has been rising significantly in the last two decades (Caballero and Jaffe, 1993). Thus the observations at long lags in Figure 1 all come from patents granted when relatively many citations were made (e.g., citations from 1993 patents to 1963 patents), whereas the observations at short lags are mixtures of many different cited cohort/citing cohort combinations (e.g., for lag=5 we have 1977 to 1972, 1987 to 1982, and so forth). We will see below that controlling for these interacting time effects yields predicted probabilities for long lags considerably lower than those shown in Figure 1.