Database doesn’t save USPTO resources – requires maintenance and info errors increase delays and confusion

With the few exceptions noted above, we reproduced the data from the XML files in their entirety, making no attempt to correct errors. Because the USPTO does not validate the accuracy of information recorded, such errors and inaccuracies are expected. Regrettably, we are unable to conduct a large-scale, retrospective validation to determine the extent to which data errors may be non-random. Moreover, substantial portions of these data are not regularized, especially in the assignor and assignee name and address fields. Wanting to provide researchers maximal flexibility in employing these data, we opted to retain the original free-form text throughout the Dataset.

Serrano (2010) provides a thorough discussion of the general limitations of patent assignment data, namely that these data do not distinguish the acquisition of a firm from the acquisition of a bundle of patents. Further they cannot distinguish parties’ intended purpose in transacting (for instance, between assignments intended to facilitate technology adoption versus those made by licensing entities interested only in monetizing the patent). Moreover, patent assignment records provide no information on the price paid for assigned patents. For these reasons, we encourage researchers to link the Dataset to other data sources to further study these transactions within the market for technology.

Dataset users should be aware of potential biases resulting from selective recording. Because the Dataset only captures those transactions voluntarily recorded with the USPTO, users should be cautious about making inferences about the entire population of assignments and other transactions. In Section 7, we discuss in more detail the potential bias from systematic non-recording. In this section, we highlight less serious, though noteworthy, complications for Dataset users.

5.1 Duplicate recording

Duplicate recordings occur when the same transaction is (apparently) recorded multiple times (when the variables – assignor and assignee names, execution dates, properties, and conveyance text – are identical across multiple rf_id observations) or under different conveyance types (assignor and assignee names, execution dates, and properties are identical, but the nature of conveyance differs across multiple rf_id observations). An example of the latter involves rf_id 17057/0787 (“assignment”) and rf_id 17057/0793 (“security interest”), both recorded with the Office on October 6, 2005 for transactions executed on October 3, 2005 between Anvil International, Inc. (assignor) and Bank of America, N.A. (assignee), involving the same 13 patents. Subsequent transactions suggest that Bank of America retained only a security interest in the patents, which was released back to Anvil in 2010.42 It is possible that these are independent transactions, involving a security interest agreement and a transfer of the collateral assets upon foreclosure, or a duplicate record of a single, complex deal. Whether the recording is a duplicate is uncertain, and raises thorny questions concerning how to think about multiple activities occurring within a single transaction. Moreover, the motivation for such dual recordings is unclear, though it may reflect uncertainty regarding perfecting title (see Section 3.2) or (overly) cautious recording practices.

Because duplicate recordings will tend to overstate the number of transactions (rf_id counts) and the number of properties involved (property counts), Dataset users may – depending on the research question under consideration – opt to remove likely duplicates. Additional rules may need to be imposed by the researcher, particularly for certain conveyance types. We identified roughly 63,600 transactions (1.0%) involving about 252,600 properties (2.3%) that may be duplicate recordings with matching conveyance text.43 We also identify 3,110 transactions (0.05%), involving roughly 31,500 properties (0.3%) that may be duplicates, but with different conveyance texts (although most are the same type of conveyance).44 Figure 2 shows annual transactions and property counts with and without duplicate recordings, by recording year.45

5.2 Recording lag

Another important consideration when using the Dataset is the recording lag between stated execution dates and recording of transactions. Short recording lags suggest diligence by the parties or possible posturing for subsequent behavior.46 Longer recording lags may suggest inattention, time spans of no foreseeable transaction activity, or abandonment of the property altogether. Figure 3 depicts the mean and 95 percent confidence interval recording lag over time, broken out by recording year and execution year to highlight censoring (which appears most acute in the post-2005 figures). Recording lags execution date by roughly 182 days, on average (21 day median), but lags appear to be increasing over time.

Technically, a timely recordation should occur within three months (or, prior to a subsequent purchase or mortgage).47 Because we anticipate that incentives to meet these requirements will differ depending on the nature of a conveyance, we examine recording lags by conveyance type in Figure 4. Assignments, employer assignments, security interests, and releases tend to have the shortest lags. Short recording lags in employer assignments may reflect incentives for employers to quickly secure rights from inventing employees. By contrast, the recording lags for mergers and name changes are much longer and grow steadily over time.48

The assignee data shows that security interests are often recorded by financial institutions and lien holders. Figure 4 suggests that such parties record in a more timely fashion, possibly to provide notice of their stake in the patent collateral. Releases from such agreements have lags that are also relatively short, consistent with debtors having an interest in clearing their title. Interestingly, some of the volatility shown in the government interest lags by recording year in Figure 4 reflects the behaviors of individual government entities. For example, the spike in average lags near recording year 2008 was primarily caused by the National Institutes of Health (NIH) recording roughly 8,600 transactions (about 98% of the government interests recorded that year). These transactions, some executed as early as 1981, had an abnormal mean lag, and may have been the result of an attempt by the agency to improve compliance with regulations (see Section 3.3). This anecdote provides users with a good example of how such housekeeping efforts may affect the data and measures derived from the Dataset.

Across nearly all conveyance types, excepting government interests, the mean recording lag in the first decade of the 21st century has increased relative to the prior decades. We leave as open questions what is driving recording lags in general, or by conveyance type in particular, and why lags vary over time and over the life of patents. Of particular interest is whether shorter recording lags for assignments and security interests reflect some motivation to provide timely notice for such transactions in general or are the result of selection bias (possibly reflecting organizations behaving differently when changes in patent value or circumstances motivate more or less prompt recording).