Appendix h — Data Card — NBER / KPSS Patent–Compustat Links

Source slug: nber-patent-link · Appendix C (Data Dictionary) · last touched 2026-05-28

Provider & what it is

This card covers the two crosswalk files that turn raw patents into firm-level finance variables — the bridge PatentsView deliberately does not build. There are two distinct public goods here, and students must not confuse them.

The patent-to-firm crosswalk (the NBER Patent Data Project lineage — Hall, Jaffe & Trajtenberg’s original assignee→firm match, and its modern successors). This is a mapping table: each patent’s assignee_id (or assignee name) ↔︎ a firm identifier (Compustat gvkey / CRSP permno). Without it, a patent is owned by “NVIDIA Corporation” the string; with it, the patent is owned by permno 86580 the tradable firm, and you can finally run a finance regression.
The KPSS patent-value file — the public dataset released alongside Kogan, Papanikolaou, Seru & Stoffman (2017), giving each patent a market-based dollar value estimated from the stock-price reaction to the patent’s grant announcement, plus a citation-based value and the firm link. This is the file behind the “innovation as a dollar number” measure the camp builds in Week 6.

The reveal-the-trick point: the value in the KPSS file is not measured, it is estimated from a market reaction inside a short event window, and the link in any crosswalk is a match, not a fact — both carry error you inherit silently if you treat the file as ground truth.

Coverage

The crosswalk and the KPSS file cover granted U.S. patents matched to publicly listed (CRSP/Compustat) firms, roughly 1926/1950s–present for the value series (the KPSS announcement-based measure needs daily returns, so it begins when CRSP daily data begin; the citation-based measure reaches further back) [CHECK exact start year of the public KPSS file and its most recent update vintage]. Only public firms with a CRSP permno are in scope — private firms, foreign assignees, and individual inventors are absent by construction, which is itself a sample-selection fact to disclose.

Key identifiers

patent_num / patent_id (links to PatentsView), permno (CRSP) and/or gvkey (Compustat) on the firm side, plus the KPSS value columns — typically a nominal dollar value (xi / xi_nominal), a real (CPI-deflated) value, and a citation-based value cites [CHECK exact column names in the current public release]. The join chain is: PatentsView patent ⟶ (crosswalk) ⟶ permno/gvkey ⟶ (CRSP/Compustat, GMU-licensed) ⟶ firm outcomes. Note the asymmetry of reproducibility: the crosswalk and KPSS value file are public and shippable, but the CRSP/Compustat tables at the end of the chain are licensed and must stay on GMU infrastructure (CONVENTIONS §5).

Access path (bulk download; no key)

These are bulk flat-file downloads, not APIs, and they require no key.

The KPSS patent-value data are posted publicly (the authors’ / Kelley School data page) as a download — typically CSV/DTA, one row per patent with the value columns and permno.
The NBER-lineage crosswalk ships as a downloadable mapping table (CSV).

import pandas as pd
# Files are downloaded once into data/raw/ and pinned by snapshot date.
kpss = pd.read_csv("data/raw/kpss_2017_public.csv",
                   dtype={"patent_num": "string", "permno": "Int64"})
xwalk = pd.read_csv("data/raw/patent_permno_xwalk.csv",
                    dtype={"patent_num": "string", "permno": "Int64"})
# Aggregate patent-level value to a firm-year panel:
panel = (kpss.merge(xwalk, on=["patent_num", "permno"], how="left")
              .groupby(["permno", "issue_year"])["xi_real"].sum()
              .rename("patent_value").reset_index())

No env secret is needed for the public files; the licensed CRSP/Compustat join at the end uses ${WRDS_USERNAME} via a pinned WRDS pull on GMU infra.

License (FREE / public) & note

The KPSS public file and the NBER-lineage crosswalk are released free for research use (academic/non-commercial; attribute the source papers) — they are public and may be redistributed inside a replication packet’s data/raw/ [CHECK the exact license/terms wording on the current KPSS download page]. The downstream CRSP/Compustat data are emphatically not public and must never be committed; a packet reconstructs them with a pinned WRDS script (the public-vs-licensed asymmetry from Appendix D.4). Cite KPSS (2017, QJE) for the value file and Hall–Jaffe–Trajtenberg / the NBER Patent Data Project for the crosswalk lineage.

Gotchas

The link is a match, not a fact. The crosswalk resolves messy assignee names to firms with error — missed matches (a patent whose firm exists but isn’t linked) and false matches (a patent assigned to the wrong subsidiary/parent). Coverage is also better for large, long-lived firms, so a “patents per firm” panel is non-randomly incomplete.
Vintage and assignee-id drift. PatentsView’s assignee_id changes across releases (see the uspto-patentsview card), so a crosswalk built against one PatentsView vintage may not key cleanly to another. Pin both vintages together.
The value is an estimate from a window. KPSS value = (stock reaction in a short window around grant) × adjustments. Any other firm news in that window contaminates it (the same confound Ch 6.1 dissects). It is an unbiased estimate of value on average, not the true value of any single patent — never read a single patent’s dollar figure as fact.
Nominal vs. real, and grant-year dating. Use the CPI-deflated column for cross-year comparisons, and aggregate to firm-year by issue/grant year (consistent with the event measured) — but remember grant-year dating imports the grant-lag truncation from the patent card.
permno vs gvkey and merge keys as strings. CRSP uses permno, Compustat uses gvkey; the file may carry one and not the other, requiring the CRSP/Compustat link table. As everywhere in this appendix, read id columns as strings/Int64, not floats, or merges silently drop.
Selection on listing. Only public firms are present; a firm that IPOs mid-sample appears only after listing, and de-listed firms vanish — survivorship and entry both bias firm-year innovation trends.

“First 10 rows” — schema sketch (ILLUSTRATIVE, not real values)

KPSS public patent-value file, one row per patent:

patent_num	permno	issue_year	filing_year	xi_nominal	xi_real	cites
6000001	10107	2000	1997	18.4	28.1	12
6000002	14593	2000	1996	540.9	826.0	47
6000003	87445	2000	1998	3.2	4.9	2
6000004	10107	2001	1998	65.7	98.3	9
6000005	(no match)	2001	1999	.	.	5

Patent→firm crosswalk, one row per (patent × firm match):

patent_num	assignee_id	permno	gvkey	match_score
6000001	a3f9…	10107	012141	0.98
6000002	b1c4…	14593	005047	0.95
6000005	c8d2…	(no match)	(no match)	0.41

(Illustrative only; do not cite. xi_real in millions of CPI-deflated dollars; row 5 shows an unmatched patent — exactly the non-random hole the gotchas warn about.)

Which chapter / lab / capstone uses it

Capstone 3 — Innovation from USPTO PatentsView — the central enabling files: the crosswalk turns PatentsView patents into a firm-year panel and the KPSS value file supplies the dollar innovation measure, so the capstone’s firm-outcome event study (ties to KPSS 2017) is literally built on this card. The selection, vintage, and value-is-an-estimate gotchas are the threats the capstone must name (CONVENTIONS §4).
Week 6, Ch 6.1 (Reader’s Guide — KPSS 2017) — Ch 6.1 reads this exact public dataset as research infrastructure; this card is the operational companion to that reading, and the “value is a windowed estimate” caveat is the chapter’s “what’s vulnerable.”
Leah’s running thread (patents, innovation) — the firm-linked patent panel is Leah’s workbench across Week 6.
Appendix D.4 (Replication Packet Standard) — the public-crosswalk-shippable / licensed-CRSP-not-shippable split is the textbook illustration of D.4’s public-vs-licensed asymmetry of reproducibility.

# Data Card — NBER / KPSS Patent–Compustat Links **Source slug:** `nber-patent-link` · Appendix C (Data Dictionary) · last touched 2026-05-28 ## Provider & what it is This card covers the two *crosswalk* files that turn raw patents into firm-level finance variables — the bridge PatentsView deliberately does not build. There are two distinct public goods here, and students must not confuse them. 1. **The patent-to-firm crosswalk** (the NBER Patent Data Project lineage — Hall, Jaffe & Trajtenberg's original assignee→firm match, and its modern successors). This is a mapping table: each patent's `assignee_id` (or assignee name) ↔ a firm identifier (Compustat `gvkey` / CRSP `permno`). Without it, a patent is owned by "NVIDIA Corporation" the *string*; with it, the patent is owned by `permno 86580` the *tradable firm*, and you can finally run a finance regression. 2. **The KPSS patent-value file** — the *public* dataset released alongside Kogan, Papanikolaou, Seru & Stoffman (2017), giving each patent a **market-based dollar value** estimated from the stock-price reaction to the patent's grant announcement, plus a citation-based value and the firm link. This is the file behind the "innovation as a dollar number" measure the camp builds in Week 6. The reveal-the-trick point: the *value* in the KPSS file is not measured, it is *estimated from a market reaction inside a short event window*, and the *link* in any crosswalk is a *match*, not a fact — both carry error you inherit silently if you treat the file as ground truth. ## Coverage The crosswalk and the KPSS file cover **granted U.S. patents matched to publicly listed (CRSP/Compustat) firms**, roughly **1926/1950s–present** for the value series (the KPSS announcement-based measure needs daily returns, so it begins when CRSP daily data begin; the citation-based measure reaches further back) [CHECK exact start year of the public KPSS file and its most recent update vintage]. Only *public* firms with a CRSP `permno` are in scope — private firms, foreign assignees, and individual inventors are absent by construction, which is itself a sample-selection fact to disclose. ## Key identifiers `patent_num` / `patent_id` (links to PatentsView), `permno` (CRSP) and/or `gvkey` (Compustat) on the firm side, plus the KPSS value columns — typically a nominal dollar value (`xi` / `xi_nominal`), a real (CPI-deflated) value, and a citation-based value `cites` [CHECK exact column names in the current public release]. The join chain is: **PatentsView patent ⟶ (crosswalk) ⟶ permno/gvkey ⟶ (CRSP/Compustat, GMU-licensed) ⟶ firm outcomes.** Note the asymmetry of reproducibility: the crosswalk and KPSS value file are *public and shippable*, but the CRSP/Compustat tables at the end of the chain are *licensed* and must stay on GMU infrastructure (CONVENTIONS §5). ## Access path (bulk download; no key) These are **bulk flat-file downloads, not APIs**, and they require **no key**. - The KPSS patent-value data are posted publicly (the authors' / Kelley School data page) as a download — typically CSV/DTA, one row per patent with the value columns and `permno`. - The NBER-lineage crosswalk ships as a downloadable mapping table (CSV). ```python import pandas as pd # Files are downloaded once into data/raw/ and pinned by snapshot date. kpss = pd.read_csv("data/raw/kpss_2017_public.csv", dtype={"patent_num": "string", "permno": "Int64"}) xwalk = pd.read_csv("data/raw/patent_permno_xwalk.csv", dtype={"patent_num": "string", "permno": "Int64"}) # Aggregate patent-level value to a firm-year panel: panel = (kpss.merge(xwalk, on=["patent_num", "permno"], how="left") .groupby(["permno", "issue_year"])["xi_real"].sum() .rename("patent_value").reset_index()) ``` No env secret is needed for the public files; the licensed CRSP/Compustat join at the end uses `${WRDS_USERNAME}` via a pinned WRDS pull on GMU infra. ## License (FREE / public) & note The **KPSS public file** and the **NBER-lineage crosswalk** are released **free for research use** (academic/non-commercial; attribute the source papers) — they are *public* and may be redistributed inside a replication packet's `data/raw/` [CHECK the exact license/terms wording on the current KPSS download page]. The downstream **CRSP/Compustat** data are emphatically **not** public and must never be committed; a packet reconstructs them with a pinned WRDS script (the public-vs-licensed asymmetry from Appendix D.4). Cite KPSS (2017, *QJE*) for the value file and Hall–Jaffe–Trajtenberg / the NBER Patent Data Project for the crosswalk lineage. ## Gotchas - **The link is a match, not a fact.** The crosswalk resolves messy assignee names to firms with error — missed matches (a patent whose firm exists but isn't linked) and false matches (a patent assigned to the wrong subsidiary/parent). Coverage is also better for large, long-lived firms, so a "patents per firm" panel is non-randomly incomplete. - **Vintage and assignee-id drift.** PatentsView's `assignee_id` changes across releases (see the `uspto-patentsview` card), so a crosswalk built against one PatentsView vintage may not key cleanly to another. Pin both vintages together. - **The value is an *estimate from a window*.** KPSS value = (stock reaction in a short window around grant) × adjustments. Any other firm news in that window contaminates it (the same confound Ch 6.1 dissects). It is an unbiased estimate of value *on average*, not the true value of any single patent — never read a single patent's dollar figure as fact. - **Nominal vs. real, and grant-year dating.** Use the CPI-deflated column for cross-year comparisons, and aggregate to firm-year by **issue/grant year** (consistent with the event measured) — but remember grant-year dating imports the grant-lag truncation from the patent card. - **`permno` vs `gvkey` and merge keys as strings.** CRSP uses `permno`, Compustat uses `gvkey`; the file may carry one and not the other, requiring the CRSP/Compustat link table. As everywhere in this appendix, read id columns as strings/`Int64`, not floats, or merges silently drop. - **Selection on listing.** Only public firms are present; a firm that IPOs mid-sample appears only after listing, and de-listed firms vanish — survivorship and entry both bias firm-year innovation trends. ## "First 10 rows" — schema sketch (ILLUSTRATIVE, not real values) KPSS public patent-value file, one row per patent: | patent_num | permno | issue_year | filing_year | xi_nominal | xi_real | cites | |------------|-------:|-----------:|------------:|-----------:|--------:|------:| | 6000001 | 10107 | 2000 | 1997 | 18.4 | 28.1 | 12 | | 6000002 | 14593 | 2000 | 1996 | 540.9 | 826.0 | 47 | | 6000003 | 87445 | 2000 | 1998 | 3.2 | 4.9 | 2 | | 6000004 | 10107 | 2001 | 1998 | 65.7 | 98.3 | 9 | | 6000005 | (no match) | 2001 | 1999 | . | . | 5 | Patent→firm crosswalk, one row per (patent × firm match): | patent_num | assignee_id | permno | gvkey | match_score | |------------|-------------|-------:|------:|------------:| | 6000001 | a3f9... | 10107 | 012141 | 0.98 | | 6000002 | b1c4... | 14593 | 005047 | 0.95 | | 6000005 | c8d2... | (no match) | (no match) | 0.41 | (Illustrative only; do not cite. `xi_real` in millions of CPI-deflated dollars; row 5 shows an unmatched patent — exactly the non-random hole the gotchas warn about.) ## Which chapter / lab / capstone uses it - **Capstone 3 — Innovation from USPTO PatentsView** — the central enabling files: the crosswalk turns PatentsView patents into a firm-year panel and the KPSS value file supplies the dollar innovation measure, so the capstone's firm-outcome event study (ties to KPSS 2017) is literally built on this card. The selection, vintage, and value-is-an-estimate gotchas are the threats the capstone must name (CONVENTIONS §4). - **Week 6, Ch 6.1 (Reader's Guide — KPSS 2017)** — Ch 6.1 reads this exact public dataset as research infrastructure; this card is the operational companion to that reading, and the "value is a windowed estimate" caveat is the chapter's "what's vulnerable." - **Leah's running thread (patents, innovation)** — the firm-linked patent panel is Leah's workbench across Week 6. - **Appendix D.4 (Replication Packet Standard)** — the public-crosswalk-shippable / licensed-CRSP-not-shippable split is the textbook illustration of D.4's public-vs-licensed asymmetry of reproducibility.