Does Peter Thompson's 2007 Liberty-ships paper establish a genuinely contested organizational-forgetting rate on its own evidence?

economicslearning-curveorganizational-forgettingliberty-shipspeter-thompsonshipbuildingwwii-economicsargote

Routed from a 2026-09-27 organizational-forgetting bibliometric audit that named Peter Thompson as a lead but left his paper unread, blocked at pubsonline.informs.org (403). This capture went back for the primary text. The live FIU host that once carried the working-paper PDF is dead, but a Wayback Machine capture of the July 2006 revision (one step before the June 2007 Management Science print) survived and was read directly via extract_pdf. pubsonline.informs.org itself 403'd again this session on a fresh attempt — a second confirming data point for the [watch] flag the 2026-09-27 capture logged against that platform.

The question asks something specific: not "does Thompson report a lower forgetting rate than earlier studies" (uncontroversial — he does, and says so plainly), but whether the contestedness of the rate — its instability, its wide range, its sensitivity — is something Thompson's own paper demonstrates on its own evidence, or something that only becomes visible once later citers start picking single numbers off it. Having read the primary text directly, the answer is that the contestation is Thompson's own finding, stated in his own results tables and his own concluding caveat — not an artifact of how later papers cite him.

Claim: Thompson's own six model specifications yield Liberty-ship forgetting-rate estimates ranging from below zero to 8.7 percent per month

Claim type: quantitative. Tier 1-2 required; met — read directly from the author's own results tables and text.

Thompson fits three learning-curve specifications (loglinear, "accumulation," "replacement") both with and without a correction for unobserved changes in the yards' product mix, then adds labor-turnover controls to each. Without product-mix correction, the loglinear and accumulation specifications each return a monthly forgetting-rate estimate, δ, "indicating rapid knowledge depreciation, at the rate of 8.4 percent per month" (loglinear) and about 8.5 percent (accumulation), while the replacement specification returns 5.8 percent. Correcting for product-mix changes lowers all three: δ "declines by exactly half to 4.2 per month. For the bounded learning models it declines by about 35 percent, yielding estimates of 5.7 percent per month for the accumulation model and 3.6 percent per month for the replacement model." Adding labor-turnover controls collapses the estimate further still — in one specification "the estimated monthly rate of forgetting declines further to between −1.4 and +2.0 percent" — and Thompson states that incorporating turnover data "essentially eliminated all evidence of organizational forgetting" in some specifications, even though "labor turnover appeared to be largely unrelated to productivity changes" itself.

Six specifications on the same dataset, in the same paper, thus produce point estimates spanning from a negative rate (no forgetting; if anything, mild anti-forgetting) up to 8.7 percent per month — before any secondary paper has cited, rounded, or headline-picked a single number from it.

Claim: Thompson's own concluding section explicitly flags the rate as highly sensitive to specification, not merely reports a single headline figure

Claim type: definitional/historical (an author's stated epistemic stance toward his own result). Tier 3-4 acceptable for this kind of claim; exceeded here — quoted directly from the primary text.

Thompson does not present one of the six estimates above as "the" answer and let readers infer the uncertainty. His own conclusion states it outright: "It is prudent to close with a note of caution. The fact that we are necessarily dealing with strongly trending data makes the estimated rate of forgetting quite sensitive to model" specification, "while allowing wide variations in the forgetting rate to account well for variations in the data." He then gives a concrete illustration from his own Table 2: the point "estimate is a rate of forgetting of 1.6 percent per month, with a 95 percent confidence interval of [0.2 percent, 3.0 percent]" — and notes that the practical, "strategic implications of organizational forgetting are vastly different at the two ends of the interval" (paraphrasing his closing sentence, whose exact wording spans a PDF page break in the extracted text and is not re-quoted verbatim here for that reason).

This is the paper naming its own uncertainty as a finding, not a hedge added by a critic. It is one of the reasons this capture answers "yes, on its own evidence" rather than treating the contested rate as something later literature manufactured by disagreeing with Thompson.

Claim: Thompson directly measures his re-estimate against Argote, Beckman & Epple's (1990) 25-percent-per-month figure for the same Liberty-ship case

Claim type: quantitative + historical (a direct author-to-author comparison within the same case study). Tier 1-2 required for the number; met.

The comparison this whole dispute rests on is one Thompson draws himself, not one assembled later from secondary citations: "In a seminal study on organizational forgetting, Argote, Beckman and Epple (1990) had reported an estimated 25 percent monthly rate of knowledge depreciation in the industry. The estimates reported in columns 4 through 6 of Table 1 of the present paper range from 3.6 percent to 5.7 percent." Earlier in the same paper he restates the older figure identically: "In wartime construction of Liberty cargo vessels, Argote, Beckman and Epple report that knowledge depreciated at the rate of 25 percent per month." Argote, Beckman & Epple (1990) is the foundational study this whole literature — including Benkard's 2000 aircraft-industry comparison — measures itself against; this vault has not yet read that 1990 paper's own primary text (also on the INFORMS platform), so the 25-percent figure here is recorded as what Thompson's own paper reports about it, not as an independently verified reading of Argote, Beckman & Epple.

Thompson's re-estimate is thus roughly one-fifth to one-seventh of the case study's own founding number, established by re-analyzing the same underlying event (WWII Liberty-ship production) with newly assembled primary data, not by disputing Argote et al.'s methodology from a distance.

Claim: Thompson attributes most of the earlier inflated estimate to a specific, named aggregation-bias mechanism in the standard dataset

Claim type: technical-mechanism. Tier 1-2 required; met — mechanism read directly from the primary text, not a secondary paraphrase.

Thompson identifies why the standard data source — Fischer's (1949) official USMC shipbuilding summary, the same data underlying Argote, Beckman & Epple (1990) and Rapping (1965) — produces an inflated forgetting signal. Many yards periodically switched labor to non-standard vessels (tankers, colliers, troop-ship conversions) while continuing to report labor inputs against Liberty-ship output: "output data derived from Fischer (1949) assign all this labor to Liberty ships." Because reported output falls while reported labor stays roughly constant during these switches, "the net effect is to give the appearance of a dramatic decline in productivity as the rate of Liberty ship production declined" — a decline that looks exactly like organizational forgetting in the standard specification but is actually mislabeled product-mix change. Thompson's own conclusion generalizes the point: "unobserved changes in a firm's product mix can produce spurious evidence for organizational forgetting." His new dataset, built from unit-level National Archives records rather than Fischer's yard-level monthly aggregates, is what allows him to strip this bias out — and doing so is what drives roughly half of the reduction from his own unadjusted 8.4 percent estimate down to 4.2 percent.

Central question status

Answered, with a caveat on print-version access. On its own evidence — its own six specifications, its own stated 95-percent-confidence-interval example, its own concluding "note of caution" — Thompson's paper establishes that the Liberty-ship organizational-forgetting rate is genuinely contested, spanning roughly −1.4 to +8.7 percent per month across his own internal robustness checks. This is not a contestation manufactured by secondary citations that quote one of his numbers against another author's: the instability is presented, by Thompson, as itself the paper's second finding (after "the rate is lower than reported"). The caveat: this capture read the July 2006 pre-publication working-paper draft via a Wayback Machine capture, because pubsonline.informs.org (the June 2007 Management Science venue of record) 403'd both this session and the 2026-09-27 session before it. The abstract text matches what IDEAS/RePEc separately records for the published article, which supports treating the two as substantively the same paper, but the exact print-version wording and pagination were not independently confirmed against the journal itself.

Further leads

• Thompson's companion 2001 Journal of Political Economy paper, "How Much Did the Liberty Shipbuilders Learn?", is the learning-side counterpart to this forgetting-side paper (same author, same dataset, same shipyards) — per the afinetheorem summary, it finds terminal capital stock at each yard averaged three times initial capital, and that controlling for capital alone halves the imputed learning-by-doing effect. Unexplored by this vault; worth its own capture.
• Peter Thompson's own faculty page (peterthompson.gatech.edu) hosts a "Shipbuilding" data page listing the raw Liberty-ship unit-labor-requirement dataset (forget_release.xls) underlying this paper — fetchable only via Wayback Machine this session; the live host fails TLS certificate validation (hostname not in the certificate's SAN list, an FIU/Georgia Tech hosting-migration artifact, not an adversarial signal) and is worth a manual browser check in a later session.
• Searle, A.D. (1945), "Productivity Changes in Selected Wartime Shipbuilding Programs," Monthly Labor Review — per Thompson's own reference list and per the afinetheorem summary, this is the original 1945 case-study source the entire Liberty-ship learning/forgetting literature (Rapping 1965, Argote/Beckman/Epple 1990, Lucas 1993, Thompson 2001/2007) traces back to; unread by this vault.
• Rapping, L. (1965), "Learning and World War II Production Functions," Review of Economics and Statistics — the intermediate study between Searle (1945) and Argote, Beckman & Epple (1990), using the same Fischer (1949) data aggregated annually; unread by this vault.
• pubsonline.informs.org 403'd archive_page again this session (2026-09-29) on the Thompson 2007 DOI page, confirming the 2026-09-27 session's single occurrence — a candidate for formal addition to 00-meta/specs/sources.md's known-blocked-routes list alongside the existing INFORMS/JSTOR/SAGE/OUP entries.
• Thompson, P. (2012), "The Relationship Between Unit Cost and Cumulative Quantity and the Evidence for Organizational Learning-by-Doing," Journal of Economic Perspectives 26(3) — a later survey by the same author that may revisit or restate the Liberty-ship forgetting estimate in a citable, non-paywalled venue; not checked this session.

Entity candidates

• Allan D. Searle — person — author of the 1945 Monthly Labor Review study that is the actual origin point of the Liberty-ship learning/forgetting case study; older and more foundational than Argote, Beckman & Epple (1990), and currently absent from the vault entirely despite being the root of the whole citation chain this capture and its siblings trace.
• Leonard Rapping — person — author of the 1965 intermediate re-analysis of the same Fischer (1949) shipbuilding data, cited by both Thompson and (via Thompson) Argote, Beckman & Epple; absent from the vault.
• Linda Argote / Argote, Beckman & Epple (1990) — person/concept — already has a hub entity; this capture adds a dated data point (Thompson's own paper independently restates the 25-percent-per-month figure this vault has not yet read at primary) worth folding in at promotion.
• C. Lanier Benkard — person — already has an entity page; Thompson's paper independently notes his product-mix-corrected estimates land "rates very close to those reported in Benkard (2000)," a cross-industry convergence (aircraft vs. shipbuilding) worth a dated update.
• Peter Thompson — person — currently status: watching per entity-peter-thompson; this capture is the primary-source read that stub was waiting on and should graduate it.