Most limitation statements acknowledge a limit; fewer than three in ten say something could not be done. Of 14,262 problems, 14.7% state an explicit impossibility and 13.7% an infeasibility within stated constraints. The rest acknowledge a limit without claiming it was infeasible.
Source: stage04 problem definitions (problems.jsonl), verified by verify_stage04.
Papers state more distinct limitations than they used to, and the kind has shifted. In 1996-2000 the average paper listed 5.87 problems; in 2021-2026, 7.62. Caveats about model assumptions fell from 19% to 7% of problems, while identification problems rose from 11% to 17% and generalizability problems from 20% to 28%.
Source: stage14 T14_1_trend_by_period.
The published record is filtered on the way to print. For 323 papers with an earlier version (dissertation chapter, conference paper or working paper), 28.9% of the early version's limitations are gone at publication (19.4% when only working papers are counted). 52.7% of the published limitations have no counterpart in the early version.
Source: stage13 T13_1 and T13_7; paper-clustered 95% intervals in the replication files.
What would resolve them is mostly measurement, new settings and data access. Problems most often call for a better measure (34%) or a new setting or population (24%). Calls for an exogenous shock or an experiment have grown several-fold since the late 1990s. Of the open problems that authors called infeasible, 2,954 are identification, data-access or generalizability problems that the paper did not resolve.
Source: stage14 T14_9 and T14_14 (controlled-vocabulary tags assigned by model).
A dataset that sounds relevant rarely supplies the missing condition. 40 post-publication datasets, platforms and shocks were checked against their official pages and set against the open problems they might address. A model judged 24 pairings a strong fit and 163 a possible fit, out of about 2,100 considered.
Source: stage15 resource index; candidate fits only, none verified.

Charts

Figure 1

Papers state more limitations than they used to; the share stated as infeasible rose in the 2000s and then held

Source: stage14 T14_1 (problem-level; paper-cluster bootstrap intervals). Hover over a mark for its value.

Show data

Period	Papers	Problems per paper	Infeasible share	95% CI
1996-2000	175	5.87	22.0%	[0.184, 0.259]
2001-2005	207	6.58	25.6%	[0.224, 0.289]
2006-2010	261	6.77	28.7%	[0.262, 0.315]
2011-2015	469	6.77	29.5%	[0.273, 0.318]
2016-2020	339	6.96	29.9%	[0.277, 0.323]
2021-2026	600	7.62	29.1%	[0.274, 0.309]

Figure 2

Model-assumption caveats faded; identification and generalizability problems grew

Source: stage14 T14_1. A problem can name more than one threat, so shares within a period sum to more than 100%. Hover over a mark for its value.

Show data

Period	Identification	Mechanism	Data access	Generalizability	Measurement	Model assumption
1996-2000	10.7%	20.6%	21.5%	19.8%	27.9%	19.1%
2001-2005	10.9%	19.4%	25.3%	22.2%	26.5%	15.9%
2006-2010	13.7%	19.1%	29.1%	22.2%	30.3%	10.0%
2011-2015	17.0%	19.1%	28.9%	22.6%	24.5%	9.4%
2016-2020	19.9%	17.6%	29.4%	25.3%	24.9%	8.5%
2021-2026	17.1%	21.3%	26.5%	28.3%	23.3%	6.5%

Figure 3

Each design admits its own kind of limitation

Source: stage14 T14_2. Row shares; a problem can name more than one threat. Hover over a mark for its value.

Show data

Design	Problems	Identification	Mechanism	Data access	Generalizability	Measurement	Model assumption
Quantitative	9178	20%	19%	30%	23%	30%	4%
Qualitative	1692	11%	18%	30%	41%	15%	3%
Mixed methods	1418	16%	19%	29%	26%	28%	6%
Simulation / formal	754	2%	9%	4%	12%	7%	77%
Theory	1131	3%	37%	11%	21%	13%	29%
Review	89	7%	42%	29%	19%	21%	10%

Figure 4

Three in ten early-version limitations are gone at publication; half of the published ones are new

Source: stage13 T13_1, T13_5, T13_6 (323 papers with a verified earlier version; model-judged mappings). Hover over a mark for its value.

Show data

Category	Count	Share
Kept as stated	992	60.1%
Kept, changed (narrowed, broadened, strengthened, softened)	181	11.0%
Dropped	477	28.9%
Dropped and not handled in the published article	255
Published problems added after the early version	1315	52.7%

Figure 5

Limitations stated as infeasible within constraints are the most likely to disappear

Source: stage13 raw mappings; 95% intervals from a paper-cluster bootstrap (1,000 draws). Filled dots: infeasibility claims. Hover over a mark for its value.

Show data

Strength in early version	Early limitations	Dropped	95% CI
Explicit impossibility	208	24.0%	16.9%–31.3%
Infeasible within constraints	349	38.1%	31.4%–45.3%
Acknowledged only	932	24.2%	19.8%–29.0%
Ambiguous	161	42.2%	33.5%–52.3%

Figure 6

Identification and measurement limits are usually handled at least in part; mechanism and generalizability limits mostly are not

Source: stage14 T14_4 (adopted resolution labels aggregated to problems). Hover over a mark for its value.

Show data

Threat	Problems	Addressed	Partly	Unresolved
Identification	2288	10.4%	66.6%	22.9%
Mechanism	2809	1.3%	40.7%	57.5%
Data access	3902	1.9%	51.9%	45.8%
Generalizability	3508	2.8%	42.6%	54.3%
Measurement	3613	6.9%	68.0%	25.0%
Model assumption	1386	8.2%	53.8%	37.4%

Figure 7

Most open problems wait on restricted data, better measures or longer panels

Source: stage14 T14_14. 2,954 open primary candidates: identification, data-access or generalizability problems stated as infeasible in empirical studies and not fully resolved in the paper. A problem is counted in every threat row it names and under every resource it needs, so rows and columns overlap. Hover over a mark for its value.

Show data

Threat	Open problems	Longitudinal data	Exogenous shock	Experiment	Micro / process data	New setting	Better measure	Restricted data access	Fieldwork access	Larger sample
Identification	844	240	295	143	62	84	185	131	20	90
Data access	2248	539	82	52	285	236	729	882	144	342
Generalizability	372	47	21	26	24	244	50	66	18	57