TWFE in author replication code
Under development — author-code evidence screen, 12 October 2026. This is a score of inspectable code evidence, not study quality, identification validity, template eligibility or successful execution.
Back to Under development · 262-paper evidence register · Counts by score
What the code establishes
At least 148 of the 245 library deposits contain a reviewed regression with entity/group and temporal-category fixed effects. 84 of those papers were among the previous 161 provisional exclusions. Thus an exclusion from the binary-adoption comparison template cannot be interpreted as absence of TWFE. The screen does not establish that every paper uses TWFE, nor that papers without a positive match do not.
| Score | Meaning | Papers |
|---|---|---|
| 3 | Confirmed entity/group + temporal-category effects | 148 |
| 2 | Possible two-way effects; coding or dimensions unresolved | 19 |
| 1 | Partial FE evidence; two-way structure not established | 66 |
| 0 | No qualifying FE evidence detected by this code screen | 12 |
| NA | Original author analysis code not inspected/available here | 17 |
Definition and interpretation
A confirmed example contains an entity or group effect and a categorical temporal dimension. This broad definition includes calendar periods, survey waves, birth/publication cohorts, two-period post indicators, and richer group-by-time effects. It does not imply a balanced longitudinal panel, full calendar-period saturation, a treatment effect, a headline regression, or a binary observational DiD. Cohort/group-level effects, nonlinear models, IV models, first stages and auxiliary residualization regressions are identified in the row notes.
One confirmed example is enough for a paper-level positive. The model-family field describes that example only, not every model in the paper. Confirmed examples: IV/2SLS: 4; Linear OLS: 141; Poisson/PPML: 3.
Score 3 requires an inspected command and resolved FE structure. Scores 2/1/0 are conservative static-search triage, not probabilities or exhaustive negative findings. Score 2 flags plausible two-way structure; score 1 records partial FE or unresolved formula evidence; score 0 means no qualifying pattern was detected. NA is reserved for the 17 additional legacy papers whose original author deposits were not inspected here. Their existing original-collection results are not downgraded.
Provenance and coverage
All 245 supplied original author ZIP deposits were inspected for Stata, R/Rmd, Python, MATLAB, SAS and Julia source extensions. This did not use the library-generated R/Stata translations as author evidence. Source files were read, not executed. Selected-command review follows a static candidate search; it is not a line-by-line semantic audit of every model in every file. The register reports package URL, member path, original source line, command, SHA-256, file count and explicit nested-archive omissions.
Bundled ado implementations, vendor packages and their demonstration/test scripts are excluded from positive evidence. A clustered standard error is not an absorbed fixed effect. An xtset declaration is not a regression. xtreg without fe is not counted as a within estimator. Continuous time trends alone do not count as categorical time effects. Simple same-file macros are expanded only for candidate selection; the published evidence retains the original command. Data-stored panel identifiers are labeled as such, rather than inferred from the clustering variable.
Nested ZIP members below 100 MB and depth 3 were searched recursively; larger nested archives and source files above 20 MB are listed as omissions. Standard command/formula search can miss manually demeaned data, first-difference equivalences, custom routines, cross-file or dynamically constructed macros and unrecognized file formats. Such absence is not used to exclude papers. No code-score calculation changes any eligibility decision, fit, original report or calibration point.
Illustrative confirmed code
Investment Banks as Corporate Monitors in the Early Twentieth Century United States
Railroad and year effects in an author-defined binary robustness specification.
Effects: RRid and year. Model family: Linear OLS.
Author deposit; 20150143_data/appendix.do, line 253.
qui areg q `newvars' bvlev13trend bvlev13post20 surplus3 i.year, absorb(RRid) cluster(RRid)E-governance, Accountability, and Leakage in Public Programs: Experimental Evidence from a Financial Management Reform in India
Nationwide rollout specification; separate from the paper’s randomized experiment.
Effects: DistrictId and Year. Model family: Linear OLS.
Author deposit; data/Do/Fig-4-DiD-Graph.do, line 21.
eststo : xtreg Exp`item' F?EFMS_`item' EFMS_`item' L?EFMS_`item' _IYear* $Weight,fe vce(cluster StateId)Public Information Is an Incentive for Politicians: Experimental Evidence from Delhi Elections
Ward fixed effects and a two-period post indicator in an experimental specification.
Effects: ward_rd2 and post. Model family: Linear OLS.
Author deposit; code/Private information analysis/tables 5 and a12.do, line 36.
eststo: qui xtreg `depvar' i.post##i.treatmentmcd_rd2 , fe i(ward_rd2) vce(cl ward_rd2)Do Consumers Exploit Commitment Opportunities? Evidence from Natural Experiments Involving Liquor Consumption
Entity within effects and year indicators despite nonbinary policy hours.
Effects: id and year. Model family: Linear OLS.
Author deposit; Bernheim_Meer_Novarro_AEJ_05_Data/Bernheim_Meer_Novarro_Tables.do, line 41.
xtreg logliq liqonsun liqoffsun liquortax cigtax unempw popover18 liqage18-liqage20 liqonOBH liqoffOBH liqonOBO liqoffOBO allow restr i.year, r fe i(id)Taxes Today, Benefits Tomorrow
State and quarter effects in an IV specification; not automatically observational DiD.
Effects: ST_ and col_quarter. Model family: IV/2SLS.
Author deposit; replication package AEJ Partial UI/do/5.triggers_tab3.do, line 487.
ivregress 2sls ${Y} (${X}=trig) ${COV} i.ST_ i.col_quarter [aweight=${WEIGHT}], vce(robust) firstOf Mice and Academics: Examining the Effect of Openness on Innovation
Entity conditional FE plus annual indicators; nonlinear count model.
Effects: o_id_global and y1993-y2006. Model family: Poisson/PPML.
Author deposit; MADKS-Replication-Files/MADKS_mouse_regs_graphs_and_robustness_11-30-2014_v2.do, line 1438.
xtpoisson f_last_new_old $year01 $age_T_001 $age_B_001 yr_T window_T shift_T window_B shift_B , fe i(o_id_global) vce(bootstrap, reps(200)) iterate(100) difficult technique(nr bfgs 1)