study,estimate_as_printed,uncertainty_as_printed,sample_as_printed,definition,locator,harmonised_estimate,harmonised_se,harmonisable,poolable,transformation,reason_if_not
S1,"ExpCon coefficient (Model 2) = -0.24. (marker '.' = p<.1) on LogToT; text: beta_1=-0.24, effect of AI = 21% (faster)","SE (0.14); 95% CI = [-0.51, 0.03]; p=0.086 (NS)",n = 96 (Table III); time-on-task means N=47 (AI) and N=46 (no AI) in Sec. VI-B,"OLS coefficient of the AI experimental-condition indicator on the natural log of time on task, controlling for developer-level covariates (Model 2); randomised, one common enterprise coding task","arxiv-2410.12944v3, Table III, row ExpCon (+ SE row below), column (2) 'Dev' [extract lines 76-78]; Section VI-C, 95% CI and p-value [line 58]",-0.24,0.14,True,True,theta = b = -0.24; SE(theta) = SE(b) = 0.14 (cross-check from printed CI: (0.03 - (-0.51))/(2*1.96) = 0.1378),
S2,+19% (allowing AI increases completion time by 19%; 'implement issues in 19% more time on average'),"none printed (the 95% CI is shown only in Figure 1, which is not reproduced in the extract)","16 developers; 246 tasks (136 issues AI-allowed, 110 AI-disallowed)","S = E[time with AI allowed]/E[time with AI disallowed] - 1, estimated as exp(beta)-1 from a log-linear OLS of total implementation time on the AI indicator and log forecast time (issue-level randomisation)","arxiv-2507.09089v2, Abstract [line 34] and Section 3 Results opening [line 62]; estimator Sections 2.3 and D.1-D.3",0.17395330712343798,,True,False,theta = ln(1 + c) = ln(1 + 0.19) = 0.173953,"Not poolable: no SE, CI or p-value for the 19% estimate is printed in the extract; the confidence interval appears only in Figure 1, which is not reproduced, so SE(theta) cannot be obtained."
S3,"All Content: Overall 53.84 min, Treatment 44.29 min, Control 63.10 min, Improvement -29.79%***",significance stars only: *** (defined as 0<p<=0.01); no SE or CI printed,181 subjects completed the task satisfactorily (randomised half to Copilot),"Percentage difference in mean task duration (minutes), Copilot vs control, holding accuracy constant (control duration predicted at treatment accuracy; Appendix B), all tasks across three IT-admin scenarios","arxiv-2411.01067v2, Section 4.2, Table 3, row 'Aggregate Results - All Content', columns Treatment/Control/Improvement [line 55]",-0.353679434958934,,True,False,theta = ln(1 + c) = ln(1 + (-0.2979)) = -0.353679 (printed ratio check: ln(44.29/63.1) = -0.353962),"Not poolable: uncertainty is reported only as significance stars (*** = p<=0.01); bootstrap SEs/CIs are not printed, so SE(theta) cannot be obtained."
S4,"Full Corpus: Aware Time per Email (s) 50.5, Control Time per Email (s) 56.9, % Time Savings 11.1",p-value 0.471,"Control N=52, Aware N=56 (Blind N=59; 167 analysts in total; 13 responses scoring <40% filtered: 3 Aware, 6 Blind, 4 Control)","Percentage time saving per processed email (mean time per email), Aware arm (agent queue ordering + verdicts/rationales) vs Control (no AI), full 25-email queue","arxiv-2511.13860v1, Section 3.3, Table 3, column 'Full Corpus', rows Aware/Control time, % Time Savings, p-value [lines 88-96]",-0.11765804346823246,0.1632205919720678,True,True,c = -11.1/100 = -0.111; theta = ln(1 + c) = ln(0.889) = -0.117658; approximate SE = |theta| / z_(1-p/2) = 0.117658 / z_(1-0.471/2) = 0.117658 / 0.720853 = 0.163221 (approximate SE from p-value only),
S5,"Treat x Post (Output, column 3, panel A) = -0.278",SE (0.295) (clustered at user level); not significant,"N = 18332 (user-day observations, users with affiliation)","Difference-in-differences coefficient on an aggregate GitHub output count measure (Push, PRs, comments, Create, Issues) for Italian vs Austrian/French users around Italy's ChatGPT ban (pre 27/28.03 vs post 03/04.04.2023)","arxiv-2304.09339v1, Section 4, Table 1, panel A, row Treat x Post (+ SE row), column (3) Output; N row [lines 80-88]",,,False,False,none (not harmonisable),"Not harmonisable: the effect is on output counts (activity volume) around the withdrawal of access, not a comparison of completion times with vs without AI on the same tasks, and it is not a randomised comparison."