study,estimate_as_printed,uncertainty_as_printed,sample_as_printed,definition,locator,harmonised_estimate,harmonised_se,harmonisable,poolable,transformation,reason_if_not
S1,"-0.24 (ExpCon, Model 2, LogToT; marker '.'); text: beta_1=-0.24, effect of AI = 21%","SE (0.14); 95% CI = [-0.51, 0.03]; p=0.086","Table III: n = 96 (Section VI-B: N=47 AI, N=46 no-AI)","OLS coefficient of the experimental condition (AI vs no-AI, randomised) on the natural log of time on task, controlling for developer-level factors (Model 2)","arxiv-2410.12944v3; Table III (L76-L78), row ExpCon, column (2) 'Dev'; Section VI-C (L58) for CI and p",-0.24,0.14,True,True,log-time coefficient: theta = b = -0.24; SE(theta) = SE(b) = 0.14 (printed SE used; CI-based SE = (0.03 - (-0.51))/3.92 = 0.1378 used only as sensitivity),
S2,19% more time on average (AI allowed increases completion time by 19%),"none printed (confidence interval shown only in Figure 1 / Figure 15, not reproduced in the extract)","16 developers, 246 tasks; 136 issues AI-allowed, 110 issues AI-disallowed","S = E[time with AI allowed]/E[time AI disallowed] - 1, estimated as exp(beta)-1 from OLS of log implementation time on AI indicator and log forecast time (issue-level randomisation)",arxiv-2507.09089v2; Section 3 Results (L62); abstract (L34); estimator in Section 2.3 (L45) and Appendix D.1-D.3,0.17395330712343798,,True,False,percentage change: c = +0.19; theta = ln(1 + 0.19) = 0.173953 (equals the regression beta since S = exp(beta) - 1),"Not poolable: no SE, CI or p-value is printed in the extract; the confidence interval appears only in a figure that is not reproduced, so SE(theta) cannot be obtained."
S3,"Treatment 44.29 min, Control 63.10 min (Overall 53.84); Improvement -29.79%***",significance stars only (***; per the paper's scheme 0<p<=0.01); no SE or CI printed,181 subjects completed the task satisfactorily (one subject dropped),"Percentage difference in mean task duration (minutes), Copilot (treatment) vs control without Copilot, holding accuracy constant (control duration predicted at treatment accuracy; Appendix B), randomised","arxiv-2411.01067v2; Section 4.2, Table 3 (L51-L55), row 'Aggregate Results / All Content', columns Treatment, Control, Improvement",-0.353679434958934,,True,False,"percentage change: c = -0.2979; theta = ln(1 + (-0.2979)) = -0.353679 (ratio of displayed means ln(44.29/63.10) = -0.353962, sensitivity only)",Not poolable: uncertainty is printed only as significance stars (p<=0.01); SE(theta) cannot be recovered from a threshold.
S4,"Aware 50.5, Control 56.9 (row labels: Time per Email (s)); % Time Savings 11.1",p-value 0.471,"Aware N=56, Control N=52 (of 167 analysts in three arms; 13 low-scoring responses filtered)","Percentage time saving per processed email (Full Corpus, whole 25-email queue without resolve-benign protocol), Aware arm (agent verdicts, rationale, agent-ordered queue) vs Control (random order, no AI), randomised","arxiv-2511.13860v1; Section 3.3, Table 3 (L86-L96), column 'Full Corpus', rows Aware/Control Time per Email, % Time Savings, p-value",-0.11765804346823246,0.1632205919720678,True,True,"percentage time saving 11.1% -> c = -0.111; theta = ln(1 + (-0.111)) = -0.117658; APPROXIMATE SE from two-sided p: SE = |theta|/z_(1-0.471/2) = 0.117658/0.720853 = 0.163221 (ratio of displayed means ln(50.5/56.9) = -0.119322, sensitivity only)",
S5,"-0.278 (Treat x Post, Output, panel A, column (3))",SE (0.295) (clustered at user level),"N = 18332 (user-day observations, users with affiliation)","Difference-in-differences coefficient (Italy vs Austria/France, post vs pre ChatGPT ban) on an aggregate count of GitHub Output events (Push, Pull Requests, comments, Create, Issues)","arxiv-2304.09339v1; Section 4, Table 1 (L80-L88), panel A 'Pre 27/28.03 - Post 03/04.04', row Treat x Post, column (3) Output (users with affiliation)",,,False,False,none: no transformation maps an effect on output counts onto ln(T_AI/T_noAI),"Not harmonisable: the effect is on output counts (GitHub events) around the withdrawal of ChatGPT access in a non-randomised difference-in-differences design, not a comparison of task-completion times with and without AI on the same tasks."