---------------------------------------------------------------------------------------------------- name: log: /Users/carlosmendez/Documents/GitHub/starter-academic-v501/content/post/stata_sdid_stag > gered/analysis.log log type: text opened on: 8 Jun 2026, 10:06:48 . . *------------------------------------------------------------------------------- . * 1. DATA -- load, persist, document the staggered structure . *------------------------------------------------------------------------------- . capture confirm file quota_example.dta . if _rc { . webuse set www.damianclarke.net/stata/ . webuse quota_example, clear . save quota_example.dta, replace . } . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . . describe womparl quota lngdp country year quotaYear Variable Storage Display Value name type format label Variable label ---------------------------------------------------------------------------------------------------- womparl float %9.0g Women in parliament quota float %9.0g Parliamentary gender quota lngdp float %9.0g log(GDP) country str30 %30s Country year int %8.0g Year quotaYear float %9.0g Year quota adopted . summarize womparl quota lngdp Variable | Obs Mean Std. dev. Min Max -------------+--------------------------------------------------------- womparl | 3,094 14.96531 10.97328 0 63.8 quota | 3,094 .0303814 .1716621 0 1 lngdp | 2,990 9.154291 1.136837 5.8701 11.61789 . encode country, gen(id) . xtset id year Panel variable: id (strongly balanced) Time variable: year, 1990 to 2015 Delta: 1 unit . xtdescribe id: 1, 2, ..., 119 n = 119 year: 1990, 1991, ..., 2015 T = 26 Delta(year) = 1 unit Span(year) = 26 periods (id*year uniquely identifies each observation) Distribution of T_i: min 5% 25% 50% 75% 95% max 26 26 26 26 26 26 26 Freq. Percent Cum. | Pattern ---------------------------+---------------------------- 119 100.00 100.00 | 11111111111111111111111111 ---------------------------+---------------------------- 119 100.00 | XXXXXXXXXXXXXXXXXXXXXXXXXX . . * Adoption cohort = first year a country is treated (quotaYear ships with the data). . * Verify it equals the first treated year, then tabulate cohort sizes. . bysort country (year): egen firsttreat = min(cond(quota==1, year, .)) (2,860 missing values generated) . gen byte evertreat = !missing(firsttreat) . capture assert firsttreat == quotaYear if evertreat // sanity check (quotaYear ships with the da > ta) . di as result "firsttreat == quotaYear check rc = " _rc firsttreat == quotaYear check rc = 0 . di as result _n "=== Adoption cohorts (one row per country) ===" === Adoption cohorts (one row per country) === . preserve . keep country firsttreat . duplicates drop Duplicates in terms of all variables (2,975 observations deleted) . tab firsttreat, missing firsttreat | Freq. Percent Cum. ------------+----------------------------------- 2000 | 1 0.84 0.84 2002 | 2 1.68 2.52 2003 | 2 1.68 4.20 2005 | 1 0.84 5.04 2010 | 1 0.84 5.88 2012 | 1 0.84 6.72 2013 | 1 0.84 7.56 . | 110 92.44 100.00 ------------+----------------------------------- Total | 119 100.00 . restore . count if quota==1 94 . di as result "treated country-year observations = " r(N) treated country-year observations = 94 . . * cohort sizes (countries per adoption year) for later merges + the web app . preserve . keep if evertreat (2,860 observations deleted) . bysort country: keep if _n==1 (225 observations deleted) . contract firsttreat, freq(n_treated) . rename firsttreat cohort . list, noobs +-------------------+ | cohort n_trea~d | |-------------------| | 2000 1 | | 2002 2 | | 2003 2 | | 2005 1 | | 2010 1 | |-------------------| | 2012 1 | | 2013 1 | +-------------------+ . tempfile csize . save `csize', replace (file /tmp/claude-501/St23259.000003 not found) file /tmp/claude-501/St23259.000003 saved as .dta format . restore . . *------------------------------------------------------------------------------- . * 2. EDA -- panelview (staggered structure) + a site-coloured trend figure . *------------------------------------------------------------------------------- . * (2a) Treatment-timing heatmap: the staggered "staircase" . panelview womparl quota, i(country) t(year) type(treat) bytiming /// > xtitle("Year") ytitle("Country (sorted by adoption timing)") /// > title("Staggered adoption of parliamentary gender quotas", size(medium)) /// > ylabdist(10) xlabdist(5) # Variable # Missing % Missing -------------------------------------------- 1 womparl 0 0.0 2 quota 0 0.0 Missing for | how many | variables? | Freq. Percent Cum. ------------+----------------------------------- 0 | 3,094 100.00 100.00 ------------+----------------------------------- Total | 3,094 100.00 Note: White cells represent missing values/observations in data. . graph export "stata_sdid_staggered_panelview_treat.png", replace width(2400) file stata_sdid_staggered_panelview_treat.png written in PNG format . . * (2b) Outcome trajectories, treated (orange) vs control (blue) . panelview womparl quota, i(country) t(year) type(outcome) /// > xtitle("Year") ytitle("% women in parliament") /// > title("Women in parliament: treated vs. control trajectories", size(medium)) now display lines of continuous outcome . graph export "stata_sdid_staggered_panelview_outcome.png", replace width(2400) file stata_sdid_staggered_panelview_outcome.png written in PNG format . . * (2c) Mean outcome: ever-adopting vs never-adopting (site colours) . preserve . collapse (mean) womparl, by(evertreat year) . reshape wide womparl, i(year) j(evertreat) (j = 0 1) Data Long -> Wide ----------------------------------------------------------------------------- Number of observations 52 -> 26 Number of variables 3 -> 3 j variable (2 values) evertreat -> (dropped) xij variables: womparl -> womparl0 womparl1 ----------------------------------------------------------------------------- . label var womparl1 "Ever-adopting countries (mean)" . label var womparl0 "Never-adopting countries (mean)" . twoway (line womparl1 year, lcolor("$TREAT") lwidth(thick)) /// > (line womparl0 year, lcolor("$CTRL") lwidth(medthick) lpattern(dash)), /// > ytitle("% women in parliament") xtitle("") xlabel(1990(5)2015) /// > legend(order(1 "Ever-adopting (mean)" 2 "Never-adopting (mean)") /// > pos(11) ring(0) cols(1) size(small)) /// > title("Raw outcome trends by treatment group", size(medium)) /// > note("Adoption is staggered (2000-2013); a single group mean blurs the timing.") . graph export "stata_sdid_staggered_raw_trends.png", replace width(2400) file stata_sdid_staggered_raw_trends.png written in PNG format . restore . . *------------------------------------------------------------------------------- . * 3. BASELINE -- static two-way fixed-effects DiD (the biased foil) . * Under staggered timing with heterogeneous effects, this TWFE coefficient is . * a contaminated weighted average that uses already-treated units as controls . * (Goodman-Bacon 2021; de Chaisemartin & D'Haultfoeuille 2020). Reported only . * as a benchmark, NOT as a credible ATT. . *------------------------------------------------------------------------------- . capture reghdfe womparl quota, absorb(id year) vce(cluster id) . if _rc { . di as error "reghdfe unavailable (rc=" _rc "); falling back to xtreg." . xtreg womparl quota i.year, fe vce(cluster id) . } . scalar twfe_att = _b[quota] . scalar twfe_se = _se[quota] . di as result "Static TWFE 'ATT' (biased foil) = " twfe_att " (cluster SE " twfe_se ")" Static TWFE 'ATT' (biased foil) = 7.961266 (cluster SE 3.77388) . . *------------------------------------------------------------------------------- . * 4. MAIN STAGGERED SDID -- bootstrap inference + cohort-specific effects . *------------------------------------------------------------------------------- . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . . sdid womparl country year quota, vce(bootstrap) seed(1213) Bootstrap replications (50). This may take some time. ----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 .................................................. 50 Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 8.03410 3.74040 2.15 0.032 0.70305 15.36516 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar sdid_att = e(ATT) . scalar sdid_se = e(se) . scalar sdid_cil = e(ATT_l) . scalar sdid_cir = e(ATT_r) . di as result "SDID staggered ATT = " sdid_att " SE = " sdid_se /// > " 95% CI = [" sdid_cil ", " sdid_cir "]" SDID staggered ATT = 8.034102 SE = 3.740402 95% CI = [.70304879, 15.365155] . . * ---- e(tau): cohort-specific ATTs (Tau, Std.Err., Time) ---- . matrix Tau = e(tau) . matrix list Tau Tau[7,3] Tau Std.Err. Time r1 8.3888685 .68278345 2000 r2 6.9677465 .64102999 2002 r3 13.952256 9.1289943 2003 r4 -3.4505431 .75603453 2005 r5 2.7490355 .44799502 2010 r6 21.762716 .91589982 2012 r7 -.82032354 .83151601 2013 . preserve . clear . svmat Tau // Tau1=tau, Tau2=se, Tau3=cohort year number of observations will be reset to 7 Press any key to continue, or Break to abort Number of observations (_N) was 0, now 7. . rename (Tau1 Tau2 Tau3) (tau se cohort) . gen lci = tau - 1.96*se . gen uci = tau + 1.96*se . gen t_post = 2015 - cohort + 1 . merge 1:1 cohort using `csize', nogen Result Number of obs ----------------------------------------- Not matched 0 Matched 7 ----------------------------------------- . * aggregation weight = treated (unit x post-period) share -> reproduces overall ATT . gen w_raw = n_treated * t_post . egen w_tot = total(w_raw) . gen agg_weight = w_raw / w_tot . gsort cohort . order cohort tau se lci uci n_treated t_post agg_weight . list cohort tau se n_treated t_post agg_weight, noobs +--------------------------------------------------------------+ | cohort tau se n_trea~d t_post agg_we~t | |--------------------------------------------------------------| | 2000 8.388868 .6827834 1 16 .1702128 | | 2002 6.967746 .64103 2 14 .2978723 | | 2003 13.95226 9.128994 2 13 .2765957 | | 2005 -3.450543 .7560346 1 11 .1170213 | | 2010 2.749036 .447995 1 6 .0638298 | |--------------------------------------------------------------| | 2012 21.76272 .9158998 1 4 .0425532 | | 2013 -.8203235 .831516 1 3 .0319149 | +--------------------------------------------------------------+ . * verification: weighted average of cohort taus = overall ATT . gen wtau = agg_weight*tau . egen check_att = total(wtau) . di as result "Sum of weighted cohort taus = " check_att " (should match overall ATT " sdid_at > t ")" Sum of weighted cohort taus = 8.0341015 (should match overall ATT 8.034102) . drop w_raw w_tot wtau check_att . export delimited cohort tau se lci uci n_treated t_post agg_weight /// > using "web_app/data/cohorts.csv", replace file web_app/data/cohorts.csv saved . restore . . * ---- cohort-ATT figure: tau_a with 95% CI, zero line, aggregate-ATT line ---- . preserve . clear . svmat Tau // Tau1=tau, Tau2=se, Tau3=cohort year number of observations will be reset to 7 Press any key to continue, or Break to abort Number of observations (_N) was 0, now 7. . rename (Tau1 Tau2 Tau3) (tau se cohort) . gen lci = tau - 1.96*se . gen uci = tau + 1.96*se . twoway (rcap lci uci cohort, lcolor("$CTRL")) /// > (scatter tau cohort, mcolor("$TREAT") msize(large) ms(d)), /// > yline(0, lcolor(gs10) lpattern(dash)) /// > yline(`=sdid_att', lcolor("$TEAL") lwidth(medthick)) /// > xlabel(2000 2002 2003 2005 2010 2012 2013, angle(45)) /// > ytitle("Adoption-cohort ATT (pp)") xtitle("Adoption year (cohort)") /// > legend(off) /// > title("Cohort-specific SDID effects", size(medium)) /// > note("Teal line: overall weighted ATT (8.0 pp). Cohorts range from -3.5 (2005) to +21.8 > (2012).") . graph export "stata_sdid_staggered_cohort_taus.png", replace width(2400) file stata_sdid_staggered_cohort_taus.png written in PNG format . restore . . * ---- e(series): treated vs synthetic outcome path per cohort -> CSV ---- . matrix S = e(series) . preserve . clear . svmat S, names(col) number of observations will be reset to 26 Press any key to continue, or Break to abort Number of observations (_N) was 0, now 26. . reshape long Yco Ytr, i(year) j(cohort) (j = 2000 2002 2003 2005 2010 2012 2013) Data Wide -> Long ----------------------------------------------------------------------------- Number of observations 26 -> 182 Number of variables 15 -> 4 j variable (7 values) -> cohort xij variables: Yco2000 Yco2002 ... Yco2013 -> Yco Ytr2000 Ytr2002 ... Ytr2013 -> Ytr ----------------------------------------------------------------------------- . rename (Yco Ytr) (y_synth y_treated) . drop if missing(y_treated) & missing(y_synth) (0 observations deleted) . order cohort year y_treated y_synth . sort cohort year . export delimited using "web_app/data/series_by_cohort.csv", replace file web_app/data/series_by_cohort.csv saved . restore . . * ---- e(lambda): pre-period time weights per cohort -> CSV ---- . matrix L = e(lambda) . local cohlist 2000 2002 2003 2005 2010 2012 2013 . preserve . clear . svmat L number of observations will be reset to 27 Press any key to continue, or Break to abort Number of observations (_N) was 0, now 27. . * last row holds adoption-year labels; last column (L8) is the calendar year . local nr = rowsof(L) . drop in `nr' (1 observation deleted) . rename L8 year . local j = 0 . foreach c of local cohlist { 2. local ++j 3. rename L`j' lam`c' 4. } . reshape long lam, i(year) j(cohort) (j = 2000 2002 2003 2005 2010 2012 2013) Data Wide -> Long ----------------------------------------------------------------------------- Number of observations 26 -> 182 Number of variables 8 -> 3 j variable (7 values) -> cohort xij variables: lam2000 lam2002 ... lam2013 -> lam ----------------------------------------------------------------------------- . rename lam lambda . drop if missing(lambda) (67 observations deleted) . order cohort year lambda . sort cohort year . export delimited using "web_app/data/lambda_by_cohort.csv", replace file web_app/data/lambda_by_cohort.csv saved . restore . . * ---- donor (unit) weights per cohort -> CSV ---- . * Use returnweights: the country name is native in the data, avoiding the . * e(omega) matrix rownames (which mattitles fills with names that can carry . * spaces). A quick noinference re-fit returns the same weights. . sdid womparl country year quota, vce(noinference) returnweights Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 8.03410 . . . . . ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . preserve . keep country quotaYear omega2000 omega2002 omega2003 omega2005 omega2010 omega2012 omega2013 . duplicates drop Duplicates in terms of all variables (2,975 observations deleted) . keep if missing(quotaYear) // donors = never-treated countries (9 observations deleted) . drop quotaYear . reshape long omega, i(country) j(cohort) (j = 2000 2002 2003 2005 2010 2012 2013) Data Wide -> Long ----------------------------------------------------------------------------- Number of observations 110 -> 770 Number of variables 8 -> 3 j variable (7 values) -> cohort xij variables: omega2000 omega2002 ... omega2013 -> omega ----------------------------------------------------------------------------- . drop if missing(omega) | omega==0 // keep nonzero donors (321 observations deleted) . order cohort country omega . gsort cohort -omega . export delimited using "web_app/data/omega_by_cohort.csv", replace file web_app/data/omega_by_cohort.csv saved . restore . . * ---- treated-vs-synthetic path for the 2002 cohort (the worked example) ---- . * SDID matches the pre-period TREND, not the level (the unit fixed effect . * absorbs the level gap). To visualise the counterfactual we anchor the . * synthetic to the treated cohort by its lambda-weighted pre-period gap . * (exactly the baseline SDID differences against; see the sdid post). . preserve . import delimited "web_app/data/lambda_by_cohort.csv", clear (encoding automatically selected: ISO-8859-9) (3 vars, 115 obs) . keep if cohort==2002 (103 observations deleted) . keep year lambda . tempfile l2002 . save `l2002', replace (file /tmp/claude-501/St23259.00000b not found) file /tmp/claude-501/St23259.00000b saved as .dta format . . import delimited "web_app/data/series_by_cohort.csv", clear (encoding automatically selected: ISO-8859-1) (4 vars, 182 obs) . keep if cohort==2002 (156 observations deleted) . keep year y_treated y_synth . merge 1:1 year using `l2002', nogen Result Number of obs ----------------------------------------- Not matched 14 from master 14 from using 0 Matched 12 ----------------------------------------- . replace lambda = 0 if missing(lambda) (14 real changes made) . gen double pg = lambda*(y_synth - y_treated) if year<2002 (14 missing values generated) . egen double offset = total(pg) . gen y_synth_anch = y_synth - offset . di as result "2002 cohort lambda-weighted pre-period gap (anchor) = " offset[1] 2002 cohort lambda-weighted pre-period gap (anchor) = 10.444628 . twoway (line y_treated year, lcolor("$TREAT") lwidth(thick)) /// > (line y_synth_anch year, lcolor("$CTRL") lwidth(medthick) lpattern(dash)), /// > xline(2001.5, lcolor(gs10)) /// > ytitle("% women in parliament") xtitle("") xlabel(1990(5)2015) /// > legend(order(1 "Treated cohort (2002)" 2 "Synthetic control (anchored)") /// > pos(11) ring(0) cols(1) size(small)) /// > title("SDID counterfactual for the 2002 cohort", size(medium)) /// > note("Synthetic anchored to the treated cohort by its {&lambda}-weighted pre-2002 gap; > the post-2002 divergence is the effect.") . graph export "stata_sdid_staggered_cohort2002_path.png", replace width(2400) file stata_sdid_staggered_cohort2002_path.png written in PNG format . restore . . * ---- 2002-cohort pre-period time weights (lambda) bar chart ---- . preserve . import delimited "web_app/data/lambda_by_cohort.csv", clear (encoding automatically selected: ISO-8859-9) (3 vars, 115 obs) . keep if cohort==2002 (103 observations deleted) . twoway (bar lambda year, color("$CTRL") barwidth(0.8)), /// > ytitle("SDID time weight ({&lambda})") xtitle("") /// > xlabel(1990(2)2001, angle(45)) legend(off) /// > title("Where SDID looks: 2002-cohort pre-period time weights", size(medium)) /// > note("Weight concentrates on the years just before 2002 -- the pre-period most like the > post-period.") . graph export "stata_sdid_staggered_lambda.png", replace width(2400) file stata_sdid_staggered_lambda.png written in PNG format . restore . . *------------------------------------------------------------------------------- . * 5. COVARIATES -- optimized (Arkhangelsky et al.) vs projected (Kranz 2022) . * sdid needs a balanced panel, so drop the 104 obs with missing lngdp first. . *------------------------------------------------------------------------------- . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . drop if missing(lngdp) (104 observations deleted) . . sdid womparl country year quota, vce(bootstrap) seed(2022) covariates(lngdp, optimized) Bootstrap replications (50). This may take some time. ----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 .................................................. 50 Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 8.05150 3.04660 2.64 0.008 2.08027 14.02273 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar att_opt = e(ATT) . scalar se_opt = e(se) . di as result "SDID + lngdp (optimized) ATT = " att_opt " SE = " se_opt SDID + lngdp (optimized) ATT = 8.051498 SE = 3.046601 . . sdid womparl country year quota, vce(bootstrap) seed(1213) covariates(lngdp, projected) Bootstrap replications (50). This may take some time. ----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 .................................................. 50 Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 8.05927 3.11913 2.58 0.010 1.94589 14.17264 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar att_prj = e(ATT) . scalar se_prj = e(se) . di as result "SDID + lngdp (projected) ATT = " att_prj " SE = " se_prj SDID + lngdp (projected) ATT = 8.059266 SE = 3.119126 . . *------------------------------------------------------------------------------- . * 6. EVENT STUDY -- sdid_event on the full staggered panel + the 2002 cohort . *------------------------------------------------------------------------------- . * (6a) Full staggered panel: aggregated ATT + cohort-aggregated dynamic effects . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . drop if missing(lngdp) (104 observations deleted) . sdid_event womparl country year quota, vce(bootstrap) brep(100) effects(8) /// > placebo(5) covariates(lngdp) Synthetic Difference-in-differences Boostrap replications (100), bootstrap mode. |0% ----------------------------------------- 100%| |.................................................| | Estimate SE LB CI UB CI Switchers -------------+------------------------------------------------------ ATT | 8.05289 3.007588 2.158017 13.94776 9 Effect_1 | 6.739649 3.174403 .5178192 12.96148 9 Effect_2 | 7.38131 2.996325 1.508514 13.25411 9 Effect_3 | 6.749378 3.048767 .7737942 12.72496 9 Effect_4 | 8.593622 2.475094 3.742437 13.44481 8 Effect_5 | 6.304031 2.458463 1.485443 11.12262 7 Effect_6 | 8.35644 3.105559 2.269545 14.44334 7 Effect_7 | 8.191521 3.514041 1.304 15.07904 6 Effect_8 | 8.248533 3.525711 1.338139 15.15893 6 Placebo_1 | 6.71234 4.419751 -1.950371 15.37505 6 Placebo_2 | 7.088328 4.635759 -1.99776 16.17442 6 Placebo_3 | 8.233008 5.426494 -2.402921 18.86894 6 Placebo_4 | 12.4084 4.672981 3.249356 21.56744 5 Placebo_5 | 11.90183 4.662348 2.763633 21.04004 5 . matrix Hfull = e(H) . di as result "sdid_event full-panel aggregated ATT = " Hfull[1,1] " SE = " Hfull[1,2] sdid_event full-panel aggregated ATT = 8.0528896 SE = 3.007588 . . * (6b) Clean event study on the 2002 cohort (the package authors' worked example). . * Effect_l = event time (l-1); Placebo_l = event time (-l). . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . keep if quotaYear==2002 | quotaYear==. (182 observations deleted) . drop if missing(lngdp) (104 observations deleted) . sdid_event womparl country year quota, vce(placebo) brep(100) placebo(all) covariates(lngdp) Synthetic Difference-in-differences Boostrap replications (100), placebo mode. |0% ----------------------------------------- 100%| |.................................................| | Estimate SE LB CI UB CI Switchers -------------+------------------------------------------------------ ATT | 6.853472 2.9215 1.127332 12.57961 2 Effect_1 | 4.086404 1.064752 1.99949 6.173319 2 Effect_2 | 9.164442 1.572465 6.082411 12.24647 2 Effect_3 | 7.938504 2.21875 3.589755 12.28725 2 Effect_4 | 7.198138 2.554 2.192299 12.20398 2 Effect_5 | 7.00087 2.797605 1.517563 12.48418 2 Effect_6 | 6.828273 3.259383 .4398817 13.21666 2 Effect_7 | 6.864062 3.63997 -.2702804 13.9984 2 Effect_8 | 6.302866 3.975074 -1.488279 14.09401 2 Effect_9 | 6.089694 3.835474 -1.427835 13.60722 2 Effect_10 | 8.778086 4.282917 .3835691 17.1726 2 Effect_11 | 8.185586 4.674373 -.9761843 17.34736 2 Effect_12 | 6.34902 5.128008 -3.701875 16.39992 2 Effect_13 | 5.587233 5.35509 -4.908743 16.08321 2 Effect_14 | 5.575424 5.452834 -5.11213 16.26298 2 Placebo_1 | -.2184166 .5740168 -1.34349 .9066564 2 Placebo_2 | .242148 1.137022 -1.986414 2.47071 2 Placebo_3 | .1393919 1.163741 -2.14154 2.420324 2 Placebo_4 | .2832749 1.334074 -2.33151 2.89806 2 Placebo_5 | .4654359 1.3908 -2.260532 3.191404 2 Placebo_6 | .5156405 1.341509 -2.113718 3.144999 2 Placebo_7 | .5274185 1.343222 -2.105296 3.160133 2 Placebo_8 | .5443867 1.208532 -1.824336 2.913109 2 Placebo_9 | .7578531 1.192515 -1.579475 3.095182 2 Placebo_10 | .4876656 1.260834 -1.983569 2.9589 2 Placebo_11 | .4131066 1.008039 -1.562651 2.388864 2 Placebo_12 | .1307408 1.61198 -3.02874 3.290222 2 . matrix H = e(H) . matrix list H H[27,5] Estimate SE LB CI UB CI Switchers ATT 6.8534716 2.9214996 1.1273324 12.579611 2 Effect_1 4.0864041 1.0647522 1.9994897 6.1733185 2 Effect_2 9.1644416 1.5724646 6.0824109 12.246472 2 Effect_3 7.9385045 2.21875 3.5897545 12.287254 2 Effect_4 7.1981381 2.5539996 2.1922989 12.203977 2 Effect_5 7.0008696 2.7976052 1.5175635 12.484176 2 Effect_6 6.828273 3.2593833 .43988175 13.216664 2 Effect_7 6.8640615 3.6399704 -.27028038 13.998403 2 Effect_8 6.3028661 3.9750741 -1.4882792 14.094011 2 Effect_9 6.0896941 3.8354741 -1.4278351 13.607223 2 Effect_10 8.7780859 4.2829167 .38356912 17.172603 2 Effect_11 8.1855863 4.6743727 -.97618427 17.347357 2 Effect_12 6.3490205 5.1280081 -3.7018753 16.399916 2 Effect_13 5.5872333 5.35509 -4.9087431 16.08321 2 Effect_14 5.5754244 5.4528338 -5.1121298 16.262979 2 Placebo_1 -.21841663 .57401684 -1.3434896 .90665636 2 Placebo_2 .24214798 1.1370217 -1.9864145 2.4707104 2 Placebo_3 .13939188 1.163741 -2.1415405 2.4203242 2 Placebo_4 .28327488 1.3340739 -2.33151 2.8980598 2 Placebo_5 .46543586 1.3908 -2.2605322 3.1914039 2 Placebo_6 .51564047 1.3415094 -2.1137179 3.1449988 2 Placebo_7 .52741851 1.3432216 -2.1052958 3.1601328 2 Placebo_8 .54438668 1.2085319 -1.8243358 2.9131091 2 Placebo_9 .75785315 1.1925146 -1.5794754 3.0951817 2 Placebo_10 .48766562 1.2608338 -1.9835685 2.9588998 2 Placebo_11 .41310657 1.0080394 -1.5626506 2.3888637 2 Placebo_12 .13074083 1.61198 -3.02874 3.2902217 2 . . local Lg = 2015 - 2002 + 1 // 14 post-treatment (dynamic) effects . local Lpl = 2002 - 1990 // 12 pre-treatment placebos . preserve . clear . svmat H number of observations will be reset to 27 Press any key to continue, or Break to abort Number of observations (_N) was 0, now 27. . rename (H1 H2 H3 H4 H5) (coef se ci_l ci_u switchers) . gen row = _n . drop if row==1 // drop the aggregate ATT row (1 observation deleted) . gen event_time = . (26 missing values generated) . gen str4 period_type = "" (26 missing values generated) . replace event_time = row-2 if row>=2 & row<=1+`Lg' (14 real changes made) . replace period_type = "post" if row>=2 & row<=1+`Lg' (14 real changes made) . replace event_time = -(row-1-`Lg') if row>=2+`Lg' & row<=1+`Lg'+`Lpl' (12 real changes made) . replace period_type = "pre" if row>=2+`Lg' & row<=1+`Lg'+`Lpl' (12 real changes made) . keep event_time coef se ci_l ci_u period_type . sort event_time . list, noobs +-------------------------------------------------------------------+ | coef se ci_l ci_u event_~e period~e | |-------------------------------------------------------------------| | .1307408 1.61198 -3.02874 3.290222 -12 pre | | .4131066 1.008039 -1.562651 2.388864 -11 pre | | .4876656 1.260834 -1.983569 2.9589 -10 pre | | .7578532 1.192515 -1.579475 3.095182 -9 pre | | .5443867 1.208532 -1.824336 2.913109 -8 pre | |-------------------------------------------------------------------| | .5274185 1.343222 -2.105296 3.160133 -7 pre | | .5156405 1.341509 -2.113718 3.144999 -6 pre | | .4654359 1.3908 -2.260532 3.191404 -5 pre | | .2832749 1.334074 -2.33151 2.89806 -4 pre | | .1393919 1.163741 -2.141541 2.420324 -3 pre | |-------------------------------------------------------------------| | .242148 1.137022 -1.986414 2.470711 -2 pre | | -.2184166 .5740168 -1.34349 .9066564 -1 pre | | 4.086404 1.064752 1.99949 6.173318 0 post | | 9.164442 1.572465 6.082411 12.24647 1 post | | 7.938505 2.21875 3.589755 12.28725 2 post | |-------------------------------------------------------------------| | 7.198138 2.554 2.192299 12.20398 3 post | | 7.00087 2.797605 1.517563 12.48418 4 post | | 6.828273 3.259383 .4398817 13.21666 5 post | | 6.864061 3.63997 -.2702804 13.9984 6 post | | 6.302866 3.975074 -1.488279 14.09401 7 post | |-------------------------------------------------------------------| | 6.089694 3.835474 -1.427835 13.60722 8 post | | 8.778086 4.282917 .3835691 17.1726 9 post | | 8.185586 4.674373 -.9761842 17.34736 10 post | | 6.34902 5.128008 -3.701875 16.39992 11 post | | 5.587233 5.35509 -4.908743 16.08321 12 post | |-------------------------------------------------------------------| | 5.575424 5.452834 -5.11213 16.26298 13 post | +-------------------------------------------------------------------+ . export delimited using "web_app/data/event_study.csv", replace file web_app/data/event_study.csv saved . . * headline event-study figure . twoway (rarea ci_l ci_u event_time, color("${CTRL}%35") lwidth(none)) /// > (line coef event_time, lcolor("$TREAT") lwidth(medthick)) /// > (scatter coef event_time, mcolor("$TREAT") msize(small) ms(O)), /// > yline(0, lcolor("$TEAL") lpattern(dash)) /// > xline(-0.5, lcolor(gs9) lpattern(solid)) /// > xlabel(-12(2)13) xtitle("Years relative to quota adoption (event time)") /// > ytitle("Effect on women in parliament (pp)") /// > legend(order(3 "Point estimate" 1 "95% CI") pos(11) ring(0) cols(1) size(small)) /// > title("Event-study SDID for the 2002 cohort (sdid_event)", size(medium)) /// > note("Pre-period placebos hug zero (parallel trends); post-period effects trace the dyn > amic ATT.") . graph export "stata_sdid_staggered_event_study.png", replace width(2400) file stata_sdid_staggered_event_study.png written in PNG format . restore . . *------------------------------------------------------------------------------- . * 7. INFERENCE -- bootstrap vs placebo vs jackknife (paper's 2-cohort subsample) . * Drop the five single-country cohorts so jackknife (needs >1 treated unit per . * period) is defined; only the 2002 & 2003 cohorts remain. . *------------------------------------------------------------------------------- . use quota_example.dta, clear (Balanced panel from Bhalotra, Clarke, Gomes & Venkataramani (2023)) . label variable quota "Parliamentary gender quota" . drop if inlist(country,"Algeria","Kenya","Samoa","Swaziland","Tanzania") (130 observations deleted) . . sdid womparl country year quota, vce(bootstrap) seed(1213) Bootstrap replications (50). This may take some time. ----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 .................................................. 50 Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 10.33066 4.72911 2.18 0.029 1.06178 19.59954 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar b_att = e(ATT) . scalar b_se = e(se) . scalar b_cil = e(ATT_l) . scalar b_cir = e(ATT_r) . sdid womparl country year quota, vce(placebo) seed(1213) Placebo replications (50). This may take some time. ----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5 .................................................. 50 Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 10.33066 2.34040 4.41 0.000 5.74356 14.91776 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar p_att = e(ATT) . scalar p_se = e(se) . scalar p_cil = e(ATT_l) . scalar p_cir = e(ATT_r) . sdid womparl country year quota, vce(jackknife) Synthetic Difference-in-Differences Estimator ----------------------------------------------------------------------------- womparl | ATT Std. Err. t P>|t| [95% Conf. Interval] -------------+--------------------------------------------------------------- quota | 10.33066 6.00560 1.72 0.085 -1.44009 22.10141 ----------------------------------------------------------------------------- 95% CIs and p-values are based on large-sample approximations. Refer to Arkhangelsky et al., (2021) for theoretical derivations. . scalar j_att = e(ATT) . scalar j_se = e(se) . scalar j_cil = e(ATT_l) . scalar j_cir = e(ATT_r) . . clear . set obs 3 Number of observations (_N) was 0, now 3. . gen str10 method = "" (3 missing values generated) . gen double att = . (3 missing values generated) . gen double se = . (3 missing values generated) . gen double ci_l = . (3 missing values generated) . gen double ci_u = . (3 missing values generated) . replace method="bootstrap" in 1 (1 real change made) . replace att=b_att in 1 (1 real change made) . replace se=b_se in 1 (1 real change made) . replace ci_l=b_cil in 1 (1 real change made) . replace ci_u=b_cir in 1 (1 real change made) . replace method="placebo" in 2 (1 real change made) . replace att=p_att in 2 (1 real change made) . replace se=p_se in 2 (1 real change made) . replace ci_l=p_cil in 2 (1 real change made) . replace ci_u=p_cir in 2 (1 real change made) . replace method="jackknife" in 3 (1 real change made) . replace att=j_att in 3 (1 real change made) . replace se=j_se in 3 (1 real change made) . replace ci_l=j_cil in 3 (1 real change made) . replace ci_u=j_cir in 3 (1 real change made) . gen tstat = att/se . gen pval = 2*(1-normal(abs(tstat))) . list, noobs +--------------------------------------------------------------------------------+ | method att se ci_l ci_u tstat pval | |--------------------------------------------------------------------------------| | bootstrap 10.33066 4.729109 1.0617767 19.599543 2.184483 .0289268 | | placebo 10.33066 2.3404 5.7435603 14.91776 4.414057 .0000101 | | jackknife 10.33066 6.005597 -1.4400938 22.101414 1.720172 .0854012 | +--------------------------------------------------------------------------------+ . export delimited using "web_app/data/inference.csv", replace file web_app/data/inference.csv saved . . * forest plot of the three inference methods (same point estimate, different SEs) . gen order = _n . twoway (rcap ci_l ci_u order, horizontal lcolor("$CTRL")) /// > (scatter order att, mcolor("$TREAT") msize(large) ms(d)), /// > xline(0, lcolor(gs10) lpattern(dash)) /// > ylabel(1 "bootstrap" 2 "placebo" 3 "jackknife", angle(0)) /// > ytitle("") xtitle("ATT on women in parliament (pp)") /// > legend(off) ysc(reverse) /// > title("Same ATT, three variance estimators (2002 & 2003 cohorts)", size(medium)) /// > note("Point estimate is identical (10.3 pp); jackknife is most conservative, placebo tighte > st.") . graph export "stata_sdid_staggered_inference.png", replace width(2400) file stata_sdid_staggered_inference.png written in PNG format . . *------------------------------------------------------------------------------- . * 8. SUMMARY TABLE -> web_app/data/atts.csv . *------------------------------------------------------------------------------- . clear . set obs 4 Number of observations (_N) was 0, now 4. . gen str28 spec = "" (4 missing values generated) . gen double att = . (4 missing values generated) . gen double se = . (4 missing values generated) . gen double ci_l = . (4 missing values generated) . gen double ci_u = . (4 missing values generated) . replace spec="Static TWFE (biased foil)" in 1 (1 real change made) . replace att=twfe_att in 1 (1 real change made) . replace se=twfe_se in 1 (1 real change made) . replace spec="SDID (no covariates)" in 2 (1 real change made) . replace att=sdid_att in 2 (1 real change made) . replace se=sdid_se in 2 (1 real change made) . replace ci_l=sdid_cil in 2 (1 real change made) . replace ci_u=sdid_cir in 2 (1 real change made) . replace spec="SDID + lngdp (optimized)" in 3 (1 real change made) . replace att=att_opt in 3 (1 real change made) . replace se=se_opt in 3 (1 real change made) . replace spec="SDID + lngdp (projected)" in 4 (1 real change made) . replace att=att_prj in 4 (1 real change made) . replace se=se_prj in 4 (1 real change made) . gen tstat = att/se . gen pval = 2*(1-normal(abs(tstat))) . replace ci_l = att - 1.96*se if missing(ci_l) (3 real changes made) . replace ci_u = att + 1.96*se if missing(ci_u) (3 real changes made) . list, noobs +-----------------------------------------------------------------------------------------------+ | spec att se ci_l ci_u tstat pval | |-----------------------------------------------------------------------------------------------| | Static TWFE (biased foil) 7.961266 3.77388 .56446111 15.358071 2.109571 .0348954 | | SDID (no covariates) 8.034102 3.740402 .70304879 15.365155 2.147925 .0317197 | | SDID + lngdp (optimized) 8.051498 3.046601 2.08016 14.022836 2.642781 .0082228 | | SDID + lngdp (projected) 8.059266 3.119126 1.945779 14.172753 2.583822 .0097712 | +-----------------------------------------------------------------------------------------------+ . export delimited spec att se ci_l ci_u pval using "web_app/data/atts.csv", replace file web_app/data/atts.csv saved . . *------------------------------------------------------------------------------- . * 9. KEY NUMBERS . *------------------------------------------------------------------------------- . di as result _n "==================== KEY NUMBERS ====================" ==================== KEY NUMBERS ==================== . di as result "Static TWFE ATT (biased foil) = " %7.2f twfe_att " SE " %5.2f twfe_se Static TWFE ATT (biased foil) = 7.96 SE 3.77 . di as result "SDID staggered ATT = " %7.2f sdid_att " SE " %5.2f sdid_se SDID staggered ATT = 8.03 SE 3.74 . di as result " 95% CI = [" %5.2f sdid_cil ", " %5.2f sdid_cir "]" 95% CI = [ 0.70, 15.37] . di as result "SDID + lngdp (optimized) = " %7.2f att_opt " SE " %5.2f se_opt SDID + lngdp (optimized) = 8.05 SE 3.05 . di as result "SDID + lngdp (projected) = " %7.2f att_prj " SE " %5.2f se_prj SDID + lngdp (projected) = 8.06 SE 3.12 . di as result "Inference subsample ATT = " %7.2f b_att Inference subsample ATT = 10.33 . di as result " bootstrap SE = " %7.2f b_se bootstrap SE = 4.73 . di as result " placebo SE = " %7.2f p_se placebo SE = 2.34 . di as result " jackknife SE = " %7.2f j_se jackknife SE = 6.01 . di as result "=====================================================" ===================================================== . . log close name: log: /Users/carlosmendez/Documents/GitHub/starter-academic-v501/content/post/stata_sdid_stag > gered/analysis.log log type: text closed on: 8 Jun 2026, 10:16:21 ----------------------------------------------------------------------------------------------------