------------------------------------------------------------------------------------------------ name: log: /Users/carlosmendez/Documents/GitHub/starter-academic-v501/content/post/stata_doubl > e_lasso/analysis.log log type: text opened on: 24 May 2026, 18:55:51 . . * Check required packages are installed (already in user's ado/plus). . foreach pkg in rlasso cvlasso lasso2 pdslasso coefplot { 2. capture which `pkg' 3. if _rc { 4. di as error "Required package not installed: `pkg'" 5. di as error "Install via: ssc install lassopack" 6. di as error " ssc install pdslasso" 7. di as error " ssc install coefplot" 8. exit 198 9. } 10. } . . * Site palette (dark theme, mirrors R version's ggplot theme_site). . * DARK_BG = #0f1729 = "15 23 41" . * DARK_PNL = #1f2b5e = "31 43 94" . * LIGHT_TX = #c8d0e0 = "200 208 224" . * LIGHTER = #e8ecf2 = "232 236 242" . * STEEL = #6a9bcc = "106 155 204" . * ORANGE = #d97757 = "217 119 87" . * TEAL = #00d4c8 = "0 212 200" . * LT_ORG = #e8956a = "232 149 106" . global C_BG "15 23 41" . global C_PANEL "31 43 94" . global C_GRID "31 43 94" . global C_TEXT "200 208 224" . global C_TXTHI "232 236 242" . global C_STEEL "106 155 204" . global C_ORG "217 119 87" . global C_TEAL "0 212 200" . global C_LTORG "232 149 106" . . * Standard graph-region options for the dark theme. . global DARKBG /// > graphregion(fcolor("$C_BG") ifcolor("$C_BG") lcolor("$C_BG") ilcolor("$C_BG")) /// > plotregion(fcolor("$C_BG") lcolor("$C_BG")) . . . * ── 1. Data loading (six CSVs from GitHub raw URLs) ─────────────── . . di _n(2) "========================================" ======================================== . di "STEP 1 - DATA LOADING (six CSVs over HTTPS)" STEP 1 - DATA LOADING (six CSVs over HTTPS) . di "========================================" ======================================== . . local BASE = "https://raw.githubusercontent.com/cmg777/starter-academic-v501/master/content/po > st/r_double_lasso/data" . . tempfile linear partialled ctrl_v ctrl_p ctrl_m . . * (a) Raw first-differenced outcomes and treatments + state IDs. . * Cols: state, Dyv, Dxv, Dyp, Dxp, Dym, Dxm . import delimited "`BASE'/levitt_linear.csv", clear varnames(1) case(preserve) (encoding automatically selected: ISO-8859-2) (7 vars, 576 obs) . gen long obs_id = _n . di " levitt_linear.csv : `=_N' obs x `=c(k)' cols" levitt_linear.csv : 576 obs x 8 cols . save "`linear'" file /var/folders/tq/t98kb27n6djgrh085g476yhc0000gn/T//S_64524.000001 saved as .dta format . . * (b) Partialled (year-FE-removed) outcomes and treatments. . * Cols: state, DxV, DyV, DxP, DyP, DxM, DyM . import delimited "`BASE'/levitt_partialled.csv", clear varnames(1) case(preserve) (encoding automatically selected: ISO-8859-1) (7 vars, 576 obs) . * drop the duplicate state col so the merge is clean . drop state . gen long obs_id = _n . di " levitt_partialled.csv : `=_N' obs x `=c(k)' cols (incl obs_id)" levitt_partialled.csv : 576 obs x 7 cols (incl obs_id) . save "`partialled'" file /var/folders/tq/t98kb27n6djgrh085g476yhc0000gn/T//S_64524.000002 saved as .dta format . . * (c)-(e) Three 284-column control matrices, one per outcome. . * Column names in source CSV use ^, *, ( ) etc. which Stata . * sanitises on import; we then rename to zv1..zv284 (etc.) . * so downstream code can address them uniformly. . foreach o in v p m { 2. local fname : word 1 of "viol prop murd" 3. if "`o'" == "v" local long "viol" 4. if "`o'" == "p" local long "prop" 5. if "`o'" == "m" local long "murd" 6. import delimited "`BASE'/levitt_controls_`long'.csv", clear varnames(1) 7. local k = 0 8. foreach var of varlist _all { 9. local ++k 10. rename `var' z`o'`k' 11. } 12. di " levitt_controls_`long'.csv : `=_N' obs x `k' cols (renamed zv1..z`o'`k')" 13. gen long obs_id = _n 14. save "`ctrl_`o''" 15. } (encoding automatically selected: ISO-8859-1) (284 vars, 576 obs) levitt_controls_viol.csv : 576 obs x 284 cols (renamed zv1..zv284) file /var/folders/tq/t98kb27n6djgrh085g476yhc0000gn/T//S_64524.000003 saved as .dta format (encoding automatically selected: ISO-8859-1) (284 vars, 576 obs) levitt_controls_prop.csv : 576 obs x 284 cols (renamed zv1..zp284) file /var/folders/tq/t98kb27n6djgrh085g476yhc0000gn/T//S_64524.000004 saved as .dta format (encoding automatically selected: ISO-8859-1) (284 vars, 576 obs) levitt_controls_murd.csv : 576 obs x 284 cols (renamed zv1..zm284) file /var/folders/tq/t98kb27n6djgrh085g476yhc0000gn/T//S_64524.000005 saved as .dta format . . * Combine into one working dataset. . use "`linear'", clear . merge 1:1 obs_id using "`partialled'", nogen Result Number of obs ----------------------------------------- Not matched 0 Matched 576 ----------------------------------------- . merge 1:1 obs_id using "`ctrl_v'", nogen Result Number of obs ----------------------------------------- Not matched 0 Matched 576 ----------------------------------------- . merge 1:1 obs_id using "`ctrl_p'", nogen Result Number of obs ----------------------------------------- Not matched 0 Matched 576 ----------------------------------------- . merge 1:1 obs_id using "`ctrl_m'", nogen Result Number of obs ----------------------------------------- Not matched 0 Matched 576 ----------------------------------------- . . * Sanity check: must be 576 obs, 48 states. . assert _N == 576 . qui levelsof state, local(states) . local nstates : word count `states' . assert `nstates' == 48 . . di _n " Merged working dataset: `=_N' obs, `nstates' clusters (states)" Merged working dataset: 576 obs, 48 clusters (states) . . . * ── 2. Convenience: outcome metadata ───────────────────────────── . . * For each outcome we have: . * raw differenced: y = Dy[v/p/m] d = Dx[v/p/m] . * year-FE partialled:y = Dy[V/P/M] d = Dx[V/P/M] . * 284 partialled controls: z[v/p/m]1..z[v/p/m]284 . * . * The partialling step (done in the original Matlab pre-processing) . * absorbs year fixed effects via Frisch-Waugh-Lovell: every variable . * v becomes v - T (T'T)^-1 T' v, where T is the matrix of year . * dummies. Regressions on the partialled vars are equivalent to . * regressing the raw differences while controlling for year dummies, . * with one less degree of freedom. . . * Will use these foreach loops repeatedly. . * Iteration uses short prefixes v/p/m (matching the renamed control vars . * zv1..zv284, zp1..zp284, zm1..zm284). Look-up helpers below map the . * short prefix to outcome label and raw/partialled variable names. . * . * prefix -> label y (raw) d (raw) y (partialled) d (partialled) . * v -> Violent crime Dyv Dxv DyV DxV . * p -> Property crime Dyp Dxp DyP DxP . * m -> Murder Dym Dxm DyM DxM . . . * ── 3. Cluster-robust SEs ───────────────────────────────────────── . . * Stata's vce(cluster state) on -regress- applies exactly the . * small-sample correction used in the Fitzgerald et al. (2026) . * replication code: . * . * V_cluster = (N-1)/(N-k) * G/(G-1) * (X'X)^-1 . S . (X'X)^-1 . * . * with G = 48 clusters of 12 observations each. No extra code needed. . . . * ── 4. Estimator A - First-difference OLS (no controls) ────────── . . * ESTIMAND. Throughout the five estimators below the parameter of . * interest is alpha, the average partial effect of (first-differenced) . * effective abortion rate on (first-differenced) state crime rate. . * Identification rests on: . * (1) conditional independence given the 284 partialled controls; and . * (2) parallel trends in levels (state FEs absorbed by differencing, . * year FEs by the partialling step). . . di _n(2) "========================================" ======================================== . di "STEP 4 - FIRST-DIFFERENCE OLS (no controls)" STEP 4 - FIRST-DIFFERENCE OLS (no controls) . di "========================================" ======================================== . . * Storage for downstream table. . tempname FD_b FD_se FD_n . mat `FD_b' = J(3, 1, .) . mat `FD_se' = J(3, 1, .) . . local row = 0 . foreach o in v p m { 2. local ++row 3. local lab = cond("`o'"=="v","Violent crime",cond("`o'"=="p","Property crime","Murder")) 4. local Y = cond("`o'"=="v","Dyv", cond("`o'"=="p","Dyp","Dym")) 5. local D = cond("`o'"=="v","Dxv", cond("`o'"=="p","Dxp","Dxm")) 6. di _n " Outcome: `lab' (y=`Y', d=`D')" 7. qui regress `Y' `D', noconstant vce(cluster state) 8. mat `FD_b'[`row',1] = _b[`D'] 9. mat `FD_se'[`row',1] = _se[`D'] 10. di as text " alpha_hat = " as result %9.4f _b[`D'] /// > as text " SE = " as result %9.4f _se[`D'] 11. } Outcome: Violent crime (y=Dyv, d=Dxv) alpha_hat = -0.1521 SE = 0.0337 Outcome: Property crime (y=Dyp, d=Dxp) alpha_hat = -0.1084 SE = 0.0219 Outcome: Murder (y=Dym, d=Dxm) alpha_hat = -0.2039 SE = 0.0667 . . . * ── 5. Estimator B - OLS with all 284 controls ─────────────────── . . di _n(2) "========================================" ======================================== . di "STEP 5 - OLS WITH ALL 284 CONTROLS (kitchen-sink)" STEP 5 - OLS WITH ALL 284 CONTROLS (kitchen-sink) . di "========================================" ======================================== . di "Feasible because p=284 < n=576: OLS technically inverts. But" Feasible because p=284 < n=576: OLS technically inverts. But . di "many controls are near-collinear, so SEs balloon. This is what" many controls are near-collinear, so SEs balloon. This is what . di "motivates LASSO: keep the controls that matter, drop the rest." motivates LASSO: keep the controls that matter, drop the rest. . . tempname OLS_b OLS_se OLS_n . mat `OLS_b' = J(3, 1, .) . mat `OLS_se' = J(3, 1, .) . mat `OLS_n' = J(3, 1, .) . . local row = 0 . foreach o in v p m { 2. local ++row 3. local lab = cond("`o'"=="v","Violent crime",cond("`o'"=="p","Property crime","Murder")) 4. local Y = cond("`o'"=="v","DyV", cond("`o'"=="p","DyP","DyM")) 5. local D = cond("`o'"=="v","DxV", cond("`o'"=="p","DxP","DxM")) 6. di _n " Outcome: `lab' (y=`Y', d=`D')" 7. qui regress `Y' `D' z`o'1-z`o'284, noconstant vce(cluster state) 8. mat `OLS_b'[`row',1] = _b[`D'] 9. mat `OLS_se'[`row',1] = _se[`D'] 10. mat `OLS_n'[`row',1] = 284 11. di as text " alpha_hat = " as result %9.4f _b[`D'] /// > as text " SE = " as result %9.4f _se[`D'] /// > as text " (using 284 controls)" 12. } Outcome: Violent crime (y=DyV, d=DxV) alpha_hat = 0.0134 SE = 0.7149 (using 284 controls) Outcome: Property crime (y=DyP, d=DxP) alpha_hat = -0.1950 SE = 0.2236 (using 284 controls) Outcome: Murder (y=DyM, d=DxM) alpha_hat = 2.3411 SE = 2.7831 (using 284 controls) . . . * ── 6. Estimator C - Post-Structural LASSO (PSL) ───────────────── . . * PSL = one LASSO on (d, X) -> y with the treatment d FORCED IN . * (penalty.factor=0 in R; pnotpen() in Stata's rlasso). The selected . * controls are passed to plain OLS for the final alpha. . * . * DESIGN NOTE: The R companion uses cv.glmnet for PSL (3-fold CV). . * Stata's cvlasso supports notpen(), but its CV path through 100+ . * lambda values with notpen-partialling is dramatically slower than . * cv.glmnet's coordinate descent. We use the Belloni-Chernozhukov- . * Hansen RIGOROUS penalty via rlasso instead, with d pinned via . * pnotpen(). This is fast, deterministic, and answers the same . * conceptual question PSL was designed to ask: "one LASSO with the . * treatment forced in, then post-OLS". The penalty rule differs . * across the two implementations but the recipe is identical. . * . * WHY POST-OLS, not LASSO coefficients? LASSO shrinks the . * coefficients of variables it keeps toward zero - that introduces . * bias in alpha. Refitting with plain OLS on the selected variables . * removes the shrinkage. Throughout this script LASSO is used for . * SELECTION only; the final alpha always comes from OLS. . . di _n(2) "========================================" ======================================== . di "STEP 6 - POST-STRUCTURAL LASSO (PSL)" STEP 6 - POST-STRUCTURAL LASSO (PSL) . di "========================================" ======================================== . di "One rlasso on (d, X) -> y with d pinned (pnotpen)," One rlasso on (d, X) -> y with d pinned (pnotpen), . di "then OLS on d + selected controls. Rigorous penalty (c=1.1, gamma=0.05)." then OLS on d + selected controls. Rigorous penalty (c=1.1, gamma=0.05). . . tempname PSL_b PSL_se PSL_n . mat `PSL_b' = J(3, 1, .) . mat `PSL_se' = J(3, 1, .) . mat `PSL_n' = J(3, 1, .) . . local row = 0 . foreach o in v p m { 2. local ++row 3. local lab = cond("`o'"=="v","Violent crime",cond("`o'"=="p","Property crime","Murder")) 4. local Y = cond("`o'"=="v","DyV", cond("`o'"=="p","DyP","DyM")) 5. local D = cond("`o'"=="v","DxV", cond("`o'"=="p","DxP","DxM")) 6. di _n " Outcome: `lab' (y=`Y', d=`D')" 7. qui rlasso `Y' `D' z`o'1-z`o'284, nocons pnotpen(`D') c(1.1) gamma(0.05) 8. local sel "`e(selected)'" 9. * e(selected) includes pnotpen variable(s) (here: DxV). Strip d out . * so we can pass clean "d + controls" to the post-OLS regression. . local sel : list sel - D 10. local nsel : word count `sel' 11. qui regress `Y' `D' `sel', noconstant vce(cluster state) 12. mat `PSL_b'[`row',1] = _b[`D'] 13. mat `PSL_se'[`row',1] = _se[`D'] 14. mat `PSL_n'[`row',1] = `nsel' 15. di as text " alpha_hat = " as result %9.4f _b[`D'] /// > as text " SE = " as result %9.4f _se[`D'] /// > as text " | `nsel' controls selected" 16. } Outcome: Violent crime (y=DyV, d=DxV) alpha_hat = -0.1553 SE = 0.0330 | 0 controls selected Outcome: Property crime (y=DyP, d=DxP) alpha_hat = -0.0665 SE = 0.0244 | 1 controls selected Outcome: Murder (y=DyM, d=DxM) alpha_hat = -0.2397 SE = 0.0635 | 1 controls selected . . . * ── 7. Estimator D - Double LASSO, rigorous penalty (rlasso) ───── . . * Belloni-Chernozhukov-Hansen Double LASSO with the rigorous penalty: . * . * 1. rlasso y on X -> I_y = selected outcome-equation indices . * 2. rlasso d on X -> I_d = selected treatment-equation indices . * 3. OLS y on d + X[I_y ∪ I_d] with state-clustered SEs . * . * The rigorous penalty is data-driven (Belloni, Chen, Chernozhukov & . * Hansen 2012) and chosen so that selection-error noise is dominated . * by the signal. It is much more parsimonious than CV. . * . * Penalty constants c=1.1 and gamma=0.05 match the JAE (2026) . * replication code (readdata_all_OLS.R lines 585, 653) and the . * R companion's hdm::rlasso call. . . di _n(2) "========================================" ======================================== . di "STEP 7 - DOUBLE LASSO, RIGOROUS PENALTY (rlasso)" STEP 7 - DOUBLE LASSO, RIGOROUS PENALTY (rlasso) . di "========================================" ======================================== . di "Two rlasso calls (y on X, d on X), union of selected, then post-OLS." Two rlasso calls (y on X, d on X), union of selected, then post-OLS. . di "'Rigorous' = lambda from Belloni et al. (2012) theory, not CV." 'Rigorous' = lambda from Belloni et al. (2012) theory, not CV. . . tempname DLR_b DLR_se DLR_n DLR_Iy DLR_Id DLR_U . mat `DLR_b' = J(3, 1, .) . mat `DLR_se' = J(3, 1, .) . mat `DLR_n' = J(3, 1, .) . mat `DLR_Iy' = J(3, 1, .) . mat `DLR_Id' = J(3, 1, .) . mat `DLR_U' = J(3, 1, .) . . local row = 0 . foreach o in v p m { 2. local ++row 3. local lab = cond("`o'"=="v","Violent crime",cond("`o'"=="p","Property crime","Murder")) 4. local Y = cond("`o'"=="v","DyV", cond("`o'"=="p","DyP","DyM")) 5. local D = cond("`o'"=="v","DxV", cond("`o'"=="p","DxP","DxM")) 6. di _n " Outcome: `lab' (y=`Y', d=`D')" 7. . * Step 1: LASSO y on X. . qui rlasso `Y' z`o'1-z`o'284, nocons c(1.1) gamma(0.05) 8. local Iy "`e(selected)'" 9. local nIy : word count `Iy' 10. . * Step 2: LASSO d on X. . qui rlasso `D' z`o'1-z`o'284, nocons c(1.1) gamma(0.05) 11. local Id "`e(selected)'" 12. local nId : word count `Id' 13. . * Step 3: union of selected, then post-OLS. . local U : list Iy | Id 14. local nU : word count `U' 15. . if `nU' > 0 { 16. qui regress `Y' `D' `U', noconstant vce(cluster state) 17. } 18. else { 19. * Fall back to univariate first-difference fit (no controls survived). . qui regress `Y' `D', noconstant vce(cluster state) 20. } 21. . mat `DLR_b'[`row',1] = _b[`D'] 22. mat `DLR_se'[`row',1] = _se[`D'] 23. mat `DLR_n'[`row',1] = `nU' 24. mat `DLR_Iy'[`row',1] = `nIy' 25. mat `DLR_Id'[`row',1] = `nId' 26. mat `DLR_U'[`row',1] = `nU' 27. . di as text " |I_y| = " as result %3.0f `nIy' /// > as text " |I_d| = " as result %3.0f `nId' /// > as text " |union| = " as result %3.0f `nU' 28. di as text " alpha_hat = " as result %9.4f _b[`D'] /// > as text " SE = " as result %9.4f _se[`D'] 29. } Outcome: Violent crime (y=DyV, d=DxV) |I_y| = 0 |I_d| = 8 |union| = 8 alpha_hat = -0.1744 SE = 0.1155 Outcome: Property crime (y=DyP, d=DxP) |I_y| = 3 |I_d| = 14 |union| = 17 alpha_hat = -0.1144 SE = 0.0470 Outcome: Murder (y=DyM, d=DxM) |I_y| = 1 |I_d| = 12 |union| = 13 alpha_hat = -0.1229 SE = 0.1404 . . . * ── 8. Estimator E - Double LASSO, CV penalty (cvlasso) ────────── . . * Same three steps as section 7, but each LASSO is tuned by 3-fold . * CV (matching Fitzgerald et al. 2026 footnote 2). Lambda is picked . * to minimise out-of-sample MSE (lopt = "lambda at MSE minimum"). . * . * RUNTIME NOTE. Stata's cvlasso at p/n ≈ 0.5 is dramatically slower . * than R's cv.glmnet: a single call with the default lcount(100) . * takes 5+ minutes per outcome-equation, which makes the 6-call DL-CV . * pipeline impractical. We use lcount(10) to get the CV grid down to . * 10 lambda values, sacrificing precision in lambda selection in . * exchange for finishing in under 60 seconds per call. The Stata-vs-R . * drift in Section 14 captures the consequence: DL-CV's α and selected . * set sizes diverge more than the other Tier-C estimators do. . . di _n(2) "========================================" ======================================== . di "STEP 8 - DOUBLE LASSO, CV PENALTY (cvlasso, 3-fold, lcount=10)" STEP 8 - DOUBLE LASSO, CV PENALTY (cvlasso, 3-fold, lcount=10) . di "========================================" ======================================== . . tempname DLC_b DLC_se DLC_n DLC_Iy DLC_Id DLC_U . mat `DLC_b' = J(3, 1, .) . mat `DLC_se' = J(3, 1, .) . mat `DLC_n' = J(3, 1, .) . mat `DLC_Iy' = J(3, 1, .) . mat `DLC_Id' = J(3, 1, .) . mat `DLC_U' = J(3, 1, .) . . local row = 0 . foreach o in v p m { 2. local ++row 3. local lab = cond("`o'"=="v","Violent crime",cond("`o'"=="p","Property crime","Murder")) 4. local Y = cond("`o'"=="v","DyV", cond("`o'"=="p","DyP","DyM")) 5. local D = cond("`o'"=="v","DxV", cond("`o'"=="p","DxP","DxM")) 6. di _n " Outcome: `lab' (y=`Y', d=`D')" 7. . qui cvlasso `Y' z`o'1-z`o'284, nfolds(3) seed(20260520) lopt lglmnet lcount(10) 8. local Iy "`e(selected)'" 9. local nIy : word count `Iy' 10. . qui cvlasso `D' z`o'1-z`o'284, nfolds(3) seed(20260520) lopt lglmnet lcount(10) 11. local Id "`e(selected)'" 12. local nId : word count `Id' 13. . local U : list Iy | Id 14. local nU : word count `U' 15. . if `nU' > 0 { 16. qui regress `Y' `D' `U', noconstant vce(cluster state) 17. } 18. else { 19. qui regress `Y' `D', noconstant vce(cluster state) 20. } 21. . mat `DLC_b'[`row',1] = _b[`D'] 22. mat `DLC_se'[`row',1] = _se[`D'] 23. mat `DLC_n'[`row',1] = `nU' 24. mat `DLC_Iy'[`row',1] = `nIy' 25. mat `DLC_Id'[`row',1] = `nId' 26. mat `DLC_U'[`row',1] = `nU' 27. . di as text " |I_y| = " as result %3.0f `nIy' /// > as text " |I_d| = " as result %3.0f `nId' /// > as text " |union| = " as result %3.0f `nU' 28. di as text " alpha_hat = " as result %9.4f _b[`D'] /// > as text " SE = " as result %9.4f _se[`D'] 29. } Outcome: Violent crime (y=DyV, d=DxV) Warning: lopt is at the limit of the lambda range. Warning: lse is at the limit of the lambda range. Warning: lopt is at the limit of the lambda range. Warning: lse is at the limit of the lambda range. |I_y| = 0 |I_d| = 0 |union| = 0 alpha_hat = -0.1553 SE = 0.0330 Outcome: Property crime (y=DyP, d=DxP) Warning: lse is at the limit of the lambda range. |I_y| = 0 |I_d| = 0 |union| = 0 alpha_hat = -0.1015 SE = 0.0218 Outcome: Murder (y=DyM, d=DxM) Warning: lopt is at the limit of the lambda range. Warning: lse is at the limit of the lambda range. Warning: lopt is at the limit of the lambda range. |I_y| = 0 |I_d| = 0 |union| = 0 alpha_hat = -0.2061 SE = 0.0514 . . . * ── 9. Build results_table2.csv (5 estimators x 3 outcomes) ────── . . di _n(2) "========================================" ======================================== . di "STEP 9 - REPLICATION OF PAPER TABLE 2" STEP 9 - REPLICATION OF PAPER TABLE 2 . di "========================================" ======================================== . . preserve . clear . set obs 15 Number of observations (_N) was 0, now 15. . gen str14 method = "" (15 missing values generated) . gen str16 outcome = "" (15 missing values generated) . gen double estimate = . (15 missing values generated) . gen double std_error = . (15 missing values generated) . gen long n_selected = . (15 missing values generated) . . local outcomes "viol prop murd" . local methods "FD OLS PSL DLR DLC" . local mlabels `""First diff" "OLS (full)" "PSL" "DL (rigorous)" "DL (CV)""' . . local i = 0 . forvalues r = 1/3 { 2. local olab : word `r' of "`labels_viol' `labels_prop' `labels_murd'" 3. } . * Easier: fill by hand row-by-row. . local i = 0 . local rownames : word count `outcomes' . forvalues oi = 1/3 { 2. local o : word `oi' of `outcomes' 3. if "`o'" == "viol" local olab "Violent crime" 4. if "`o'" == "prop" local olab "Property crime" 5. if "`o'" == "murd" local olab "Murder" 6. forvalues mi = 1/5 { 7. local ++i 8. local m : word `mi' of `methods' 9. local mlab : word `mi' of `mlabels' 10. qui replace method = "`mlab'" in `i' 11. qui replace outcome = "`olab'" in `i' 12. if "`m'" == "FD" { 13. qui replace estimate = `FD_b'[`oi',1] in `i' 14. qui replace std_error = `FD_se'[`oi',1] in `i' 15. qui replace n_selected = 0 in `i' 16. } 17. else if "`m'" == "OLS" { 18. qui replace estimate = `OLS_b'[`oi',1] in `i' 19. qui replace std_error = `OLS_se'[`oi',1] in `i' 20. qui replace n_selected = `OLS_n'[`oi',1] in `i' 21. } 22. else if "`m'" == "PSL" { 23. qui replace estimate = `PSL_b'[`oi',1] in `i' 24. qui replace std_error = `PSL_se'[`oi',1] in `i' 25. qui replace n_selected = `PSL_n'[`oi',1] in `i' 26. } 27. else if "`m'" == "DLR" { 28. qui replace estimate = `DLR_b'[`oi',1] in `i' 29. qui replace std_error = `DLR_se'[`oi',1] in `i' 30. qui replace n_selected = `DLR_n'[`oi',1] in `i' 31. } 32. else if "`m'" == "DLC" { 33. qui replace estimate = `DLC_b'[`oi',1] in `i' 34. qui replace std_error = `DLC_se'[`oi',1] in `i' 35. qui replace n_selected = `DLC_n'[`oi',1] in `i' 36. } 37. } 38. } . . gen double ci_lo = estimate - 1.96 * std_error . gen double ci_hi = estimate + 1.96 * std_error . order method outcome estimate std_error n_selected ci_lo ci_hi . list, sepby(outcome) abbreviate(20) +-----------------------------------------------------------------------------------+ 1. | method | outcome | estimate | std_error | n_selected | ci_lo | | First diff | Violent crime | -.15209745 | .03367644 | 0 | -.21810328 | |-----------------------------------------------------------------------------------| | ci_hi | | -.08609161 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 2. | method | outcome | estimate | std_error | n_selected | ci_lo | | OLS (full) | Violent crime | .01338395 | .71491436 | 284 | -1.3878482 | |-----------------------------------------------------------------------------------| | ci_hi | | 1.4146161 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 3. | method | outcome | estimate | std_error | n_selected | ci_lo | | PSL | Violent crime | -.15529 | .03297352 | 0 | -.2199181 | |-----------------------------------------------------------------------------------| | ci_hi | | -.0906619 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 4. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (rigorous) | Violent crime | -.17439471 | .11554708 | 8 | -.40086699 | |-----------------------------------------------------------------------------------| | ci_hi | | .05207757 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 5. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (CV) | Violent crime | -.15529 | .03297352 | 0 | -.2199181 | |-----------------------------------------------------------------------------------| | ci_hi | | -.0906619 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 6. | method | outcome | estimate | std_error | n_selected | ci_lo | | First diff | Property crime | -.10837626 | .02193445 | 0 | -.15136778 | |-----------------------------------------------------------------------------------| | ci_hi | | -.06538474 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 7. | method | outcome | estimate | std_error | n_selected | ci_lo | | OLS (full) | Property crime | -.19499381 | .22356196 | 284 | -.63317526 | |-----------------------------------------------------------------------------------| | ci_hi | | .24318764 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 8. | method | outcome | estimate | std_error | n_selected | ci_lo | | PSL | Property crime | -.06654344 | .02443351 | 1 | -.11443311 | |-----------------------------------------------------------------------------------| | ci_hi | | -.01865376 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 9. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (rigorous) | Property crime | -.11441552 | .04695949 | 17 | -.20645612 | |-----------------------------------------------------------------------------------| | ci_hi | | -.02237493 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 10. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (CV) | Property crime | -.10146893 | .02184183 | 0 | -.14427892 | |-----------------------------------------------------------------------------------| | ci_hi | | -.05865895 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 11. | method | outcome | estimate | std_error | n_selected | ci_lo | | First diff | Murder | -.20386472 | .06672788 | 0 | -.33465136 | |-----------------------------------------------------------------------------------| | ci_hi | | -.07307808 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 12. | method | outcome | estimate | std_error | n_selected | ci_lo | | OLS (full) | Murder | 2.3411482 | 2.7830742 | 284 | -3.1136773 | |-----------------------------------------------------------------------------------| | ci_hi | | 7.7959736 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 13. | method | outcome | estimate | std_error | n_selected | ci_lo | | PSL | Murder | -.23974556 | .06349433 | 1 | -.36419445 | |-----------------------------------------------------------------------------------| | ci_hi | | -.11529667 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 14. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (rigorous) | Murder | -.12290333 | .14041769 | 13 | -.398122 | |-----------------------------------------------------------------------------------| | ci_hi | | .15231533 | +-----------------------------------------------------------------------------------+ +-----------------------------------------------------------------------------------+ 15. | method | outcome | estimate | std_error | n_selected | ci_lo | | DL (CV) | Murder | -.20613219 | .05137252 | 0 | -.30682232 | |-----------------------------------------------------------------------------------| | ci_hi | | -.10544206 | +-----------------------------------------------------------------------------------+ . export delimited "results_table2.csv", replace (file results_table2.csv not found) file results_table2.csv saved . di _n "Wrote results_table2.csv" Wrote results_table2.csv . restore . . . * ── 9b. Selection diagnostic CSV ───────────────────────────────── . . preserve . clear . set obs 6 Number of observations (_N) was 0, now 6. . gen str16 outcome = "" (6 missing values generated) . gen str14 method = "" (6 missing values generated) . gen long n_Iy = . (6 missing values generated) . gen long n_Id = . (6 missing values generated) . gen long n_union = . (6 missing values generated) . local i = 0 . forvalues oi = 1/3 { 2. local o : word `oi' of viol prop murd 3. if "`o'" == "viol" local olab "Violent crime" 4. if "`o'" == "prop" local olab "Property crime" 5. if "`o'" == "murd" local olab "Murder" 6. local ++i 7. qui replace outcome = "`olab'" in `i' 8. qui replace method = "DL (rigorous)" in `i' 9. qui replace n_Iy = `DLR_Iy'[`oi',1] in `i' 10. qui replace n_Id = `DLR_Id'[`oi',1] in `i' 11. qui replace n_union = `DLR_U'[`oi',1] in `i' 12. local ++i 13. qui replace outcome = "`olab'" in `i' 14. qui replace method = "DL (CV)" in `i' 15. qui replace n_Iy = `DLC_Iy'[`oi',1] in `i' 16. qui replace n_Id = `DLC_Id'[`oi',1] in `i' 17. qui replace n_union = `DLC_U'[`oi',1] in `i' 18. } . list, sepby(outcome) abbreviate(20) +--------------------------------------------------------+ | outcome method n_Iy n_Id n_union | |--------------------------------------------------------| 1. | Violent crime DL (rigorous) 0 8 8 | 2. | Violent crime DL (CV) 0 0 0 | |--------------------------------------------------------| 3. | Property crime DL (rigorous) 3 14 17 | 4. | Property crime DL (CV) 0 0 0 | |--------------------------------------------------------| 5. | Murder DL (rigorous) 1 12 13 | 6. | Murder DL (CV) 0 0 0 | +--------------------------------------------------------+ . export delimited "selection_diagnostic.csv", replace (file selection_diagnostic.csv not found) file selection_diagnostic.csv saved . di "Wrote selection_diagnostic.csv" Wrote selection_diagnostic.csv . restore . . . * ── 10. Figures (4 dark-theme PNGs) ────────────────────────────── . . di _n(2) "========================================" ======================================== . di "STEP 10 - FIGURES" STEP 10 - FIGURES . di "========================================" ======================================== . . * For figures we work in a small auxiliary dataset. . preserve . clear . set obs 15 Number of observations (_N) was 0, now 15. . gen byte outcome_id = ceil(_n/5) // 1..3 . gen byte method_id = mod(_n-1,5)+1 // 1..5 . . label define olab 1 "Violent crime" 2 "Property crime" 3 "Murder" . label values outcome_id olab . label define mlab 1 "First diff" 2 "OLS (full)" 3 "PSL" 4 "DL (rigorous)" 5 "DL (CV)" . label values method_id mlab . . gen double estimate = . (15 missing values generated) . gen double std_error = . (15 missing values generated) . . * Fill in from the saved matrices. . forvalues i = 1/15 { 2. local oi = ceil(`i'/5) 3. local mi = mod(`i'-1,5)+1 4. if `mi' == 1 { 5. qui replace estimate = `FD_b'[`oi',1] in `i' 6. qui replace std_error = `FD_se'[`oi',1] in `i' 7. } 8. if `mi' == 2 { 9. qui replace estimate = `OLS_b'[`oi',1] in `i' 10. qui replace std_error = `OLS_se'[`oi',1] in `i' 11. } 12. if `mi' == 3 { 13. qui replace estimate = `PSL_b'[`oi',1] in `i' 14. qui replace std_error = `PSL_se'[`oi',1] in `i' 15. } 16. if `mi' == 4 { 17. qui replace estimate = `DLR_b'[`oi',1] in `i' 18. qui replace std_error = `DLR_se'[`oi',1] in `i' 19. } 20. if `mi' == 5 { 21. qui replace estimate = `DLC_b'[`oi',1] in `i' 22. qui replace std_error = `DLC_se'[`oi',1] in `i' 23. } 24. } . gen double ci_lo = estimate - 1.96*std_error . gen double ci_hi = estimate + 1.96*std_error . . * Plot positions: invert so "First diff" is at top. . gen byte y = 6 - method_id . . * === Figure 1: forest plot of all five estimates === . twoway /// > (rspike ci_lo ci_hi y if method_id==1, horizontal lcolor("$C_STEEL") lwidth(medthick)) /// > (scatter y estimate if method_id==1, mcolor("$C_STEEL") msymbol(O) msize(medlarge)) /// > (rspike ci_lo ci_hi y if method_id==2, horizontal lcolor("$C_TEXT") lwidth(medthick)) /// > (scatter y estimate if method_id==2, mcolor("$C_TEXT") msymbol(O) msize(medlarge)) /// > (rspike ci_lo ci_hi y if method_id==3, horizontal lcolor("$C_ORG") lwidth(medthick)) /// > (scatter y estimate if method_id==3, mcolor("$C_ORG") msymbol(O) msize(medlarge)) /// > (rspike ci_lo ci_hi y if method_id==4, horizontal lcolor("$C_TEAL") lwidth(medthick)) /// > (scatter y estimate if method_id==4, mcolor("$C_TEAL") msymbol(O) msize(medlarge)) /// > (rspike ci_lo ci_hi y if method_id==5, horizontal lcolor("$C_LTORG") lwidth(medthick)) /// > (scatter y estimate if method_id==5, mcolor("$C_LTORG") msymbol(O) msize(medlarge)) /// > , by(outcome_id, cols(3) /// > title("Treatment-effect estimates: abortion -> crime, 1985-1997", color("$C_TXTHI") > size(medsmall)) /// > subtitle("Each panel: 95% CIs from state-clustered SEs.", color("$C_TEXT") size(smal > l)) /// > note("Replication of Table 2 in Fitzgerald et al. (2026). Dashed line at zero." , co > lor("$C_TEXT") size(vsmall)) /// > graphregion(fcolor("$C_BG") lcolor("$C_BG")) /// > plotregion(fcolor("$C_BG") lcolor("$C_BG")) /// > ${DARKBG} legend(off) /// > imargin(small)) /// > subtitle(, fcolor("$C_BG") lcolor("$C_BG") size(small) color("$C_TXTHI")) /// > ylabel(1 "DL (CV)" 2 "DL (rigorous)" 3 "PSL" 4 "OLS (full)" 5 "First diff", /// > labcolor("$C_TEXT") angle(0) noticks nogrid) /// > xlabel(, labcolor("$C_TEXT")) /// > xtitle("alpha hat (effect of effective abortion rate)", color("$C_TEXT") size(small)) /// > ytitle("") /// > xline(0, lpattern(dash) lcolor("$C_TEXT")) /// > ${DARKBG} /// > name(fig_forest, replace) . . graph export "stata_double_lasso_estimates.png", replace width(2400) height(1100) file /Users/carlosmendez/Documents/GitHub/starter-academic-v501/content/post/stata_double_lasso > /stata_double_lasso_estimates.png saved as PNG format . di "Wrote stata_double_lasso_estimates.png" Wrote stata_double_lasso_estimates.png . restore . . . * === Figure 2: selection-count bar chart === . preserve . clear . set obs 12 Number of observations (_N) was 0, now 12. . * 3 outcomes x 2 methods (rig, CV) x 2 metrics (Iy, Id) = 12 rows. . gen byte outcome_id = ceil(_n/4) . gen byte block = mod(_n-1,4) + 1 // 1=rig Iy, 2=rig Id, 3=cv Iy, 4=cv Id . label define olab 1 "Violent crime" 2 "Property crime" 3 "Murder" . label values outcome_id olab . . gen str20 metric = "" (12 missing values generated) . gen str14 mtype = "" (12 missing values generated) . gen long count = . (12 missing values generated) . forvalues i = 1/12 { 2. local oi = ceil(`i'/4) 3. local bl = mod(`i'-1,4) + 1 4. if `bl' == 1 { 5. qui replace metric = "|I_y|" in `i' 6. qui replace mtype = "DL (rigorous)" in `i' 7. qui replace count = `DLR_Iy'[`oi',1] in `i' 8. } 9. if `bl' == 2 { 10. qui replace metric = "|I_d|" in `i' 11. qui replace mtype = "DL (rigorous)" in `i' 12. qui replace count = `DLR_Id'[`oi',1] in `i' 13. } 14. if `bl' == 3 { 15. qui replace metric = "|I_y|" in `i' 16. qui replace mtype = "DL (CV)" in `i' 17. qui replace count = `DLC_Iy'[`oi',1] in `i' 18. } 19. if `bl' == 4 { 20. qui replace metric = "|I_d|" in `i' 21. qui replace mtype = "DL (CV)" in `i' 22. qui replace count = `DLC_Id'[`oi',1] in `i' 23. } 24. } . . * x positions: 1,2,5,6,9,10 etc. - grouped by metric, dodged by method. . gen double xpos = . (12 missing values generated) . * rigorous Iy at x=1, CV Iy at x=2; rigorous Id at x=4, CV Id at x=5. . qui replace xpos = 1 if metric == "|I_y|" & mtype == "DL (rigorous)" . qui replace xpos = 2 if metric == "|I_y|" & mtype == "DL (CV)" . qui replace xpos = 4 if metric == "|I_d|" & mtype == "DL (rigorous)" . qui replace xpos = 5 if metric == "|I_d|" & mtype == "DL (CV)" . . twoway /// > (bar count xpos if mtype == "DL (rigorous)", barwidth(0.85) fcolor("$C_TEAL") lcolor("$C_T > EAL")) /// > (bar count xpos if mtype == "DL (CV)", barwidth(0.85) fcolor("$C_LTORG") lcolor("$C > _LTORG")) /// > (scatter count xpos, mcolor("$C_TXTHI") msymbol(none) mlabel(count) mlabcolor("$C_TXTHI") > mlabsize(small) mlabposition(12)) /// > , by(outcome_id, cols(3) /// > title("Variable selection: rigorous vs CV penalty", color("$C_TXTHI") size(medsmall) > ) /// > subtitle("Out of 284 candidate controls per outcome.", color("$C_TEXT") size(small)) > /// > note("|I_y| = controls selected when LASSOing y on X; |I_d| = controls selected when > LASSOing d on X.", color("$C_TEXT") size(vsmall)) /// > ${DARKBG}) /// > subtitle(, fcolor("$C_BG") lcolor("$C_BG") size(small) color("$C_TXTHI")) /// > xlabel(1.5 "|I_y|" 4.5 "|I_d|", noticks labcolor("$C_TEXT")) /// > ylabel(, labcolor("$C_TEXT")) /// > xtitle("Selection step", color("$C_TEXT") size(small)) ytitle("Number of controls", color( > "$C_TEXT") size(small)) /// > legend(order(1 "DL (rigorous)" 2 "DL (CV)") region(fcolor("$C_BG") lcolor("$C_BG")) textco > lor("$C_TEXT") size(small) cols(2)) /// > ${DARKBG} /// > name(fig_select, replace) option textcolor() not allowed r(198); end of do-file r(198);