ivreg2 1

*! ivreg2 1.0.27 02Nov2002 *! authors cfb & mes *! cloned from official ivreg version 5.0.9 19Dec2001 * 1.0.2: add logic for reg3. Sargan test * 1.0.3: add prunelist to ensure that count of excluded exogeneous is correct * 1.0.4: revise option to exog(), allow included exog to be specified as well * 1.0.5: switch from reg3 to regress, many options and output changes * 1.0.6: fixed treatment of nocons in Sargan and C-stat, and corrected problems * relating to use of nocons combined with a constant as an IV * 1.0.7: first option reports F-test of excluded exogenous; prunelist bug fix * 1.0.8: dropped prunelist and switched to housekeeping of variable lists * 1.0.9: added collinearity checks; C-stat calculated with recursive call; * added ffirst option to report only F-test of excluded exogenous * from 1st stage regressions * 1.0.10: 1st stage regressions also report partial R2 of excluded exogenous * 1.0.11: complete rewrite of collinearity approach - no longer uses calls to * _rmcoll, does not track specific variables dropped; prunelist removed * 1.0.12: reorganised display code and saved results to enable -replay()- * 1.0.13: -robust- and -cluster- now imply -small- * 1.0.14: fixed hascons bug; removed ivreg predict fn (it didn't work); allowed * robust and cluster with z stats and correct dofs * 1.0.15: implemented robust Sargan stat; changed to only F-stat, removed chi-sq; * removed exog option (only orthog works) * 1.0.16: added clusterised Sargan stat; robust Sargan handles collinearities; * predict now works with standard SE options plus resids; fixed orthog() * so it accepts time series operators etc. * 1.0.17: fixed handling of weights. fw, aw, pw & iw all accepted. * 1.0.18: fixed bug in robust Sargan code relating to time series variables. * 1.0.19: fixed bugs in reporting ranks of X'X and Z'Z * fixed bug in reporting presence of constant * 1.0.20: added GMM option and replaced robust Sargan with (equivalent) J; * added saved statistics of 1st stage regressions * 1.0.21: added Cragg HOLS estimator, including allowing empty endog list; * -regress- syntax now not allowed; revised code searching for "_cons" * 1.0.22: modified cluster output message; fixed bug in replay for Sargan/Hansen stat; * exactly identified Sargan/Hansen now exactly zero and p-value not saved as e(); * cluster multiplier changed to 1 (from buggy multiplier), in keeping with * eg Wooldridge 2002 p. 193. * 1.0.23: fixed orthog option to prevent abort when restricted equation is underid. * 1.0.24: fixed bug if 1st stage regressions yielded missing values for saving in e(). * 1.0.25: Added Shea version of partial R2 * 1.0.26: Replaced Shea algorithm with Godfrey algorithm * 1.0.27: Main call to regress is OLS form if OLS or HOLS is specified; error variance * in Sargan and C statistics use small-sample adjustment if -small- option is * specified; dfn of S matrix now correctly divided by sample size * * Variable naming: * lhs = LHS endogenous * endo = RHS endogenous (instrumented) * inexog = included exogenous (instruments) * exexog = excluded exogenous (instruments) * iv = {inexog exexog} = all instruments * rhs = {endo inexog} = RHS regressors * 1 at the end of the name means the original varlist before duplicates and * collinearities removed program define ivreg2, eclass byable(recall) version 7.0 local version 01.0.27 if replay() { if `"`e(cmd)'"' != "ivreg2" { error 301 } else { syntax [, Level(integer $S_level) NOHEader NOFOoter /* */ EForm(string) PLUS] if "`eform'"!="" { local efopt "eform(`eform')" } } } else { local n 0 gettoken lhs 0 : 0, parse(" ,[") match(paren) IsStop `lhs' if `s(stop)' { error 198 } while `s(stop)'==0 { if "`paren'"=="(" { local n = `n' + 1 if `n'>1 { capture noi error 198 di in red `"syntax is "(all instrumented variables = instrument variables)""' exit 198 } gettoken p lhs : lhs, parse(" =") while "`p'"!="=" { if "`p'"=="" { capture noi error 198 di in red `"syntax is "(all instrumented variables = instrument variables)""' di in red `"the equal sign "=" is required"' exit 198 } local endo`n' `endo`n'' `p' gettoken p lhs : lhs, parse(" =") } * To enable Cragg HOLS estimator, allow for empty endo1 list if "`endo1'" != "" { tsunab endo`n' : `endo`n'' } tsunab exexog`n' : `lhs' } else { local inexog1 `inexog1' `lhs' } gettoken lhs 0 : 0, parse(" ,[") match(paren) IsStop `lhs' } local 0 `"`lhs' `0'"' if "`endo1'`inexog1'"=="" { di in red "no regressors specified" exit 102 } if "`endo1'`exexog1'"=="" { di in red "too few variables specified" exit 102 } tsunab inexog1 : `inexog1' tokenize `inexog1' local lhs "`1'" local 1 " " local inexog1 `*' syntax [if] [in] [aw fw pw iw] [, /* */ FIRST FFIRST SMall Beta Robust CLuster(varname) hc2 hc3 /* */ GMM WMATRIX(string) ORTHOG(string) NOConstant Hascons /* */ Level(integer $S_level) NOHEader NOFOoter /* */ DEPname(string) EForm(string) PLUS *] if "`orthog'" != "" { tsunab orthog : `orthog' } if "`hc2'`hc3'" != "" { if "`hc2'"!="" { di in red "option `hc2' invalid" } else di in red "option `hc3' invalid" exit 198 } if "`beta'" != "" { di in red "option `beta' invalid" exit 198 } * Weights * fweight and aweight accepted as is * iweight not allowed with robust or gmm and requires a trap below when used with summarize * pweight is equivalent to aweight + robust if "`weight'" == "fweight" | "`weight'"=="aweight" { local wgt `"[`weight'`exp']"' } if "`weight'" == "iweight" { if "`robust'`cluster'`gmm'" !="" { di in red "iweights not allowed with robust or gmm" exit 101 } else { local wgt `"[`weight'`exp']"' } } if "`weight'" == "pweight" { local wgt `"[aweight`exp']"' local robust "robust" } marksample touse markout `touse' `lhs' `inexog1' `exexog1' `endo1' `cluster', strok local insts1 `inexog1' `exexog1' local rhs1 `endo1' `inexog1' local iv1_ct : word count `insts1' local rhs1_ct : word count `rhs1' local endo1_ct : word count `endo1' * Quick check; see also regress "bug" below if `rhs1_ct' > `iv1_ct' { di in red "equation not identified; must have at " /* */ "least as many instruments not in" di in red "the regression as there are " /* */ "instrumented variables" exit 481 } * If depname not provided (default) name is lhs variable if "`cluster'"!="" { local clopt "cluster(`cluster')" } if "`eform'"!="" { local efopt "eform(`eform')" } if "`depname'"=="" { local depname `lhs' } *********** Main call to regress *************** * A "bug" in regress means that if the system is identified after removal * of collinearities but appears unidentified by counting variables * before removal, regress will refuse to estimate. * If no included endogenous vars, call using OLS syntax. if `endo1_ct' > 0 { quietly regress `lhs' `rhs1' (`insts1') if `touse' `wgt', /* */ `noconstant' `hascons' `options' `robust' `clopt' } else { quietly regress `lhs' `rhs1' if `touse' `wgt', /* */ `noconstant' `hascons' `options' `robust' `clopt' } ************************************************ if "`first'`ffirst'" != "" & (`endo1_ct' > 0) { * For Godfrey method of Shea partial R2, need IV and OLS estimates without robust vces tempname olsV ivV godfrey ols_s2 iv_s2 sols siv vrat godfrey ivest if "`robust'`clopt'" == "" { mat `ivV' = e(V) scalar iv_s2 = e(rmse)^2 } else { estimates hold ivest quietly regress `lhs' `rhs1' (`insts1') /* */ if `touse' `wgt', `noconstant' `hascons' /* */ `options' mat `ivV' = e(V) scalar iv_s2 = e(rmse)^2 estimates unhold ivest } estimates hold ivest qui regress `lhs' `rhs1' /* */ if `touse' `wgt', `noconstant' `hascons' /* */ `options' mat `olsV' = e(V) scalar ols_s2 = e(rmse)^2 estimates unhold ivest scalar vrat = iv_s2/ols_s2 mat `godfrey' = J(1,`endo1_ct',0) mat colnames `godfrey' = `endo1' mat rownames `godfrey' = "sheapr2" local i 1 foreach w of local endo1 { mat `sols'=`olsV'["`w'","`w'"] mat `siv'=`ivV'["`w'","`w'"] mat `godfrey'[1,`i'] = vrat*`sols'[1,1]/`siv'[1,1] local i = `i'+1 } } tempvar iota y2 tempname ivresid Nprec ysum yy yyc r2u r2c b V ivest qui predict double `ivresid' if `touse', resid qui gen `iota'=1 qui gen double `y2'=`lhs'^2 * Stata summarize won't work with iweights, so must use matrix cross-product qui matrix vecaccum `ysum' = `iota' `y2' `lhs' `wgt' if `touse' * Nprec is ob count from mat accum. Use this rather than `N' in calculations * here and below because `N' is rounded if iweights are used. scalar `Nprec'=`ysum'[1,3] scalar `yy'=`ysum'[1,1] scalar `yyc'=`yy'-`ysum'[1,2]^2/`Nprec' * Store results of IV regress local N=e(N) if "`cluster'"!="" { local N_clust=e(N_clust) } local df_m=e(df_m) local df_r=e(df_r) tempname r2 r2_a rss mss rmse F scalar `r2'=e(r2) scalar `r2_a'=e(r2_a) scalar `rss'=e(rss) scalar `mss'=e(mss) scalar `rmse'=e(rmse) scalar `F'=e(F) if "`gmm'`robust'`cluster'" != "" { tempname S Sinv Ztu A XZ XZa XZb ZY B1 B V uZWZu j tempvar gresid gresid2 yhat * Bread of the sandwich qui mat vecaccum `Ztu'=`ivresid' `exexog1' `inexog1' /* */ `wgt' if `touse', `noconstant' * Optimal weighting matrix * If user-supplied wmatrix is used, use it, otherwise use _robust (hence weights allowed) if "`wmatrix'" != "" { matrix `Sinv'=`wmatrix' } else { * Has same results as mat accum `S'=`exexog1' `inexog1' [iweight=`ivresid'^2] * Create identity matrix with correct col/row names mat `S'=I(colsof(`Ztu'))+0*diag(`Ztu') _robust `ivresid' `wgt' if `touse', variance(`S') `clopt' minus(0) * Above doesn't work properly with iweights (i.e. yield same matrix as fw), * hence iweight trap at start mat `S'=`S'*1/`Nprec' mat `Sinv'=syminv(`S') } * Generate cross-products of y, X, Z qui matrix accum `A' = `lhs' `endo1' `exexog1' `inexog1' /* */ if `touse', `noconstant' if "`noconstant'"=="" { matrix rownames `A' = `lhs' `endo1' `exexog1' /* */ `inexog1' _cons matrix colnames `A' = `lhs' `endo1' `exexog1' /* */ `inexog1' _cons } else { matrix rownames `A' = `lhs' `endo1' `exexog1' `inexog1' matrix colnames `A' = `lhs' `endo1' `exexog1' `inexog1' } if `endo1_ct' > 0 { * X'Z is [endo1 inexog1]'[exexog1 inexog1] mat `XZ'=`A'[2..`endo1_ct'+1,`endo1_ct'+2...] * Append portion corresponding to included exog if they (incl constant) exist if 2+`endo1_ct'+`iv1_ct'-(`rhs1_ct'-`endo1_ct') /* */ <= rowsof(`A') { mat `XZ'=`XZ' \ /* */ `A'[2+`endo1_ct'+`iv1_ct'- /* */ (`rhs1_ct'-`endo1_ct')..., /* */ `endo1_ct'+2...] } mat `ZY'=`A'[`endo1_ct'+2...,1] } else { * Cragg HOLS estimator with no endogenous variables mat `XZ'= `A'[2+`iv1_ct'-(`rhs1_ct'-`endo1_ct')..., /* */ 2...] mat `ZY'=`A'[2...,1] } mat `B1'=syminv(`XZ'*`Sinv'*`XZ'') mat `B'=(`B1'*`XZ'*`Sinv'*`ZY')' mat `V' = syminv(`XZ'*`Sinv'*`XZ'')*`Nprec' * Rank of Z'Z is same as rank of S=Z'ee'Z local iv_ct = rowsof(`Sinv') - diag0cnt(`Sinv') * Save IV estimates if only J is needed if "`gmm'"=="" {estimates hold `ivest'} est post `B' `V' qui _predict double `yhat', xb qui gen double `gresid'=`lhs'-`yhat' qui gen double `gresid2'=`gresid'^2 * J statistic qui mat vecaccum `Ztu'=`gresid' `exexog1' `inexog1' /* */ `wgt' if `touse', `noconstant' mat `uZWZu'= `Ztu'*`Sinv'*`Ztu'' * 1/`Nprec' scalar `j' = `uZWZu'[1,1] * Restore iv estimates if only J was needed, otherwise calc new R2s if "`gmm'"=="" {estimates unhold `ivest'} else { qui matrix vecaccum `ysum' = `iota' `gresid2' /* */ `wgt' if `touse', `noconstant' scalar `rss'= `ysum'[1,1] if "`noconstant'"=="" { scalar `mss'=`yyc'-`rss' } else { scalar `mss'=`yy'-`rss' } } } matrix `b'=e(b) matrix `V'=e(V) * Counts modified to include constant if appropriate if "`noconstant'"=="" { local iv1_ct = `iv1_ct' + 1 local rhs1_ct = `rhs1_ct' + 1 } * Correct count of rhs variables accounting for dropped collinear vars * Count includes constant * Will take account of extraneous _cons (with zero coeff) if hascons local rhs_ct=colsof(`b') local i 1 while `i' <= colsof(`b') { if `b'[1,`i']==0 { local rhs_ct = `rhs_ct'-1 } local i = `i'+1 } * Now do small sample correction for GMM var-cov matrix. * Stata's IV with robust has a small sample correction to the var-cov matrix * so do the same for the GMM var-cov matrix. Corrections are exactly Stata's * corrections at this point (see removal of small sample corrections below). if "`gmm'" != "" { if "`cluster'"=="" { matrix `V'=`V'*`Nprec'/(`Nprec'-`rhs_ct') } else { matrix `V'=`V'*(`Nprec'-1)/(`Nprec'-`rhs_ct') /* */ * `N_clust'/(`N_clust'-1) } } scalar `r2u'=1-`rss'/`yy' scalar `r2c'=1-`rss'/`yyc' if "`gmm'" != "" { if "`noconstant'"=="" { scalar `r2'=`r2c' scalar `r2_a'=1-(1-`r2')*(`Nprec'-1)/(`Nprec'-`rhs_ct') } else { scalar `r2'=`r2u' scalar `r2_a'=1-(1-`r2')*`Nprec'/(`Nprec'-`rhs_ct') } } * hascons code - check if true/false * Strip out extraneous _cons in last col if hascons true * Report error if hascons false and reset macro to empty string * True if _cons in last column and coeff==0, false if !=0 if "`hascons'"!="" { local cn : colnames `b' local cn : subinstr local cn "_cons" "_cons" , word count(local hc) if `hc' == 1 { if `b'[1,colsof(`b')]==0 { matrix `b'=`b'[1,1..colsof(`b')-1] matrix `V'=`V'[1..rowsof(`V')-1,1..colsof(`V')-1] } else { di in gr "nb: hascons false" local hascons "" } } } * Sargan statistic if overid, J if robust or GMM * Error variance in denominator uses small adjustment if -small- option specified. * Code also calculates correct count of instruments (including constant) * after possible collinearities => dropped instruments if "`robust'`cluster'`gmm'" == "" { tempname ZZ Ztu ZZinv uZWZu sargan sargmse sarganp if "`small'" == "" { scalar `sargmse'=`rss'/`Nprec' } else { scalar `sargmse'=`rss'/(`Nprec'-`rhs_ct') } qui mat accum `ZZ' = `insts1' `wgt' if `touse', `noconstant' qui mat vecaccum `Ztu' = `ivresid' `insts1' `wgt' /* */ if `touse', `noconstant' matrix `ZZinv' = syminv(`ZZ') mat `uZWZu' = `Ztu'*`ZZinv'*`Ztu'' local iv_ct = rowsof(`ZZ') - diag0cnt(`ZZinv') local sargandf = `iv_ct' - `rhs_ct' if `sargandf' == 0 { scalar `sargan' = 0 } else { scalar `sargan' = `uZWZu'[1,1]/`sargmse' scalar `sarganp' = chiprob(`sargandf',`sargan') } } else { tempname jp local jdf = `iv_ct' - `rhs_ct' if `jdf' == 0 { scalar `j' = 0 } else { scalar `jp' = chiprob(`jdf',`j') } } * orthog option: C statistic (difference of Sargan statistics) if "`orthog'"!="" { * Initialize cstat local cstat 0 * Each variable listed must be in instrument list. * To avoid overwriting, use cendo, cinexog1, cexexog, cendo_ct, cex_ct local cendo1 "`endo1'" local cinexog1 "`inexog1'" local cexexog1 "`exexog1'" local cinsts1 "`insts1'" local crhs1 "`rhs1'" local clist1 "`orthog'" local clist_ct : word count `clist1' * Check to see if c-stat vars are in original list of all ivs * cinexog1 and cexexog1 are after c-stat exog list vars have been removed * cendo1 is endo1 after included exog being tested has been added foreach x of local clist1 { local llex_ct : word count `cexexog1' Subtract cexexog1 : "`cexexog1'" "`x'" local cex1_ct : word count `cexexog1' local ok = `llex_ct' - `cex1_ct' if (`ok'==0) { * Not in excluded, check included and add to endog list if it appears local llin_ct : word count `cinexog1' Subtract cinexog1 : "`cinexog1'" "`x'" local cin1_ct : word count `cinexog1' local ok = `llin_ct' - `cin1_ct' if (`ok'==0) { * Not in either list di in r "Error: `x' listed in orthog() but does not appear as exogenous." error 198 } else { local cendo1 "`cendo1' `x'" } } } * If robust or GMM, create optimal weighting matrix to pass to ivreg2 * by extracting the submatrix from the full S and then inverting. * This guarantees the C stat will be non-negative. See Hayashi (2000), p. 220. if "`robust'`cluster'`gmm'" != "" { tempname CSa CS CSinv if "`noconstant'" !="" { local cexin "`cexexog1' `cinexog1'" } else { local cexin "`cexexog1' `cinexog1' _cons" } local v1 1 foreach x of local cexin { if `v1'==1 { mat `CSa' = `S'[`"`x'"',.] local v1 0 } else { mat `CSa' = `CSa' \ `S'[`"`x'"',.] } } local v1 1 foreach x of local cexin { if `v1'==1 { mat `CS' = `CSa'[.,`"`x'"'] local v1 0 } else { mat `CS' = `CS', `CSa'[.,`"`x'"'] } } mat `CSinv'=syminv(`CS') } * Calculate C statistic with recursive call to ivreg2 * Collinearities may cause problems, hence -capture-. capture { tempname csargan cj cstat cstatp cj estimates hold `ivest', restore if "`robust'`cluster'`gmm'" == "" { capture ivreg2 `lhs' `cinexog1' /* */ (`cendo1'=`cexexog1') if `touse' /* */ `wgt', `noconstant' `hascons' /* */ `small' `options' local rc = _rc if `rc' == 481 { local cstat = 0 local cstatdf = 0 } else { * Use MSE from original Sargan test to ensure p.d. scalar `csargan'=e(sargan)*e(rss)/`rss' local csargdf=e(sargandf) scalar `cstat' = `sargan' - `csargan' local cstatdf = `sargandf' - `csargdf' } } else { capture ivreg2 `lhs' `cinexog1' /* */ (`cendo1'=`cexexog1') /* */ if `touse' `wgt', `noconstant' /* */ `hascons' `small' `options' `robust' /* */ `clopt' `gmm' wmatrix("`CSinv'") local rc = _rc if `rc' == 481 { local cstat = 0 local cstatdf = 0 } else { scalar `cj'=e(j) local cjdf=e(jdf) scalar `cstat' = `j' - `cj' local cstatdf = `jdf' - `cjdf' } } estimates unhold `ivest' scalar `cstatp'= chiprob(`cstatdf',`cstat') * Collinearities may cause C-stat dof to differ from the number of variables in orthog() * If so, set cstat=0 if `cstatdf' != `clist_ct' { local cstat = 0 } } } * Check if adequate number of observations and clusters if `N' <= `iv_ct' { di in r "Error: number of observations must be greater than number of instruments" di in r " including constant." error 2001 } if "`cluster'" != "" { if `N_clust' <= `iv_ct' { if "`gmm'" != "" { di in r "Error: number of clusters insufficient to calculate optimal weighting matrix" error 498 } else { * Need to caculate correct rank of Z'Z matrix = # of IVs qui _rmcoll `insts1' if `touse', `noconstant' local iv_ct : word count `r(varlist)' if "`noconstant'"=="" { local iv_ct = `iv_ct' + 1 } * Re-check if adequate number of observations using correct rank of Z'Z if `N' <= `iv_ct' { di in r "Error: number of observations must be greater than number of instruments" di in r " including constant." error 2001 } * VCV matrix, J, F stat, and C stat (if present) all meaningless mat `V'=e(V)*0 scalar `j' = . scalar `F' = . if "`orthog'"!="" { `cstat' = . } } } } * First stage regression option * Code here because it relies on proper count of (non-collinear) IVs * generated earlier * Note that nocons option + constant in instrument list means first-stage * regressions are reported with nocons option. First-stage F-stat therefore * correctly includes the constant as an explanatory variable. if ("`first'`ffirst'"!="") & (`endo1_ct' > 0) { tempname firstmat sheapr2 pr2 pr2F pr2p di in gr _newline "First-stage regressions" di in smcl in gr "{hline 23}" if `iv1_ct' > `iv_ct' { di in gr "Warning: collinearities detected among instruments" di in gr "1st stage tests of excluded exogenous variables may be incorrect" } di doFirst "`endo1'" "`inexog1'" "`exexog1'" /* */ `touse' `"`wgt'"' `"`noconstant'"' `"`ffirst'"' capture mat `firstmat'=`godfrey' \ r(firstmat) if _rc==0 { di in gr "Summary results for first-stage regressions:" di di in gr " Shea" di in gr _c "Variable Partial R2 Partial R2 F(" di in gr %3.0f `firstmat'[4,1] "," %6.0f `firstmat'[5,1] ") P-value" local i = 1 local nrvars : word count `endo1' while `i' <= `nrvars' { local vn : word `i' of `endo1' scalar `sheapr2'=`firstmat'[1,`i'] scalar `pr2'=`firstmat'[2,`i'] scalar `pr2F'=`firstmat'[3,`i'] scalar `pr2p'=`firstmat'[6,`i'] di in y %-12s "`vn'" _col(15) %8.4f `sheapr2' _col(31) %8.4f `pr2' /* */ _col(49) %8.2f `pr2F' _col(64) %8.4f `pr2p' local i = `i' + 1 } di } else { di in ye "Warning: missing values encountered; first stage regression results not saved" } } * Post results * If robust, the finite sample correction is N/(N-K), and with no small * we change this to 1 (a la Davidson & MacKinnon 1993, p. 554, HC0). * If cluster, the finite sample correction is (N-1)/(N-K)*M/(M-1), and with no small * we change this to 1 (a la Wooldridge 2002, p. 193), where M=number of clusters. * In the adj of the V matrix for non-small, we use Nprec instead of N because * iweights rounds off N. Note that iweights are not allowed with robust * but the robust section uses Nprec anyway to maintain consistency of code. local NminusK = `N'-`rhs_ct' if "`small'"!="" { est post `b' `V', dep(`depname') obs(`N') esample(`touse') /* */ dof(`NminusK') } else { * Remove finite sample corrections if "`cluster'"=="" { matrix `V'=`V'*(`Nprec'-`rhs_ct')/`Nprec' scalar `rmse'=sqrt(`rss'/`Nprec') } else { matrix `V'=`V'*(`Nprec'-`rhs_ct')/(`Nprec'-1) /* */ * (`N_clust'-1)/`N_clust' scalar `rmse'=sqrt(`rss'/`Nprec') } est post `b' `V', dep(`depname') obs(`N') esample(`touse') } if "`gmm'"=="" { est scalar F=`F' est scalar p=Ftail(`df_m',`df_r',`F') est scalar Fdf2=`df_r' } else { est matrix W `Sinv' } if "`small'"!="" { est scalar df_r=`df_r' est local small "small" } if "`cluster'"!="" { est scalar N_clust=`N_clust' est local clustvar `cluster' } if "`robust'`cluster'`gmm'" != "" { estimates local vcetype "Robust" } est scalar df_m=`df_m' est scalar r2=`r2' est scalar rmse=`rmse' est scalar rss=`rss' est scalar mss=`mss' est scalar r2_a=`r2_a' est scalar yy=`yy' est scalar yyc=`yyc' est scalar r2u=`r2u' est scalar r2c=`r2c' est scalar rankzz=`iv_ct' est scalar rankxx=`rhs_ct' if "`gmm'`robust'`cluster'"=="" { est scalar sargan=`sargan' if `sargan' != 0 & `sargan' != . { est scalar sarganp=`sarganp' est scalar sargandf=`sargandf' } } else { est scalar j=`j' if `j' != 0 & `j' != . { est scalar jp=`jp' est scalar jdf=`jdf' } } if "`orthog'"!="" { est scalar cstat=`cstat' if `cstat'!=0 & `cstat' != . { est scalar cstatp=`cstatp' est scalar cstatdf=`cstatdf' est local clist `clist1' } } if "`first'`ffirst'" != "" & `endo1_ct'>0 { * Capture here because firstmat empty if mvs encountered in 1st stage regressions capture est matrix first `firstmat' } est local depvar `lhs' if "`gmm'"=="" { if "`endo1'" == "" { est local model "ols" } else { est local model "iv" } } else { est local model "gmm" } if `"`exp'"' != "" { est local wexp "`exp'" est local wtype `weight' } est local cmd ivreg2 est local version `version' if "`noconstant'`hascons'"!="" {est scalar cons=0} else {est scalar cons=1} est local instd `endo1' est local insts `insts1' est local predict "ivreg2_p" } *************************************************************** * Display results * Prepare for insuff # of clusters problem * Macro `clustprob' is either "1" (problem) or "0" or "" (no problem) capture local clustprob = string(e(N_clust) <= e(rankzz)) if "`noheader'"=="" { if "`e(model)'"=="gmm" { if "`e(instd)'"=="" { di in gr _n "HOLS-GMM estimation" di in gr "{hline 19}" } else { di in gr _n "GMM estimation" di in gr "{hline 14}" } } else { if "`e(instd)'"=="" { if "`robust'"!="" | "`cluster'"!="" { di in gr _n "OLS regression with robust standard errors" di in gr "{hline 42}" } else { di in gr _n "Ordinary Least Squares (OLS) regression" di in gr "{hline 39}" } } else { if "`robust'"!="" | "`cluster'"!="" { di in gr _n "IV (2SLS) regression with robust standard errors" di in gr "{hline 48}" } else { di in gr _n "Instrumental variables (2SLS) regression" di in gr "{hline 40}" } } } di if "`e(clustvar)'"!="" { di in gr "Number of clusters (" "`e(clustvar)'" ") = " in ye %-4.0f e(N_clust) _continue } else { di in gr " " _continue } di in gr _col(55) "Number of obs = " in ye %8.0f e(N) local Fdf1=e(df_m) if "`e(clustvar)'"=="" { local Fdf2=e(N)-e(rankzz) } else { local Fdf2=e(N_clust)-1 } * No Stata-generated F stat for GMM, or if insuff # of clusters problem if "`e(model)'" != "gmm" & "`clustprob'" != "1" { di in gr _col(55) "F(" %3.0f `Fdf1' "," %6.0f `Fdf2' ") = " in ye %8.2f e(F) di in gr _col(55) "Prob > F = " in ye %8.4f e(p) } di in gr "Total (centered) SS = " in ye %12.0g e(yyc) _continue di in gr _col(55) "Centered R2 = " in ye %8.4f e(r2c) di in gr "Total (uncentered) SS = " in ye %12.0g e(yy) _continue di in gr _col(55) "Uncentered R2 = " in ye %8.4f e(r2u) di in gr "Residual SS = " in ye %12.0g e(rss) _continue di in gr _col(55) "Root MSE = " in ye %8.1g e(rmse) di } * Display coefficients etc. * Unfortunate but necessary hack here: to suppress message about cluster adjustment of * standard error, clear e(clustvar) and then reset it after display local cluster `e(clustvar)' est local clustvar est di, `plus' `efopt' level(`level') est local clustvar `cluster' * Display footer if "`nofooter'"=="" { * Report either (a) Sargan-Hansen-C stats, or (b) insufficient # of clusters problem if "`clustprob'" == "0" | "`clustprob'" == "" { * Display Sargan-Hansen statistic if "`e(vcetype)'" == "Robust" { if "`e(instd)'" != "" { di in gr _c "Hansen J statistic (overidentification test of all instruments): " } else { di in gr _c "Hansen J statistic (Lagrange mulitplier test of excluded instruments): " } di in ye _col(72) %7.3f e(j) local overiddf e(jdf) local overidp e(jp) } else { if "`e(instd)'" != "" { di in gr _c "Sargan statistic (overidentification test of all instruments): " } else { di in gr _c "Sargan statistic (Lagrange mulitplier test of excluded instruments): " } di in ye _col(72) %7.3f e(sargan) local overiddf e(sargandf) local overidp e(sarganp) } if e(rankxx) < e(rankzz) { di in gr _col(52) "Chi-sq(" in ye `overiddf' /* */ in gr ") P-val = " in ye _col(72) %6.5f `overidp' } else { di in gr _col(50) "(equation exactly identified)" } * Display orthog option: C statistic (difference of Sargan statistics) if e(cstat) != . { * If C-stat = 0 then warn, otherwise output if e(cstat) > 0 { di in gr _c "C statistic (exogeneity/orthogonality of specified instruments): " di in ye _col(73) %6.3f e(cstat) di in gr _col(52) "Chi-sq(" in ye e(cstatdf) /* */ in gr ") P-val = " in ye _col(72) %6.5f e(cstatp) di in gr "Instruments tested: `e(clist)'" } if e(cstat) == 0 { di in gr _n "Collinearity/identification problems in restricted equation:" di in gr " C statistic not calculated for orthog option" } } } else { di in r "Warning: insufficient number of clusters to calculate robust var-cov matrix." di in r " Test statistics not reported." } di in smcl in gr "{hline 78}" * Warn about dropped instruments if any * (Re-)calculate number of user-supplied instruments local iv1_ct : word count `e(insts)' local iv1_ct = `iv1_ct' + `e(cons)' if `iv1_ct' > e(rankzz) { di in gr "Collinearities detected among instruments: " _c di in gr `iv1_ct'-e(rankzz) " instrument(s) dropped" } if "`e(instd)'" != "" { di in gr "Instrumented: " _c Disp `e(instd)' } di in gr "Instruments: " _c Disp `e(insts)' di in smcl in gr "{hline 78}" } end program define IsStop, sclass /* sic, must do tests one-at-a-time, * 0, may be very large */ if `"`0'"' == "[" { sret local stop 1 exit } if `"`0'"' == "," { sret local stop 1 exit } if `"`0'"' == "if" { sret local stop 1 exit } if `"`0'"' == "in" { sret local stop 1 exit } if `"`0'"' == "" { sret local stop 1 exit } else sret local stop 0 end program define Disp local first "" local piece : piece 1 64 of `"`0'"' local i 1 while "`piece'" != "" { di in gr "`first'`piece'" local first " " local i = `i' + 1 local piece : piece `i' 64 of `"`0'"' } if `i'==1 { di } end * Performs first-stage regressions * Modified by MES to produce an F-test of the excluded instruments program define doFirst, rclass /* /* */ */ args endog /* variable list (including depvar) */ inexog /* list of included exogenous */ exexog /* list of excluded exogenous */ touse /* touse sample */ weight /* full weight expression w/ [] */ nocons /* noconstant option */ ffirst /* display f-stat only */ tokenize `endog' tempname statmat statmat1 local i 1 while "``i''" != "" { * di in gr "First-stage regression of ``i'':" if "`ffirst'" != "" { quietly regress ``i'' `inexog' `exexog' `weight' /* */ if `touse', `nocons' } else { di in gr "First-stage regression of ``i'':" regress ``i'' `inexog' `exexog' `weight' if `touse', `nocons' } tempvar y2 iota xhat tempname ysum yy rssall rssinc pr2 F p quietly predict double `xhat' if `touse', xb local endoghat "`endoghat' `xhat'" quietly test `exexog' scalar `F'=r(F) scalar `p'=r(p) local df=r(df) local df_r=r(df_r) scalar `rssall'=e(rss) qui gen double `y2'=``i''^2 * Stata summarize won't work with iweights, so must use matrix cross-product qui gen `iota'=1 qui matrix vecaccum `ysum' = `iota' `y2' `weight' if `touse', noconstant scalar `yy'=`ysum'[1,1] * 1st stage regression without excluded exogenous quietly regress ``i'' `inexog' `weight' if `touse', `nocons' scalar `rssinc'=e(rss) * NB: uncentered R2 for main regression is 1-rssall/yy; for restricted is 1-rssinc/yy; * squared semipartial correlation=(rssinc-rssall)/yy=diff of 2 R2s * Squared partial correlation (="partialled-out R2") scalar `pr2'=(`rssinc'-`rssall')/`rssinc' if "`ffirst'" == "" { di in gr "Partial R-squared of excluded instruments: " _c di in ye %8.4f `pr2' di in gr "Test of excluded instruments:" di in gr " F(" %3.0f `df' "," %6.0f `df_r' ") = " in ye %8.2f `F' di in gr " Prob > F = " in ye %8.4f `p' di } capture { mat `statmat1' = (`pr2' \ `F' \ `df' \ `df_r' \ `p') mat colname `statmat1' = ``i'' if `i'==1 {mat `statmat'=`statmat1'} else {mat `statmat' = `statmat' , `statmat1'} } local i = `i' + 1 } capture mat rowname `statmat' = pr2 F df df_r pvalue if _rc==0 { return matrix firstmat `statmat' } end * Remove all tokens in dirt from full * Returns "cleaned" full list in cleaned program define Subtract /* : */ args cleaned /* macro name to hold cleaned list */ colon /* ":" */ full /* list to be cleaned */ dirt /* tokens to be cleaned from full */ tokenize `dirt' local i 1 while "``i''" != "" { local full : subinstr local full "``i''" "", word all local i = `i' + 1 } tokenize `full' /* cleans up extra spaces */ c_local `cleaned' `*' end