BUGFIX: DoF - nested-within-cluster

reghdfe was not checking if the errors were nested within the cluster
variable
This means the reported DoF were wrong on those cases and we were being
too drastic when computing the variance estimates

package/hdfe.ado

@@ -1,4 +1,4 @@
-*! hdfe 2.1.6 02apr2015
+*! hdfe 2.1.8 03apr2015
 *! Sergio Correia ([email protected])
 * (built from multiple source files using build.py)
 // -------------------------------------------------------------
@@ -1020,7 +1020,7 @@ end
 // -------------------------------------------------------------
 
 program define Version, eclass
-    local version "2.1.6 02apr2015"
+    local version "2.1.8 03apr2015"
     ereturn clear
     di as text "`version'"
     ereturn local version "`version'"
@@ -1320,6 +1320,7 @@ syntax, [DOFadjustments(string) group(name) uid(varname) groupdta(string)]
 					assert inlist(_rc, 0, 498)
 					if (!_rc) {
 						Debug, level(1) msg("(categorical variable " as result "`varlabel'" as text " is nested within cluster variable " as result "`clustervar_original'" as text ", so it doesn't count towards DoF)")
+						local absvar_in_clustervar 1
 						continue, break
 					}
 				}

package/hdfe.pkg

@@ -26,7 +26,7 @@ d KW: within
 d
 d Requires: Stata version 11.2
 d
-d Distribution-Date: 20150402
+d Distribution-Date: 20150403
 d
 
 f hdfe.ado

package/reghdfe.ado

@@ -1,4 +1,4 @@
-*! reghdfe 2.1.6 02apr2015
+*! reghdfe 2.1.8 03apr2015
 *! Sergio Correia ([email protected])
 * (built from multiple source files using build.py)
 // -------------------------------------------------------------
@@ -861,7 +861,7 @@ end
 // -------------------------------------------------------------
 
 program define Version, eclass
-    local version "2.1.6 02apr2015"
+    local version "2.1.8 03apr2015"
     ereturn clear
     di as text "`version'"
     ereturn local version "`version'"
@@ -3526,6 +3526,7 @@ syntax, [DOFadjustments(string) group(name) uid(varname) groupdta(string)]
 					assert inlist(_rc, 0, 498)
 					if (!_rc) {
 						Debug, level(1) msg("(categorical variable " as result "`varlabel'" as text " is nested within cluster variable " as result "`clustervar_original'" as text ", so it doesn't count towards DoF)")
+						local absvar_in_clustervar 1
 						continue, break
 					}
 				}

package/reghdfe.pkg

@@ -16,7 +16,7 @@ d KW: instrumental variables
 d
 d Requires: Stata version 11.2
 d
-d Distribution-Date: 20150402
+d Distribution-Date: 20150403
 d
 
 f reghdfe.ado

source/_hdfe/EstimateDoF.ado

@@ -121,6 +121,7 @@ syntax, [DOFadjustments(string) group(name) uid(varname) groupdta(string)]
 					assert inlist(_rc, 0, 498)
 					if (!_rc) {
 						Debug, level(1) msg("(categorical variable " as result "`varlabel'" as text " is nested within cluster variable " as result "`clustervar_original'" as text ", so it doesn't count towards DoF)")
+						local absvar_in_clustervar 1
 						continue, break
 					}
 				}

source/hdfe.pkg

@@ -26,7 +26,7 @@ d KW: within
 d
 d Requires: Stata version 11.2
 d
-d Distribution-Date: 20150402
+d Distribution-Date: 20150403
 d
 
 f hdfe.ado

source/reghdfe.pkg

@@ -16,7 +16,7 @@ d KW: instrumental variables
 d
 d Requires: Stata version 11.2
 d
-d Distribution-Date: 20150402
+d Distribution-Date: 20150403
 d
 
 f reghdfe.ado

source/version.txt

@@ -1 +1 @@
-2.1.6 02apr2015
+2.1.8 03apr2015

test/super.do

@@ -21,6 +21,7 @@
 	run test-cluster
 	run test-ivreg2
 	run test-cluster-same-as-absvar
+	run test-cluster-nested // Prevent regression of bug reported by Michael Wittry (see email)
 	run test-vce-complex-bw
 	run test-weights
 	run test-mwc

test/test-cluster-nested.do

@@ -0,0 +1,132 @@
+cscript "reghdfe fixed effect nested in cluster" adofile reghdfe
+
+* Setup
+	discard
+	clear all
+	set more off
+	* cls
+
+* Convenience: "Trim <size>" will trim e(b) and e(V)
+	capture program drop TrimMatrix
+	program define TrimMatrix, eclass
+	args size
+		assert `size'>0
+		matrix trim_b = e(b)
+		matrix trim_V = e(V)
+		matrix trim_b = trim_b[1, 1..`size']
+		matrix trim_V = trim_V[1..`size',1..`size']
+		ereturn matrix trim_b = trim_b
+		ereturn matrix trim_V = trim_V
+	end
+
+* Create fake dataset
+	sysuse auto
+	bys turn: gen t = _n
+
+* [TEST]
+	local lhs price
+	local rhs weight length
+	local absvars turn // make it the same as panelvar
+	local clustervar `absvars'
+	fvunab tmp : `rhs'
+	local K : list sizeof tmp
+
+	tsset `absvars' t
+
+	* Custom adjustment: in this simple case we can compare _cons
+	drop if missing(rep)
+
+// -------------------------------------------------------------------------------------------------
+
+	* Check that _xtreg_chk_cl2 is working
+
+	reghdfe `lhs' `rhs', abs(`absvars') vce(cluster `clustervar') dof(clusters)
+	assert e(df_a)==1
+	reghdfe `lhs' `rhs', abs(`absvars') vce(cluster `clustervar')
+	assert e(df_a)==1
+
+// -------------------------------------------------------------------------------------------------
+
+	* Match against xtreg
+
+	* 1. Run benchmark
+	xtreg `lhs' `rhs', cluster(`clustervar') fe
+	TrimMatrix `K'
+	di e(df_a)
+	di e(df_m)
+	local bench_df_a = e(df_a)
+	local bench_within = e(r2_w)
+	storedresults save benchmark e()
+
+	* 2. Run reghdfe
+	reghdfe `lhs' `rhs', absorb(`absvars') vce(cluster `clustervar')
+	di e(df_a)
+	di e(df_m)
+	* NOTE: See statalist post for discussion on df_m discrepancy
+	* http://www.stata.com/statalist/archive/2010-03/msg00941.html	
+	* "So I think some explanation is necessary. I see no reason, conceptually, why xtreg,fe with small-sample statistics should not be exactly equivalent to areg"
+	TrimMatrix `K'
+
+	* 3. Compare
+	storedresults compare benchmark e(), tol(1e-12) include( ///
+		scalar: N tss rss F df_r ///  
+		matrix: trim_b trim_V ///
+		macros: wexp wtype )
+
+	* WILL NOT HOLD B/C xtreg still include the nested ones in e(df_a)
+	* assert `bench_df_a'==e(df_a)-1
+
+	assert abs(`bench_within'-e(r2_within))<1e-6
+	storedresults drop benchmark
+
+// -------------------------------------------------------------------------------------------------
+
+	* Check that _xtreg_chk_cl2 is working
+
+	reghdfe `lhs' `rhs', abs(`absvars') vce(cluster `clustervar') dof(clusters)
+	assert e(df_a)==1
+	reghdfe `lhs' `rhs', abs(`absvars') vce(cluster `clustervar')
+	assert e(df_a)==1
+
+// -------------------------------------------------------------------------------------------------
+// Now repeat the above but with a different clustervar name
+// -------------------------------------------------------------------------------------------------
+	gen samecluster = `absvars'
+	local clustervar samecluster
+
+	* Match against xtreg
+
+	* 1. Run benchmark
+	xtreg `lhs' `rhs', cluster(`clustervar') fe
+	TrimMatrix `K'
+	di e(df_a)
+	di e(df_m)
+	local bench_df_a = e(df_a)
+	local bench_within = e(r2_w)
+	storedresults save benchmark e()
+
+	* 2. Run reghdfe
+	reghdfe `lhs' `rhs', absorb(`absvars') vce(cluster `clustervar')
+	di e(df_a)
+	di e(df_m)
+	* NOTE: See statalist post for discussion on df_m discrepancy
+	* http://www.stata.com/statalist/archive/2010-03/msg00941.html	
+	* "So I think some explanation is necessary. I see no reason, conceptually, why xtreg,fe with small-sample statistics should not be exactly equivalent to areg"
+	TrimMatrix `K'
+
+	* 3. Compare
+	storedresults compare benchmark e(), tol(1e-12) include( ///
+		scalar: N tss rss F df_r ///  
+		matrix: trim_b trim_V ///
+		macros: wexp wtype )
+
+	* WILL NOT HOLD B/C xtreg still include the nested ones in e(df_a)
+	* assert `bench_df_a'==e(df_a)-1
+
+	assert abs(`bench_within'-e(r2_within))<1e-6
+	storedresults drop benchmark
+
+
+// -------------------------------------------------------------------------------------------------
+cd "D:/Github/reghdfe/test"
+exit