Potential outcomes and experiments
causality
u I treated
Let u betreatment
y u 0 nottreated
outcome w treatmentforperson i
y bethefaggoting Yai
f
is
onlyoneofyo or y isrealised
I ESTIMANDSOFINTEREST
AugTreatmentEffect ATE1
CAUSAL
E Yi yo
howmuchydiffers iftreated
forsamepersonacrossstates averagedacrosspop
Treated ATTI
AugTreatmentEffecton
E yi yo u 1 1
howmuchydiffersamongthosechoosingtoparticipate
, REGRESSIONESTIMANDS
y Bo t pin t a
Bois E y u 1l E y u o
Coutuyl
Boes var ul
Yi i
Pli pl E y u 1l E y u of
pl i pl
Bois E y u 1l E y u o
Bois E ly u 1 E yo u o
E ly u 11 E yo u 11 E yo1 u 11 E yo u 01
AugTreatmentEffect att
selectionbias
Treated
on
negativeif
selfselected c untreated
y y
i Given
yoismean independentof u
E yo u 1 E yo u 0 O
holdsittreat s
Bois E ly Yo u 1 I randomlyassignedin
randomizedcontrolledtrial Rct
AugTreatmentEffect Att
onTreated
v
i u h yo y l
ye
treatment u control
Bois E ly yo u 11 similaronpotential
outcomes
E ly yoI yo y
ATT ATEI
n is notindependent
, eg
CONTROLVARIABLES
y potBasu paw tu
w is predetermined or G w causallyaffectedby u
u k yo ya w u yo yall W
Boos still Ely yo I
justu andEInt
residuals
OVERCONTROLLING
controllingforw t Var a o vB
um
Yi Yai ni t yo l mil
Yo t Yai Yailui
Be t Baini t ai
EEyo I causaleffect
outcome
wo
treatment
ifcausaleffectsareconstantacrosspop Bai B1
if E ui ui 0 ELYoilwit ELYoil
NB onlyforATEW HETEROGENOUSEFFECTS
, ISATIONBASEDONCO
RANDOM VARIATES
controlVARIABLES covariates w I are
moreimportantif randomisation isbased
on w'svalues
reg y on u has
selectionbias
yo mean independentof u isn'treq
u k y mayberelatedto w
similarly
unconfoundness holds
u Yo yall W
reggroupbygroup
t
Ely yo I w
groupspecificATE
removesselectionbias
OR
controlforallnon experimentaldeterminants oftreatment u
I
addsgroupdummies
groupfixed
effects
groupspecificintercepts
if wi u
I wi u I
if wi t U
B1 weightedaverageofgroupspecificATE
Ely yo I w u
causality
u I treated
Let u betreatment
y u 0 nottreated
outcome w treatmentforperson i
y bethefaggoting Yai
f
is
onlyoneofyo or y isrealised
I ESTIMANDSOFINTEREST
AugTreatmentEffect ATE1
CAUSAL
E Yi yo
howmuchydiffers iftreated
forsamepersonacrossstates averagedacrosspop
Treated ATTI
AugTreatmentEffecton
E yi yo u 1 1
howmuchydiffersamongthosechoosingtoparticipate
, REGRESSIONESTIMANDS
y Bo t pin t a
Bois E y u 1l E y u o
Coutuyl
Boes var ul
Yi i
Pli pl E y u 1l E y u of
pl i pl
Bois E y u 1l E y u o
Bois E ly u 1 E yo u o
E ly u 11 E yo u 11 E yo1 u 11 E yo u 01
AugTreatmentEffect att
selectionbias
Treated
on
negativeif
selfselected c untreated
y y
i Given
yoismean independentof u
E yo u 1 E yo u 0 O
holdsittreat s
Bois E ly Yo u 1 I randomlyassignedin
randomizedcontrolledtrial Rct
AugTreatmentEffect Att
onTreated
v
i u h yo y l
ye
treatment u control
Bois E ly yo u 11 similaronpotential
outcomes
E ly yoI yo y
ATT ATEI
n is notindependent
, eg
CONTROLVARIABLES
y potBasu paw tu
w is predetermined or G w causallyaffectedby u
u k yo ya w u yo yall W
Boos still Ely yo I
justu andEInt
residuals
OVERCONTROLLING
controllingforw t Var a o vB
um
Yi Yai ni t yo l mil
Yo t Yai Yailui
Be t Baini t ai
EEyo I causaleffect
outcome
wo
treatment
ifcausaleffectsareconstantacrosspop Bai B1
if E ui ui 0 ELYoilwit ELYoil
NB onlyforATEW HETEROGENOUSEFFECTS
, ISATIONBASEDONCO
RANDOM VARIATES
controlVARIABLES covariates w I are
moreimportantif randomisation isbased
on w'svalues
reg y on u has
selectionbias
yo mean independentof u isn'treq
u k y mayberelatedto w
similarly
unconfoundness holds
u Yo yall W
reggroupbygroup
t
Ely yo I w
groupspecificATE
removesselectionbias
OR
controlforallnon experimentaldeterminants oftreatment u
I
addsgroupdummies
groupfixed
effects
groupspecificintercepts
if wi u
I wi u I
if wi t U
B1 weightedaverageofgroupspecificATE
Ely yo I w u
