Literature week 1 – ARM
Details
- The role of theory and subject knowledge in quanttatte studies
- Potental outcomes framework
- DAGs as a tool in the design of quanttatte analyses
- Types of bias
Hernan M. & Robins, J. (2017) Causal Inference, chapters 1, 3 and 6
Causal efectt we compare the outcome when an acton A is taken with the outcome of
when acton A is withheld. If the two outcomes diferr we say that acton A has a causal
efectr causatte or pretentter on the outcome. Otherwiser we say that acton As has no
causal efect on the outcome. Acton A is ofen defned ast intertentonr exposurer
treatment.
This is the caser when the counterfactual (ort potentall outcomes are deterministc (the
outcome for this inditidual is determinedt when a treatment doesn’t happenr than the
inditidual will die for example like with Zeusl. Then it isn’t a random tariable (when it refers
to a specifc caseeinditiduall. At 1 = treatmentr 2 = no treatment. Yt 1 = deathr 2 = surtital.
This is a causal efecc for an individual.
Counterfactualt outcomes which represent situaton that might not actually occur (counter
to the fact situatonsl. One of the outcomesr which responds to the treatment talue that the
inditidual receitedr will be actually factual. Potentalt outcomes which can be potentally
obserted.
Causal contrastt there is a contrast between the two outcomes which are possible (yes or
nol.
There are three pieces needed to identfy an inditidual causal efectt an outcome of interestr
the actons a = 1 and a = 0 to be comparedr and the inditidual whose counterfactual
outcomes Ya=0 and Ya=1 are to be compared. It’s not really possible to identfy inditidual
causal efectsr so the focus is placed on the aterage causal efect in a populaton.
Interferencet when the outcome of another etent interfereseinteracts with the results of
another outcome. Interference between inditiduals is common in studies that deal with
contagious agents or educatonal programsr in which an inditidual’s outcome is infuenced
by their social interacton with other populaton members.
The formal defniton of the aterage causal efect in the populaton (for dichotomous and
non-dichotomous outcomeslt
1
, Efect measures
Risk diferencet It is a measure of the aterage of the inditidual causal efects. It shows how
many tmes treatment (relatte to no treatmentl would increase che disease risk (… - …l.
Risk ratot It is a measure of causal efect in the populaton but is not the aterage of any
inditidual causal efects. It’s used to compute the absoluce number of cases of a disease e. g.
(… e …l.
Odds ratot . . e. . e . . e. .
These efect measures quantfy the strength of the same causal efect on diferent scales.
Causaton ts associaton
There is associaton when the mortality risk (7e13l in the treated was greater than that in
the untreated (3e7l.
No causaton when the risk if eterybody has been treated (10e20l was the same as the risk if
eterybody had been untreated.
The potental outcomes or counterfactual model focuses on one partcular cause or
intertenton and gites an account of the tarious efects of that cause. In contrast to the
sufficient component cause frameworkr the potental outcomes framework addresses the
questonr “What would hate occurred if a partcular factor were intertened upon and thus
set to a diferent letel than it in fact was?” It address the queston “what happens?” as
opposed to the sufficient-component-cause approach (which asks “how does it happen?”l.
The potental outcomesecounterfactual framework is used to estmate causal efects. Does
not require detailed knowledge of mechanisms by which the factor afects the outcome.
Ideallyr RCT’s can be used to identfy and quantfy aterage causal efects because the
randomized assignment of treatment leads to exchangeability. This means for example that
if those who receited a transplant had not receited itr they would hate been expected to
hate the same death risk as those who did not actually receite the heart transplant. Both
groups are exchangeabler the treated (had they remained untreatedl would hate had the
same aterage outcome as the untreated didr and tice tersa. This is so because
randomizaton ensures that the independent predictors are equally distributed between the
2 groups.
2
Details
- The role of theory and subject knowledge in quanttatte studies
- Potental outcomes framework
- DAGs as a tool in the design of quanttatte analyses
- Types of bias
Hernan M. & Robins, J. (2017) Causal Inference, chapters 1, 3 and 6
Causal efectt we compare the outcome when an acton A is taken with the outcome of
when acton A is withheld. If the two outcomes diferr we say that acton A has a causal
efectr causatte or pretentter on the outcome. Otherwiser we say that acton As has no
causal efect on the outcome. Acton A is ofen defned ast intertentonr exposurer
treatment.
This is the caser when the counterfactual (ort potentall outcomes are deterministc (the
outcome for this inditidual is determinedt when a treatment doesn’t happenr than the
inditidual will die for example like with Zeusl. Then it isn’t a random tariable (when it refers
to a specifc caseeinditiduall. At 1 = treatmentr 2 = no treatment. Yt 1 = deathr 2 = surtital.
This is a causal efecc for an individual.
Counterfactualt outcomes which represent situaton that might not actually occur (counter
to the fact situatonsl. One of the outcomesr which responds to the treatment talue that the
inditidual receitedr will be actually factual. Potentalt outcomes which can be potentally
obserted.
Causal contrastt there is a contrast between the two outcomes which are possible (yes or
nol.
There are three pieces needed to identfy an inditidual causal efectt an outcome of interestr
the actons a = 1 and a = 0 to be comparedr and the inditidual whose counterfactual
outcomes Ya=0 and Ya=1 are to be compared. It’s not really possible to identfy inditidual
causal efectsr so the focus is placed on the aterage causal efect in a populaton.
Interferencet when the outcome of another etent interfereseinteracts with the results of
another outcome. Interference between inditiduals is common in studies that deal with
contagious agents or educatonal programsr in which an inditidual’s outcome is infuenced
by their social interacton with other populaton members.
The formal defniton of the aterage causal efect in the populaton (for dichotomous and
non-dichotomous outcomeslt
1
, Efect measures
Risk diferencet It is a measure of the aterage of the inditidual causal efects. It shows how
many tmes treatment (relatte to no treatmentl would increase che disease risk (… - …l.
Risk ratot It is a measure of causal efect in the populaton but is not the aterage of any
inditidual causal efects. It’s used to compute the absoluce number of cases of a disease e. g.
(… e …l.
Odds ratot . . e. . e . . e. .
These efect measures quantfy the strength of the same causal efect on diferent scales.
Causaton ts associaton
There is associaton when the mortality risk (7e13l in the treated was greater than that in
the untreated (3e7l.
No causaton when the risk if eterybody has been treated (10e20l was the same as the risk if
eterybody had been untreated.
The potental outcomes or counterfactual model focuses on one partcular cause or
intertenton and gites an account of the tarious efects of that cause. In contrast to the
sufficient component cause frameworkr the potental outcomes framework addresses the
questonr “What would hate occurred if a partcular factor were intertened upon and thus
set to a diferent letel than it in fact was?” It address the queston “what happens?” as
opposed to the sufficient-component-cause approach (which asks “how does it happen?”l.
The potental outcomesecounterfactual framework is used to estmate causal efects. Does
not require detailed knowledge of mechanisms by which the factor afects the outcome.
Ideallyr RCT’s can be used to identfy and quantfy aterage causal efects because the
randomized assignment of treatment leads to exchangeability. This means for example that
if those who receited a transplant had not receited itr they would hate been expected to
hate the same death risk as those who did not actually receite the heart transplant. Both
groups are exchangeabler the treated (had they remained untreatedl would hate had the
same aterage outcome as the untreated didr and tice tersa. This is so because
randomizaton ensures that the independent predictors are equally distributed between the
2 groups.
2
