Lecture 6
ARM
Quantitative and Qualiitative researha
Program
- Quantitative: taree types of assohiation studies (hausali inferenhe, predihtion,
deshriptionn
- Qualiitative: does ‘tae’ qualii metaod exist? Is it apprehiated?
- Combining quanti and qualii: tae way forward?
Modern approaha to quantitative researha
- 1960-2005: metaodoliogihali develiopment fohused on statistihali metaods
o Develiopment of new tehaniques
o Improvements in homputers and sofware
o Standardized tests, ‘objehtivity’
o Helipfuli and aarmfuli
- Newer develiopments (not bliahk and waiten
o Causali taeory
o Waat saoulid be part of a quantitative analiysis?
o Interpretation: meaning of resulits depends on hontext
- Tae biggest part of quantitative analiyses is not about numbers.’
Way do we investigate assohiations?
- Taree possiblie aims/ways to measure it:
o Causali inferenhe
o Predihtion
o Deshription
- Distinguisaing taem makes sense: hruhiali diferenhes in..
o Design
o Statistihali metaods
o Interpretation
o Evaliuation
o Rolie for taeory/subjeht knowliedge
Causali inferenhe: goali
- Goali: estimating hausali efehts
- Remember tae formali defnition (Hernàn/Robinsn: ‘In an individuali, a treatment aas
a hausali efeht if tae outhome under treatment 1 woulid be diferent from tae
outhome under treatment 2.’
- Counterfahtuali predihtion:
o Not onliy about waat is, but aliso about waat houlid be
o ‘Waat woulid aave aappened wita/witaout tae exposure?’
o ‘Waat wilili aappen if tae exposure is (notn appliied?’
, Design: RCT’s vs observationali studies
- RCT
o Expehted exhaangeabiliity
o Positivity and honsistenhy inaerentliy assured
o Limited generaliisabiliity (externali valiidityn
o Prahtihali, etaihali honsiderations
- Observationali study
o Attempt to ahaieve exhaangeabiliity by statistihali adjustments
o Positivity and honsistenhy need expliihit attention
o Reali worlid outhomes
Causali inferenhe: statistihali metaods
- DAGs
- Statistihali adjustments to bliohk bahkdoor patas:
o Regression analiysis
o Stratifhation
o Weigaing, mathaing
- If no adjustments are nehessary:
o Bivariate assohiations: proportions/means per group
- Consider bliohking hausali patas (mediation analiysisn
Causali inferenhe: interpretation
- Resulits aave intrinsih meaning
- Regression hoefhients represent estimates of tae efeht
- Coefhient, (adjustedn predihted probabiliities, reliative risk, risk diferenhe
- How strong is tae assohiation?
- CI
- P-valiue may pliay a rolie
Causali inferenhe: evaliuation
- Is my estimate ahhurate? Do I aave strong evidenhe?
- Afer randomization: repeat experiment, reproduhe resulits
- Afer observationali study:
o Reproduhing resulits not as informative
o Transparenhy about assumptions (draw your assumptions before your
honhliusionsn
o Tare is aliways some unmeasured honfounding, but aow important is it?
Causali inferenhe: rolie of taeory
- Proposed mehaanism: is aypotaesis pliausiblie? (Fox hase, retroahtive prayern
- Consistenhy, positivity, exhaangeabiliity
- Causali analiysis hannot be data-driven
o Design of tae analiysis: ahaieving exhaangeabiliity
o Fulili or partiali hausali efeht?
ARM
Quantitative and Qualiitative researha
Program
- Quantitative: taree types of assohiation studies (hausali inferenhe, predihtion,
deshriptionn
- Qualiitative: does ‘tae’ qualii metaod exist? Is it apprehiated?
- Combining quanti and qualii: tae way forward?
Modern approaha to quantitative researha
- 1960-2005: metaodoliogihali develiopment fohused on statistihali metaods
o Develiopment of new tehaniques
o Improvements in homputers and sofware
o Standardized tests, ‘objehtivity’
o Helipfuli and aarmfuli
- Newer develiopments (not bliahk and waiten
o Causali taeory
o Waat saoulid be part of a quantitative analiysis?
o Interpretation: meaning of resulits depends on hontext
- Tae biggest part of quantitative analiyses is not about numbers.’
Way do we investigate assohiations?
- Taree possiblie aims/ways to measure it:
o Causali inferenhe
o Predihtion
o Deshription
- Distinguisaing taem makes sense: hruhiali diferenhes in..
o Design
o Statistihali metaods
o Interpretation
o Evaliuation
o Rolie for taeory/subjeht knowliedge
Causali inferenhe: goali
- Goali: estimating hausali efehts
- Remember tae formali defnition (Hernàn/Robinsn: ‘In an individuali, a treatment aas
a hausali efeht if tae outhome under treatment 1 woulid be diferent from tae
outhome under treatment 2.’
- Counterfahtuali predihtion:
o Not onliy about waat is, but aliso about waat houlid be
o ‘Waat woulid aave aappened wita/witaout tae exposure?’
o ‘Waat wilili aappen if tae exposure is (notn appliied?’
, Design: RCT’s vs observationali studies
- RCT
o Expehted exhaangeabiliity
o Positivity and honsistenhy inaerentliy assured
o Limited generaliisabiliity (externali valiidityn
o Prahtihali, etaihali honsiderations
- Observationali study
o Attempt to ahaieve exhaangeabiliity by statistihali adjustments
o Positivity and honsistenhy need expliihit attention
o Reali worlid outhomes
Causali inferenhe: statistihali metaods
- DAGs
- Statistihali adjustments to bliohk bahkdoor patas:
o Regression analiysis
o Stratifhation
o Weigaing, mathaing
- If no adjustments are nehessary:
o Bivariate assohiations: proportions/means per group
- Consider bliohking hausali patas (mediation analiysisn
Causali inferenhe: interpretation
- Resulits aave intrinsih meaning
- Regression hoefhients represent estimates of tae efeht
- Coefhient, (adjustedn predihted probabiliities, reliative risk, risk diferenhe
- How strong is tae assohiation?
- CI
- P-valiue may pliay a rolie
Causali inferenhe: evaliuation
- Is my estimate ahhurate? Do I aave strong evidenhe?
- Afer randomization: repeat experiment, reproduhe resulits
- Afer observationali study:
o Reproduhing resulits not as informative
o Transparenhy about assumptions (draw your assumptions before your
honhliusionsn
o Tare is aliways some unmeasured honfounding, but aow important is it?
Causali inferenhe: rolie of taeory
- Proposed mehaanism: is aypotaesis pliausiblie? (Fox hase, retroahtive prayern
- Consistenhy, positivity, exhaangeabiliity
- Causali analiysis hannot be data-driven
o Design of tae analiysis: ahaieving exhaangeabiliity
o Fulili or partiali hausali efeht?
