De Bias
finition: any syste atic e m rror that distorts thetruth and re sults in an incorre e ateof association (non-randome ct stim rror) always a pote ntial re gardle of how rigorously de ss signe the d inve stigation is Type of bias in e ating causal e cts in e m s stim ffe pide iologic studie s S le e ction bias I nform ation bias C onfounding
S le e ction bias
Occurs whe there n lations for thesam pling validity of thecausal contrast study de do not hold sign S le e ction biase aredistortions (of e m s pide iologic studie that re s) sult fromproce s use to se ct subje and fromfactors that dure d le cts influe study participation nce
How we participants se cte for thestudy? re le d Did the havetheopportunity to de y clineparticipation? What factors m haveinflue d participation? ay nce
S le e ction bias
Re sults fromproce s use to se ct subje and from dure d le cts factors that influe study participation such that: nce
there lation be e e twe n xposureand dise is diffe nt for those ase re who participatein thestudy and all thosewho should be e ligiblefor study, including non-participants
Doe not re fromdiffe ntial se ction by e s sult re le xposure status aloneor by outcom status alone e
S le e ction bias
Occurs if:
cohort study dise status influe s se ction or ase nce le participation to a diffe nt de ein e re gre xpose and une d xpose d com parison groups case -control study e xposurestatus influe s se ction or nce le participation to a diffe nt de eam case and controls re gre ong s cross-se ctional study e r variableinflue s se ction or ithe nce le participation to a diffe nt de ein com re gre parison groups for study, including non-participants
S le e ction bias
S ource of se ction bias: s le
S le e ctiveloss to follow-up in cohort studie s
bias is introduce if thedise of inte st influe s theloss of subje d ase re nce cts to follow-up diffe ntly for e re xpose and une d xpose cohort m m rs d e be theam ount of bias m diffe for a give am ay r n ount of loss-to-follow-up but thegre r theabsoluteloss, thegre r them umpossiblebias ate ate axim
S le e ction bias
S ource of se ction bias: s le
I ncom teasce ple rtainm nt of case / Diagnostic bias e s
se ction bias is introduce if thee le d xposureof inte st influe s the re nce de ction of dise diffe ntly in e te ase re xpose and une d xpose d
S le e ction bias
S ource of se ction bias: s le
I nappropriatecontrol group
se ction bias is introduce if thecontrols who participatein thestudy le d do not re se thee pre nt xposuree rie of thepopulation fromwhich xpe nce thecase aroseand thepre nceof e s vale xposuream case is not ong s distorte to sam de e d e gre
S le e ction bias
S ource of se ction bias: s le
Diffe ntial m re otivation to participate
se ction bias is introduce if e le d xpose pe d rsons with thedise arem ase ore highly m otivate to participatein re arch than e d se xpose pe d rsons without thedise and une ase xpose pe d rsons with or without thedise ase
S le e ction bias
How doe se ction bias e ct m asure of dise s le ffe e s ase fre ncy? que
Distortion fromse ction bias can re in ove stim le sult re ation or unde stim re ation of m asure of dise fre ncy such as e s ase que pre nceand incide vale nce This m occur without le ay ading to bias in them asureof e association
S le e ction bias
How doe se ction bias e ct m asure of association? s le ffe e s
Usually distortion fromse ction bias re le sults in a downward or upward bias in a m asureof association that m cross thenull (1.0) e ay For e pleif thetruerisk ratio is 1.5, a downward bias m re in a xam ay sult risk ratio of 1.1 or 0.85. If thetruerisk ratio is 0.5 an upward bias m ay re in a risk ratio of 0.85 or 2.0 sult This m occur without le ay ading to bias in m asure of dise fre ncy e s ase que
S le e ction bias
How doe se ction bias e ct m asure of association? s le ffe e s
The aretwo type of bias that ge rally causem asure of re s ne e s association to bebiase toward thenull but not to cross the d null (in contrast to thepre vious slide )
non-diffe ntial m re isclassification (the arese ral e ptions to this) re ve xce bias in S fromusing ge ral population as a re re group MR ne fe nt
S le e ction bias
S pling validity = lack of se ction bias am le
= sam pling fraction of e xpose case d s = sam pling fraction of e xpose de inator d nom = sam pling fraction of une xpose case d s = sam pling fraction of une xpose de inator d nom
S le e ction bias
S pling validity in cohort study am
=e xpose case in sam / e d s ple xpose case in targe d s t =e xpose pe d rsons-at-risk1 in sam /e ple xpose pe d rsons-at-risk1 in targe t or =e xpose pe d rson-tim in sam /e e ple xpose pe d rson-tim in targe e t = une xpose case in sam / une d s ple xpose case in targe d s t = une xpose pe d rsons-at-risk1 in sam / une ple xpose pe d rsons-at-risk1 in targe t or = une xpose pe d rson-tim in sam /une e ple xpose pe d rson-tim in targe e t
1
num r of pe be rsons at risk in cohort at start of follow-up for RR and RD; num r of noncase for OR be s
S le ction e bias
S pling validity in case am -control study
=e xpose case in sam / e d s ple xpose case in targe d s t =e xpose non-case in sam /e d s ple xpose non-case in targe d s t = une xpose case in sam / une d s ple xpose case in targe d s t = une xpose non-case in sam / une d s ple xpose non-case in targe d s t
S le e ction bias: sam pling
A A C C B D D =C /C =D/D =A/A =B/B B
Actual
Targe t
S le e ction bias
Ratio m asure arefre of se ction bias if: e s e le
=1
Diffe ncem asure arefre of se ction bias if: re e s e le
=1 & =1
S le e ction bias
For random d trials and cohort studie m asure of association ize s, e s arefre fromse ction bias if: e le
= & =
For case -control studie m asure of association arefre from s, e s e se ction bias if: le
=
S le e ction bias: C -Control S ase tudy
Hypothe tical study
I nclude all case and non-case of a de d population s s s fine Assum no confounding or m ing isclassification
Total population Risk factor Pre nt se Abse nt Oddsexposure Case s 500 500 1.0:1.0 C ontrols 1800 7200 1.0:4.0
OR=4.0
S le e ction bias: C -Control S ase tudy
Hypothe tical study
Unbiase sam d pling of 50%case and 10%controls s Assum no confounding or m ing isclassification
Total population Risk factor Pre nt se Abse nt Oddsexposure Case s 250 250 1.0:1.0 C ontrols 180 720 1.0:4.0
OR=4.0
Unbiase Oddsexposure for case and controls; d s unbiase OR d
S le e ction bias: C -Control S ase tudy
Hypothe tical study
S pling of 50%case varie (unknown to inve am s s stigator) by e xposurestatus (sam pling fraction .60 vs .40); whe as 10%control sam is unbiase re ple d Assum no confounding or m ing isclassification
Total population Risk factor Pre nt se Abse nt Oddsexposure Case s 300 200 1.5:1.0 C ontrols 180 720 1.0:4.0
OR=6.0
Biase Oddsexposure for case but not controls; d s biase OR d
S le e ction bias: C -Control S ase tudy
Hypothe tical study
S pling of 50%case and 10%controls by sam sam am s e pling bias (1.5) Assum no confounding or m ing isclassification
Total population Risk factor Pre nt se Abse nt Oddsexposure Case s 300 200 1.5:1.0 C ontrols 245 655 1.0:2.67
OR=4.0
Biase Oddsexposure to sam e nt for case and controls, d e xte s unbiase OR d
S le e ction bias: C -Control S ase tudy
C pe om nsation bias (se pre e vious e ple xam )
Bias factor m bepre nt in both num rator (Oddsexposure-cases and the ust se e de inator (Oddsexposure-controls) nom
(1.01.0) x 1.5 = (1.5 1.0) = 4.0 OR = (1.0 4.0) x 1.5 (1.0 2.67)
S le e ction bias
Be rkson's-typebias (Be rksonian bias)
First de scribe by Be d rkson (Mayo C linic) Burde of sym n ptom incre s like s ase lihood of hospitalization for individuals with m than onecondition ore C -control study using hospitalize patie m find an association ase d nts ay be e two dise s that is spurious twe n ase Ge rally, som usethete Be ne e rm rksonian bias to re r to a bias that arise in fe s a case -control study of hospitalize patie d nts
S le e ction bias
Avoiding se ction bias le
I n de signing ...