Co nstruc ting I nve rse Pro b a b ility We ig hts fo r Sta tic I - PowerPoint PPT Presentation

Co nstruc ting I nve rse Pro b a b ility We ig hts fo r Sta tic I nte rve ntio ns K unja l Pa te l, DSc MPH Se nio r Re se a rc h Sc ie ntist Ha rva rd T .H. Cha n Sc ho o l o f Pub lic He a lth

Ac kno wle dg e me nt  Slide s c o ntrib ute d b y Mig ue l He rná n o r a da pte d fro m Causal I nfe re nc e (Cha pma n & Ha ll/ CRC, 2017) b y Mig ue l He rná n a nd Ja mie Ro b ins  Any mista ke s a re my o wn  Cha pte rs o f b o o k a nd SAS, ST AT A, a nd R c o de fre e ly a va ila b le a t http:/ / www.hsph.ha rva rd.e d u/ mig ue l- he rna n/ c a usa l-infe re nc e -b o o k/  Yo u c a n “like ” Ca usa l I nfe re nc e a t https:/ / www.fa c e b o o k.c o m/ c a usa linfe re nc e

Ca se Study

I ntro duc tio n/ Ba c kg ro und  Ra ndo mize d c o ntro lle d tria ls (RCT s) in HI V-infe c te d a dults ha ve sho wn HAART to b e hig hly e ffe c tive in re duc ing the risk o f mo rta lity  Na tura l pro g re ssio n o f HI V infe c tio n in c hildre n is diffe re nt fro m a dults:  HI V RNA le ve ls re ma in pe rsiste ntly hig he r tha n a dults fo r first 2-3 ye a rs o f life , de c re a sing to ste a dy sta te le ve ls in a dults a fte r a ppro xima te ly five -ye a rs  Ge ne ra liza b ility o f a dult tria l re sults?

Studie s o f HAART in Childre n  RCT s o f HAART in c hildre n ha ve fo c use d o n inte rme dia te immuno lo g ic a nd viro lo g ic e ndpo ints  L o ng -te rm studie s o f HAART o n mo rta lity re lia nt o n o b se rva tio na l studie s:  I ta lia n c o ho rt study: triple c o mb ina tio n the ra py vs. no the ra py – HR=0.29 (0.13-0.67) (De Martino e t al. 2000)  PACT G 219 study: c o mb ina tio n the ra py with PI vs. the ra py witho ut PI – HR=0.33 (0.19-0.58) (Go rtmake r e t al. 2000)

Ne e d to Re -E va lua te E ffe c t o f HAART o n Mo rta lity  Pre vio us studie s e nde d fo llo w-up in 1999:  Use o f ne w a ntire tro vira l drug s ha s inc re a se d  Cha ng e s in initia l HAART re g ime ns o ve r time Va n Dyke e t a l. JAIDS 2011; 57:165-173

Study q ue stio n  Wha t is the e ffe c t o f HAART o n mo rtality a mo ng pe rina ta lly HI V-infe c te d c hildre n?  I s this a g o o d study q ue stio n?

F o rmula tio n o f a we ll-de fine d study q ue stio n  We ll-de fine d c a usa l infe re nc e q ue stio ns c a n b e ma ppe d into a ta rg e t tria l  Ca se e xa mple : Wha t is the e ffe c t o f initiating HAART o n mo rta lity a mo ng pe rina ta lly HI V-infe c te d c hildre n?  Spe c ify the pro to c o l o f the ta rg e t tria l inc luding :  E lig ib ility c rite ria  T re a tme nt stra te g ie s  Ra ndo mize d tre a tme nt a ssig nme nt  F o llo w-up pe rio d  Outc o me  Ca usa l c o ntra st o f inte re st  Ana lysis Pla n He rna n, Ro b ins Am J E pid e mio l. 2016;183(8):758–764

Pe dia tric AI DS Clinic a l T ria ls Gro up (PACT G) Pro to c o ls 219 & 219C  Pro spe c tive c o ho rt studie s o f HI V-e xpo se d c hildre n (infe c te d a nd uninfe c te d) fro m mo re tha n 80 study site s in the US  Asse ss the lo ng -te rm e ffe c ts o f HI V infe c tio n a nd in- ute ro a nd po stna ta l e xpo sure to a ntire tro vira l the ra py  PACT G 219: April 1993-Se pte mb e r 2000  PACT G 219C: Se pte mb e r 2000-2006  E xte nsive c linic a l, ne uro psyc ho lo g ic a l, a nd la b o ra to ry e va lua tio ns

Study Po pula tio n, E xpo sure , F o llo w-up  1,236 pe rina ta lly HI V-infe c te d c hildre n e nro lle d in PACT G 219 a nd 219C b e twe e n Ja nua ry 1, 1996 a nd June 30, 2006  E xc lude s tho se with pre vio us o r c urre nt use o f HAART a t time o f study e ntry  HAART de fine d a s the use o f a t le a st 3 drug s fro m a t le a st 2 diffe re nt c la sse s o f HI V the ra py (NRT I s, NNRT I s, o r PI s)  Onc e sub je c ts initia te d HAART the y we re a ssume d to re ma in o n HAART fo r the dura tio n o f the ir fo llo w-up  F o llo w-up fo r a ma ximum o f te n ye a rs to the la st visit a t whic h sub je c t wa s se e n a live o r the la st visit b e fo re June 30, 2006 (i.e . “c o mple tio n o f study”)

Cla ssific a tio n o f tre a tme nt stra te g ie s a c c o rding to the ir time c o urse  Point inte rve ntio ns  I nte rve ntio n o c c urs a t a sing le time  E xa mple s: o ne -do se va c c ina tio n, sho rt-live d tra uma tic e ve nt, surg e ry…  I nte ntio n-to -tre a t e ffe c ts in RCT s a re a b o ut po int inte rve ntio ns  Sustaine d stra te g ie s  I nte rve ntio ns o c c ur a t se ve ra l time s  E xa mple s: me dic a l tre a tme nts, life style , e nviro nme nta l e xpo sure s…  Ma ny (mo st? ) q ue stio ns a re a b o ut susta ine d e xpo sure s

Cla ssific a tio n o f susta ine d tre a tme nt stra te g ie s  Static  a fixe d str ate gy for e ve r yone  E xample : tr e at with 150mg of daily aspir in dur ing 5 ye ar s  Case e xample : initiate HAART  Dyna mic  a stra te g y tha t a ssig ns diffe re nt va lue s to diffe re nt individua ls a s a func tio n o f the ir e vo lving c ha ra c te ristic s  E xa mple : sta rt a spirin tre a tme nt if c o ro na ry he a rt dise a se , sto p if stro ke  Ca se e xa mple : initia te HAART if CD4 dro ps b e lo w 500 c e lls/ mm 3

Ra ndo mize d tre a tme nt a ssig nme nt  Ca usa l infe re nc e me tho ds a re me tho ds tha t e mula te ra ndo miza tio n  Why is ra ndo miza tio n impo rta nt?

De finitio n o f a n a ve ra g e c a usa l e ffe c t  E a c h pe rso n ha s two c o unte rfa c tua l o utc o me s:  Outc o me Y if tre a te d - Y i, a=1  Outc o me Y if untre a te d – Y i, a=0  I ndividua l c a usa l e ffe c t:  Y i, a=1 ≠ Y i, a=0  Ca nno t b e de te rmine d e xc e pt unde r e xtre me ly stro ng a ssumptio ns  Ave ra g e (po pula tio n) c a usa l e ffe c t: [ Y a=1 = 1] ≠ E  E [ Y a=0 = 1]  Ca n b e e stima te d unde r:  No a ssumptio ns (ide a l ra ndo mize d e xpe rime nts)  Stro ng a ssumptio ns (o b se rva tio na l studie s) l

Ca usa tio n ve rsus Asso c ia tio n Pr[ Y a=0 = 1] Pr[ Y a=1 = 1] Pr[ Y= 1 |A= 0] Pr[ Y =1| A =1] l

Ca usa tio n ve rsus Asso c ia tio n  Pr[ Y a = 1]  pro po rtio n o f sub je c ts tha t wo uld ha ve de ve lo pe d the o utc o me Y ha d a ll sub je c ts in the po pula tio n re c e ive d e xpo sure va lue a  (Co unte rfa c tua l) risk o f Y a  Unc o nditio na l o f ma rg ina l pro b a b ility – “c a lc ula te d” using da ta fro m the who le po pula tio n  Ca usa tio n: Pr[ Y a=1 = 1] ≠ Pr[ Y a=0 = 1]  Pr[ Y =1 |A = a ]  Pro po rtio n o f sub je c ts tha t de ve lo pe d o utc o me Y a mo ng tho se tha t re c e ive d e xpo sure va lue a in the po pula tio n  Risk o f Y a mo ng tho se e xpo se d/ une xpo se d  Co nditio na l pro b a b ility – c a lc ula te d b y using da ta fro m a sub se t o f the po pula tio n  Asso c ia tio n: Pr[ Y =1 |A =1] ≠ Pr[ Y =1 |A =0]

I de a l Ra ndo mize d E xpe rime nt  L a rg e (ne a r-infinite ) po pula tio n  No lo ss to fo llo w-up  F ull c o mplia nc e (a dhe re nc e ) to a ssig ne d e xpo sure o r tre a tme nt  Do ub le b lind a ssig nme nt

Ra ndo miza tio n (I )  Assume two e xpo sure g ro ups (tre a te d a nd untre a te d)  Me mb e rship in e a c h g ro up is ra ndo mly a ssig ne d  e .g ., b y a flip o f a c o in  F irst o ptio n:  T re a t sub je c ts in g ro up 1, do n’ t tre a t sub je c ts in g ro up 2  Pr[ Y =1 |A =1] is, sa y, 0.57  Se c o nd o ptio n:  T re a t sub je c ts in g ro up 2, do n’ t tre a t sub je c ts in g ro up 1  Wha t is the risk? Pr[ Y =1 |A =1] is ? 0.57

Ra ndo miza tio n (I I )  Whe n g ro up me mb e rship is ra ndo mly a ssig ne d, risks a re the sa me  Bo th g ro ups a re c o mpa ra b le o r e xc hange able  E xc ha ng e a b ility is the c o nse q ue nc e o f ra ndo miza tio n

E xc ha ng e a b ility  Sub je c ts in g ro up 1 wo uld ha ve ha d the sa me risk a s tho se in g ro up 2 ha d the y re c e ive d the tre a tme nt o f tho se in g ro up 2  T he c o unte rfa c tua l risk a mo ng the tre a te d e q ua ls the c o unte rfa c tua l risk a mo ng the untre a te d unde r the sa me e xpo sure le ve l   Pr[ Y a =1 |A =1] = Pr[ Y a =1 |A =0] A Y a Y a A   I mplie s la c k o f c o nfo unding

I n ide a l ra ndo mize d e xpe rime nts  Pr[ Y =1 |A =1] is e q ua l to Pr[ Y a=1 = 1]  Pr[ Y =1 |A =0] is e q ua l to Pr[ Y a=0 = 1]  T he re fo re the a sso c ia tio na l risk ra tio  Pr[ Y =1 |A =1]/ Pr[ Y =1 |A =0] is e q ua l to the c a usa l risk ra tio  Pr[ Y a=1 = 1]/ Pr[ Y a=0 = 1]

Why is Pr[ Y =1/ A =1] is e q ua l to Pr[ Y a=1 = 1]  A two ste p pro o f: 1. Pr[ Y =1 |A =1] = Pr[ Y a=1 = 1 | A=1]  b y de finitio n o f a c o unte rfa c tua l va ria b le (i.e ., c o nsiste nc y) 2. Pr[ Y a=1 = 1 | A=1] = Pr[ Y a=1 = 1 | A=0] = Pr[ Y a=1 = 1]  b y ra ndo miza tio n – (i.e ., e xc ha ng e a b ility)  Ste p 2 no t g e ne ra lly true in the a b se nc e o f ra ndo miza tio n