Selective Ignorability Assumptions in Causal Inference

TL;DR: This paper outlines selective ignorability assumptions mathematically and sketches how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models to derive valid causal inferences.

Abstract: Most attempts at causal inference in observational studies are based on assumptions that treatment assignment is ignorable. Such assumptions are usually made casually, largely because they justify the use of available statistical methods and not because they are truly believed. It will often be the case that it is plausible that conditional independence holds at least approximately for a subset but not all of the experience giving rise to one's data. Such selective ignorability assumptions may be used to derive valid causal inferences in conjunction with structural nested models. In this paper, we outline selective ignorability assumptions mathematically and sketch how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models. We also consider use of these assumptions in the presence of selective measurement error or missing data when the missingness is not at random. We motivate and illustrate our development by considering an analysis of an observational database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.

Summary

Summary

• Most attempts at causal inference in observational studies are based on assumptions that treatment assignment is ignorable.
• Such assumptions are usually made casually, largely because they justify the use of available statistical methods and not because they are truly believed.
• It will often be the case that it is plausible that conditional independence holds at least approximately for a subset but not all of the experience giving rise to one's data.
• Such selective ignorability assumptions may be used to derive valid causal inferences in conjunction with structural nested models.
• The authors outline selective ignorability assumptions mathematically and sketch how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models.
• The authors also consider use of these assumptions in the presence of selective measurement error or missing data when the missingness is not at random.

Full text

Volume 6, Issue 2 2010 Article 11
The International Journal of
Biostatistics
CAUSAL INFERENCE
Selective Ignorability Assumptions in Causal
Inference
Marshall M. Joffe, University of Pennsylvania School of
Medicine
Wei Peter Yang, University of Pennsylvania School of
Medicine
Harold I. Feldman, University of Pennsylvania School of
Medicine
Recommended Citation:
Joffe, Marshall M.; Yang, Wei Peter; and Feldman, Harold I. (2010) "Selective Ignorability
Assumptions in Causal Inference," The International Journal of Biostatistics: Vol. 6: Iss. 2,
Article 11.
DOI: 10.2202/1557-4679.1199

Selective Ignorability Assumptions in Causal
Inference
Marshall M. Joffe, Wei Peter Yang, and Harold I. Feldman
Abstract
Most attempts at causal inference in observational studies are based on assumptions that
treatment assignment is ignorable. Such assumptions are usually made casually, largely because
they justify the use of available statistical methods and not because they are truly believed. It will
often be the case that it is plausible that conditional independence holds at least approximately for
a subset but not all of the experience giving rise to one's data. Such selective ignorability
assumptions may be used to derive valid causal inferences in conjunction with structural nested
models. In this paper, we outline selective ignorability assumptions mathematically and sketch
how they may be used along with otherwise standard G-estimation or likelihood-based methods to
obtain inference on structural nested models. We also consider use of these assumptions in the
presence of selective measurement error or missing data when the missingness is not at random.
We motivate and illustrate our development by considering an analysis of an observational
database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.
KEYWORDS: causal inference, ignorability, end-stage renal disease, anemia
Author Notes: This work was supported by an unrestricted grant from Amgen.




 
    !    "   
#$
%
%& 
"
$                             !  $                                                  
#!$
!!'$
$$()
       *     



 
#               !  $                             $   '                            
* $$
!+)
                           
,                              -          .          / 0                   
$'%1!
"&!!
2 $"$
"       3   !   
  ! 3!
          
!  ,
 !          2
1
Joffe et al.: Selective Ignorability Assumptions



,                                                                       
                 $                                                                   !
!,$
4                                                   $                $               
   (  56  7.. 
%!
8          !  $    !   
 !!
         !                                                                3             !
                                                                     $          
 !
                         
1        )                   !    9                                                    
2                                                     ! 
! ! 
  )   *       !
$          *     
       ! 
! * 
           !

:                                             $         %               
            ,    !  $   $                       $                 
                                                   ;


2
The International Journal of Biostatistics, Vol. 6 [2010], Iss. 2, Art. 11
DOI: 10.2202/1557-4679.1199

  $
,! $
                 $                                                    
9 ! 
$"
$9
<            +          $   '                          
!9!
                     !  =       !             *            7                  
!$',9
$"9+#
!!)'"&
!+
",$
,>
?   7   7!   $  


                                                  
;                                                                            
9           ,
  %      
          (
   5            7 . . 7           )          =       ' !  6      !  6                 
7 . . 0        !        + '                  7 . . .                <                       
%     ' .' 5 
  
@!


3
Joffe et al.: Selective Ignorability Assumptions

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Selective ignorability assumptions in causal inference" ?

In this paper, the authors outline selective ignorability assumptions mathematically and sketch how they may be used along with otherwise standard G-estimation or likelihood-based methods to obtain inference on structural nested models. The authors also consider use of these assumptions in the presence of selective measurement error or missing data when the missingness is not at random. The authors motivate and illustrate their development by considering an analysis of an observational database to estimate the effect of erythropoietin use on mortality among hemodialysis patients.