Discussion: View Thread

  • 1.  propensity score methods when some subjects have a 0 probability of receiving one of the treatments

    Posted 20 days ago

    I may be making my first foray into propensity score methods. Probably in over my head--but I suppose that's a good way to learn.

    The context is a planned observational study based on existing electronic medical record data. There are multiple treatments, each one a different antibiotic for laboring mothers known to be colonized by S. agalactiae, otherwise known as "group B strep". The outcome is binary: occurence of S. agalactiae sepsis in the newborn. 

    I find in a number of sources the caution/admonition/requirement that every subject must have a non-zero probability of receiving each of the treatments. For example, this excerpt from Austin et al (in turn from Rosenbaum and Rubin) in the case of a binary outcome:

    "Rosenbaum and Rubin (1983a) defined treatment assignment to be strongly ignorable if the following two conditions hold: (a) (Y (1), Y(0)) ╩ Z|X and (b) 0 < P(Z = 1|X) < 1. The first condition says that treatment assignment is independent of the potential outcomes conditional on the observed baseline covariates. The second condition says that every subject has a nonzero probability to receive either treatment."

    https://pmc.ncbi.nlm.nih.gov/articles/PMC3144483/#R13

    In my study, most laboring patients will receive pencillin or ampicillin. Some will receive a different agent, and the most likely reason for that will be a stated penicillin allergy. In practice, those patients have a 0 probability of receiving penicillin or ampicillin, absent a mishap. How much of a problem will this be? A deal-breaker?  Is there a concept of "weakly ignorable?"

    I would include penicillin allergy as one of the predictors in the model to produce the propensity scores.

    Thanks.



    ------------------------------
    Christopher Ryan
    Agency Statistical Consulting, LLC
    ------------------------------


  • 2.  RE: propensity score methods when some subjects have a 0 probability of receiving one of the treatments

    Posted 19 days ago

    You're asking a good question, and, to answer it, it's helpful to understand the purpose of propensity scores in this context. Propensity scores are intended to account for potential confounding by mimicking a randomized experiment. Confounding occurs if and only if a variable is associated with both the outcome and the treatment assignment. In a randomized experiment, treatments are assigned at random to remove confounding. Propensity scores attempt to estimate the probability of treatment assignments such that each treatment is assigned at random with a specific probability conditional on the propensity score. We're kind of pretending that every patient was randomized, but all with different probabilities for each treatment.

    With that context in mind, you can't address confounding with propensity scores if the probability of a treatment assignment is zero. Keep in mind though that two things have to be true for confounding to occur. The confounder has to be associated with the treatment assignment and the confounder has to be associated with the outcome. If there is substantial biological and/or experimental data to suggest that allergy is unrelated to the outcome, you could consider leaving allergy out of your propensity score and admit that you can't rule it out as a confounding factor. The alternative is to exclude those patients from your analysis. In other words, either ignore allergy completely (while admitting you are doing so), or use it as an exclusion. Don't attempt to make an adjustment for it that isn't possible.

    My two cents (reasonable practitioners may differ for good reasons, and reviewers at an average medical journal are even more likely to differ for poor reasons)



    ------------------------------
    David Miller
    Principal
    DPM Biostatistics LLC
    ------------------------------



  • 3.  RE: propensity score methods when some subjects have a 0 probability of receiving one of the treatments

    Posted 16 days ago
    Thank you David. This was very insightful. You've helped me understand
    that maternal penicillin allergy is unlikely to influence the risk of
    neonatal sepsis, *except* through the choice of intra-partum antibiotic
    given to the mother. So no confounding with allergy; it's just the
    reason why a mother might be given a non-penicillin antibiotic in labor.

    Thanks.

    --Chris Ryan

    David Miller via American Statistical Association wrote:
    > You're asking a good question, and, to answer it, it's helpful to
    > understand the purpose of propensity scores in this context. Propensity
    > scores...
    > Please note that using your email client's "Reply" option will send your
    > response to all members of the discussion group. To send a private
    > message to one person, click the blue "Reply to Sender" button at the
    > upper right of their post.
    > ------------------------------------------------------------------------
    >
    >
    > Statistical Consulting Section
    > <https: community.amstat.org cnsl ourdiscussiongroup digestviewer?communitykey=f77c549a-69cc-4a92-9d74-714fcacae535>
    >
    >
    > Post New Message Online
    > <http: community.amstat.org participate postmessage?groupid=1777&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21>
    >
    >
    > Re: propensity score methods when some subjects have a 0 probability of
    > receiving one of the treatments
    > <https: community.amstat.org cnsl discussion propensity-score-methods-when-some-subjects-have-a-0-probability-of-receiving-one-of-the-treatments#bm704146a3-54a3-468a-81da-aea932301a1f>
    >
    > Reply to Group Online
    > <https: community.amstat.org cnsl ourdiscussiongroup postreply?messagekey=704146a3-54a3-468a-81da-aea932301a1f&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21>
    > Reply to Sender
    > <https: community.amstat.org cnsl ourdiscussiongroup postreply?messagekey=704146a3-54a3-468a-81da-aea932301a1f&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21&SenderKey=128795b8-780f-4871-acf1-45bb4b8b7016>
    >
    >
    > David Miller
    > <https: community.amstat.org profile?userkey=128795b8-780f-4871-acf1-45bb4b8b7016>
    >
    > Aug 24, 2026 3:22 PM
    > David Miller
    > <https: community.amstat.org profile?userkey=128795b8-780f-4871-acf1-45bb4b8b7016>
    >
    >
    > You're asking a good question, and, to answer it, it's helpful to
    > understand the purpose of propensity scores in this context. Propensity
    > scores are intended to account for potential confounding by mimicking a
    > randomized experiment. Confounding occurs if and only if a variable is
    > associated with both the outcome and the treatment assignment. In a
    > randomized experiment, treatments are assigned at random to remove
    > confounding. Propensity scores attempt to estimate the probability of
    > treatment assignments such that each treatment is assigned at random
    > with a specific probability conditional on the propensity score. We're
    > kind of pretending that every patient was randomized, but all with
    > different probabilities for each treatment.
    >
    > With that context in mind, you can't address confounding with propensity
    > scores if the probability of a treatment assignment is zero. Keep in
    > mind though that two things have to be true for confounding to occur.
    > The confounder has to be associated with the treatment assignment and
    > the confounder has to be associated with the outcome. If there is
    > substantial biological and/or experimental data to suggest that allergy
    > is unrelated to the outcome, you could consider leaving allergy out of
    > your propensity score and admit that you can't rule it out as a
    > confounding factor. The alternative is to exclude those patients from
    > your analysis. In other words, either ignore allergy completely (while
    > admitting you are doing so), or use it as an exclusion. Don't attempt to
    > make an adjustment for it that isn't possible.
    >
    > My two cents (reasonable practitioners may differ for good reasons, and
    > reviewers at an average medical journal are even more likely to differ
    > for poor reasons)
    >
    >
    >
    > ------------------------------
    > David Miller
    > Principal
    > DPM Biostatistics LLC
    > ------------------------------
    >
    >   *Reply to Group Online
    > <https: community.amstat.org cnsl ourdiscussiongroup postreply?messagekey=704146a3-54a3-468a-81da-aea932301a1f&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21>*
    >   *Reply to Sender
    > <https: community.amstat.org cnsl ourdiscussiongroup postreply?messagekey=704146a3-54a3-468a-81da-aea932301a1f&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21&SenderKey=128795b8-780f-4871-acf1-45bb4b8b7016>*
    >   *View Thread
    > <https: community.amstat.org cnsl discussion propensity-score-methods-when-some-subjects-have-a-0-probability-of-receiving-one-of-the-treatments#bm704146a3-54a3-468a-81da-aea932301a1f>*
    >   *Recommend
    > <https: community.amstat.org:443 cnsl discussion propensity-score-methods-when-some-subjects-have-a-0-probability-of-receiving-one-of-the-treatments?messagekey=704146a3-54a3-468a-81da-aea932301a1f&cmd=rate&cmdarg=add#bm704146a3-54a3-468a-81da-aea932301a1f>*
    >   *Forward
    > <https: community.amstat.org cnsl ourdiscussiongroup forwardmessages?messagekey=704146a3-54a3-468a-81da-aea932301a1f&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21>*
    >   *Flag as Inappropriate
    > <https: community.amstat.org cnsl discussion propensity-score-methods-when-some-subjects-have-a-0-probability-of-receiving-one-of-the-treatments?markappropriate=704146a3-54a3-468a-81da-aea932301a1f#bm704146a3-54a3-468a-81da-aea932301a1f>*
    >   *Post New Message Online
    > <http: community.amstat.org participate postmessage?groupid=1777&ListKey=ac0f6215-000e-4179-801f-d62beb5b8a21>*
    >  
    > -------------------------------------------
    > Original Message:
    > Sent: 08-23-2026 23:33
    > From: Christopher Ryan
    > Subject: propensity score methods when some subjects have a 0
    > probability of receiving one of the treatments
    >
    > I may be making my first foray into propensity score methods. Probably
    > in over my head--but I suppose that's a good way to learn.
    >
    > The context is a planned observational study based on existing
    > electronic medical record data. There are multiple treatments, each one
    > a different antibiotic for laboring mothers known to be colonized by S.
    > agalactiae, otherwise known as "group B strep". The outcome is binary:
    > occurence of S. agalactiae sepsis in the newborn. 
    >
    > I find in a number of sources the caution/admonition/requirement that
    > every subject must have a non-zero probability of receiving each of the
    > treatments. For example, this excerpt from Austin et al (in turn from
    > Rosenbaum and Rubin) in the case of a binary outcome:
    >
    > "Rosenbaum and Rubin (1983a) defined treatment assignment to be strongly
    > ignorable if the following two conditions hold: (a) (Y (1), Y(0)) ╩ Z|X
    > and (b) 0 < P(Z = 1|X) < 1. The first condition says that treatment
    > assignment is independent of the potential outcomes conditional on the
    > observed baseline covariates. The second condition says that every
    > subject has a nonzero probability to receive either treatment."
    >
    > pmc.ncbi.nlm.nih.gov/articles/PMC3144483/#R13
    > <https: pmc.ncbi.nlm.nih.gov articles pmc3144483 #r13>
    >
    > In my study, most laboring patients will receive pencillin or
    > ampicillin. Some will receive a different agent, and the most likely
    > reason for that will be a stated penicillin allergy. In practice, those
    > patients have a 0 probability of receiving penicillin or ampicillin,
    > absent a mishap. How much of a problem will this be? A deal-breaker?  Is
    > there a concept of "weakly ignorable?"
    >
    > I would include penicillin allergy as one of the predictors in the model
    > to produce the propensity scores.
    >
    > Thanks.
    >
    >
    >
    > ------------------------------
    > Christopher Ryan
    > Agency Statistical Consulting, LLC
    > ------------------------------
    >
    >
    >
    >
    >  
    >
    > You are subscribed to "Statistical Consulting Section" as
    > cwr@agencystatistical.com. To change your subscriptions, go to My
    > Subscriptions
    > <http: community.amstat.org preferences?section=Subscriptions>. To
    > unsubscribe from this community discussion, go to Unsubscribe
    > <https: community.amstat.org higherlogic egroups unsubscribe.aspx?userkey=b16385a2-8ae1-4389-baca-0d70145c3500&sKey=KeyRemoved&GroupKey=ac0f6215-000e-4179-801f-d62beb5b8a21>.
    >
    >




  • 4.  RE: propensity score methods when some subjects have a 0 probability of receiving one of the treatments

    Posted 15 days ago

    great question. You wrote "In practice, those patients have a 0 probability...  absent a mishap"

    and by mishap, does that mean a mistake that could cause harm? 

    IN other words, those patients could be said to have a very small probability of getting treatment? 

    How large or small is your sample size? 

    You do have some choices.

    Curiosity may have killed the cat, but it would be interesting to include in the model and see what happens. And some sensitivity analyses around those patients. 

    And to identify the circumstances leading to the errors. 

    • If those patients are harmed or at risk of harm, then proceed with caution - set those aside and report those patients separately. 

    One other option is whatever you decide to do, and following a paper by Rubin. 

    Specify the entire propensity analysis in advance 

    Rubin, Donald B. "Propensity score methods." American journal of ophthalmology 149.1 (2010): 7.

    excerpting

    ....is discussed by
    Rubin with emphasis on the imperative of finalizing
    propensity score estimation before evaluating outcomes.

    -good luck! 



    ------------------------------
    Chris Barker, Ph.D.
    Adjunct Professor of Biostatistics
    University of Illinois Chicago, UIC-SPH
    www.barkerstats.com


    ---
    "In composition you have all the time you want to decide what to say in 15 seconds, in improvisation you have 15 seconds."
    -Steve Lacy
    ------------------------------