Invitation
This page is a companion to our recent
editorial in which we invite readers to suggest a case study illustrating “a
problem in medical science that can be usefully solved by propensity scores,
but not by other methods” (Stevens & Oke, 2022; Colorectal Disease Volume
24, in press) .
For context, consider our role as statistics
teachers in evidence based medicine.
Our students are health professionals: anaesthetists, dentists, general
practitioners, haematologists, internists, midwives, nurses, psychiatrists,
radiologists, surgeons, veterinarians, etc. When we teach propensity scores, we
can expect to be asked, “what are the advantages of this over other methods
taught on the course?” We would like to
be able to give a good example.
We think answers about marginal vs.
conditional effect measures are unlikely to persuade students on our evidence
based health care programme. There is no
consensus in evidence based medicine to prefer either marginal or conditional
odds ratios: on the contrary, we prefer the absolute risk difference, and this
happens to be a collapsible measure (that is, there is no difference between
the marginal and the conditional risk difference).
This and other requirements were
stated in our article:
Entrants
should be careful that their proposed case study is not a comparison of one
propensity score method to another; nor a comparison restricted to limited
competing methods (e.g. propensity score matching compared to multivariate
adjustment); nor an argument based on marginal vs. conditional odds ratios,
unless you can first persuade us that the distinction matters.
(Stevens
& Oke, 2022; Colorectal Disease Volume 24, in press).
Entries may be sent to Dr Jason Oke
at jason.oke@phc.ox.ac.uk. Before submitting your entry, please consider
the paper (not the title and abstract alone) in the light of the requirements
listed above.
Here is a list of some of the papers
that our colleagues have previously referred to us.
Kurth et al. (2006). Results of multivariable logistic
regression, propensity matching, propensity adjustment, and propensity-based
weighting under conditions of nonuniform effect.
American Journal of Epidemiology, 163(3), 262-270.
In this fascinating paper, the
different propensity score methods show more differences, in the results, to
each other than to the comparator, multivariate logistic regression. The ability to target different estimands (e.g. ATE vs ATT) by choosing different weighting
methods is promising. On the other hand,
Kurth et al. write that their findings “... should
not be taken as evidence that, compared with other multivariable outcome
models, these two methods are a better tool to adjust for covariates in
observational research”, noting that if analyses are restricted to the
sub-population most likely to be treated, “all adjustment methods gave fairly
similar results.”
Martens et al. (2008). Systematic differences
in treatment effect estimates between propensity score methods and logistic
regression. International Journal of Epidemiology, 37(5), 1142-1147.
As an argument for propensity scores
over multivariate logistic regression, this paper rests on a preference for
marginal odds ratios over conditional odds ratios. See discussion above and in Stevens and
Oke. Interestingly, the simulations in
Martens et al. consider only the case that there are no confounders.
Payet et al. (2021) High-dimensional
propensity scores improved the control of indication bias in surgical
comparative effectiveness studies. Journal of Clinical Epidemiology, 130,
78-86.
This paper is a comparison of
different propensity score approaches to each other: specifically,
“High-dimensional propensity scores (HdPS)” to
propensity scores (PS) of lower dimension.