Choosing propensity matching or weighting
Propensity-score matching and weighting both balance measured confounders, but each targets the effect in a particular population, so decide whose effect you want before choosing.
Fit
When it matters, and when it does not
It matters when
- You must choose whose treatment effect matters before choosing how to balance measured confounders.
- Matching may discard treated patients, or weighting may give a few patients too much influence.
It matters less when
- Treatment changes in response to earlier outcomes. Use a method for treatment-confounder feedback rather than a baseline comparison. Read: Marginal structural models with IPTW
- The important confounders were never measured. Plan sensitivity analyses; neither matching nor weighting removes that uncertainty. Read: Sensitivity analyses
Before you start
What to check in your study
- Whose effect you wantThe treated, called the ATT. Everyone, called the ATE. Or patients for whom either treatment was plausible, called the ATO.
- How much the groups overlapPlot the propensity scores by group. Little overlap means unmatched patients or extreme weights, whichever you choose.
- The treatment, the comparison and time zeroassignmentConfounders are measured before time zero, whichever method you choose.
- How the interval will be estimatedMatched data need errors that allow for the pairs. Weighted data need errors that allow for the estimated weights.
- The measured factors explain who got the treatment.How to check: This cannot be tested. Report how strong a missed confounder would need to be to change the answer.
- Patients like each one could have received either treatment.How to check: Check support for the chosen target. Full-population inverse weights can be extreme; overlap weights instead downweight the tails and change the target.
- After matching or weighting, the groups are alike on every confounder.How to check: Report standardized mean differences before and after, with spreads as well as means. Under 0.1 is a common target.
Plain language
How it works
Matching treated patients usually describes those who find a match. Full-population IPTW asks about everyone eligible. Overlap weighting asks about patients whose profiles support either treatment. AIPW combines weights and outcome predictions for its stated target. These choices need measured confounders and credible overlap.
A lack of comparable patients is a limit of the data. Changing to overlap weights can answer a narrower question; it does not recover an effect for everyone. Match losses, trimming and weight limits must be reported. Choose the target and uncertainty method before looking at outcomes. The linked single-method articles give the procedures.
After weighting, scale both drug groups to 100 users. Admissions were 8 of 100 with drug A and 10 of 100 with drug B: a risk difference of -2 per 100 and risk ratio of 0.8.
Worked example
A worked example
- Setting
- Records of 5,000 older adults admitted with pneumonia, 1,000 of whom got early mobilization. Frail patients rarely got it.
- Question
- Does early mobilization reduce readmission within 30 days?
- Approach
- Each mobilized patient is matched to one similar patient on a propensity score within a caliper. 150 mobilized patients have no match and are set aside.
- Result
- In the 850 matched pairs, 85 of 850 mobilized patients (10 in 100) and 119 of 850 others (14 in 100) were readmitted. That is a risk difference of -4 per 100, a risk ratio of 0.71.
- What it means
- This is the effect in the 850 mobilized patients who found a match, not all 1,000. Weights for everyone would be extreme for frail patients, who were almost never mobilized.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Matching and weighting address measured confounding under assumptions; neither removes unmeasured confounding or makes a lack of overlap disappear. Trimming and discarded matches can change the target population.
- Trying both methods and reporting whichever gives the result you hoped for.Instead: Fix the target population and the method before opening the outcomes.
- Using the default standard errors from a regression on matched or weighted data.Instead: Allow for the pairs after matching. Allow for the estimated weights after weighting, for example with a bootstrap.
- Judging the propensity model by how well it predicts treatment, its c-statistic.Instead: That is not a check of balance. Check standardized differences on every confounder after matching or weighting.
Inside Mallard
How Mallard uses this
For a causal, nonrandomized plan with a propensity cue, Mallard offers the matching-or-weighting concept. It explains the choice without selecting a method.
concept.choosing_matching_weighting, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies (opens in a new tab) · Multivariate Behav Res · doi 10.1080/00273171.2011.568786
- Principles of confounder selection (opens in a new tab) · Eur J Epidemiol · doi 10.1007/s10654-019-00494-6
- Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies (opens in a new tab) · Stat Med · doi 10.1002/sim.6607