Multi-period Inverse Probability Weighting
Multi-period Inverse Probability Weighting Estimator · Also known as: longitudinal IPW, multi-period IPW, time-varying IPW, sequential IPW
Multi-period Inverse Probability Weighting (IPW) estimates the causal effect of a treatment that varies across multiple time periods by reweighting observations according to the probability of receiving each period's treatment given past treatment history and time-varying confounders. It creates a pseudo-population where treatment at each period is independent of measured confounders, enabling unbiased estimation of sustained treatment strategies.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use multi-period IPW when (1) treatment status varies across multiple measurement occasions, (2) time-varying covariates both confound treatment and are themselves affected by prior treatment (treatment-confounder feedback), and (3) you wish to compare sustained treatment strategies or dynamic regimes over time. It is appropriate for longitudinal panel data with a continuous or binary outcome and sufficiently large samples (typically n > 200 per arm) to estimate propensity models reliably at each period. Do not use it when treatment is assigned only once (standard IPW suffices), when positivity is violated (some covariate profiles have near-zero probability of one treatment level), when propensity models are highly misspecified, or when the panel is very short (fewer than three periods adds minimal value over a static approach).
Strengths & limitations
- Handles treatment-confounder feedback that invalidates conventional regression adjustment in longitudinal settings.
- Stabilized weights produce a pseudo-population with finite variance, making estimates more reliable than unstabilized alternatives.
- Non-parametric in the outcome model: any functional form can be used after weighting, including flexible machine-learning models.
- Directly targets the effect of sustained or dynamic treatment strategies across time, matching policy-relevant research questions.
- Combines naturally with doubly robust estimators (e.g., TMLE) to gain protection against one model being misspecified.
- Requires correct specification of the propensity model at every time period; misspecification at any stage propagates through the product of weights.
- Extreme weights arise when positivity is near-violated, inflating variance and potentially biasing estimates; weight truncation or trimming is often needed.
- Relies on the sequential ignorability (no unmeasured time-varying confounders) assumption, which cannot be tested from the data.
- Computational burden grows with the number of periods, and multiple propensity models must be estimated and validated.
- Requires relatively large samples to estimate stable propensity models at each time point, limiting applicability in small cohorts.
Frequently asked
How does multi-period IPW differ from standard (single-period) IPW?
Standard IPW handles a single fixed treatment assignment and weights by the inverse of one propensity score. Multi-period IPW handles a sequence of treatment decisions, computing a propensity score at each period and forming the weight as the product of period-specific ratios. This product structure accounts for the temporal dependency between treatment, covariates, and outcomes across periods.
What is treatment-confounder feedback and why does it matter?
Treatment-confounder feedback occurs when earlier treatment affects later confounders (e.g., a drug lowers a biomarker that in turn affects the next treatment decision). Standard regression that conditions on the intermediate biomarker blocks the treatment effect path, while multi-period IPW avoids conditioning on post-treatment variables, preserving the full causal pathway.
Why use stabilized rather than unstabilized weights?
Unstabilized weights are the inverse of the product of conditional probabilities; their variance grows rapidly with the number of periods, often producing extremely large individual weights. Stabilized weights include a marginal probability in the numerator that typically brings the mean weight near one and substantially reduces variance, yielding more stable estimates.
How many time periods are needed?
At minimum three periods are recommended so that treatment-confounder feedback can meaningfully occur between the first treatment, an intermediate confounder, and the second treatment decision. In practice, the method is most useful and most commonly applied in studies with four or more measurement occasions.
What happens if the positivity assumption is violated?
If some covariate profiles have near-zero probability of one treatment level at any period, the corresponding weights approach infinity, inflating variance and biasing estimates. Diagnostics include examining weight distributions and checking empirical positivity. Truncating extreme weights at a percentile threshold (e.g., the 99th) is a common pragmatic remedy, though it introduces slight bias.
Sources
- Robins, J. M., Hernan, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550-560. DOI: 10.1097/00001648-200009000-00011 ↗
- Hernan, M. A., & Robins, J. M. (2020). Causal Inference: What If. Chapman and Hall/CRC. link ↗
How to cite this page
ScholarGate. (2026, June 3). Multi-period Inverse Probability Weighting Estimator. ScholarGate. https://scholargate.app/en/causal-inference/multi-period-inverse-probability-weighting
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Doubly Robust EstimationCausal inference↔ compare
- Dynamic Inverse Probability WeightingCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Panel Data Inverse Probability WeightingCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare