Stepped Wedge Design

A longitudinal experimental design in which all clusters are initially in the control condition and treatment is introduced in successive waves, allowing causal estimation using within-cluster before-after comparisons while ensuring every cluster eventually receives the intervention.

What it is

The stepped wedge design is a variant of the cluster randomized trial in which treatment is rolled out sequentially to clusters over multiple time periods. At each step (time period), a randomly selected group of clusters crosses from control to treatment. All clusters are eventually treated; they differ only in when treatment begins. This means the design uses within-cluster comparisons (each cluster is observed in both control and treatment phases) rather than relying solely on cross-cluster comparisons, substantially improving efficiency.

The design gets its name from the staircase pattern it produces: at each time step, more clusters are in the treated condition. The “wedge” shape represents the expanding treated group over time.

When to use it

The stepped wedge design is appropriate when: (1) the intervention is considered beneficial and it would be unethical to permanently withhold it from any cluster; (2) operational constraints prevent simultaneous rollout to all clusters; or (3) the research team wants to exploit within-cluster variation to control for time-invariant cluster-level confounders. It is widely used in health systems research and is increasingly applied in social protection and agriculture extension evaluations.

The design is not appropriate when treatment effects are expected to grow substantially over time (because later-treated clusters have less exposure time) or when contamination across clusters during the rollout period is a concern.

How it works

Design structure. With J clusters and T time periods, randomise the order in which clusters receive treatment. A balanced design with one cluster per step has T = J + 1 periods (period 1: all clusters in control; period J+1: all clusters in treatment; intermediate periods: some treated, some not). Each cluster is observed in all T periods.

Estimation. The standard analysis uses a two-way fixed effects model:

y_ij = α_i + β_t + δ·X_it + ε_it

where α_i are cluster fixed effects, β_t are time period fixed effects, X_it is a treatment indicator (0 before the cluster crosses, 1 after), and δ is the average treatment effect. Cluster fixed effects control for any time-invariant cluster-level characteristics; time fixed effects control for secular trends common to all clusters.

Time-varying treatment effects. If treatment effects grow or decay over time (duration effects), the basic model is misspecified. Include interaction terms between time-since-treatment and treatment status, or estimate separate effects for each period-since-treatment.

Calendar time vs. exposure time. The design confounds calendar time (secular trends) and exposure time (time since treatment). If outcomes trend over calendar time for reasons unrelated to treatment, the calendar time fixed effects absorb this — but only if the trend is parallel across clusters. Differential time trends by cluster violate the parallel trends assumption.

Key decisions

Number of steps and clusters per step. More steps with fewer clusters per step produce finer estimation of time-varying effects but increase survey costs (more data collection waves). A design with 3–5 steps is common in practice. Having multiple clusters cross at each step (rather than one) increases power and allows estimation of within-step variation.

Baseline period. Including a baseline period (all clusters in control) before treatment rollout begins allows estimation of pre-treatment trends and supports parallel trends testing. A minimum of one baseline period before the first treated cluster is strongly recommended.

Complete vs. incomplete designs. Incomplete stepped wedge designs (where not all cells are observed) can arise from attrition or survey budgets. These require careful attention in the analysis to ensure that the remaining data are still informative for causal identification.

Caveats & common mistakes

Assuming parallel trends without testing. The key identifying assumption is that, absent treatment, all clusters would have followed the same time trend. With multiple pre-treatment observations per cluster, this can be partially tested by checking whether the pre-treatment trends are parallel. A significant pre-treatment difference in trends is a threat to validity.

Ignoring within-cluster autocorrelation. Repeated observations on the same cluster over time are correlated. Standard errors must account for this autocorrelation. Cluster-robust standard errors clustered at the cluster level are a minimum requirement; using a GLS or mixed model approach that models the autocorrelation structure is preferable.

Conflating secular trends with treatment effects. If outcomes improve nationally during the study period (due to economic growth, rainfall, policy changes), the two-way fixed effects estimator will attribute this to treatment if the rollout correlates with the secular trend. Running the analysis using only between-cluster variation (as a robustness check) tests whether the time fixed effects are adequately absorbing secular trends.

Analysis Guide

import pandas as pd
import numpy as np
from linearmodels.panel import PanelOLS

# 1. Set the panel index — PanelOLS needs a MultiIndex of (entity, time) to know
#    which dimensions to absorb; cluster_id is the entity, period is the time axis
panel = df.set_index(["cluster_id", "period"]).sort_index()

# 2. Two-way fixed effects: EntityEffects absorbs all time-invariant cluster
#    differences (geography, demographics) and TimeEffects absorbs secular trends
#    common to all clusters; treatment is identified from within-cluster before-
#    after variation, net of these confounders; cluster-robust SEs at cluster_id
#    handle the autocorrelation across repeated observations on the same cluster
fit = PanelOLS.from_formula(
  "outcome ~ treatment + EntityEffects + TimeEffects",
  data=panel).fit(cov_type="clustered", cluster_entity=True)
print(fit.summary)

# 3. Time-since-treatment exposure variable — if the treatment effect grows or
#    decays after crossover (duration effects), the pooled coefficient above is a
#    weighted average; computing periods_since for each observation lets the next
#    model estimate each exposure horizon separately
df["periods_since"] = np.where(
  df["treatment"] == 1,
  df["period"] - df.groupby("cluster_id")["period"].transform(
      lambda s: s[df.loc[s.index, "treatment"] == 1].min()),
  0)

# 4. Event-study style model: dummies for each exposure horizon, with pre-period
#    as the reference; flat post-coefficients suggest no build-up, rising ones
#    indicate effects that accumulate over the post-crossover horizon; the
#    pre-period coefficients double as a parallel-trends diagnostic
panel = df.set_index(["cluster_id", "period"]).sort_index()
fit_dyn = PanelOLS.from_formula(
  "outcome ~ C(periods_since) + EntityEffects + TimeEffects",
  data=panel).fit(cov_type="clustered", cluster_entity=True)
print(fit_dyn.summary)

library(lfe)      # felm for high-dimensional FE
library(estimatr)

# 1. Two-way FE: the cluster_id and period FEs are absorbed (not estimated as
#    explicit coefficients) — this is equivalent to demeaning by cluster and period
#    jointly; clustering SEs at cluster_id accounts for within-cluster autocorrelation
#    across the repeated observations
fit <- felm(outcome ~ treatment | cluster_id + period | 0 | cluster_id,
          data = df)
summary(fit)

# 2. fixest::feols is preferred for large datasets — it uses the same two-way
#    absorb but is faster and handles unbalanced panels without extra steps
library(fixest)
feols(outcome ~ treatment | cluster_id + period,
    data = df, cluster = ~cluster_id)

# 3. Time-since-treatment: compute exposure duration for each observation —
#    periods_since = 0 in pre-treatment periods so i() can use ref=0 to compare
#    each post-crossover period to the pre-crossover baseline within clusters
df <- df %>%
group_by(cluster_id) %>%
mutate(periods_since = ifelse(treatment == 1, period - min(period[treatment==1]), 0)) %>%
ungroup()

# 4. Event-study plot equivalent for stepped wedge: ref=0 normalises the crossover
#    period to zero; pre-treatment coefficients should be near zero (parallel trends);
#    post-treatment coefficients reveal whether effects grow, plateau, or reverse
feols(outcome ~ i(periods_since, ref=0) | cluster_id + period,
    data = df, cluster = ~cluster_id)

Reading the output

The coefficient on treatment is the average treatment effect pooled across all post-crossover periods; a positive value indicates the intervention raised the outcome after clusters crossed into treatment.
Cluster-robust standard errors are the minimum requirement; if the number of clusters is below 30, switch to wild cluster bootstrap for reliable inference.
In the time-since-treatment specification, coefficients on i.periods_since_treat should be near zero for period 0 (the crossover period itself) and grow if effects accumulate; a flat profile after period 1–2 indicates no dynamic build-up.
Pre-treatment trends: run the time-since-treatment model on pre-crossover periods only; coefficients significantly different from zero indicate the parallel trends assumption is violated.
If the two-way FE coefficient differs substantially from a simple first-differences estimate, inspect whether secular time trends are parallel across clusters.

References

Hemming, K., Haines, T. P., Chilton, P. J., Girling, A. J., & Lilford, R. J. (2015). The stepped wedge cluster randomised trial: Rationale, design, analysis, and reporting. BMJ, 350, h391. https://doi.org/10.1136/bmj.h391

Brown, C. A., & Lilford, R. J. (2006). The stepped wedge trial design: A systematic review. BMC Medical Research Methodology, 6, 54. https://doi.org/10.1186/1471-2288-6-54

Hussey, M. A., & Hughes, J. P. (2007). Design and analysis of stepped wedge cluster randomized trials. Contemporary Clinical Trials, 28(2), 182–191. https://doi.org/10.1016/j.cct.2006.05.007

Mdege, N. D., Man, M.-S., Taylor, C. A., & Torgerson, D. J. (2011). Systematic review of stepped wedge cluster randomized trials shows that design is particularly used to evaluate interventions during routine implementation. Journal of Clinical Epidemiology, 64(9), 936–948. https://doi.org/10.1016/j.jclinepi.2010.12.003