Update (Sept 10, 2013): I believe it would be more correct to say that increasing baseline or endline measurements are ways to decrease the design effect, thereby making the SW design efficient, rather than stating that the maximum number of measurements is necessary. Woertman et al. (2013).
The stepped wedge design (pdf) is a nice alternative to parallel-group designs when, for logistical reasons, the intervention must be rolled-out in stages. A potential downside of this design, however, can be the number of measurement rounds. Even though SW designs can have increased power (thus reducing required sample size to detect the same effect), every unit is observed/measured before and after every treatment round (step). If you have five steps, there are six measurement rounds, including the initial measurement round when all units are in the control group. So if you have n=1000, this is 1000 x 6 = 6000 observations/measurements.
I'm writing to ask about a possible alternative (see pic below):
- Stratifying by community (let's say 5 communities overall; increase to N = 1500 because design has less power than SW)
- Randomize strata (communities) to order of intervention (first, second, third, fourth, fifth)
- Within the first community strata, Community A, conduct baseline surveys with all n=300 and then randomize units to treatment or control
- Deliver the intervention to n/2 units randomized to treatment
- Conduct endline survey with all n=300 in Community A (treatment and wait-list control)
- Conduct baseline survey with all n=300 in Community B (could be at same time as #5) and then randomize units to treatment or control
- Deliver the intervention to n/2 units in Community B randomized to treatment AND n/2 wait-list control units from Community A (optional, but this is what we would do)
In the alternative design, every unit is surveyed twice, just at different times. With a total sample of n=1500, this is 1500 x 2 = 3000 surveys. Compared to the SW design, this is 6000 - 3000 = 3000 fewer surveys, which has big cost implications.
SW works because we observe every unit before and after every step and then model time.
In the alternative design, we only have 2 measurements (baseline and endline) for every unit assigned to treatment (n=750) and wait-list control (n=750).
- Baseline for Community A conducted in month 1
- Endline for Community A conducted in month 3
- Baseline for Community B conducted in month 3
- Endline for Community B conducted in month 6
- Baseline for Community C conducted in month 6
- Endline for Community C conducted in month 9
- Baseline for Community D conducted in month 9
- Endline for Community D conducted in month 12
- Baseline for Community E conducted in month 12
- Endline for Community E conducted in month 15
- (would not measure post-treatment for Community E wait-list control; just deliver program)
In the alternative design, can we account for the fact that the observations are made at different times? In SW, every unit is measured before and after every round, which makes it easier to model time effects.
Could we regress endline DVs on assignment to treatment (0/1), a vector of baseline controls, dummy variables for community strata, and month of endline measurement? Better alternatives?
Assuming there is a solution, how to think through the implications for power?