We are currently evaluating a government Microenterprise program using a cluster RCT design. The treatment involves the provision of a grant to a poor household to start a micro-enterprise. We are looking at outcomes such as household income and expenditures.

An issue is that program households received their grant at different times. Implementation is devolved to the municipalities, so the timing of the release of the grants depend on whether the local program officers start recruitment early or late. Some households received the grant as early as July 2018, while some as late as December 2018. The baseline survey took place in August 2018 - December 2018, and the endline survey will take place in May-June 2019. This would mean exposure of households to the treatment would vary in length when the endline comes. Some would have had their microenterprises running longer than others.

Our estimation of the treatment effect involves a simple OLS regression with municipal fixed effects and cluster robust errors. It has the following specification:

$y_{ij} = a + \delta T_j+ \gamma_k y_{ij_{-1}} + \Sigma_h \gamma_{h} x_{hij_{-1}} + \Sigma_m \gamma_{m}c_m + \epsilon_{ij} $


  • $y_{ij}$ is the outcome for household $i$ in municipality $j$
  • $T_j$ is the treatment dummy
  • $y_{ij_{-1}}$ if the baseline value of the outcome variable
  • $x_{hij_{-1}}$ are $h$ household-specific covariates measures at baseline
  • $c_m$ is a municipal dummy for the $m^{th}$ cluster
  • $\epsilon_{ij}$ is the random error.

However I am not sure how we should take into account the varying exposure lengths of the households to the treatment. I have come across papers on how to do so when using a DID strategy but not in the context of an RCT.


This involves making some assumptions on how the continuous outcome evolves over time. First question: do you expect it to make a meaningful difference and vary across groups? Otherwise one could perhaps ignore it?

If you can specify that in terms of some function (or something very flexible like a Gaussian process), for which parameters could differ between groups that would be one approach.

Alternatively, assume that the function can be well approximated by a piecewise constant function and put some number of time windows into the model. Additionally, you'll have to think about whether the intervention effect changes over time (presumably to since extent - surely, there's no effect after 5 min, but perhaps is negligible for your assessment times?) and then a time window by intervention interaction may become necessary.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.