2
$\begingroup$

Question: What are my best options for handling group-level correlations in a panel differences-in-differences model using StataSE 14.2?

Background: I have a Differences-in-Differences (DiD) design where a policy (subsidised primary healthcare) was rolled out to different age-groups (k=4) in different years. Like many DiD designs, my model faces the issue of individual-level outcomes (unmet need for a doctor in past 12 months; binary) and group-level treatment, which means there will be (1) correlations within age*year clusters, and (2) serial correlation between age-groups across years.

Ie. UNMETigt = αi+ αg + αt + γgt + βTREATgt + δCOVARSigt + εigt

While I haven't yet obtained access to the panel data, I imagine my Stata code might start from a basis of something like:

xtset id year
xtreg unmet L.i.age##L.i.year + L.treat + L.covars

Currently, this code does not account for the fact that treatment does not vary by i. I am aware that clustering by age*year will bias least-squares standard errors due to serial correlation for age-groups across years. I am also mindful that clustering solely by age-group to remove the time component will only leave me with four clusters, which is insufficient to estimate the correlation without bias. The literature I have consulted appears to cast doubt on the suitability of Stata's cluster option given serial correlation and few clusters (ie, Angrist & Pischke, 2008)

Question: What are my best options for handling group-level correlations in a panel differences-in-differences model using StataSE 14.2?

(1) One option in the literature is to correct the standard errors for first-order serial correlation at the group-level, using an AR(p) structure along the lines of Hansen (2007). I haven't seen any Stata modules that do this, but does anyone else know of one or something similar?

(2) Another option I've seen is block bootstrapping, which bootstraps the SEs by age-group akin to Cameron, Gelbach & Miller (2008). Would this involve simply specifying the option: vce(bootstrap, cluster(age)), or have I misunderstood?

(3) Is there a better approach I'm missing?

Apologies if I've missed something self-evident! I'm new to this form of analysis.

$\endgroup$

Your Answer

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

Browse other questions tagged or ask your own question.