How to handle missing department-year observations after collapsing ACLED data

[marie carmel]

Hi everyone,

I’m working with the ACLED dataset to study the effect of terrorist attacks at the department-year level.

In my raw data, each row is an event, with variables for depart_id, year, and some event characteristics (e.g., success, fatalities). I collapse it to the department-year level as follows:
collapse (sum) fatalities (max) successful = success (count) n_events = event_id, by(depart_id year) As expected, this produces one observation per department-year only if at least one event occurred. So, if a department had no attacks in a given year, that (depart_id, year) combination does not appear in the collapsed dataset. To create a full panel for my analysis (e.g., event-study or DiD), I then merge this collapsed file onto a complete list of all department-year combinations and replace missing values with zeros, assuming that if a department-year is not recorded in ACLED, then no event occurred. My question is: Is this a good practice? Or should I treat these as missing instead of zeros? From my understanding, ACLED provides full coverage for all regions and years within my study period, so the absence of an event means that no attack was recorded, not that data are missing. But I’d like to confirm that this is the right way to construct the panel before proceeding. Thank you!

[Clyde Schechter]

Caveat: I have never myself used ACLED data, and I have only a passing knowledge of it. This might be complicated.

But here's the problem. ACLED has coverage of different regions and countries starting in different years. So an absent observation for Country A in year Y might reliably mean zero, but for Country B in that same year Y, it might just be that Country B was not yet part of ACLED's universe in that year. As far as I know (but, see caveat above), if an observation for Country A in year Y is absent, but there is data for Country A in some year(s) preceding Y, then you can safely create an observation with 0 events to fill that gap. But if there is no data for Country A in any year preceding Y, then it is unclear whether year Y precedes the start of ACLED data collection for that country, or whether the data was collected but there were no events that year.

I suspect that when you downloaded the ACLED data you also got (or could and should have gotten) documentation of the data set which, hopefully, would provide information about the range of years of coverage for each country/region. If so, then referring to that documentation would safely allow you to distinguish zero events from non-coverage. If you don't have such documentation, I recommend going back to the ACLED site and searching for it.

[marie carmel]

Hi Clyde,
Thank you for your response. I understand your point, and I’ve looked at the ACLED coverage list as well. However, since my analysis focuses on departments (or counties) within a specific country, what really matters is whether ACLED covers all departments in that country.

In my case, I’m working with departments in Burkina Faso. ACLED lists Burkina Faso as being covered from “1/1997 – Present.” That said, coverage at the national level doesn’t necessarily imply uniform or complete coverage across all departments. From what I’ve seen in their documentation, they mainly emphasize that their data depend on publicly available sources, local media, and reports, which may vary by region.