In discrete time models where events are daily, we have double censoring Data that is on a daily scale where two events are linked by a delay are double censored. - ThreadSky

seabbs.bsky.social • 7 days ago

In discrete time models where events are daily, we have double censoring

Data that is on a daily scale where two events are linked by a delay are double censored.

Comments

seabbs.bsky.social•7 days ago

Nearly all epi surveillance data is censored. The bias is absolute (about a day) so as distributions get longer relative to the censoring it matters less.

However, many delays such as the incubation period for COVID-19 are short enough this matters

seabbs.bsky.social•7 days ago

Nearly all implementations I have seen in the literature both in applied studies and in tools for use in applied studies don't account for double censoring when using a PMF. This is both in new and old work and as far as I can tell the field is largely unaware of the problem.

seabbs.bsky.social•7 days ago

A notable exception to this is @anne-cori.bsky.social epiestim gamma si function

https://mrc-ide.github.io/EpiEstim/reference/discr_si.html

Which is the only applied cas I know of getting it right.

seabbs.bsky.social•7 days ago

Does this matter? What is your quarantine period was 1 day shorter? Would that matter both individually or on a population level?

seabbs.bsky.social•7 days ago

You can use `primarycensored` to do this https://primarycensored.epinowcast.org/ if you want to use PMF and have a setting like above

seabbs.bsky.social•7 days ago

A good case in point here is EpiNow2. This currently handles double censoring incorrectly (like nearly everywhere else): https://epiforecasts.io/EpiNow2/dev/articles/EpiNow2.html

Hopefully not for much longer: https://github.com/epiforecasts/EpiNow2/issues/776

seabbs.bsky.social•7 days ago

I (with @adamjkucharski.bsky.social) made a little simulator script so people can play about with this.

https://gist.github.com/seabbs/e1fcd58f0e749d48a3e98df23396dd98

seabbs.bsky.social•7 days ago

(note another point here is we don't expect our discretised and continuous dists to line up if everything has been done right so plotting them over the top of each other might just be confusing)

Comments

Posting Rules

Reply