seabbs
commented
1 month ago This looks really nice from reading the pitch but...
My high-level concern is the proposed interaction between missingness and accumulation. The current design says that something is accumulated until an observation occurs which seems potentially error prone to me? Why did you discount the alternative where these are separable? The reason I think it concerning is that it seems very easy to make a mistake or to get unexpected behaviour. My impression is that having missing data in your weekly example would look like this:
| date | confirm | accumulate |
|---|---|---|
| 2024-10-18 | NA | 1 |
| 2024-10-19 | NA | 1 |
| 2024-10-20 | NA | 1 |
| 2024-10-21 | NA | 1 |
| 2024-10-22 | NA | 1 |
| 2024-10-23 | NA | 1 |
| 2024-10-24 | NA | 0 |
| 2024-10-25 | NA | 1 |
| 2024-10-26 | NA | 1 |
| 2024-10-27 | NA | 1 |
| 2024-10-28 | NA | 1 |
| 2024-10-29 | NA | 1 |
| 2024-10-30 | NA | 1 |
| 2024-10-31 | 17 | 1 |
| 2024-11-01 | NA | 1 |
| 2024-11-02 | NA | 1 |
| 2024-11-03 | NA | 1 |
| 2024-11-04 | NA | 1 |
| 2024-11-05 | NA | 1 |
| 2024-11-06 | NA | 1 |
| 2024-11-07 | 11 | 1 |
(see the 24th) but would this reset the accumulation counter as you might expect? From the outline the mechanism to reset accumulation is implicit and I think occurs when data is present?
A more restrictive separable approach would be:
Example
| date | confirm | accumulate |
|---|---|---|
| 2024-10-23 | NA | 1 |
| 2024-10-24 | 10 | 0 |
| 2024-10-26 | NA | 1 |
| 2024-10-27 | NA | 1 |
| 2024-10-28 | 17 | 0 |
The likelihood is evaluated on two days, 24 October and 28 October. On 24 October the data (10) is compared to (modelled value on 23 October) + (modelled value on 24 October). On 28 October the data (17) is compared to (modelled value on 26 October) + (modelled value on 27 October) + (modelled value on 28 October).
Whilst removing the interaction this design switches to accumulate being logical. It loses the partial ascertainment and the ability to skip accumulation (i.e adding a delay effect as you did on the 27th). I am unclear as to the utility of those but happy to hear? You could relax this to make accumulate numeric and so allow for partial ascertainment.
sbfnk
jamesmbaazam
Following discussion in #547 here's a proposed design for the future accumulation / missingness interface. As proposed this would also accommodate changes in reporting. As default behaviour would be unchanged it would require fairly minimal deprecation. All comments welcome.
Supporting missing data
We want to support reporting patterns that include accumulation (i.e. batch reporting of data from multiple dates, for example weekly) and missingness (dates which are lacking reports) in incidence and prevalence data.
Proposed interface
Missing data
Any dates between the minimum date and maximum date in the data that is either absent, or present with an
NAvalue (currently calledconfirm) is interpreted as missing and ignored in the likelihood. All other data points are used in the likelihood. This matches the current default behaviour, introduced in version 1.5.0.Accumulation
If instead modelled values on these days should be accumulated onto the next reporting date, the passed
data.framemust have an additional logical column,accumulate. Ifaccumulateis TRUE then the modelled value of the observed latent variable on that day is added to any existing accumulation and stored for later observation. Ifaccumulateis FALSE the modelled value is added to any stored accumulated variables before potentially being used in the likelihood on that day (if notNA). Subsequently the stored accumulated variable is reset to zero.Example
The likelihood is evaluated on two days, 24 October and 28 October. On 24 October the data (10) is compared to (modelled value on 23 October) + (modelled value on 24 October). On 28 October the data (17) is compared to (modelled value on 26 October) + (modelled value on 27 October) + (modelled value on 28 October).
Helper functions
A helper function,
fill_missing_dates()can be used to convert weekly data to the required format, i.e.can be converted with
fill_missing_dates(missing = "accumulate", initial = 7)to