---
title: "Incorporating observation data into models"
output:
  html_document:
    toc: true
    theme: null
vignette: >
  %\VignetteIndexEntry{Incorporating observation data}
  %\VignetteEngine{knitr::rmarkdown}
  \usepackage[utf8]{inputenc}
---

```{css, echo=FALSE}
.hide {
  display: none!important;
}

/* https://bookdown.org/yihui/rmarkdown-cookbook/html-css.html */
.modelScript pre.sourceCode, .modelScript pre.sourceCode code {
  background-color: #f0f8ff !important;
}
```


```{r, message=FALSE, echo=FALSE}
library(gadget3)
library(dplyr)
if (nzchar(Sys.getenv('G3_TEST_TMB'))) options(gadget3.tmb.work_dir = gadget3:::vignette_base_dir('work_dir'))
set.seed(123)
```

Each run of a model will generate a likelihood score, by summing the result of any *likelihood actions* in the model together.
To fit a model to observation data, we need to add a *likelihood action* that will compare observation data to the model.

To do this you either need `g3l_catchdistribution()` or `g3l_abundancedistribution()`.
As the names suggest, `g3l_catchdistribution()` will compare catch data of stocks from the provided fleets/predators,
`g3l_abundancedistribution()` will compare total abundance of stocks.
Otherwise they are identical.

This comparison is a 2 stage process:

1. Aggregation: The columns in the observation data are inspected, and converted into an array. Model data is then aggregated into an identically-sized array ready for the next step
2. Comparison: A comparison function (`function_f`) is supplied to `g3l_*distribution()`, to convert to a likelihood score.
   For example `g3l_distribution_sumofsquares()`, which compares the relative abundance of each grouping vs. the model, and sums the square of these

We saw the following example in `vignette("introduction-single-stock")`:

```{r, message=FALSE, echo=FALSE}
library(dplyr)
area_names <- g3_areas(c('IXa', 'IXb'))
fish <- g3_stock("fish", seq(5L, 25L, 5)) |>
  g3s_livesonareas(area_names["IXa"]) |>
  g3s_age(1L, 5L)
```

<div class="modelScript">
```{r}
# TODO: This isn't a brilliantly-named example, something else?
# Generate random data
expand.grid(year = 1990:1994, step = 3, area = 'IXa') |>
    # Fill in a number column with total numbers in that year/step/area combination
    mutate(number = runif(n(), min = 10000, max = 100000)) ->
    dist_si_acoustic

actions_likelihood_si_acoustic <- list(

  g3l_abundancedistribution(
    "dist_si_acoustic",
    obs_data = dist_si_acoustic,
    
    stocks = list(fish),
    function_f = g3l_distribution_surveyindices_log(alpha = NULL, beta = 1),
    area_group = area_names,
    report = TRUE,
    nll_breakdown = TRUE),
  NULL)
```
</div>

The first step generates random observation data. We have the following columns:

* *year*: We want to do comparisons in 5 years, 1990..1994. Other years nothing will happen.
* *step*: We compare abundance in the autumn step
* *area*: We only compare with stock within the 'IXa' area
* *number*: We compare against number individuals.

There are no *age* or *length* columns, so this is total abundance within that area/time.

We use `g3l_distribution_surveyindices_log()`, which calculates the likelihood score by performing a linear fit using the log scale,
the slope (*beta*) is fixed.


The model optimisation process will minimise the total likelihood score,
and in the process fit the model to the observation data.

## gadget2<->gadget3 translation

The function of many of gadget2's likeihood components has been collapsed down into these 2 functions.
Here is a summary of how to translate:

* [CatchDistribution](https://gadget-framework.github.io/gadget2/userguide/chap-like.html#sec-catchdist) likelihood component
  * Use ``g3l_catchdistribution()``
  * Set *function_f* parameter to the matching ``g3l_distribution_*()`` function, e.g. ``g3l_distribution_sumofsquares()``
  * Supply data with columns *year*, *step*, *area*, *age*, *length*, *number*
* [SurveyIndices](https://gadget-framework.github.io/gadget2/userguide/chap-like.html#sec-surveyindices) likelihood component
  * Use ``g3l_abundancedistribution()``
  * Set *function_f* parameter to ``g3l_distribution_surveyindices_log()`` or ``g3l_distribution_surveyindices_linear()``
  * Supply data with columns *year*, *step*, *area*, *length*, *number*
* [CatchInKilos](https://gadget-framework.github.io/gadget2/userguide/chap-like.html#sec-catchinkilos) likelihood component
  * Use ``g3l_catchdistribution(function_f = g3l_distribution_sumofsquares())``
  * Supply data with columns *year*, *step*, *area*, *fleet*, *weight*

## Observation data format

gadget3 bases it's decision on how to aggregate on the incoming data.
Whilst it tries to do the "right thing" in most cases,
it's important to get the shape of this data to match what you require.

In doing so, the incoming `data.frame()` is converted into an `array()`.
We can use the `g3_distribution_preview()` function to see what that array will look like,
and in doing so see how gadget3 will treat the observation data.

Critically, your data needs to have column names it recognises.
The following breaks down what can be by each column.

### *weight* and *number* columns

At least a *weight* or *number* column needs to be supplied.
If a *number* column is present then each value will be compared to number of individuals in that group.
If a *weight* column is present then each value will be compared to total biomass in that group.

The *weight* is suitable for total catches in kilos, otherwise the *number* column will generally be used.

### *year* and *step* columns: Landings, abundance indices

If a *year* column is given, then catch/abundance will be aggregated by year.
Gaps are allowed, if so then no comparisons will be made for that year/step.

The following observations will be compared against the total number of indivduals caught in years 1999, 2000, 2002, 2003:

```{r}
g3_distribution_preview(read.table(header = TRUE, text="
year  number
1999    1000
2000    1002
2002    1004
2003    1008
"))
```

If *step* is given, then it will restrict it to that step within the year (see `?g3a_time` for how to define steps in a year).

In the following example, we will aggregate **spring** of 1999 & 2000, **autumn** of 2001.
Any catch in spring of 2001, or other periods not mentioned, will be ignored:

```{r}
g3_distribution_preview(read.table(header = TRUE, text="
year step number
1999    2   1020
2000    2   2040
2001    3   1902
"))
```

### *length* column: Length distribution data

Adding a length column will aggregate catch/abundance data by the same length bins as in the observation data.

For instance:

```{r}
g3_distribution_preview(read.table(header = TRUE, text="
year  length number
1999  [0,10)  1023
1999 [10,20)  2938
1999 [20,30)  3948
1999 [30,40)  3855
2000  [0,10)  1023
2000 [10,20)  2938
# NB: No [10,30)
2000 [30,40)  3855
"))
```

Note that unlike with the *year* & *step* columns, here gaps in data are preserved,
in the preview output we see ``NA`` for the missing *year* & *length* combination.
By default we will compare to ``0`` at this point,
this behaviour is controlled with the *missing_val* parameter to `?g3l_catchdistribution`.

Length aggregations do not have to be hand-crafted like we do above,
a length column could be generated using ``dplyr::group_by()`` and ``cut()``:

```{r}
# Generate an unaggregated length distribution
ldist.lln.raw <- data.frame(
    year = c(1999, 2000),
    length = sample(10:75, 100, replace = TRUE),
    number = 1,
    stringsAsFactors = FALSE)

# Group length into 10-long bins
ldist.lln.raw |> dplyr::group_by(
  year = year,
  length = cut(length, breaks = seq(10, 100, by = 10), right = FALSE)
) |> dplyr::summarise(number = sum(number), .groups = 'keep') -> ldist.lln.agg
ldist.lln.agg

# NB: The last 2 bins are empty, but because cut() creates a factor column,
#     gadget3 knows about them even though they don't appear in the data.
g3_distribution_preview(ldist.lln.agg)
```                        

gadget3 will also automatically read the aggregation attributes used by `?mfdb::mfdb_sample_count`.

```{r, eval=nzchar(Sys.getenv('G3_TEST_TMB'))}
# Import data into a temporary database
library(mfdb)
mdb <- mfdb(tempfile(fileext=".duckdb"))
ldist.lln.raw$month <- 1
ldist.lln.raw$areacell <- 'all'  # NB: We have to have an areacell mapping for MFDB
mfdb_import_area(mdb, data.frame(name = c('all'), size = c(5)))
mfdb_import_survey(mdb, ldist.lln.raw)

# Use mfdb_sample_count to extract & group in the same manner as above
ldist.lln.agg <- mfdb_sample_count(mdb, c('length'), list(
    year=1999:2000,
    length = mfdb_interval("len", seq(10, 100, by = 10)) ))[[1]]
g3_distribution_preview(ldist.lln.agg, area_group = c(all=1))

mfdb_disconnect(mdb)
```

### *age* column: Age-Length distribution data

Age-length aggregations can be performed by adding an *age* column in a very similar manner to the *length* column:

We can both group by individual age values:

```{r}
g3_distribution_preview(read.table(header = TRUE, text="
year  age  length number
1999    1  [0,10)  1026
1999    1 [10,20)  2936
1999    1 [20,30)  3962
1999    1 [30,40)  3863
1999    2  [0,10)  1026
1999    2 [10,20)  2936
1999    2 [20,30)  3962
1999    2 [30,40)  3863
"))
```

...or group ages together:

```{r}
g3_distribution_preview(read.table(header = TRUE, text="
year  age    length number
1999  [1,1]    [0,10)  1026
1999  [1,1]   [10,20)  2936
1999  [1,1]   [20,30)  3962
1999  [1,1]   [30,40)  3863
1999  [2,4]  [0,10)  1011
# Missing [2,4] [10,20)
1999  [2,4] [20,30)  3946
1999  [2,4] [30,40)  3872
"))
```

As before, gaps in data are preserved, and ``missing_val`` is used to decide what to do with them.

Again, gadget3 will also interpret aggregation generated by ``group_by(age = cut(...))`` or `?mfdb::mfdb_sample_count`.

### *area* column

If a stock is divided up into multiple areas, then data can be broken down by area

```{r}
area_names <- g3_areas(c('IXa', 'IXb', 'IXc'))

g3_distribution_preview(read.table(header = TRUE, text="
year    area    number
1999   IXa   1000
1999   IXb   4305
2000   IXa   7034
2000   IXb   2381
2001   IXb   3913
"), area_group = area_names)[length = '0:Inf',,]
```

As before, gaps in data are preserved, and ``missing_val`` is used to decide what to do with them.

However, if an area isn't mentioned at all (note that ``IXc`` does not figure in the above data), then it won't be compared.

### *stock* column: Maturity stage distribution

If you have multiple stocks, for example because you have divided up your species into mature and immature substocks,
you can use this division in likelihood components also:

```{r}
st_imm <- g3_stock(c(species = 'fish', 'imm'), 1:10)
st_mat <- g3_stock(c(species = 'fish', 'mat'), 1:10)

g3_distribution_preview(read.table(header = TRUE, text="
year    stock    number
1999   fish_imm   1000
1999   fish_mat   4305
2000   fish_imm   7034
2000   fish_mat   2381
2001   fish_mat   3913
"), stocks = list(st_imm, st_mat))[length = '0:Inf',,]
```

The stock names have to match what gadget3 uses, or an error will be generated.

You can also use partial stock names, for example:

```{r}
stocks <- list(
   g3_stock(c(species = 'fish', 'imm', 'f'), 1:10),
   g3_stock(c(species = 'fish', 'imm', 'm'), 1:10),
   g3_stock(c(species = 'fish', 'mat', 'f'), 1:10),
   g3_stock(c(species = 'fish', 'mat', 'm'), 1:10) )

drop(g3_distribution_preview(read.table(header = TRUE, text="
year    stock    number
1999   imm   1000
1999   mat   4305
2000   imm   7034
2000   mat   2381
"), stocks = stocks))
```

The `imm` columns will compare to the sum of `fish_imm_f` & `fish_mat_f`.

The parts do not have to be in order, the following is also valid:

```{r}
drop(g3_distribution_preview(read.table(header = TRUE, text="
year    stock    number
1999   fish_f   1000
1999   fish_m   4305
2000   fish_f   7034
2000   fish_m   2381
"), stocks = stocks))
```

*NB:* A stock can only appear in one grouping:

```{r}
# NB: Wrong!
drop(g3_distribution_preview(read.table(header = TRUE, text="
year    stock    number
1999   f         1000
1999   imm       4305
2000   f         7034
2000   imm       2381
"), stocks = stocks))
```

`fish_imm_f` will only appear in the `f` rows, not `imm` rows.
In this case, multiple likelhood components would be a better approach.

As before, gaps in data are preserved, and ``missing_val`` is used to decide what to do with them.

### *fleet* column

```{r}
fleets <- list(
    g3_fleet(c('comm', country = 'se')),
    g3_fleet(c('comm', country = 'fi')),
    g3_fleet(c('surv', country = 'se')) )

g3_distribution_preview(read.table(header = TRUE, text="
year    fleet    number
1999   comm      1000
1999   surv_se   4305
2000   comm      7034
2000   surv_se   2381
2001   surv_se   3913
"), fleets = fleets)[length = '0:Inf',,]
```

The name matching works in the same way as stocks above, and should be either the name gadget3 uses or parts of it.

As before, gaps in data are preserved, and ``missing_val`` is used to decide what to do with them.

### *predator* column

```{r}
predators <- list(
    g3_stock(c('seal', 'imm', 'f'), 10:20),
    g3_stock(c('seal', 'mat', 'f'), 10:20),
    g3_stock(c('seal', 'imm', 'm'), 10:20),
    g3_stock(c('seal', 'mat', 'm'), 10:20) )

drop(g3_distribution_preview(read.table(header = TRUE, text="
year    predator    number
1999   seal_f   1000
1999   seal_m   4305
2000   seal_f   7034
2000   seal_m   2381
"), predators = predators))
```

The name matching works in the same way as stocks above, and should be either the name gadget3 uses or parts of it.

As before, gaps in data are preserved, and ``missing_val`` is used to decide what to do with them.