---
title: "RieszCML"
---

I think a good syntax we could be going after would involve something like this:

```{r}
# select a riesz representer from the catalog 
rc <- riesz_representer_catalog$ate

# or any of
rc <- riesz_representer_catalog$cfmean_a
rc <- riesz_representer_catalog$cfmean_a1
rc <- riesz_representer_catalog$cfmean_a0
rc <- riesz_representer_catalog$att
rc <- riesz_representer_catalog$cde # a joint-intervention -- similar to censoring -- joint on A and mediator 
rc <- riesz_representer_catalog$nde 
rc <- riesz_representer_catalog$mtp
rc <- riesz_representer_catalog$subgroup_mean
rc <- riesz_representer_catalog$longitudinal_ate
rc <- riesz_representer_catalog$ade # average derivative effect
# something for missingness mechanisms

# what's inside the riesz representer? 
rc$name 
rc$description 
rc$mapping # maps causal labels to data ( e.g., A |-> 'trt' )
rc$required_nuisances # a character vector like c('Q_a', 'g') 
rc$representer_weights(data, ) 
rc$eif(data, representer_weights, nuis)   # efficient influence function (depends on a vector riesz representer values [weights] and nuisances) 
rc$estimate # function to produce an estimate of the parameter

# what we can do with it:

riesz_estimate(data = data, representer = rc, nuisances = nuis)
riesz_tmle(data = data, representer = rc, nuisances = nuis)
# where nuis is a named list of functions that take data |-> estimate nuisance
# parameter

# aspirationally, we want to be able to compose representers per Sal's 
# theorem 
riesz_representer_subgroup_mtp <- riesz_compose(
  list(
    riesz_representer_catalog$subgroup_mean,
    riesz_representer_catalog$mtp
  )
)

# or subgroup-specific average derivative effects
riesz_representer_subgroup_ade <- riesz_compose(
  list(
    riesz_representer_catalog$subgroup_mean,
    riesz_representer_catalog$ade
  )
)

# I think the right way to think about this will be in "layers" of representers; 
# e.g., the indexing that Sal uses in Theorem 2. 
# 
# And after the required nuisance functions and estimating the representer 
# from one layer, it becomes available to the next layer?  and the subsequent
# layers treat the prior representers as named 'required_nuisances'?

# riesz_estimate should handle:
#  - cross fitting required nuisances
#    - could be nice to use origami::make_folds if we anticipate non-iid settings(?)
# 
#  ***** 
#   Sal suggests keeping as simple as possible, and just not doing cross-fitting;
#   AND 
#   If we set it up so that the user can pass values for `nuisances` then they can
#   figure out cross-fitting themselves. 
#  *****
# 
#  - if passed a ComposedRieszRepresenter (e.g., a list of RieszRepresenters) 
#    then perform estimation sequentially 
#    - perhaps that means calling riesz_estimate_single_representer (an internal fn)
#      either once or repeatedly 
#
# riesz_tmle should handle
#  - fluctuation at every layer of a ComposedRieszRepresenter or 
#    the one fluctuation for the single-timepoint estimators
# 
# both riesz_estimate and riesz_tmle should return psi_hat, EIFs, std errors, 
# confidence intervals, etc. 
```


# Need to Figure Out 

 - A data <-> labels interface -- probably similar to what I've already done 
 in the CausalData R6 TargetedLearning implementation -- 
   - or maybe I don't need this with the way nuis is described above 
   - an alternative might be to store inside the RieszRepresenter objects 
   something called
   `rc$mapping` that defines a mapping like 
     A |-> 'trt'
     L |-> c('confounder1', 'confounder2')
     Y |-> 'outcome'
     Delta |-> 'mis'

 - I think this implies that behind the scenes longitudinal_ate will be a 
 `ComposedRieszRepresenter` object. 
 
 - Haven't thought at all about vector-valued parameters ... is that
 something I need to think about?