growth_prediction_example.Rmd

---
title: "Example HAL counterfactuals"
output: html_document
---

```{r setup, include=FALSE}

rm(list=ls())
source(paste0(here::here(), "/0-config.R"))
source(paste0(here::here(), "/functions/0_calc_empirical_IC.R"))
library(hal9001)
library(washb)
birthweight_shift <- readRDS(paste0(here::here(),"/figures/birthweight_shift_example.RDS"))
waz_shift <- readRDS(paste0(here::here(),"/figures/waz_shift_example.RDS"))

d <- readRDS(paste0(data_dir,"analysis_data.RDS"))

y_name = "whz_24" 
x_names = c(covars, "arm", "waz_birth")
dfull <- d %>% ungroup() %>% select(all_of(c("studyid","country",y_name, x_names)))

#temp: complete cases
dim(dfull)
dfull <- dfull[complete.cases(dfull),]
dfull <- droplevels(dfull)
dim(dfull)

dfull$cohort <- paste0(dfull$studyid,"-",dfull$country)

```


## Data description

```{r}

#Number children:
nrow(dfull)

#cohort sizes:
table(dfull$cohort)

#predictors:
x_names



```


## HAL model

```{r, eval=FALSE}

  fit_init <- fit_hal(X = X,
                      Y = Y,
                      smoothness_orders = 0,
                      return_x_basis = TRUE,
                      family = "gaussian",
                      num_knots = hal9001:::num_knots_generator(
                        max_degree = ifelse(ncol(X) >= 20, 2, 3),
                        smoothness_orders = 0,
                        base_num_knots_0 = max(100, ceiling(sqrt(n)))
                      ))

```


## Counterfactual 1: Shifting all low birthweight babies to the low-birthweight threshold (2500g)

```{r, echo=F}

waz_shift$p

d <- waz_shift$plotdf

diff <- round( d$est[d$estimate=="counterfactual"& d$study=="Pooled"] - d$est[d$estimate=="observed" & d$study=="Pooled"], 3)


#mean(d$est[d$estimate=="counterfactual" & !grepl(d$study,"Pooled")] - d$est[d$estimate=="observed" & !grepl(d$study,"Pooled")])


```

WHZ increase at 6 months: `r diff``


## Counterfactual 2: LNS suppliments for all moms during pregnancy (40g birthweight gain based on RCT meta-analysis)



```{r, echo=FALSE}

birthweight_shift$p

d <- birthweight_shift$plotdf

diff <- round( d$est[d$estimate=="counterfactual"& d$study=="Pooled"] - d$est[d$estimate=="observed" & d$study=="Pooled"], 3)



```


WHZ increase at 6 months: `r diff``



## Next steps:

  -  Make shiny interface
  -  Estimate outcome difference directly
  -  Get prevalence difference of wasting/stunting
  -  Fit models for all growth exposures/outcomes
  -  Tune HAL models
  -  Add HAL model accuracy
  -  Add diagnostics (covariates used, N's for each study, percent of kids shifted)
  -  Ability to subset by subgroups (region, sex, children intervened on)
  -  Add maternal size shifting
  -  Add variable importance ranking of predictors
  -  Question? Should I be using the shift TMLE to be able to target the impact of shifting early growth, etc?