vanAmsterdam

library(data.table)
library(purrr)
set.seed(12345)
x <- rnorm(100)
x[sample(1:100, size = 50)] = NA
df <- data.table(matrix(x, nrow=10))
df[, columnA:=rep(letters[1:5], 2)]
df[, list(n_missings=map_dbl(.SD, ~(mean(is.na(.x))<0.33))%>%sum), by='columnA']

vanAmsterdam

'''
define and run a latent variable model
'''

import numpyro
from jax import numpy as np, random
from jax.scipy.special import logsumexp
from numpyro import distributions as dist
from numpyro.distributions import constraints
from numpyro.handlers import mask, substitute, trace, seed

vanAmsterdam

#' Latent confounder model

# dag: 

#' W1 <- U -> W2
#' tx -> y
#' U -> y
#' U -> tx

library(lavaan)

vanAmsterdam

'''
define and run a latent variable model, everything gaussian, with posterior predictive on newdata
for a description of the model see twitter thread: https://twitter.com/WvanAmsterdam/status/1251214875394740226?s=20

DAG:

W1 <- U -> W2  # latent confounder with 2 proxies
U -> tx
U -> y
tx -> y

vanAmsterdam

# Ordinal regression with ImproperUniform

from jax import numpy as np, random
import numpyro
from numpyro import sample
from numpyro.distributions import constraints, Normal, ImproperUniform, Categorical, OrderedLogistic
from numpyro.infer.mcmc import NUTS, MCMC
import pandas as pd

num_chains = 4