README.md

In audreyrenson/didgformula: The difference-in-differences g-formula

didgformula

The R package didgformula implements inverse-probability weighted, iterated conditional g-computation, and doubly robust targeted maximum likelihood estimators for sustained intervention effects under parallel trends assumptions.

Installation

Only the development version is available so far. You can install it from GitHub with:

# install.packages("devtools")
devtools::install_github("audreyrenson/didgformula")

Examples

Here is a basic example using simulated data:

library(didgformula)

set.seed(10)

time_periods = 5
N_obs        = 1e4
parameters   = generate_parameters(Tt=time_periods)
df           = generate_data(N=N_obs, Tt=time_periods, Beta=parameters, ylink = 'rnorm_identity')

head(df)
#>   uid U0 L0         W0 A0 L1         W1 A1 L2          W2 A2 L3         W3 A3
#> 1   1  1  1 0.42889213  0  0 -2.7811919  0  0  0.18963549  0  0 -2.4548948  1
#> 2   2  1  0 1.53171732  0  1  1.0080488  0  1  0.06726136  0  0  1.4825164  0
#> 3   3  0  1 0.65063303  0  1 -0.3276539  1  0 -1.68579853  1  1  1.5884367  1
#> 4   4  1  0 0.92350947  0  0 -1.3764935  0  0 -0.19326923  0  0 -1.0013806  1
#> 5   5  1  1 1.40689657  0  0 -0.5103987  0  0  0.57436194  0  0 -1.0926517  0
#> 6   6  1  0 0.07771018  0  1 -0.3072492  0  0 -0.35724493  0  0  0.4300903  0
#>   L4         W4 A4 L5          W5 A5         Y0        Y1          Y2
#> 1  0 -0.2439675  1  1  2.09577923  1 -2.1574128 -1.766963  0.11996427
#> 2  0 -0.6986895  0  0 -0.08951730  0 -0.7277109 -3.525773 -1.24276996
#> 3  0 -0.7389423  1  0  0.03679411  1 -2.9713932 -4.313408 -0.04549021
#> 4  0  1.0629917  1  0 -0.85402635  1 -1.8168327 -3.201612 -0.58378946
#> 5  0 -0.3215103  0  0 -2.07722722  0 -1.0881852 -3.542649  0.45214506
#> 6  1 -0.7852357  0  0 -0.07780930  0 -1.5003276 -2.923101 -1.09694604
#>           Y3         Y4          Y5
#> 1 -2.4577156 -0.4874623  3.61649314
#> 2  0.4199241  0.2020285  1.22008428
#> 3 -0.7302781  1.1281042  1.50973592
#> 4 -0.0680779  0.8496794 -0.73340196
#> 5 -1.9242997  0.1638116  0.04507994
#> 6 -0.5373790 -0.5839260  3.11250228

We can calculate the true parameters by generating a large number of potential outcomes under the same data-generating mechanism:

df_po = generate_data(N=N_obs*10, Tt=time_periods, Beta=parameters, ylink='rnorm_identity', potential_outcomes = TRUE)

truth = colMeans(calc_ydiffs(df_po, Tt=time_periods)) #calc_ydiffs simply takes Y_t-Y_{t-1} for t=1,...,T
truth
#> [1] -1.84302146  2.95303511 -0.01427988  0.92387834  0.74953343

We can estimate this using IPTW:

estimates_iptw = iptw_pipeline(data = df, den_formula = '~W{t}', Tt=time_periods)
estimates_iptw
#> # A tibble: 5 x 2
#>       t estimate
#>   <int>    <dbl>
#> 1     1  -1.89  
#> 2     2   2.98  
#> 3     3   0.0177
#> 4     4   0.923 
#> 5     5   0.728

ICE:

estimates_ice = ice_pipeline(data = df, inside_formula_t = '~W{t}', inside_formula_tmin1='~W{t-1}', outside_formula = '~W{k}', Tt=time_periods)
estimates_ice
#> # A tibble: 5 x 2
#>       t estimate
#>   <int>    <dbl>
#> 1     1  -1.90  
#> 2     2   2.99  
#> 3     3   0.0186
#> 4     4   0.923 
#> 5     5   0.749

audreyrenson/didgformula documentation built on Oct. 9, 2022, 11:45 a.m.