Week 3, Session 5 — G-methods, IV, DiD, RDD; heterogeneity of treatment effect

Course 3 — #courses

Author

R. Heller

Note

Workflow lab: Goal → Approach → Execution → Check → Report.

Learning objectives

  • Use inverse-probability weighting as a g-method for time-varying confounding.
  • Recognise the data-generating situations that justify an instrumental-variable (IV), difference-in-differences (DiD), or regression-discontinuity (RDD) design.
  • Estimate and interpret heterogeneity of treatment effect (HTE) with a subgroup-agnostic method.

Prerequisites

Propensity-score and IPTW basics (W3 S4).

Background

Traditional regression adjustment fails for time-varying confounders that are affected by prior treatment — adjusting for them blocks part of the treatment effect, failing to adjust leaves residual confounding. Robins’s g-methods solve this by reweighting or standardising in a way that respects the temporal order. IV methods exploit a variable that affects treatment but not outcome directly; DiD compares changes over time between groups sharing a counterfactual trend; RDD exploits a sharp assignment rule around a threshold.

Heterogeneity of treatment effect is the shift from “is the average effect non-zero?” to “for whom is it largest?” Model-based HTE uses causal forests, meta-learners (S-, T-, X-learner), or Bayesian hierarchical models to estimate conditional average treatment effects (CATEs) while controlling for multiple testing.

Setup

library(tidyverse)
library(broom)
set.seed(42)
theme_set(theme_minimal(base_size = 12))

1. Goal

Simulate a time-varying-confounding scenario, estimate the causal effect naively and then with IPTW (a marginal structural model), and finish with a simple HTE estimate.

2. Approach

n <- 2000
df <- tibble(
  l0 = rnorm(n),
  a0 = rbinom(n, 1, plogis(0.2 * l0)),
  l1 = rnorm(n, mean = 0.5 * a0 + 0.3 * l0),
  a1 = rbinom(n, 1, plogis(0.3 * l1 + 0.2 * a0)),
  y  = 1.0 * a0 + 0.8 * a1 + 0.5 * l0 + 0.5 * l1 + rnorm(n)
)

The true effect of each treatment is positive and additive. l1 is a time-varying confounder because it is affected by a0 and predicts both a1 and y.

3. Execution

Naive regression adjusting for l1 blocks part of the a0 effect.

naive <- lm(y ~ a0 + a1 + l0 + l1, data = df)
tidy(naive, conf.int = TRUE) |>
  filter(term %in% c("a0", "a1"))
# A tibble: 2 × 7
  term  estimate std.error statistic  p.value conf.low conf.high
  <chr>    <dbl>     <dbl>     <dbl>    <dbl>    <dbl>     <dbl>
1 a0       0.956    0.0459      20.8 1.91e-87    0.866     1.05 
2 a1       0.804    0.0450      17.9 2.55e-66    0.716     0.893

IPTW with stabilised weights:

num0 <- glm(a0 ~ 1, data = df, family = binomial)
den0 <- glm(a0 ~ l0, data = df, family = binomial)
num1 <- glm(a1 ~ a0, data = df, family = binomial)
den1 <- glm(a1 ~ a0 + l0 + l1, data = df, family = binomial)

w0 <- ifelse(df$a0 == 1, fitted(num0), 1 - fitted(num0)) /
      ifelse(df$a0 == 1, fitted(den0), 1 - fitted(den0))
w1 <- ifelse(df$a1 == 1, fitted(num1), 1 - fitted(num1)) /
      ifelse(df$a1 == 1, fitted(den1), 1 - fitted(den1))
df$sw <- w0 * w1

msm <- lm(y ~ a0 + a1, data = df, weights = sw)
tidy(msm, conf.int = TRUE) |>
  filter(term %in% c("a0", "a1"))
# A tibble: 2 × 7
  term  estimate std.error statistic  p.value conf.low conf.high
  <chr>    <dbl>     <dbl>     <dbl>    <dbl>    <dbl>     <dbl>
1 a0       1.23     0.0585      21.1 2.51e-89    1.12      1.35 
2 a1       0.806    0.0587      13.7 4.19e-41    0.691     0.921

4. Check

Balance and weight sanity:

summary(df$sw)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
 0.5802  0.8982  0.9813  0.9997  1.0849  1.9544

Stabilised weights should centre on 1 with no extreme tail. If they do not, a term in the treatment model is likely misspecified.

5. Report

In a simulated cohort with time-varying confounding (n = 2000), naive regression underestimated the effect of initial treatment. An inverse-probability-weighted marginal structural model recovered effects close to the true data-generating values. Instrumental variable, difference-in-differences, and regression-discontinuity designs exploit alternative identification strategies, each with its own assumptions. Heterogeneity of treatment effect can be explored with causal forests or meta-learners.

Brief sketches

  • IV: regress y on an instrument z to recover β = cov(y,z)/cov(a,z).
  • DiD: fit lm(y ~ time*group) on repeated measures.
  • RDD: fit local linear regression on each side of the assignment cutoff.

Flag the identifying assumption for each method — IV: exclusion restriction; DiD: parallel trends; RDD: continuity at the cutoff.

Common pitfalls

  • Using ordinary regression adjustment for time-varying confounders.
  • Trimming extreme weights silently, hiding model misspecification.
  • Reporting subgroup p-values without multiplicity control as HTE evidence.

Further reading

  • Hernán MA, Robins JM. Causal Inference: What If.
  • Athey S, Imbens GW. (2019). Machine Learning Methods That Economists Should Know About.

Session info

sessionInfo()
R version 4.4.1 (2024-06-14)
Platform: x86_64-pc-linux-gnu
Running under: Ubuntu 24.04.4 LTS

Matrix products: default
BLAS:   /usr/lib/x86_64-linux-gnu/openblas-pthread/libblas.so.3 
LAPACK: /usr/lib/x86_64-linux-gnu/openblas-pthread/libopenblasp-r0.3.26.so;  LAPACK version 3.12.0

locale:
 [1] LC_CTYPE=C.UTF-8       LC_NUMERIC=C           LC_TIME=C.UTF-8       
 [4] LC_COLLATE=C.UTF-8     LC_MONETARY=C.UTF-8    LC_MESSAGES=C.UTF-8   
 [7] LC_PAPER=C.UTF-8       LC_NAME=C              LC_ADDRESS=C          
[10] LC_TELEPHONE=C         LC_MEASUREMENT=C.UTF-8 LC_IDENTIFICATION=C   

time zone: UTC
tzcode source: system (glibc)

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
 [1] broom_1.0.12    lubridate_1.9.5 forcats_1.0.1   stringr_1.6.0  
 [5] dplyr_1.2.1     purrr_1.2.2     readr_2.2.0     tidyr_1.3.2    
 [9] tibble_3.3.1    ggplot2_4.0.3   tidyverse_2.0.0

loaded via a namespace (and not attached):
 [1] gtable_0.3.6       jsonlite_2.0.0     compiler_4.4.1     tidyselect_1.2.1  
 [5] scales_1.4.0       yaml_2.3.12        fastmap_1.2.0      R6_2.6.1          
 [9] generics_0.1.4     knitr_1.51         backports_1.5.1    htmlwidgets_1.6.4 
[13] pillar_1.11.1      RColorBrewer_1.1-3 tzdb_0.5.0         rlang_1.2.0       
[17] utf8_1.2.6         stringi_1.8.7      xfun_0.57          S7_0.2.2          
[21] otel_0.2.0         timechange_0.4.0   cli_3.6.6          withr_3.0.2       
[25] magrittr_2.0.5     digest_0.6.39      grid_4.4.1         hms_1.1.4         
[29] lifecycle_1.0.5    vctrs_0.7.3        evaluate_1.0.5     glue_1.8.1        
[33] farver_2.1.2       rmarkdown_2.31     tools_4.4.1        pkgconfig_2.0.3   
[37] htmltools_0.5.9

Related labs

Course 3 — Design & CausalDAGs with dagitty
Course 3 — Design & CausalPropensity scores and IPTW

Explore the full label network →