Causal effect estimators: IPW regression, g-computation, 2SLS, PLIV, sequential matching, SNFTM

R arm of morie.fn.ate, g_comp, late, pliv, prsmtd and snmcox.

Usage

Ate(data, outcome, treatment, weights_col)

GComp(data, treatment, outcome, covariates, outcome_model = "linear")

Late(
  data,
  treatment,
  outcome,
  instrument,
  covariates = NULL,
  se_type = "homoskedastic"
)

Pliv(
  data,
  treatment,
  outcome,
  instrument,
  covariates,
  n_folds = 5,
  random_state = 42
)

Prsmtd(A, H)

Arguments

data: A data frame (or list of columns).
outcome, treatment, instrument: Column names.
weights_col: Column of analytic weights.
covariates: Covariate column names.
outcome_model: "linear" or "logistic".
se_type: "homoskedastic" or "robust" (HC0).
n_folds: Cross-fitting folds.
random_state: Philox seed of the fold assignment.
A, H: Treatment-initiation indicators and covariate, n by T matrices (or vectors).

Value

A named list (the Python result's fields). Prsmtd reports 1-based periods and rows.

Details

Ate: weighted least squares of the outcome on \((1, T)\); the treatment coefficient is the Hajek IPW contrast, with the HC3 sandwich of the weighted fit as its standard error.

GComp: g-computation. An outcome model \(E[Y | T, X]\) (least squares, or unpenalised logistic regression by Newton's method) is fitted and the ATE is the mean of \(\hat Y(1) - \hat Y(0)\); the standard error is the standard deviation of 500 bootstrap replicates (replicate \(b\) resamples rows from Philox stream \(b\) of seed 42; single-arm resamples are skipped), the interval their type-7 2.5\

Late: two-stage least squares of \(Y\) on \((1, W, T)\) with instruments \((1, W, Z)\); homoskedastic or HC0 standard error and the first-stage F statistic of the excluded instrument.

Pliv: partially linear IV by double machine learning. Ridge nuisances (penalty chosen from 0.1, 1, 10 by leave-one-out PRESS) are cross-fitted over folds dealt from the Philox order of seed random_state; \(\theta = \sum w u / \sum w v\) with the partialling-out score.

Prsmtd: sequential risk-set propensity matching. In each period a logistic propensity of initiation on the current covariate is fitted in the risk set, and initiators (in decreasing propensity) take the nearest unused not-yet-treated control.

References

Hernan, M. A. and Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC.

MacKinnon, J. G. and White, H. (1985). Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties. Journal of Econometrics 29(3), 305-325.

Robins, J. M. (1986). A new approach to causal inference in mortality studies with a sustained exposure period. Mathematical Modelling 7, 1393-1512.

Imbens, G. W. and Angrist, J. D. (1994). Identification and estimation of local average treatment effects. Econometrica 62(2), 467-475.

Chernozhukov, V. et al. (2018). Double/debiased machine learning for treatment and structural parameters. Econometrics Journal 21(1), C1-C68.

Lu, B. (2005). Propensity score matching with time-dependent covariates. Biometrics 61(3), 721-728.

Robins, J. M. (1992). Estimation of the time-dependent accelerated failure time model in the presence of confounding factors. Biometrika 79(2), 321-334.

Examples

d <- data.frame(
  y = c(1, 2.2, 1.7, 3.1, 2.8, 3.9),
  t = c(0, 0, 0, 1, 1, 1),
  w = c(1.2, 2, 1.5, 1.1, 3, 1.4)
)
Ate(d, "y", "t", "w")
#> $ate
#> [1] 1.405957
#> 
#> $se
#> [1] 0.6454726
#> 
iv <- list(z = c(0, 0, 0, 0, 1, 1, 1, 1), t = c(0, 0, 1, 0, 1, 1, 0, 1),
           y = c(1, 1.4, 3.1, 0.8, 3.3, 2.9, 1.2, 3.6))
Late(iv, "t", "y", "z")$late
#> [1] 2.35