Summary

This document details how tidyhte constructs estimates of heterogeneous treatment effects. It will highlight a variety of the features of the package and discuss the mathematics which undergird them.

After a brief introduction to the methods of HTE estimation of Kennedy (2020), the structure will generally follow the estimation API of tidyhte: it will begin by discussing the creation of cross-validation folds, then highlight nuisance function estimation, proceed to the construction of “pseudo outcomes”, a concept from Kennedy (2020), and conclude by demonstrating the calculation of a few varieties of Quantities of Interest: the actual statistics which are desired by the end-user.

Preliminaries

penguins <- within(penguins, {
  id <- 1:n
  propensity_score <- 0.5
  treatment <- rbinom(n, 1, propensity_score)
  tau <- 0.2 * (species == "Gentoo") - 0.2 * (species == "Adelie") + rnorm(n, sd = 0.05)
  food_consumed_g <- rnorm(n, 500, 5) * (1 + tau * treatment)
  body_mass_g <- body_mass_g * (1 + tau * treatment)
})

Recipe API

In order to build up definitions around how HTE should be estimated, tidyhte provides an API to progressively build up a configuration object. A basic_config function creates a bare-bones configuration consisting of only linear models for the respective nuisance functions and a number of diagnostics.

cfg <- basic_config() %>%
    add_known_propensity_score("propensity_score") %>%
    add_outcome_model("SL.glmnet", alpha = c(0.0, 1.0)) %>%
    add_moderator("Stratified", species, island, sex, year) %>%
    add_moderator("KernelSmooth", bill_length_mm, bill_depth_mm, flipper_length_mm) %>%
    add_vimp(sample_splitting = FALSE)

Since the subject of interest is an experimental intervention, the propensity score is known. Using this known propensity score provides unbiasedness for many quantities of interest (although this may leave some efficiency on the table CITE). In this case, in addition to the (default) linear model included in the SuperLearner ensemble, we add an elastic-net regression (Zou and Hastie 2005). We sweep over a variety of mixing parameters between LASSO and ridge, throwing each of these models into the ensemble. This means that SuperLearner will perform model selection and averaging to identify the best hyper-parameter values (Van der Laan, Polley, and Hubbard 2007). Furthermore, we define all of the moderators of interest and how their results should be collected and displayed. Discrete moderators will just take stratified averages at each level of the moderator, while continuous moderators will use local polynomial regression (Fan and Gijbels 2018; Calonico, Cattaneo, and Farrell 2019). Finally, we add a variable importance measure from Williamson et al. (2021).

After the configuration is completed, it can be attached to the dataset.

penguins %<>% attach_config(cfg)

Cross-validation

The first step in an analysis of heterogeneous treatment effects following this procedure is to define how to construct splits to be used for cross-validation. tidyhte accomplishes this by using blocking methods to construct lower-variance splits than a purely randomized splitting procedure would entail. Existing methods for generating splits often provide options for stratifying based on a binary outcome to ensure there is variance in the outcome in all splits, even when the outcome is very sparse. For instance, SuperLearner::SuperLearner.CV.control provides such an option.

The appropriate function in tidyhte is make_splits. This function takes in a dataframe and determines a set of splits which accord with the provided unit identifier and number of splits. If any covariates are provided to this function, it will include them in a blocking design for constructing randomized splits using the methods of Higgins, Sävje, and Sekhon (2016) as implemented in the quickblock package.

There are a few relevant methodological notes about how this blocking for cross-validation strata works. In short, blocks are constructed which are sized to be at least as large as the number of splits to use. These blocks are constructed to minimize the within-block distance between units (where distances are Euclidean based on the provided covariates). When more than one row shares an identifier, blocking is performed on the average covariate value within each identifier. Rows with the same identifier are then assigned to the same split. Precise details on the construction of blocks may be found in higgins2016improving. Within each block, a vector of split IDs is constructed which has a marginal distribution as close to the uniform distribution over splits as possible (up to integer division errors). This set of split IDs is then randomly permuted within blocks.

Consistent with tidy semantics, the original dataframe is returned, but with the addition of a column .split_id representing these newly constructed splits, and a few attributes for bookkeeping (e.g. the column name of the identifier). Since the object returned is the same as what was passed in, this makes for easy chaining of commands using dplyr.

penguins %<>% make_splits(id, species, sex, flipper_length_mm, .num_splits = 3)

## `num_splits` must be even if VIMP is requested as a QoI. Rounding up.

## `quickblock` is not installed, so falling back to un-stratified CV.

## Dropped 11 of 344 rows (3.2%) through listwise deletion.

Note that tidyhte gracefully handles missing data via listwise deletion. More advanced imputation methods are not yet supported.

Nuisance function estimation

Estimation of nuisance functions such as the propensity score and the outcome regression are typically handled by the SuperLearner library. Specifying a full array of models with diverse hyperparameters is much simplified through the tidyhte API. To specify a cross-validated learner using SuperLearner syntax requires substantially more boilerplate:

learners <- create.Learner(
    "SL.glmnet",
    tune = list(
        alpha = c(0.05, 0.15, 0.2, 0.25, 0.5, 0.75)
    ),
    detailed_names = TRUE,
    name_prefix = paste0("SLglmnet")
)

CV.SuperLearner(label, covariates, SL.library = learners$names)

In contrast, the tidyhte Recipe API requires only the following one line:

add_outcome_model(cfg, "SL.glmnet", alpha = c(0.0, 0.25, 0.5, 0.75, 1.0))

penguins %<>% produce_plugin_estimates(
  # outcome
  food_consumed_g,
  # treatment
  treatment,
  # covariates
  species, island, sex, year, bill_length_mm, bill_depth_mm, flipper_length_mm
)

Pseudo-outcome construction

Once nuisance functions are estimated, it is simply a matter of combining these results together into the appropriate pseudo-outcome. For typical HTE estimation, the pseudo-outcome of interest is the one analyzed by Kennedy (2020): the uncentered influence function of the average treatment effect (Robins, Rotnitzky, and Zhao 1995).

penguins %<>% construct_pseudo_outcomes(food_consumed_g, treatment)

Quantities of interest

Finally, it comes to the most glamorous part of the analysis, when effects are estimated and put into charts.

The design of tidyhte chooses to leave the charting to the end-user, but merely returns a tidy tibble with all of the requested quantities of interest. The package focuses on a few types of quantities of interest: - Marginal Conditional Average Treatment Effects (MCATEs): the standard Conditional Average Treatment Effects (CATEs) in the literature, in which all covariates except one are marginalized over, providing one average effect for one level of a single covariate. This is in contrast to a “Partial” CATE, in which other variables are “controlled for” in some way. This does not provide a satisfying causal interpretation without assumptions of joint randomization of treatment and covariates. - Variable Importance (VIMP): Using Williamson et al. (2021), tidyhte calculates how much each moderator contributes to the overall reduction in mean-squared-error in a joint model of the heterogeneous effects. - Diagnostics: A wide variety of diagnostics are provided for all models fit as part of the tidyhte estimation process. These include single-number summaries like mean-squared-error or AUC, as well as entire receiver operating characteristic curves and coefficients in the SuperLearner ensembles.

penguins %>%
  estimate_QoI(species, island, sex, year, bill_length_mm, bill_depth_mm, flipper_length_mm) ->
  results

The resulting tibble provides looks like the following:

results

The estimand column denotes the class of Quantities of Interest to be estimated, and include such values as “MCATE” and “VIMP”. The term column denotes the covariate being referred to (if relevant) in the results. For example, a calculated MCATE refers to a particular covariate, indicated by this column. The columns value and level refer to quantities which indicate a more granular division within a particular term, with the former representing a numeric value while the latter indicates a categorical one. For example, in the case of the MCATE, if the covariate is discrete, the value at which the MCATE was calculated would be in the level column, while if it were continuous, it would be in the value column. Each of these columns is, therefore, type-stable. The final two columns for estimate and std_error are self explanatory. The details on their calculation lie with the particular Quantity of Interest requested.

filter(results, estimand == "MCATE", is.na(value)) %>%
ggplot(aes(level, estimate)) +
geom_point() +
geom_linerange(aes(ymin = estimate - 1.96 * std_error, ymax = estimate + 1.96 * std_error)) +
geom_hline(yintercept = 0, linetype = "dashed") +
coord_flip() +
facet_wrap(~term, scales = "free_y")

filter(results, estimand == "MCATE", is.na(level)) %>%
ggplot(aes(value, estimate)) +
geom_line() +
geom_ribbon(
  aes(ymin = estimate - 1.96 * std_error, ymax = estimate + 1.96 * std_error),
  alpha = 0.5
) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_x_continuous("Covariate value") +
scale_y_continuous("CATE") +
coord_flip() +
facet_wrap(~term, scales = "free_y")

filter(results, estimand == "SL risk") %>%
ggplot(aes(reorder(term, estimate), estimate)) +
geom_point() +
geom_linerange(
  aes(ymin = estimate - 1.96 * std_error, ymax = estimate + 1.96 * std_error),
  alpha = 0.5
) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_x_discrete("") +
scale_y_continuous("Risk") +
facet_wrap(~level) +
coord_flip()

filter(results, estimand == "SL coefficient") %>%
ggplot(aes(reorder(term, estimate), estimate)) +
geom_point() +
geom_linerange(
  aes(ymin = estimate - 1.96 * std_error, ymax = estimate + 1.96 * std_error),
  alpha = 0.5
) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_x_discrete("") +
scale_y_continuous("Coefficient") +
facet_wrap(~level) +
coord_flip()

filter(results, estimand == "VIMP") %>%
ggplot(aes(reorder(term, estimate), estimate)) +
geom_point() +
geom_linerange(
  aes(ymin = estimate - 1.96 * std_error, ymax = estimate + 1.96 * std_error),
  alpha = 0.5
) +
geom_hline(yintercept = 0, linetype = "dashed") +
scale_x_discrete("") +
scale_y_continuous("Reduction in R²") +
coord_flip()

Replicating analyses

One of the biggest benefits of tidyhte is its ability to repeat similar analyses.

As an example, to perform the same analysis used above on a different outcome would require the following code:

penguins %<>% produce_plugin_estimates(
  # outcome
  body_mass_g,
  # treatment
  treatment,
  # covariates
  species, island, sex, year, bill_length_mm, bill_depth_mm, flipper_length_mm
) %>%
construct_pseudo_outcomes(body_mass_g, treatment) %>%
estimate_QoI(
  species, island, sex, year, bill_length_mm, bill_depth_mm, flipper_length_mm
) -> results_mass

All the same quantities can be easily plotted for this outcome as well, and results may be joined together conveniently.

results_all <- bind_rows(
  results %>% mutate(outcome = "food_consumed_g"),
  results_mass %>% mutate(outcome = "body_mass_g")
)

By allowing the user to flexibly compose HTE estimators, it drastically reduces the amount of work necessary for a typical HTE analysis which by nature tends to involve multiple moderators, models and outcomes.

Conclusion

This paper has introduced the concepts underlying the tidyhte package and given examples as to how the package can be used. In general, the package is written in a sufficiently general way that some features can be added with relatively little work. For instance, adding new plugin models is as simple as providing some information on configuration as well as standardizing train / predict methods. This is similarly true for providing new ways to summarize CATEs for plotting, which is handled in much the same way. This makes it relatively easy for methodologists to fit in their preferred fitting methods and thereby extending tidyhte’s functionality.

Acknowledgements

We gratefully acknowledge the collaboration with the US 2020 Facebook and Instagram Election Project for being a testbed for the initial versions of tidyhte, particularly Pablo Barberá at Meta. We received no financial support for this software project.

References