Fit a meta-analysis model with explicit control over design assumptions

A flexible interface for controlling how nuisance parameters are handled for non-DiD study designs. By default, behaves identically to meta_did(). Users can independently control whether time trends and baseline imbalances are estimated or fixed to zero for RCT and pre-post studies.

Usage

meta_did_general(
  summary_data = NULL,
  individual_data = NULL,
  normalise_by_baseline = TRUE,
  robust_heterogeneity = FALSE,
  design_effects = FALSE,
  hierarchical_rho = TRUE,
  covariates = NULL,
  center_covariates = TRUE,
  priors = set_priors(),
  time_trend = c("pooled", "fixed_zero"),
  baseline_imbalance = c("estimated", "fixed_zero"),
  method = c("sample", "optimize"),
  chains = 4L,
  iter_warmup = 1000L,
  iter_sampling = 1000L,
  seed = NULL,
  allow_no_did = FALSE,
  ...
)

Arguments

summary_data

A data frame with one row per study containing summary statistics. Must include columns study_id and design. See validate_summary_data() for the full column specification per design. Valid designs: "did", "did_change", "rct", "pp".

individual_data

A data frame in long format with one row per observation. Must include columns study_id, design, group, time, and value. Valid designs: "did", "rct", "pp". No study_id may appear in both summary_data and individual_data.

normalise_by_baseline

Logical. If TRUE (default), all means and SDs are divided by each study's pre-treatment control mean (or the grand mean for change-only studies), placing outcomes on a common fractional scale.

robust_heterogeneity

Logical. If TRUE, study-level treatment effects are drawn from a Student-t distribution rather than a normal, providing robustness to outlier studies. The degrees-of-freedom parameter is estimated with the prior specified in priors$nu.

design_effects

Logical. If TRUE, additive offsets on the population treatment effect mean are estimated for RCT and Pre-Post studies relative to DiD (the reference). Useful for testing whether designs yield systematically different effect estimates.

hierarchical_rho

Logical. If TRUE (default), the pre-post correlation is modelled hierarchically across studies. Studies with a reported correlation inform the population distribution; studies without one have their correlation imputed.

covariates

An optional one-sided formula specifying study-level covariates for meta-regression on the treatment effect (e.g., ~ dose + year). The named columns must be numeric and present in both summary_data and individual_data (whichever are provided). For individual-level data, covariate values must be constant within each study. Default NULL (no meta-regression).

center_covariates

Logical. If TRUE (default), covariates are mean-centered across all studies before fitting. This ensures that treatment_effect_mean is the population treatment effect at the average covariate values. Set to FALSE to use raw covariate values, in which case treatment_effect_mean is the effect when all covariates equal zero. The covariate coefficients (beta_cov) have the same interpretation regardless of centering: the change in expected treatment effect per unit increase in the covariate.

priors

A did_priors object from set_priors(). Controls the prior distributions on all population-level parameters.

time_trend

How to handle the time trend $\beta_i$ for non-DiD studies. One of:

"pooled": (Default) Estimate study-level time trends with a hierarchical prior shared across designs. Information from DiD studies informs the RCT and pre-post time trends. This is the same behaviour as meta_did().
"fixed_zero": Fix $\beta_i = 0$ for RCT and pre-post studies. For pre-post studies, this means the pre-post change is attributed entirely to treatment. For RCT studies, the reparameterised time trend correction is bypassed.

baseline_imbalance

How to handle the baseline difference $\gamma_i$ between treatment and control groups for RCT studies. One of:

"estimated": (Default) Estimate $\gamma_i$, borrowing information from DiD studies when baseline-normalised. This is the same behaviour as meta_did().
"fixed_zero": Fix $\gamma_i = 0$, assuming randomisation eliminates baseline imbalances. This is the standard RCT assumption.

method

Inference method. "sample" (default) runs full MCMC via Stan's HMC-NUTS sampler and returns a posterior distribution. "optimize" finds the maximum a posteriori (MAP) estimate via L-BFGS and is substantially faster, but returns only a point estimate with no uncertainty quantification.

chains

Number of MCMC chains. Ignored when method = "optimize". Default 4.

iter_warmup

Number of warmup iterations per chain. Ignored when method = "optimize". Default 1000.

iter_sampling

Number of sampling iterations per chain. Ignored when method = "optimize". Default 1000.

seed

Integer random seed for reproducibility. Default NULL.

allow_no_did

Logical. If FALSE (default), meta_did() will stop with an error when no DiD studies are present, because the treatment effect is not identified from the data without the double-difference structure. Set to TRUE to override this check if you understand the limitation (the posterior will be prior-driven).

...

Additional arguments passed to the underlying CmdStanModel method: $sample() when method = "sample" (e.g., parallel_chains, adapt_delta) or $optimize() when method = "optimize" (e.g., algorithm, iter).

Value

A meta_did_fit object, identical in structure to the return value of meta_did().

Details

DiD studies always estimate both time trends and baseline differences regardless of these settings, since they provide the identifying information for these parameters.

Examples