Instrumental-variables regression on a database backend

Description

Estimates a linear IV model via exact sufficient statistics on a database backend. The current implementation supports three exact strategies:

• strategy = “moments” for models without absorbed fixed effects.

• strategy = “compress” for repeated no-FE design rows.

• strategy = “demean” for one or more absorbed fixed effects.

Multi-way fixed effects use exact alternating projections when there are more than two fixed effects, or when two-way fixed effects are unbalanced or weights are supplied.

Usage

dbivreg(
  fml,
  conn = NULL,
  table = NULL,
  data = NULL,
  path = NULL,
  weights = NULL,
  vcov = c("iid", "hc1"),
  strategy = c("auto", "moments", "demean", "within", "compress", "mundlak"),
  compress_ratio = NULL,
  compress_nmax = 1e+06,
  cluster = NULL,
  ssc = c("full", "nested"),
  sql_only = FALSE,
  data_only = FALSE,
  drop_missings = TRUE,
  verbose = getOption("dbreg.verbose", FALSE),
  ...
)

Arguments

fml	A fixest-style IV formula. Use y ~ x1 + x2 | d ~ z for models without absorbed fixed effects, and y ~ x1 + x2 | fe1 + fe2 + fe3 | d ~ z when fixed effects are present. The IV block must always be the final pipe-separated block. Regressors and instruments can include interactions and fixest-style i() terms (see sql_model_matrix()).
conn	Database connection, or NULL to create an ephemeral DuckDB connection.
table, data, path	Mutually exclusive data source arguments with precedence table > data > path.
weights	Optional character string naming an observation-weight column.
vcov	Character string or one-sided clustering formula. Character options are “iid” and “hc1”.
strategy	Character string selecting the IV backend. Supported values are “auto” (default), “moments”, “demean” (alias “within”), “compress”, and “mundlak”. The latter currently errors because it is not yet implemented for IV. As in dbreg(), “auto” chooses “compress” when the data compress well (see compress_ratio and compress_nmax), and otherwise “moments” (no FE) or “demean” (FE). Since IV compression does not yet support fixed effects, models with FE always use “demean”.
compress_ratio, compress_nmax	Thresholds (compression ratio and maximum compressed rows) that decide whether strategy = “auto” chooses “compress”. Same defaults and meaning as in dbreg(); see its Acceleration Strategies section.
cluster	Optional clustering variable, supplied as a one-sided formula or character string.
ssc	Character string controlling the small-sample correction for clustered standard errors. Options are “full” (default) or “nested”. With “nested”, fixed effects nested within the cluster variable are excluded from K, matching the default behavior of fixest::feols. See dbreg() for details.
sql_only	Logical indicating whether to return only the underlying SQL query.
data_only	Logical indicating whether to return only the aggregated sufficient statistics.
drop_missings	Logical indicating whether incomplete cases should be dropped.
verbose	Logical indicating whether progress messages are printed.
…	Additional unused arguments.

Value

A list of class “dbivreg” and “dbreg”.