Skip to contents

Fits a difference-in-differences (DiD) model comparing the yearly outcome series of a treated institution set against a control set, before and after intervention_year. Group membership is resolved via the institution-attribution columns from sm_affiliation_match() / sm_attribute_institution().

Usage

sm_did(
  corpus,
  treated,
  control,
  intervention_year,
  outcome = c("count", "share_q1", "cnci", "leadership"),
  family = NULL,
  lag = NULL,
  institution_aware = FALSE,
  lead_role = c("corresponding", "first", "last", "first_or_last", "any_lead"),
  shared = c("keep", "exclude", "lead"),
  min_cell = NULL,
  call = rlang::caller_env()
)

# S3 method for class 'sm_did'
print(x, ...)

# S3 method for class 'sm_did'
summary(object, ...)

# S3 method for class 'sm_did'
autoplot(object, ...)

Arguments

corpus

An sm_corpus with institution-attribution columns.

treated, control

Character vectors of institution labels (matched against institution_name / institution_match / institution_id). The two may not share a label.

intervention_year

Year the intervention takes effect: one whole number such as 2020, or text holding one.

outcome

One of "count", "share_q1", "cnci", "leadership" (see sm_its()).

family

Optional GLM family; defaults to Gaussian (DiD is an additive contrast). Override for, e.g., a Poisson count model; a Poisson fit warns when the counts are overdispersed.

lag

Citation-maturity lag in years (default 2), as in sm_its(): for "cnci" the most recent lag years are citation-immature and excluded.

institution_aware

For outcome = "leadership": if TRUE, each group's share counts the works whose author in the role lead_role belongs to the group's own institutions (the treated series uses institutions = treated, the control series institutions = control, as in sm_its()). FALSE (default) counts the works with a corresponding author.

lead_role

The author role that leads a work, as in sm_its(), for outcome = "leadership" and for shared = "lead". For the leadership outcome, roles other than "corresponding" need institution_aware = TRUE.

shared

What to do with works that belong to both groups, such as works co-authored by a treated and a control institution: "keep" (default) keeps them in both series, "exclude" leaves them out of both, and "lead" puts each in the series of the group whose institution holds its author in the role lead_role (a work led from both groups or from neither leaves both series).

min_cell

For outcome = "cnci": NULL (default) or a positive integer, passed to sm_metric_fnci() when the impact is derived from citation counts, as in sm_its().

call

Caller environment for error reporting.

x

An sm_did object.

...

Ignored.

object

An sm_did object.

Value

An sm_did S3 object with components: model, did_estimate (a one-row tibble: estimate, std.error, conf.low, conf.high, p.value), series (group-by-year tibble: year, group, value, n_works), dropped_years (tibble: group, year; the group-years left out for lack of a usable value), provisional_years (integer; the citation-immature years excluded for "cnci"), shared_works (tibble: work_id, series; the works of both groups and where they went: "both", "treated", "control" or "none"), shared (the policy), leadership (the leadership definition, as in sm_its(), with institutions the labels of each group when institution_aware; NULL for the other outcomes), impact_source (where the impact values of "cnci" come from, as in sm_its(); NA for the other outcomes), and metadata (outcome, intervention_year, family, lag).

print returns x invisibly.

summary returns the DiD estimate tibble.

autoplot returns a ggplot (parallel trends + divergence).

Details

Each group's yearly series is built as in sm_its(). For "count" a group's years without works count 0, over every year from the first to the last publication year of the corpus. For the other outcomes a year in which a group has works but no usable value is dropped from that group's series, with a warning, and listed in dropped_years. For "cnci" the citation-immature years are then excluded from both series, by the same rule as in sm_its(), and listed in provisional_years; up to scimapR 0.4.0 sm_did() kept them. Each group needs 4 or more years, with at least one before intervention_year and one from it on; otherwise sm_did() aborts. The confidence interval uses the reference distribution of the p-value (t with the residual degrees of freedom for the default Gaussian family); no correction for autocorrelation is applied.

A work belongs to a group when one of its authorships carries one of the group's labels, so a work co-authored by both groups belongs to both. Such shared works pull the two series together and the estimate toward 0. sm_did() warns when there are shared works and shared is not set, and records them in shared_works. For "leadership" the denominators leave out the works whose lead is unknown, as in sm_its().

See also

Examples

corpus <- sm_example_corpus(n_works = 120, seed = 1)
# Tag two institution groups for illustration: each work belongs to one
a <- corpus$authorships
group <- rep(c("Inst A", "Inst B"), length.out = nrow(corpus$works))
corpus$authorships$institution_name <-
  group[match(a$work_id, corpus$works$work_id)]
did <- sm_did(corpus, treated = "Inst A", control = "Inst B",
              intervention_year = 2020, outcome = "count")
did
#> 
#> ── <sm_did> ────────────────────────────────────────────────────────────────────
#> Outcome: count Family: gaussian
#> Intervention year: 2020
#> 
#> DiD estimate: 2.4 [-2.0366, 6.8366] (p = 0.27)