TTE

Design and Analysis Tools for Target Trial Emulation

TTE is an R package providing design and analysis tools for target trial emulation with longitudinal observational data. Its functions cover person-period data checks, sequentially nested trial construction, inverse probability weighting, diagnostics, weighted pooled discrete-time outcome models, standardization, and weighted Kaplan-Meier and Aalen-Johansen estimation.

This companion repository accompanies TTE version 1.1.2. It provides the worked-example scripts and a PDF version of the accompanying tutorial and methodological guide.

• CRAN package: https://CRAN.R-project.org/package=TTE
• CRAN reference manual: https://cran.r-project.org/web/packages/TTE/TTE.pdf
• GitHub repository: https://github.com/nomahi/TTE/
• Tutorial and Methodological Guide: http://arxiv.org/abs/2608.01625

Install the stable CRAN release with:

install.packages("TTE")

The package includes two pairs of fully synthetic datasets. They contain no actual participant records from the motivating studies.

An active-comparator new-user target trial emulation comparing initiation of an SGLT2 inhibitor with initiation of a DPP-4 inhibitor among older adults with type 2 diabetes. The primary outcome is all-cause death.

Sequentially nested new-user trials comparing ARB and CCB treatment strategies among people with chronic kidney disease. The primary outcome is heart-failure hospitalization, and death is a competing event.

The full tutorials estimate all analysis weights from the observed synthetic data. This compact example shows the principal ITT workflow.

library("TTE")

data(SGLT2_baseline)
data(SGLT2)

baseline <- SGLT2_baseline
long <- SGLT2

# Baseline stabilized treatment weight
w_a <- est_wt(
  A ~ age + I(age^2) + female + bmi +
    hba1c + I(hba1c^2) + egfr + proteinuria +
    prior_heart_failure + prior_stroke +
    recent_hospitalization + trial_period,
  data = baseline,
  type = "treatment",
  stabilize = TRUE,
  truncate = c(0, 1)
)

# Map the baseline treatment weight to all person-month records
key_baseline <- paste(baseline$id, baseline$trial, sep = ":")
key_long <- paste(long$id, long$trial, sep = ":")

long$w_a <- weights(w_a, which = "untruncated")[
  match(key_long, key_baseline)
]

# Stabilized loss-to-follow-up weight
w_c <- est_wt(
  stay_ltfu ~ A + splines::ns(time, df = 3) +
    age + female + bmi + hba1c + egfr + proteinuria +
    prior_heart_failure + prior_stroke +
    recent_hospitalization + trial_period,
  numerator = stay_ltfu ~
    A + splines::ns(time, df = 3) + trial_period,
  data = long,
  type = "censoring",
  id = id,
  trial = trial,
  time = time,
  cumulative = TRUE,
  lag = 1,
  stabilize = TRUE,
  truncate = c(0, 1)
)

# Final ITT weight, truncated after combining components
w_itt <- combine_wt(
  long$w_a,
  w_c,
  truncate = c(0.01, 0.99),
  normalize = "none"
)

long$w_itt_est <- weights(w_itt)

# Weighted pooled discrete-time outcome model
fit <- discsurvreg(
  Y_death ~ A + splines::ns(time, df = 3) + trial_period,
  data = long,
  id = id,
  weights = w_itt_est,
  family = quasibinomial(link = "cloglog"),
  var_method = "standard"
)

confint(fit, parm = "A", eform = TRUE)

# Standardized absolute risks
std <- std_tte(
  fit,
  data = baseline,
  treatment = A,
  time = time,
  times = 0:59,
  values = c(0, 1),
  labels = c("DPP-4i", "SGLT2i")
)

summary(std, horizon = 60)
plot(std, measure = "risk")

The public scripts provide complete ITT and PP workflows. The main analyses estimate treatment and censoring weights from the observed synthetic data rather than relying on the reference weights bundled with the datasets.

• Example code

This example covers baseline IPTW, loss-to-follow-up IPCW, weighted pooled discrete-time modeling, standardized absolute risks, weighted Kaplan-Meier curves, and per-protocol artificial-censoring analyses.

• Example code

This example extends the workflow to sequentially nested trials, cause-specific outcome models, standardized cumulative incidence functions, weighted Aalen-Johansen curves, and per-protocol estimation.

Person-period interval indices are stored from zero internally, whereas public summaries use elapsed follow-up time. Thus, times = 0:59 represents 60 follow-up intervals, and

summary(std, horizon = 60)

reports a model-based standardized estimate after 60 follow-up intervals.

For a weighted Kaplan-Meier or Aalen-Johansen curve, the requested summary time must also lie within the follow-up represented in the fitted curve object. For example,

summary(km, time = 60)

reports the estimate after 60 follow-up intervals only when the curve contains follow-up through that time.

Calling

summary(km)

without specifying time returns the estimate at the latest available follow-up represented in the curve.

This distinction is particularly relevant for per-protocol analyses. Artificial censoring for deviation from the assigned treatment strategy can shorten the observed per-protocol risk set, even when model-based standardization is evaluated on a longer prespecified time grid. Consequently, a standardized estimate may be available at a later horizon than a direct weighted Kaplan-Meier or Aalen-Johansen estimate. Late per-protocol estimates should therefore be interpreted together with the corresponding risk-set and weight diagnostics.

Weighted Kaplan-Meier and Aalen-Johansen curves are displayed as right-continuous step functions.

TTE supplies computational tools; it does not automatically identify a causal effect. Applied analyses still require defensible definitions of:

• eligibility criteria;
• treatment strategies;
• assignment procedures;
• time zero;
• induction or grace periods;
• outcomes and competing events;
• follow-up and censoring;
• confounders;
• estimands; and
• analysis populations.

Users must assess consistency, exchangeability, positivity, model specification, missing data, measurement timing, and the clinical meaning of the intervention strategies.

For publication-quality uncertainty intervals for standardized risks or weighted curves, an individual-cluster bootstrap that re-estimates all treatment, censoring, adherence, and outcome models within each bootstrap sample is generally preferable to treating estimated weights as fixed.

• Noma H, Goto A, Sugimoto T, Sunada H, Oda F, Maeda M, Fukuda H. Real-world effectiveness of SGLT2 inhibitors in adults aged 75 years or older: a target trial emulation. Age and Ageing. 2026; 55, afag246. https://doi.org/10.1093/ageing/afag246

• Noma H, Kurita N, Fukuma S, Fujisawa T, Oda F, Maeda M, Fukuda H. Heart failure and renal outcomes with angiotensin receptor blockers compared with calcium channel blockers in patients with chronic kidney disease: a target trial emulation. Heart. 2026. https://doi.org/10.1136/heartjnl-2026-328193

Hisashi Noma
The Institute of Statistical Mathematics
ORCID: https://orcid.org/0000-0002-2520-9949