Package 'mlumr'

Description

Generate quasi-Monte Carlo integration points using Sobol sequences and a Gaussian copula to account for correlations between covariates in the AgD.

Usage

add_integration(
  data,
  n_int = 64,
  cor = NULL,
  cor_adjust = NULL,
  verbose = TRUE,
  ...
)

Arguments

Details

The correlation structure is assumed to transport. Aggregate data report marginal summaries only, so with cor = NULL the within-row correlation is estimated from the IPD and applied to every comparator row, as in ML-NMR (Phillippo et al. 2020). The assumption is untestable from the data; supply cor from an external source to vary it, and use check_integration() to confirm the realized moments and correlations.

cor_adjust maps the covariate-scale correlation onto the Gaussian copula: the Spearman map is exact for continuous margins, the Pearson map holds for Gaussian margins, and pairs involving a binary margin use prevalence-independent heuristics. A nonbinary discrete margin (a count or an ordered category) is treated as continuous and its realized association need not match the target; add_integration() warns when it sees one. "none" passes a latent Gaussian-copula matrix through unchanged.

Value

An mlumr_data object with integration points added

Examples

## Not run: 
dat <- add_integration(
  dat,
  n_int = 64,
  x1 = distr(qnorm, mean = x1_mean, sd = x1_sd),
  x2 = distr(qbern, prob = x2_mean)
)

## End(Not run)

Description

Computes the Deviance Information Criterion using the variance-based effective-parameters formula from Gelman et al. (2004): pD = 0.5 * Var(D). This is more stable than the plug-in alternative for multimodal posteriors.

Usage

calculate_dic(object)

Arguments

Details

DIC is retained for backward compatibility and rough comparison. For principled Bayesian model comparison, prefer calculate_loo() or calculate_waic() (Vehtari, Gelman, Gabry 2017).

Value

A list of class mlumr_dic with components DIC, pD, D_bar, n_obs and model.

Examples

## Not run: 
dic_spfa <- calculate_dic(fit_spfa)
dic_relaxed <- calculate_dic(fit_relaxed)

## End(Not run)

Description

Computes approximate leave-one-out cross-validation (PSIS-LOO, Vehtari, Gelman, Gabry 2017) using the pointwise log-likelihoods stored by the Stan models. Returns a loo object from the loo package.

Usage

calculate_loo(
  object,
  survival_unit = c("observation", "arm", "aggregate"),
  ...
)

Arguments

Details

Pareto-k diagnostics: values > 0.7 indicate observations for which the PSIS approximation is unreliable; the printed output flags these. Typical remedies are running more iterations or, for highly influential AgD rows, refitting without the offending observation to check sensitivity. Moment matching (loo::loo_moment_match()) needs the fitted model rather than a log-likelihood matrix, so moment_match is refused.

Value

An object of class psis_loo (see loo::loo()).

Note

AgD rows are treated as independent observations. Subgroup rows from one study share no clustering term, so their effective sample sizes are inflated and Pareto-k warnings understated; corroborate with prior_sensitivity() or by refitting without suspect rows.

Survival fits. The comparator enters as reconstructed pseudo-IPD, so the default pointwise unit is one pseudo-individual and the criteria are optimistic relative to leaving out the comparator arm. Set survival_unit = "arm" or "aggregate" to hold out whole comparator arms or all of the external evidence instead.

Examples

## Not run: 
loo_spfa <- calculate_loo(fit_spfa)
print(loo_spfa)

## End(Not run)

Description

Watanabe-Akaike Information Criterion (Watanabe 2010) based on the pointwise log-likelihoods. WAIC is asymptotically equivalent to LOO-CV; prefer calculate_loo() when Pareto-k is well-behaved.

Usage

calculate_waic(
  object,
  survival_unit = c("observation", "arm", "aggregate"),
  ...
)

Arguments

Value

An object of class waic (see loo::waic()).

Note

As with calculate_loo(), each AgD row is treated as an independent observation. WAIC will be optimistic for AgD rows that share a study (see the note on calculate_loo()). For survival fits the AgD pointwise unit is a reconstructed pseudo-individual, so WAIC is at the pseudo-individual level (see the survival note on calculate_loo()).

Examples

## Not run: 
waic_spfa <- calculate_waic(fit_spfa)

## End(Not run)

Description

Aggregate summary of the comparator treatment (nano-silver fluoride, NSF) for the count ML-UMR example; pair with caries_ipd. Not real patient data.

Usage

caries_agd

Format

A data frame with 1 row and 10 columns: study, treatment, n, count outcome r (total dmft) with exposure E (n children), and covariate summaries (age_mean/age_sd, gender_prop (a proportion), log_cfu_mean/log_cfu_sd).

Source

Simulated; see caries_ipd for the modeling basis.

Description

Simulated individual patient data for the index treatment (silver diamine fluoride, SDF) in a count-outcome ML-UMR example; pair with caries_agd. Not real patient data.

Usage

caries_ipd

Format

A data frame with 103 rows and 8 columns:

study

study label ("Ammar 2025")

treatment

index treatment label ("SDF")

subject

row identifier

age

age in years

gender

1/0 indicator, as coded in the source trial, which does not record which level is which. Only the covariate's distribution matters for the adjustment, so the labeling does not affect the estimand.

log_cfu

log baseline salivary S. mutans count, log(CFU/mL + 1)

dmft

decayed-missing-filled-teeth count (the count outcome)

exposure

Poisson offset, 1 per child. dmft is a whole-mouth count with no time at risk and no per-tooth denominator, so the rate is "dmft per child" and the offset is 1. The column is present because set_ipd requires exposure for the Poisson family; log(1) = 0, so it contributes nothing to the linear predictor. The comparator's E is the matching total, one unit per child.

log_cfu is bimodal: 9 of the 103 children (8.7\ detectable baseline count and so sit at exactly 0, and the rest are concentrated near 10.9 (SD 0.8). The comparator arm declares this covariate as a single normal distribution (log_cfu_mean 11.0, log_cfu_sd 0.7) because the source trial's comparator arm contains no such zeros. Integrating a normal over the comparator therefore puts almost no weight where the index arm's zero mode sits, so effects adjusted for log_cfu extrapolate over that part of the covariate space. This is a property of the source trial, not of the synthesis; treat it as a worked illustration of a covariate whose arms are not fully overlapping.

Details

The index and comparator arms come from the same trial, so both carry the study label "Ammar 2025". Splitting one randomized trial into a single-arm IPD source and a single-arm aggregate source is what makes this an unanchored example that still has a full-data comparison to be checked against: fitting the two arms together on the complete synthetic records gives the quantity the unanchored methods are trying to recover once one arm has been reduced to summaries. That is a data-reduction benchmark. It is not a randomized reference and not a known population causal effect. The generator draws treatment first and then synthesizes the covariates conditional on it, so the synthetic arms are not a fresh randomized assignment independent of the generated baseline variables; and the full-data estimate carries its own sampling error. Validating an estimator against a causal target would need a specified covariate distribution, potential-outcome mechanism, assignment rule and estimand, which these records do not provide. Here the treatment contrast itself was imposed: dmft was a balanced baseline characteristic in the source trial and the index arm's synthetic counts were thinned binomially to create an effect, so the benchmark recovers a contrast this package manufactured, not one the trial reported. Because the label is shared, combine_data warns that IPD and AgD come from the same study; that warning is expected here and is the honest reading of the data. It does not fire for psoriasis_ipd or ndmm_ipd, whose arms really do come from different trials.

Source

Simulated (not real patient data), generated with the synthpop package (sequential CART; Nowok, Raab and Dibben 2016, doi:10.18637/jss.v074.i11) from the silver-diamine-fluoride vs nano-silver-fluoride dental caries RCT (Ammar et al., BMC Oral Health 2025;25:945, CC BY 4.0; data Synapse syn43185346), preserving the covariate relationships; fidelity validated with the syntheticdata package. A plausible caries-arresting SDF effect was imposed for illustration (dmft was a balanced baseline characteristic in the source trial). See data-raw/simulate_external_data.R.

Description

Warns about divergent transitions, iterations that hit the maximum tree depth, chains that did not come back, split-Rhat above 1.01 or 1.05, and bulk or tail effective sample sizes below 400. A diagnostic the backend did not supply is reported as unavailable rather than read as clean. mlumr() runs this check after sampling; call it again on a stored fit to see the same verdict.

Usage

check_diagnostics(fit)

Arguments

Value

NULL, invisibly; called for its warnings and messages.

See Also

mlumr(), prior_sensitivity().

Description

In the relaxed model the comparator coefficients beta_comparator are informed only by the aggregate rows. With K covariates there are K + 1 comparator parameters, so at least K + 1 distinct aggregate rows are needed, and under an identity link the rows must also differ in every covariate direction.

Usage

check_identification(x, verbose = TRUE, link = NULL)

Arguments

Details

The subgroup mean profiles are centered, divided by the IPD covariate SDs and decomposed. cond_inv is the ratio of the smallest to the largest singular value and goes to 0 as the rows collapse onto a lower-dimensional set. eff_dim is the participation ratio of the squared singular values, the number of directions the rows effectively spread along, from 1 to K; it is 0 when the rows do not vary or cannot be decomposed. spread is the RMS distance of the rows from their center along the dominant direction, in IPD SDs; it supplies the absolute scale cond_inv lacks. For a normal identity-link model the subgroup means are the aggregate design and the screen flags cond_inv < 0.2 or spread < 0.05, which are package heuristics. For other links the integrated response also depends on each row's covariate distribution, so the geometry is descriptive only and flagged is NA unless there are too few rows. Reconstructed survival curves are refused, since a curve is not one scalar summary per row. Neither measure sees subgroup sizes or outcome precision, so confirm any verdict with the coefficient posterior and prior_sensitivity(). The subgroup-identification vignette works through the cases.

Value

Invisibly, a list with n_rows, n_distinct (rows that do not repeat another's integration grid), n_cov, n_rows_needed (K + 1), cond_inv, eff_dim, spread, singular_values, means (the scaled, centered subgroup mean matrix), diagnostic_scope ("identity" or "descriptive") and flagged.

See Also

mlumr() for model = "relaxed"; prior_sensitivity().

Examples

## Not run: 
dat <- add_integration(combine_data(ipd, agd), n_int = 64, ...)
check_identification(dat)

## End(Not run)

Description

Compare integration results at the current n_int against a doubled resolution to assess numerical accuracy. Large discrepancies indicate that n_int should be increased. Because the Sobol sequence is nested (the doubled set contains the current set), this current-vs-doubled difference is a convergence heuristic, not an error bound. Agreement between the two grids does not establish accuracy for rare discrete margins or for a final treatment-effect estimand.

Usage

check_integration(
  data,
  ...,
  cor = NULL,
  cor_adjust = NULL,
  check_joint = TRUE,
  verbose = TRUE
)

Arguments

Value

A list with marginals, a data frame of grid means and SDs at the current and doubled n_int against the declared targets, and verdict, whose entries are "stable" or "close" when a comparison met the heuristic, "review" when it did not, "partial" when the measured correlation pairs passed but some pair could not be measured, and "unavailable" when there was nothing finite to compare (a latent matrix under cor_adjust = "none" is never compared). With check_joint = TRUE and two or more covariates it also holds correlations, the pairwise correlations per AgD row on both grids, and correlation_pairs, which counts the pairs measured and names the rest. A binary margin's target SD is sqrt(p * (1 - p)) from the declared mean, and grid SDs are population SDs.

Description

Combine IPD and AgD for unanchored comparison

Usage

combine_data(ipd, agd)

Arguments

Value

An object of class mlumr_data

Examples

## Not run: 
dat <- combine_data(ipd, agd)

## End(Not run)

Description

Compare two or more mlumr_fit objects by DIC (default), LOO, or WAIC. For LOO/WAIC, loo::loo_compare() is used under the hood; the output is the standard loo_compare table. The pointwise scores are paired observation by observation, so the fits must be of the same data in the same row order; fits whose stored outcomes differ are refused. For DIC the return is a data frame ordered by DIC.

Usage

compare_models(
  ...,
  criterion = c("dic", "loo", "waic"),
  survival_unit = c("observation", "arm", "aggregate")
)

Arguments

Details

DIC is the default for backward compatibility and because it has no additional package dependencies. For principled Bayesian model comparison, LOO (Vehtari, Gelman, Gabry 2017) is preferred and requires the optional loo package.

Value

For "loo" / "waic": a compare.loo table from loo::loo_compare(). For "dic": a data frame (invisibly) with columns Model, DIC, pD, Delta_DIC.

Description

Compute conditional (individual-level) treatment effects at specific covariate values from a fitted ML-UMR model. Unlike marginal_effects(), which averages over a population's covariate distribution, conditional effects evaluate the treatment effect at a particular covariate profile.

Usage

conditional_effects(
  object,
  newdata = NULL,
  effect = "all",
  summary = TRUE,
  probs = c(0.025, 0.5, 0.975)
)

Arguments

Details

For SPFA models, the conditional link-scale treatment effect is constant across all covariate values because the shared beta cancels in the treatment contrast on the fitted link scale. However, risk difference (RD) and risk ratio (RR) still vary with covariates because they depend on absolute probability levels. For relaxed models, all conditional effects vary with covariate values because the index and comparator treatments have different regression coefficients.

For binomial, normal and Poisson the conditional link-scale effect is eta_index - eta_comparator; for survival the contrast is exponentiated (null 1) and labeled "EXP_ETA_CONTRAST" when the baseline shapes differ. A conditional effect is evaluated at one profile, so unlike marginal_effects() and predict.mlumr_fit() there is no averaging over a population and no gap between E[g^{-1}(eta)] and g^{-1}(E[eta]).

Value

A data frame. If summary = TRUE, contains columns profile, effect, mean, sd and quantile columns. If summary = FALSE, returns a single combined data frame of full posterior draws with a profile column indicating which covariate profile each draw belongs to.

See Also

marginal_effects() for population-averaged treatment effects; conditional_predict() for absolute predictions at specific profiles; predict.mlumr_fit() for population-level predictions.

Examples

## Not run: 
# Conditional effects at IPD covariate means (default)
conditional_effects(fit)

# At specific covariate values
conditional_effects(fit, newdata = data.frame(age = 60, sex = 1))

# Multiple profiles
profiles <- data.frame(age = c(50, 60, 70), sex = c(0, 0, 1))
conditional_effects(fit, newdata = profiles)

## End(Not run)

Description

Generate absolute predictions at specific covariate values for both treatments.

Usage

conditional_predict(
  object,
  newdata = NULL,
  type = NULL,
  summary = TRUE,
  probs = c(0.025, 0.5, 0.975)
)

Arguments

Value

A data frame with predictions for each treatment at each profile. For survival fits the summary has one row per profile, treatment, and time; with summary = FALSE the draws are rows and the prediction times are ⁠t_*⁠ columns.

See Also

conditional_effects() for covariate-conditional treatment effects; predict.mlumr_fit() for population-level predictions.

Examples

## Not run: 
conditional_predict(fit)
conditional_predict(fit, newdata = data.frame(age = 60, sex = 1))

## End(Not run)

Description

Bernoulli PMF

Usage

dbern(x, prob, log = FALSE)

Arguments

Value

Numeric vector

Default priors used by mlumr()

Description

These accessors return the current default priors used by mlumr(), tagged with ⁠$default = TRUE⁠ and the package version. For family = "normal" their scales are multiples of the IPD outcome SD wherever the parameter is in outcome units; see prior_normal(). prior_summary() prints the version so cross-release reproducibility is diagnosable: if a later release changes a default, fits produced with an older version will still carry the correct ⁠$version⁠ tag.

Usage

default_prior_intercept()

default_prior_beta()

default_prior_sigma()

default_prior_aux()

default_prior_smooth()

Details

default_prior_aux() and default_prior_smooth() apply to the survival family only. prior_aux is a half-normal(0, 2) on the shape/scale parameter(s) of parametric survival distributions (Weibull/Gompertz/gamma shape, log-normal sdlog, generalized-gamma shapes). prior_smooth is a half-normal(0, 1) on the random-walk smoothing SD of the M-spline / piecewise-exponential baseline hazard.

Value

A prior list (see prior_normal()).

Examples

default_prior_intercept()
default_prior_beta()
default_prior_sigma()

Description

Used to specify marginal distributions for covariates when adding integration points. Wraps an inverse CDF (quantile) function with its parameters.

Usage

distr(qfun, ...)

Arguments

Value

A list with class "mlumr_distr" containing the distribution specification

Examples

# Normal distribution
distr(qnorm, mean = 0, sd = 1)

# Bernoulli distribution with probability 0.3
distr(qbern, prob = 0.3)

Description

Density, distribution, and quantile functions for the gamma distribution, accepting either the native shape / rate (or scale) parameters or a mean and sd, which override them. Useful with distr() and add_integration(), where covariate moments come straight from a published baseline table. The moment parameterization uses shape = (mean / sd)^2 and rate = mean / sd^2.

Usage

qgamma(
  p,
  shape,
  rate = 1,
  scale = 1/rate,
  lower.tail = TRUE,
  log.p = FALSE,
  ...,
  mean,
  sd
)

pgamma(
  q,
  shape,
  rate = 1,
  scale = 1/rate,
  lower.tail = TRUE,
  log.p = FALSE,
  ...,
  mean,
  sd
)

dgamma(x, shape, rate = 1, scale = 1/rate, log = FALSE, ..., mean, sd)

Arguments

Value

A numeric vector, as the corresponding stats function.

See Also

distr(), add_integration(), qbern()

Examples

# Equivalent specifications
qgamma(0.5, mean = 65, sd = 8)
qgamma(0.5, shape = (65 / 8)^2, rate = 65 / 8^2)

Description

Returns ggplot2 layers drawing the observed Kaplan-Meier step curves (the index IPD and the reconstructed comparator pseudo-IPD) of a survival mlumr_data, colored by treatment so they line up with a survival prediction plot. The mlumr analogue of multinma's geom_km(), so a predicted-versus-observed figure is just plot(predict(fit, type = "survival")) + geom_km(data).

Usage

geom_km(
  data,
  treatments = NULL,
  population = NULL,
  marks = TRUE,
  linewidth = 0.4,
  ...
)

Arguments

Value

A list of ggplot2 layers (a step layer, plus a censoring-mark layer when marks = TRUE) to add to a plot with +. The layers carry the population each observed arm was measured in, so on a plot faceted by population each curve appears only in its own panel. A plot standardized to a newdata target therefore shows no observed curve, which is correct: no arm was observed in that population.

See Also

plot.mlumr_prediction()

Examples

## Not run: 
plot(predict(fit, type = "survival")) + geom_km(dat)
plot(predict(fit, type = "survival")) + geom_km(dat, population = "Comparator")

## End(Not run)

Description

Density, distribution, and quantile functions for the logit-normal distribution: the distribution of plogis(z) where z is normal with mean mu and standard deviation sigma. It is the natural marginal for a covariate reported as a proportion on ⁠(0, 1)⁠, such as percent body surface area.

Usage

dlogitnorm(x, mu = 0, sigma = 1, log = FALSE, ..., mean, sd)

plogitnorm(q, mu = 0, sigma = 1, ..., mean, sd)

qlogitnorm(p, mu = 0, sigma = 1, ..., mean, sd)

Arguments

Details

For convenience the distribution may be given by its mean and sd on the natural ⁠(0, 1)⁠ scale instead of mu and sigma on the logit scale. There is no closed form for that reparameterization, so mu and sigma are found numerically; supply mu / sigma directly if you have them.

Value

A numeric vector.

See Also

distr(), add_integration(), GammaDist

Examples

qlogitnorm(0.5, mean = 0.34, sd = 0.19)

Description

Place boundary and internal knots for the M-spline baseline hazard used by the flexible survival models (distribution = "mspline" or "pexp"). Knots are chosen from the pooled event/censoring times of the index IPD and the reconstructed comparator pseudo-IPD.

Usage

make_knots(data, n_knots = 7, type = c("quantile", "equal"))

Arguments

Details

The lower boundary knot is fixed at 0 (not the minimum delayed-entry time), so the cumulative hazard is anchored at H(0) = 0 and stays continuous with the backward constant-hazard extrapolation; delayed-entry times (> 0) are evaluated on the basis. The upper boundary knot is the maximum observed time across both data sources. The M-spline basis is normalized so that the baseline cumulative hazard equals 1 at the upper boundary; the hazard scale is carried by the model intercepts.

Value

A list with internal (internal knot locations), boundary (lower and upper boundary knots) and n_knots (the realized number of internal knots after dropping any that coincide with the boundaries).

See Also

mlumr() with distribution = "mspline".

Examples

## Not run: 
# Build a survival network, then choose M-spline baseline knots:
dat <- combine_data(index_ipd, comparator_agd)
knots <- make_knots(dat, n_knots = 5)
knots$boundary  # lower (0) and upper boundary knots
knots$internal  # internal knot locations

## End(Not run)

Description

Extract marginal treatment effects from a fitted ML-UMR model. For binomial: log odds ratio, risk difference, risk ratio. For normal: mean difference. For poisson: rate ratio. For survival: the hazard ratio (proportional-hazards distributions, labeled HR), the time ratio (accelerated-failure-time distributions with one shared shape and one shared coefficient vector, labeled TR), or the exponentiated linear-predictor contrast (EXP_DELTA_ETA, where neither of those holds); all natural-scale, null 1, like the poisson rate ratio. Plus the restricted-mean-survival-time difference (RMSTD, null 0) and the RMST ratio (RMSTR, null 1), both reported with the restriction time in a horizon column. For the time-varying log hazard ratio curve (null 0) use predict(type = "loghr").

Usage

marginal_effects(
  object,
  population = c("both", "index", "comparator"),
  effect = "all",
  summary = TRUE,
  probs = c(0.025, 0.5, 0.975),
  newdata = NULL,
  at_time = NULL
)

Arguments

Details

For survival proportional-hazards fits the scalar "hr" is always a marginal hazard ratio at one time, recorded in the at_time column: the t -> 0 limit under a shared baseline shape (where an SPFA fit's value coincides with the conditional hazard ratio, since the shared coefficients cancel), and the value at the first prediction time, or at at_time, under study-specific shapes. Hazard ratios are non-collapsible, so the marginal ratio is time-varying; predict(type = "loghr") gives the curve, and the RMST effects are collapsible within a population.

For accelerated-failure-time fits the scalar is exp(E_X[eta_index(X)] - E_X[eta_comparator(X)]). With one shared shape and SPFA coefficients it is a population time ratio (TR): every individual's survival time is accelerated by the same factor. With differing shapes or treatment-specific coefficients no single acceleration factor exists and it is labeled EXP_DELTA_ETA; use RMST effects or explicitly indexed survival quantiles there. Neither carries an evaluation time. See vignette("survival-outcomes", "mlumr").

For binomial fits "lor" is always a logit-scale marginal odds ratio, whatever the fitted link, so for a probit or cloglog fit it is on a different scale than naive() and stc() ⁠$estimate⁠.

Relaxed-model index-population estimands average beta_comparator over the IPD covariate distribution, outside the support it was identified on, so they are wider and more prior-sensitive than the comparator-population ones. marginal_effects() says so once per call; suppress with options(mlumr.quiet_relaxed_index = TRUE).

Value

A data frame. With summary = FALSE the raw posterior draws are returned as a plain data frame whose column names carry the effect scale (⁠lor_*⁠, ⁠rr_*⁠, ⁠delta_*⁠, ⁠hr_*⁠ / ⁠tr_*⁠ / ⁠exp_delta_eta_*⁠, ⁠rmst*⁠); with summary = TRUE the effect column names the measure. For survival, RMST rows carry a horizon column and the scalar rows an at_time column (attributes of the same names on the raw-draw frame).

See Also

predict.mlumr_fit() for absolute predictions; conditional_effects() for covariate-conditional effects at specific profiles; prior_sensitivity() to check how strongly the marginal effect depends on prior_beta.

Examples

## Not run: 
# All effect measures for both populations
marginal_effects(fit)

# Only the log odds ratio in the index population
marginal_effects(fit, population = "index", effect = "lor")

# Transport the effect to an external (e.g. jurisdiction-specific) population
marginal_effects(fit, newdata = target_population_covariates)

# Full posterior draws rather than summary statistics
marginal_effects(fit, summary = FALSE)

## End(Not run)

Description

Fit a Bayesian multilevel unanchored meta-regression model using individual patient data (IPD) and aggregate data (AgD). Supports binary, continuous, count, and time-to-event outcomes.

Usage

mlumr(
  data,
  model = c("spfa", "relaxed"),
  link = NULL,
  prior_intercept = default_prior_intercept(),
  prior_beta = default_prior_beta(),
  prior_sigma = default_prior_sigma(),
  distribution = NULL,
  prior_aux = NULL,
  prior_smooth = NULL,
  n_knots = 7L,
  knots = NULL,
  mspline_degree = NULL,
  aux_by = ".study",
  pred_times = NULL,
  rmst_horizon = NULL,
  n_rmst_grid = 100L,
  center = TRUE,
  qr = FALSE,
  chains = 4,
  iter = 2000,
  warmup = 1000,
  seed = NULL,
  adapt_delta = 0.95,
  max_treedepth = 15,
  refresh = 200,
  engine = NULL,
  verbose = TRUE,
  prior_beta_comparator = NULL,
  prior_aux2 = NULL,
  ...
)

Arguments

Details

The model assumes that all AgD rows come from the same comparator treatment and that, conditional on covariates, there is no between-study heterogeneity. No random effects for study-level heterogeneity are included.

AgD scale (family = "normal"). The AgD likelihood is y_agd ~ normal(E[exp(eta)], se_agd) under link = "log" and y_agd ~ normal(E[eta], se_agd) under link = "identity"; in both cases set_agd() expects outcome_mean and outcome_se on the arithmetic scale.

The comparator population is the size-weighted mixture of its aggregate rows. Comparator-population predictions weight each row by the population it represents: n_agd for binomial, outcome_n for normal (required for more than one row) and E_agd for poisson. These are mixing weights, separate from the likelihood's precision weights, so splitting a comparator population into subgroup rows leaves the estimand unchanged.

Identifying the relaxed model. beta_comparator is informed only by the aggregate likelihood. Jointly defined subgroup rows, one set_agd() row per stratum, are what can separate it from the comparator intercept; a single aggregate summary constrains one combination of them. Use check_identification() before fitting and prior_sensitivity() after; the subgroup-identification vignette works through the geometry.

Normal outcomes. A normal fit whose IPD covariates reproduce the outcome exactly has an improper posterior for the residual SD, so mlumr() refuses a constant outcome and a least-squares fit that is exact to numerical precision (residual sum of squares at most 1e-12 of the total) before sampling, and warns when the design is saturated. Everything else is left to the sampler and its diagnostics.

Value

An object of class mlumr_fit

See Also

prior_sensitivity(), check_identification(), set_agd(), prior_summary().

Examples

## Not run: 
# Binary SPFA model
fit_spfa <- mlumr(dat, model = "spfa")

# Relaxed SPFA (allows effect modification)
fit_relaxed <- mlumr(dat, model = "relaxed")

## End(Not run)

Description

mlumr fits its models through rstan by default. mlumr_engine("cmdstanr") switches to cmdstanr for the session; the choice is stored in options(mlumr.stan_engine), so a permanent default belongs in .Rprofile. When cmdstanr or CmdStan is missing, an interactive session is offered their installation (cmdstanr from stan-dev's maintained repository); otherwise the install commands are printed and the engine is left unchanged.

Usage

mlumr_engine(engine = NULL)

Arguments

Value

The current engine, invisibly when setting.

Examples

mlumr_engine()
## Not run: 
mlumr_engine("cmdstanr")

## End(Not run)

Description

A one-call forest plot for comparing a handful of estimates supplied as a data frame, e.g. several methods (naive / STC / ML-UMR) or several covariate profiles. Rows are drawn top-to-bottom in the order given. Returns a ggplot object, so further ggplot2 layers compose with +. This keeps the method-comparison forests in the vignettes to a single line instead of a hand-built ggplot() stack.

Usage

mlumr_forest(
  data,
  ref_line = NULL,
  log_x = FALSE,
  x = NULL,
  title = NULL,
  subtitle = NULL,
  color = "#3B6B9A",
  clip = TRUE,
  ...
)

Arguments

Details

All rows share one axis, so they must be on one effect scale: a frame with an effect column naming more than one measure is rejected rather than drawn against a single reference that cannot be right for both.

Value

A ggplot object.

See Also

marginal_effects(), naive(), stc()

Examples

## Not run: 
forest_df <- data.frame(
  label = c("Naive", "STC", "ML-UMR SPFA"),
  est = c(res_naive$estimate, res_stc$estimate, me$mean),
  lo = c(res_naive$ci_lower, res_stc$ci_lower, me$q2.5),
  hi = c(res_naive$ci_upper, res_stc$ci_upper, me$q97.5)
)
mlumr_forest(forest_df, ref_line = 0, x = "Log odds ratio")

## End(Not run)

Description

Compute an unadjusted (naive) indirect treatment comparison by comparing crude outcomes from the IPD and AgD without any covariate adjustment. The index outcome remains marginal over the index-study population and the comparator outcome remains marginal over the comparator population. The contrast therefore has no single standardized target population. It returns the link-scale contrast plus the two observed marginal outcomes and available natural-scale contrasts.

Usage

naive(data, link = NULL, conf_level = 0.95)

Arguments

Details

The two arms are observed directly, so their intervals are exact: the Clopper-Pearson interval for a binomial proportion and the Garwood interval for a Poisson rate, pooled across aggregate rows. The arm standard errors and the contrasts use the boundary pseudo-count (r + 0.5) / (n + 1) when an arm has zero or all events, or 0.5 events for a zero Poisson count; the reported crude proportions and rates are unchanged. The link-scale contrast, the log risk ratio and the risk difference get Wald intervals, whose coverage is approximate.

Scale note: ⁠$estimate⁠ (and the binomial ⁠$log_rr⁠) is on the link / log scale, where the null is 0. To compare against the natural-scale risk ratio or rate ratio from marginal_effects() (where the null is 1), exponentiate it (e.g. exp(result$estimate)).

Value

An object of class mlumr_naive

Normal-family weighting

Across multiple AgD rows the normal-family comparator mean here is population weighted using outcome_n, matching the Bayesian ML-UMR comparator-population estimand. outcome_n is required when there is more than one row; a single row has weight one. The comparator-mean variance combines independent, mutually exclusive strata as sum(w^2 * se^2) using normalized population weights. The same weighting applies to stc().

Examples

## Not run: 
result <- naive(dat)
print(result)

## End(Not run)

Description

Reconstructed Kaplan-Meier pseudo-IPD for the comparator arm (Morgan 2012, thalidomide), the survival AgD comparator; pair with ndmm_ipd and ndmm_agd_covs.

Usage

ndmm_agd

Format

A data frame with 408 rows and 5 columns: study, treatment, subject, eventtime, status.

Source

The Morgan-2012 thalidomide arm of multinma::ndmm_agd (GPL-3; Phillippo, multinma, 2024); event/censoring times reconstructed from published Kaplan-Meier curves. See ndmm_ipd for trial sources.

Description

Published covariate summaries for the comparator arm (Morgan 2012, thalidomide), supplying the covariate moments for the AgD integration points.

Usage

ndmm_agd_covs

Format

A data frame with 1 row and 7 columns: study, treatment, age_mean / age_sd, and the proportions iss_stage3_prop, response_cr_vgpr_prop, male_prop.

Source

The Morgan-2012 thalidomide row of multinma::ndmm_agd_covs (GPL-3; Phillippo, multinma, 2024), subset to the columns used in the survival vignette. See ndmm_ipd for trial sources.

Description

Individual patient data for the index arm (McCarthy 2012, lenalidomide), the survival (progression-free survival) ML-UMR worked example; pair with ndmm_agd and ndmm_agd_covs.

Usage

ndmm_ipd

Format

A data frame with 231 rows and 9 columns:

study, treatment

study and treatment labels

subject

row identifier

age, iss_stage3, response_cr_vgpr, male

patient covariates

eventtime, status

progression-free survival time and event indicator (1 = event, 0 = censored)

Details

The two source trials are not genuinely disconnected: McCarthy 2012 and Morgan 2012 both include a placebo arm. Those arms are omitted here deliberately, so that this pair of datasets poses the unanchored problem ML-UMR exists for.

McCarthy 2012 is the index arm because population overlap, not sample size, governs how far an unanchored comparison has to extrapolate. Unanchored MAIC effective sample size against the Morgan 2012 thalidomide arm is 35.8 of 231 (15.5\ lenalidomide arm, the other candidate for this comparison. See data-raw/prepare_multinma_subsets.R for the full pair ranking.

Source

The McCarthy-2012 lenalidomide arm of multinma::ndmm_ipd (GPL-3; Phillippo, multinma, 2024), subset to the columns used in the survival vignette. multinma provides simulated IPD resembling published newly-diagnosed-multiple-myeloma trials; the vignette compares lenalidomide (McCarthy et al. 2012) with thalidomide (Morgan et al. 2012). See data-raw/prepare_multinma_subsets.R.

Description

Bernoulli CDF

Usage

pbern(q, prob, lower.tail = TRUE, log.p = FALSE)

Arguments

Value

Numeric vector

Description

Plots the posterior density of the named parameters with their prior density overlaid, reading the prior from the fit. The mlumr analogue of multinma::plot_prior_posterior(). Intended for parameters with a known, un-autoscaled prior (the treatment intercepts mu_index / mu_comparator under the default prior_intercept).

Usage

plot_prior_posterior(object, pars = c("mu_index", "mu_comparator"), ...)

Arguments

Value

A ggplot object.

See Also

prior_summary(), mlumr()

Examples

## Not run: 
plot_prior_posterior(fit)
plot_prior_posterior(fit, pars = c("mu_index", "mu_comparator"))

## End(Not run)

Description

Point-interval of the conditional effect at each covariate profile.

Usage

## S3 method for class 'mlumr_conditional_effects'
plot(x, ref_line = NULL, ...)

Arguments

Value

A ggplot object.

See Also

conditional_effects()

Description

Plots the point estimate and credible interval for each effect measure, grouped by target population (index / comparator). The interval's coverage is read from the quantile columns present, so it matches the probs the result was summarized with. Panels showing only ratio measures are drawn on a log axis, where reciprocal effects sit at equal distances from the null. The mlumr analogue of plot(multinma::relative_effects(fit)).

Usage

## S3 method for class 'mlumr_marginal_effects'
plot(x, ref_line = NULL, ...)

Arguments

Value

A ggplot object.

See Also

marginal_effects()

Examples

## Not run: 
plot(marginal_effects(fit, effect = "all"))

## End(Not run)

Description

Dispatches on the prediction type: time-indexed types ("survival", "hazard", "cumhaz", "loghr") are drawn as curves with credible bands at the coverage the result was summarized with; scalar types ("rmst", "median", "response") as point-intervals. The mlumr analogue of plot(predict(multinma_fit)).

Usage

## S3 method for class 'mlumr_prediction'
plot(x, ref_line = NULL, ...)

Arguments

Value

A ggplot object (compose further layers, e.g. a KM overlay, with +).

See Also

predict.mlumr_fit(), geom_km()

Examples

## Not run: 
plot(predict(fit, type = "survival")) + geom_km(dat)
plot(predict(fit, type = "response"), ref_line = c(0, 1))
plot(predict(fit, type = "loghr"))

## End(Not run)

Description

Generate population-average absolute-outcome predictions in the index and comparator populations.

Usage

## S3 method for class 'mlumr_fit'
predict(
  object,
  population = c("both", "index", "comparator"),
  type = NULL,
  summary = TRUE,
  probs = c(0.025, 0.5, 0.975),
  times = NULL,
  newdata = NULL,
  ...
)

Arguments

Details

Marginalization on non-identity links. For type = "response" the reported values are E[g^{-1}(eta)], the population-average prediction for an individual drawn from that population, not g^{-1}(E[eta]); the expectation runs over the IPD individuals for the index population and over the integration points for the comparator one. type = "link" applies the fitted link after that marginalization.

Value

A data frame with predictions. type = "rmst" adds a horizon column with the restriction time actually integrated to, since RMST at different horizons is a different estimand. The plot methods require summary = TRUE; with summary = FALSE the raw posterior draws are returned as a plain data frame. For type = "median" the summary is conditional on the median being reached, and p_not_reached gives the posterior probability that it is not.

See Also

marginal_effects() for treatment-effect summaries; conditional_predict() and conditional_effects() for predictions at specific covariate profiles.

Examples

## Not run: 
# Absolute predictions for both populations:
predict(fit, population = "both")
# Survival RMST, and transport to a target covariate distribution:
predict(fit, type = "rmst")
predict(fit, newdata = target_population)

## End(Not run)

Description

Cauchy is Student-t with df = 1, a wrapper around prior_student_t(). Its very heavy tails can slow sampling; a Student-t with 3 to 7 degrees of freedom is usually preferred for regression coefficients.

Usage

prior_cauchy(mean = 0, sd = 2.5, autoscale = FALSE)

Arguments

Value

A list with components distribution = "student_t", df = 1, mean, sd, autoscale.

Examples

prior_cauchy(mean = 0, sd = 2.5)

Description

Exponential prior on a positive scalar. Currently supported for prior_sigma (normal-family residual SD) only; rejected for unconstrained intercepts and regression coefficients. prior_exponential(rate = 1) has mean 1 and is a reasonable weakly-informative choice when the outcome is standardized.

Usage

prior_exponential(rate = 1)

Arguments

Value

A list with components distribution = "exponential", rate, and placeholders (mean = 0, sd = 1 / rate) so the same Stan-field translation works.

Examples

prior_exponential(rate = 1)

Description

Construct a normal prior for passing to mlumr() via prior_intercept, prior_beta, or prior_sigma. The resulting list is consumed by the Stan models.

Usage

prior_normal(mean = 0, sd = 10, autoscale = FALSE)

Arguments

Value

A list with components distribution, mean, sd, df, autoscale.

Choosing a scale

The default intercept prior normal(0, 10) is very weak on the link scale, and the data usually constrain the intercept strongly. The coefficient default normal(0, 2.5) is a generic starting value on the link scale per unit of covariate, not a calibrated choice; use autoscale = TRUE for predictors on different scales and calibrate with prior predictive checks (Gelman et al., 2008; the Stan prior-choice wiki). prior_sigma is a normal truncated at zero through the Stan ⁠<lower=0>⁠ constraint, a half-normal at the default mean of 0.

For family = "normal" the identity link is not unit-free: the intercepts and coefficients are in the outcome's units, and so is the residual SD under either link. There the package defaults are read in units of the IPD outcome SD, sd(y): normal(0, 10 * sd(y)) for the intercepts, normal(0, 2.5 * sd(y)) for the coefficients (identity link) and a half-normal with scale 2.5 * sd(y) for the residual SD (either link), and autoscale = TRUE gives a coefficient scale of sd * sd(y) / sd(x). A prior written out by the user is used as given. prior_summary() prints the scales the model used. Run prior_sensitivity() for the relaxed model, whose beta_comparator is identified only by the aggregate likelihood.

References

Gelman, A., Jakulin, A., Pittau, M. G., & Su, Y.-S. (2008). A weakly informative default prior distribution for logistic and other regression models. Annals of Applied Statistics, 2(4), 1360-1383.

Vehtari, A. et al. Prior Choice Recommendations (Stan wiki): https://github.com/stan-dev/stan/wiki/Prior-Choice-Recommendations.

Examples

# Default weakly-very-weak intercept prior
prior_normal(mean = 0, sd = 10)

# Package starting value for regression coefficients
prior_normal(mean = 0, sd = 2.5)

# Autoscaled coefficient prior (dividing 2.5 by each covariate's SD)
prior_normal(mean = 0, sd = 2.5, autoscale = TRUE)

Description

Refit an mlumr() model across a grid of prior_beta scales (keeping the family, mean, and df fixed) and summarize how the posterior for the marginal treatment effects (delta_index, delta_comparator) moves. This is the workflow recommended by Vehtari et al.'s prior-choice wiki for judging how much of the posterior is driven by the data versus the prior.

Usage

prior_sensitivity(
  fit,
  prior_beta_scales = c(0.5, 1, 2.5, 5, 10),
  prior_beta_comparator_scales = NULL,
  probs = c(0.025, 0.5, 0.975),
  verbose = TRUE,
  ...
)

Arguments

Details

The design-matrix controls (center, qr) are taken from the original fit and replayed, so a refit reproduces the original parameterization instead of reverting to the defaults. A fit made with center = FALSE or qr = TRUE is a different parameterization, and replaying the defaults would vary the model as well as the prior.

Only the scale of prior_beta is varied; its family and mean, and every other prior and setting, come from the original fit. Each scale is applied to every coefficient, so the sweep reflects one level of prior informativeness per refit; an exponential prior_beta is swapped for prior_normal(0, scale). For a relaxed fit the comparator prior is swept alongside, because the index-population estimand is driven by the comparator coefficients; holding their prior fixed would report a flat curve for exactly the quantity most exposed to the prior.

Value

A data frame with one row per (prior scale, summarized parameter) pair, and columns scale, scale_comparator (dropped when the model has no comparator coefficient prior), parameter, effect, at_time (present only when the summarized effect has an evaluation time, so absent for every non-survival family and for survival scalars that carry none), mean, sd, and one column per requested quantile, named q followed by the percentage (the default probs give q2.5, q50, q97.5), matching marginal_effects(). Quantiles are columns, not a row dimension. Side effect: prints a summary table at the end when verbose = TRUE, which is the default; verbose = FALSE returns the same data frame and prints nothing.

See Also

prior_summary() for a one-shot description of the priors on a fit; marginal_effects() for the posterior summary quantities this sweep tracks.

Examples

## Not run: 
sens <- prior_sensitivity(fit_spfa, prior_beta_scales = c(1, 2.5, 5))

## End(Not run)

Description

Heavier-tailed alternative to prior_normal(). A Student-t with moderate degrees of freedom can be a robust weakly informative starting family, but its scale still requires calibration to the link, outcome, and predictor scaling.

Usage

prior_student_t(df = 5, mean = 0, sd = 2.5, autoscale = FALSE)

Arguments

Value

A list with components distribution = "student_t", df, mean, sd, autoscale.

Examples

# A moderately heavy-tailed coefficient prior
prior_student_t(df = 5, mean = 0, sd = 2.5)

Description

Print a human-readable summary of every prior that was used to fit an mlumr() model, including the effective per-coefficient scales after autoscaling. Mirrors the spirit of rstanarm::prior_summary().

Usage

prior_summary(object, ...)

## Default S3 method:
prior_summary(object, ...)

## S3 method for class 'mlumr_fit'
prior_summary(object, digits = 3, ...)

Arguments

Value

Invisibly returns a list describing the priors; the side effect is printing a formatted summary.

See Also

prior_sensitivity() to quantify how much the posterior moves under alternative prior_beta scales; prior_normal(), prior_student_t(), prior_cauchy(), prior_exponential() for the prior constructors themselves.

Examples

## Not run: 
fit <- mlumr(dat)
prior_summary(fit)

## End(Not run)

Description

Aggregate (arm-level) summary for the comparator arm (FIXTURE, secukinumab 300 mg), the AgD comparator in the binary ML-UMR example; pair with psoriasis_ipd.

Usage

psoriasis_agd

Format

A data frame with 1 row and 11 columns: study, treatment; response counts pasi75_r / pasi75_n; and covariate summaries age_mean / age_sd, bsa_mean / bsa_sd, weight_mean / weight_sd, and prevsys_prop (a proportion).

Source

The FIXTURE secukinumab-300 arm of multinma::plaque_psoriasis_agd (GPL-3; Phillippo, multinma, 2024), subset to the columns used in the binary vignette. See psoriasis_ipd for the original trial sources.

Description

Individual patient data for the index arm (UNCOVER-2, ixekizumab Q4W), the binary (PASI 75 response) ML-UMR worked example; pair with psoriasis_agd.

Usage

psoriasis_ipd

Format

A data frame with 347 rows and 8 columns:

study, treatment

study and treatment labels

subject

row identifier

pasi75

binary PASI 75 response indicator

age, bsa, weight, prevsys

patient covariates (age in years, body-surface area in percent, weight in kg, and previous systemic treatment as a 1/0 indicator)

weight is missing for 2 of the 347 rows, as it is in the multinma source. set_ipd drops incomplete rows with a warning, so a fit adjusting for weight uses 345 of them.

Details

The two source trials are not genuinely disconnected: UNCOVER-2 and FIXTURE both include placebo and etanercept arms. Those arms are omitted here deliberately, so that this pair of datasets poses the unanchored problem ML-UMR exists for.

Source

The UNCOVER-2 ixekizumab-Q4W arm of multinma::plaque_psoriasis_ipd (GPL-3; Phillippo, multinma, 2024, doi:10.5281/zenodo.3904454), subset to the columns used in the binary vignette. multinma provides simulated IPD resembling the UNCOVER-2 / UNCOVER-3 (ixekizumab; Griffiths et al. 2015, doi:10.1016/s0140-6736(15)60125-8) and secukinumab (Langley et al. 2014, doi:10.1056/nejmoa1314258) plaque-psoriasis trials. See data-raw/prepare_multinma_subsets.R.

Description

Bernoulli quantile function

Usage

qbern(p, prob, lower.tail = TRUE, log.p = FALSE)

Arguments

Value

Integer vector of 0s and 1s

Description

Prepare AgD from the comparator treatment for an unanchored indirect comparison.

Usage

set_agd(
  data,
  treatment,
  family = c("binomial", "normal", "poisson"),
  outcome_n = NULL,
  outcome_r = NULL,
  outcome_mean = NULL,
  outcome_se = NULL,
  outcome_E = NULL,
  cov_means,
  cov_sds = NULL,
  cov_types = NULL,
  study = NULL
)

Arguments

Details

Rows must partition the aggregate sample. Each row contributes its own likelihood factor, as if the rows were disjoint sets of patients. One arm, or one set of mutually exclusive subgroup cells, is right; several overlapping subgroup tables of the same participants count every patient once per table and overstate the precision. Nothing in the data reveals the overlap, so it is not checked. See vignette("subgroup-identification", "mlumr") for how many rows the relaxed model needs.

Scales. For family = "normal", outcome_mean and outcome_se are on the arithmetic scale under both links; a geometric mean or a log-scale summary is a different quantity and cannot be converted by the delta method. For family = "poisson", outcome_r is the total count and outcome_E the total person-time, and the covariate distribution the rate is averaged over has to describe the covariates weighted by exposure; person-level moments stand in for that only when exposure carries no information about the rate within the row. For family = "binomial", outcome_r and outcome_n are counts of events and trials.

Value

An object of class mlumr_agd. As for set_ipd(), the internal column names cannot be used as column names in data.

Examples

## Not run: 
# Binary outcome
agd <- set_agd(
  data = trial_b,
  treatment = "trt",
  outcome_n = "n_total",
  outcome_r = "n_events",
  cov_means = c("age_mean", "sex_prop"),
  cov_sds = c("age_sd", NA),
  cov_types = c("continuous", "binary")
)

# Continuous outcome
agd <- set_agd(
  data = trial_b,
  treatment = "trt",
  family = "normal",
  outcome_mean = "mean_score",
  outcome_se = "se_score",
  outcome_n = "n_total",
  cov_means = c("age_mean", "sex_prop")
)

# Count outcome
agd <- set_agd(
  data = trial_b,
  treatment = "trt",
  family = "poisson",
  outcome_r = "n_events",
  outcome_E = "person_years",
  cov_means = c("age_mean", "sex_prop")
)

## End(Not run)

Description

Prepare comparator aggregate survival data for an unanchored indirect comparison. The comparator arm is supplied as reconstructed pseudo-IPD (event/censoring times digitized from a published Kaplan-Meier curve, e.g. via the Guyot algorithm) together with summary covariate moments (means/SDs). The Stan model integrates the comparator likelihood over the covariate distribution implied by those moments.

Usage

set_agd_surv(
  data,
  treatment,
  Surv = NULL,
  time = NULL,
  status = NULL,
  entry_time = NULL,
  cov_means,
  cov_sds = NULL,
  cov_types = NULL,
  study = NULL,
  arm = NULL
)

Arguments

Details

Under delayed entry the comparator likelihood conditions each integration point on survival to its entry time and then averages, so cov_means, cov_sds and the distributions add_integration() builds must describe the population observed at entry (those in the risk set), not a baseline population before selection. With varying entry times that population can differ by entry time while the model has one distribution per arm; pooled summaries are right only under a common entry time or when the covariate distribution among those observed at entry is the same at every entry time. Neither condition is checkable from the summaries supplied. Delayed entry in the individual arm is unaffected.

Value

An object of class mlumr_agd_surv (also inheriting mlumr_agd). The internal column names cannot be used as column names in data.

Reconstruction uncertainty is not propagated

The pseudo-individual records enter the likelihood as observed data, so credible intervals are conditional on this one reconstruction and narrower than the evidence supports. Treat the reconstruction as an analysis choice: digitize the curve more than once or perturb the points within their reading error, refit, and report the spread across refits beside the within-fit interval.

See Also

set_agd() for non-survival aggregate data; multinma::set_agd_surv() is the ML-NMR equivalent.

Examples

## Not run: 
agd <- set_agd_surv(
  data = comparator_km,
  treatment = "trt",
  time = "time", status = "status",
  cov_means = c("age_mean", "male_prop"),
  cov_sds = c("age_sd", NA),
  cov_types = c("continuous", "binary")
)

## End(Not run)

Description

Prepare IPD from the index treatment for an unanchored indirect comparison.

Usage

set_ipd(
  data,
  treatment,
  outcome = NULL,
  covariates,
  family = c("binomial", "normal", "poisson", "survival"),
  exposure = NULL,
  study = NULL,
  Surv = NULL,
  time = NULL,
  status = NULL,
  entry_time = NULL
)

Arguments

Value

An object of class mlumr_ipd. Its ⁠$data⁠ holds the treatment, study, outcome and covariate columns under internal names, which cannot be used as column names in data.

Examples

## Not run: 
# Binary outcome
ipd <- set_ipd(
  data = trial_a,
  treatment = "trt",
  outcome = "response",
  covariates = c("age", "sex")
)

# Continuous outcome
ipd <- set_ipd(
  data = trial_a,
  treatment = "trt",
  outcome = "score",
  covariates = c("age", "sex"),
  family = "normal"
)

# Count outcome with exposure
ipd <- set_ipd(
  data = trial_a,
  treatment = "trt",
  outcome = "events",
  covariates = c("age", "sex"),
  family = "poisson",
  exposure = "person_years"
)

## End(Not run)

Description

Aggregate summary of the comparator treatment (exercise therapy, ET) for the continuous ML-UMR example; pair with shoulder_ipd. The two arms are treated as separate single-arm sources to illustrate an unanchored comparison. Not real patient data.

Usage

shoulder_agd

Format

A data frame with 1 row and 10 columns: study, treatment, n, outcome summary (y_mean, y_se) and covariate summaries (age_mean/age_sd, sex_prop (a proportion), baseline_vas_mean/baseline_vas_sd).

Source

Simulated; see shoulder_ipd for the modeling basis.

Description

Simulated individual patient data for the index treatment (arthroscopic subacromial decompression, ASD) in a continuous-outcome ML-UMR example. Pair with shoulder_agd (the comparator). Not real patient data.

Usage

shoulder_ipd

Format

A data frame with 147 rows and 7 columns:

study

study label ("FIMPACT")

treatment

index treatment label ("ASD")

subject

row identifier

age

age in years

sex

1 = male, 0 = female

baseline_vas

baseline shoulder pain on activity (VAS 0-100)

pain_vas_activity

shoulder pain on activity at 24 months (VAS 0-100)

Details

The index and comparator arms come from the same trial, so both carry the study label "FIMPACT". Splitting one randomized trial into a single-arm IPD source and a single-arm aggregate source is what makes this an unanchored example that still has a full-data comparison to be checked against: fitting the two arms together on the complete synthetic records gives the quantity the unanchored methods are trying to recover once one arm has been reduced to summaries. That is a data-reduction benchmark. It is not a randomized reference and not a known population causal effect. The generator draws treatment first and then synthesizes the covariates conditional on it, so the synthetic arms are not a fresh randomized assignment independent of the generated baseline variables; and the full-data estimate carries its own sampling error. Validating an estimator against a causal target would need a specified covariate distribution, potential-outcome mechanism, assignment rule and estimand, which these records do not provide. Because the label is shared, combine_data warns that IPD and AgD come from the same study; that warning is expected here and is the honest reading of the data. It does not fire for psoriasis_ipd or ndmm_ipd, whose arms really do come from different trials.

Source

Simulated (not real patient data), generated with the synthpop package (sequential CART; Nowok, Raab and Dibben 2016, doi:10.18637/jss.v074.i11) from the FIMPACT 10-year trial (BMJ 2025;391:e086201; dataset CC BY 4.0, University of Helsinki / Finnish Ministry of Education open-data portal, doi:10.23729/fd-d323a34b-f698-3bc6-b38d-9c93aeadbe74), preserving its covariate and covariate-outcome relationships; fidelity validated with the syntheticdata package. See data-raw/simulate_external_data.R.

Description

Perform an unanchored simulated treatment comparison (STC) with marginal standardization. Fits an outcome regression to the single IPD treatment arm, standardizes its predictions over the comparator-population covariate distribution, and contrasts that outcome with the reported comparator-arm outcome in the same population. It does not standardize either treatment to the index population.

Usage

stc(
  data,
  link = NULL,
  conf_level = 0.95,
  distribution = "weibull",
  n_boot = 200L,
  seed = NULL,
  rmst_horizon = NULL
)

Arguments

Details

For binomial outcomes the result carries the link-scale effect, the standardized and observed event probabilities, the risk difference and the log risk ratio, with delta-method Wald intervals; the observed comparator proportion gets the exact Clopper-Pearson interval, as in naive(). An observed comparator arm with zero or all events uses the pseudo-count (r + 0.5) / (n + 1) in transformed measures. For Poisson outcomes ⁠$rd⁠ is a rate difference per unit exposure and ⁠$estimate⁠ the log rate ratio, with a 0.5 continuity correction for a zero comparator count.

Scale note: ⁠$estimate⁠ (and the binomial ⁠$log_rr⁠) is on the link / log scale, where the null is 0. To compare against the natural-scale risk ratio or rate ratio from marginal_effects() (where the null is 1), exponentiate it (e.g. exp(result$estimate)).

Normal-family weighting note: across multiple AgD rows, the normal STC comparator-population prediction and observed mean use sample-size (outcome_n) weights, matching the Bayesian ML-UMR comparator-population estimand. outcome_n is required when there is more than one row; a single row has weight one. The observed comparator-mean variance combines independent, mutually exclusive strata as sum(w^2 * se^2) using normalized population weights.

A GLM is fitted to the IPD, its predictions are averaged over the comparator covariate distribution (the integration points, or the AgD means for the identity-link normal case) on the response scale, and the average is contrasted with the reported comparator outcome, as in Ren et al.'s unanchored STC. Standard errors are first-order delta-method values, conditional on the integration grid and the reported comparator summaries. The estimand is E_B[m_A(X)] - E_B[Y_B] in the comparator population (⁠$rd⁠ for binomial, ⁠$md⁠ for normal); ⁠$estimate⁠ is the link-scale contrast of the two standardized quantities. Survival STC contrasts the index RMST standardized to the comparator covariates with the RMST of an intercept-only flexsurv::flexsurvreg() fit to the pseudo-IPD. The effect is not transported to the index population; mlumr(model = "relaxed") estimates comparator-specific covariate effects for any other target.

Value

An object of class mlumr_stc. Its separation component records the outcome of the binomial separation check: status is "not_separated" when the exact check ran and found none, "unknown" when only the fitted-value screen ran (it cannot see quasi-complete separation), and "not_applicable" for other outcome models. A separated fit, and a Poisson fit with no events, are refused rather than returned.

References

Ren S, Ren S, Welton NJ, Strong M (2024). Advancing unanchored simulated treatment comparisons: A novel implementation and simulation study. Research Synthesis Methods, 15(4), 657-670. doi:10.1002/jrsm.1718

Remiro-Azocar A, Heath A, Baio G (2022). Parametric G-computation for compatible indirect treatment comparisons with limited individual patient data. Research Synthesis Methods, 13(6), 716-744. doi:10.1002/jrsm.1565

Examples

## Not run: 
result <- stc(dat)
print(result)

## End(Not run)

Description

Expand integration points into a long-format data frame

Usage

unnest_integration(data)

Arguments

Value

A data frame with columns for each covariate plus .int_id and .agd_row