iivw

Version 1.0.0 | 2026-10-26

Quick Start · Installation · Worked Examples · Inference · Stata-Tools · Report a bug

iivw corrects the over-representation of frequently seen patients in irregular longitudinal data by weighting each visit by the inverse of its estimated visit intensity, optionally combined with treatment-propensity weights. It gives Stata users one workflow to build the weights, check their leverage, fit the weighted outcome model, and test whether visit timing depends on the outcome.

Quick Start

After installing, create a small irregular-visit panel, compute IIW weights, check them, and fit a weighted population-average model:

clear
set seed 20260417
set obs 240
gen long id = ceil(_n/4)
bysort id: gen byte visit = _n
gen double time = 3 * (visit - 1) + runiform() * .05
replace time = 0 if visit == 1
gen double x = rnormal()
bysort id: replace x = x[1]
gen byte treated = runiform() < invlogit(.4 + .5 * x)
bysort id: replace treated = treated[1]
gen double y = 1 + .1 * time + x + rnormal()

iivw_weight, id(id) time(time) visit_cov(x) lagvars(y) maxfu(12) nolog
iivw_balance
iivw_fit y x, vce(fixed) nolog

vce(fixed) is explicit here so the short example uses the weights-known analytic sandwich. For weighted GEE fits using IIW or IPTW, the supported default when no variance option is supplied is a 999-draw subject-level refit bootstrap; see Inference.

Features

• Three weight types from one command. iivw_weight builds inverse intensity weights (IIW) from an Andersen-Gill recurrent-event Cox model, stabilized inverse treatment-propensity weights (IPTW) from a one-row-per-subject logistic model, or their product (FIPTIW), and stores the panel contract with the data.
• An observation window you must state. IIW and FIPTIW require exactly one of censor(), maxfu(), or endatlastvisit; visit-measured variables enter the visit model only as lags through lagvars().
• Weights checked before they are used. iivw_balance reports weight leverage, effective sample size, composition shifts, and a target standardized mean difference for the visit model, with a separate verdict for covariates the model left out. Nonconvergence and missing weights stop the analysis unless you opt in.
• Inference that says what it covers. Weighted IIW and IPTW fits default to a 999-draw subject-level bootstrap that refits the weight models in every draw; fixed-weight alternatives are available and labelled as ignoring weight-estimation uncertainty, and every fit records its status in e(iivw_inference_status).
• Sampling versus measurement movement. iivw_exogtest tests whether lagged outcomes predict the next visit, and iivw_diagnose splits the change in a coefficient into a sampling gap and a measurement-process gap across three stored models.
• Bootstraps split across processes. saving() and rngstream() let several Stata processes draw one bootstrap, and iivw_bspool pools the shards through Stata's bstat.

Requirements

• Stata 16 or later.
• Stata 17 or later for the iivw_fit collect option.
• No external runtime dependency is required for the core iivw commands or their direct xlsx() reporting exports.
• Optional: tabtools for regtab model-table exports, and psdash for treatment-propensity diagnostics after IPTW/FIPTIW weighting.
• The repository demo additionally uses tc_schemes for its graph scheme.

Installation

Install the released package from the public Stata-Tools distribution:

capture ado uninstall iivw
net install iivw, from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/iivw") replace

Install optional companion packages only when you need their workflows:

net install tabtools, from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/tabtools") replace
net install psdash, from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/psdash") replace

Commands

Command	Description
iivw	Display the package overview and return the available command list
iivw_weight	Compute IIW, IPTW, or FIPTIW weights
iivw_balance	Report weight leverage, composition shifts, and the visit-model target-SMD diagnostic
iivw_fit	Fit weighted or unweighted GEE outcome models
iivw_exogtest	Test whether lagged outcomes or disease activity predict subsequent visit timing
iivw_diagnose	Decompose movement in a stored marginal or reference-slope estimate across three models
iivw_bspool	Pool bootstrap replicate files from concurrent iivw_fit shards into one result

How It Works

The package addresses a visit-level sampling problem: when patients who are sicker, or otherwise different, are seen more often, a row-per-visit analysis can over-represent them. The visit model estimates the conditional intensity of observation, and the resulting weights make the outcome analysis less dominated by differential visit frequency.

Weight type	Use when	Construction
IIW	Visit timing is informative but treatment weighting is not needed	Inverse intensity weights from an Andersen–Gill recurrent-event Cox model
IPTW	Treatment assignment is confounded but visit timing is not being corrected	Stabilized inverse treatment-propensity weights from a one-row-per-subject logistic model
FIPTIW	Both visit timing and treatment assignment are informative or confounded	The product of the IIW and IPTW components

The usual workflow is:

1. iivw_weight estimates the visit and, when requested, treatment models, creates the final weight, and stores the panel contract in dataset characteristics.
2. iivw_balance reports leverage, effective sample size, descriptive composition shifts, and the target standardized mean difference for the visit model.
3. iivw_fit reads the stored weight and panel metadata and fits the outcome model.
4. iivw_exogtest and iivw_diagnose are optional diagnostics for separating observation-process movement from residual measurement-process movement.

For IIW and FIPTIW, give the study's observation window with exactly one of censor(), maxfu(), or endatlastvisit. A subject-specific or common administrative end is usually the design-relevant choice; endatlastvisit is available for studies whose follow-up genuinely ends at the last observed visit.

Visit-model covariates must be measured before the interval whose visit they explain. Put baseline and externally updated variables in visit_cov(); use lagvars() for visit-measured outcomes, disease activity, or recent events. The weight-model varlists accept plain numeric variables, not factor-variable or inline spline notation, so expand such terms into physical columns before weighting.

Choosing a Workflow

Study question	Starting workflow
Irregular registry or EHR visits distort a disease trajectory	Use IIW, inspect iivw_balance, and compare timespec(linear) with a flexible time specification
Treatment is binary and time-invariant, with confounding by baseline severity	Use IPTW or FIPTIW with treat() and treat_cov(); inspect treatment components with optional psdash
Both treatment and visit timing are confounded	Use FIPTIW; treat() is included in the visit-intensity model by default
Repeated measurement may create practice or test-count artifact	Fit unweighted, weighted, and measurement-adjusted models, then use iivw_exogtest and iivw_diagnose

Worked Examples

Run the Quick Start setup first if you want to reuse its synthetic data. The examples below are sequenced so that the same panel can support IIW, FIPTIW, inference, exogeneity, and decomposition workflows.

1. IIW only

Correct informative visit timing without a treatment-propensity component. The default baseline(entry) treats each subject's first row as study entry.

iivw_weight, id(id) time(time) visit_cov(x) lagvars(y) maxfu(12) replace nolog
iivw_balance
iivw_fit y x, timespec(linear) vce(fixed) nolog

2. FIPTIW for treatment confounding and informative visits

Adding treat() makes the default weight type FIPTIW and adds treatment to the visit-intensity model. stabcov(treated) is valid here because treated is also in the outcome design shown below.

iivw_weight, id(id) time(time) visit_cov(x) lagvars(y) maxfu(12) ///
    treat(treated) treat_cov(x) stabcov(treated) replace nolog
iivw_balance, component(final) nolog
iivw_fit y treated x, timespec(linear) vce(fixed) nolog

3. Refit-bootstrap inference

For a weighted IIW or IPTW analysis, the recommended variance route refits the nuisance models inside each subject-level bootstrap replicate. FIPTIW intervals requested explicitly are nominal and are labeled in the stored inference status.

iivw_fit y treated x, ///
    vce(bootstrap, reps(999) seed(20260417)) ///
    citype(percentile) nolog

4. Exogeneity diagnostic for visit timing

The diagnostic creates one-visit lags of the tested variables and fits counting-process Cox models for subsequent visits. Use the same end-of-follow-up contract as the weighting call.

iivw_exogtest y, id(id) time(time) maxfu(12) ///
    adjust(x) by(treated) replace nolog

5. Sampling movement versus measurement-process movement

Store three comparable models for the coefficient of interest, then ask iivw_diagnose to report the sampling gap, artifact gap, and descriptive shares. Use estimand(contrast) when the coefficient is a treatment contrast; that reports movement but suppresses the share decomposition.

Recompute unstabilized IIW weights first: the FIPTIW weights from example 2 were stabilized on treated, which these three models leave out, and iivw_fit refuses a stabilized weight whose numerator the outcome model does not contain.

iivw_weight, id(id) time(time) visit_cov(x) lagvars(y) maxfu(12) replace nolog

iivw_fit y time x, unweighted id(id) time(time) ///
    timespec(none) nolog
estimates store M_unweighted

iivw_fit y time x, timespec(none) vce(fixed) replace nolog
estimates store M_weighted

gen double log_visit = log(visit + 1)
iivw_fit y time x log_visit, timespec(none) ///
    vce(fixed) replace nolog
estimates store M_adjusted

iivw_diagnose time, unweighted(M_unweighted) ///
    weighted(M_weighted) adjusted(M_adjusted) ///
    estimand(marginal) exogeneity(unknown)

The direct reporting commands can write styled workbook sheets without tabtools:

iivw_balance, xlsx("iivw_reporting_exports.xlsx") sheet("Balance") replace
iivw_exogtest y, id(id) time(time) maxfu(12) adjust(x) ///
    by(treated) replace nolog xlsx("iivw_reporting_exports.xlsx") ///
    sheet("Exogeneity")
iivw_diagnose time, unweighted(M_unweighted) ///
    weighted(M_weighted) adjusted(M_adjusted) ///
    xlsx("iivw_reporting_exports.xlsx") sheet("Diagnostics") replace

Demo

The repository checkout workflow is demo/demo_iivw.do. It creates synthetic SDMT-like data, installs the optional companion packages needed for graph and model-table output, and is the source for the checked-in assets below; the demo script is not part of the net install payload.

Asset	Contents
	psdash combined treatment-propensity overlap, support, balance, and treatment-weight diagnostics
	Final analysis-weight distribution from psdash weights, iivwcomponent(final)
Model-comparison and visit-wave workbook	collect/regtab output for model comparison and categorical-time interactions
Direct reporting workbook	iivw_balance, iivw_exogtest, and iivw_diagnose export sheets

Command Reference

iivw

Syntax:

iivw

With no arguments, iivw displays the package overview and returns r(version), r(commands), and r(n_commands).

iivw_weight

Syntax:

iivw_weight, id(varname) time(varname) [options]

Option	Default	Purpose
id(varname)	Required	Subject identifier in long panel data
time(varname)	Required	Numeric, nonnegative visit time, unique within subject
visit_cov(varlist)	Required for IIW/FIPTIW unless lagvars() supplies covariates	Numeric covariates for the visit-intensity Cox model; ignored for IPTW-only with a note
treat(varname)	None	Binary 0/1 treatment, constant within subject; required for IPTW/FIPTIW
treat_cov(varlist)	None	Numeric covariates for the treatment logistic model; required when treatment weighting is requested
wtype(iivw|iptw|fiptiw)	Auto-detected	IIW when treat() is absent, FIPTIW when it is present; use iptw to skip the visit model
stabcov(varlist)	None	Covariates for the IIW stabilization numerator
lagvars(varlist)	None	Raw visit-measured sources to lag by one visit before use in the visit model
entry(varname)	0	Subject-specific study entry time
censor(varname)	None	Subject-specific end of follow-up; mutually exclusive with maxfu() and endatlastvisit
maxfu(#)	None	Common end of follow-up; mutually exclusive with the other end-of-follow-up options
endatlastvisit	None	End each subject's risk window at their last visit; use only when that is the study design
baseline(entry|event)	entry	Treat the first visit as study entry or as a modeled recurrent event
truncvisit(# #)	None	Trim the IIW component at row-level percentiles
trunctreat(# #)	None	Trim the IPTW component at subject-level percentiles
truncfinal(# #)	None	Trim the final analysis weight at row-level percentiles
generate(name)	_iivw_	Prefix for generated weights and metadata-linked variables
replace	Off	Overwrite variables owned by a prior iivw call
nolog	Off	Suppress model iteration logs
efron	On	Use Efron handling for tied visit times
breslow	Off	Use Breslow handling, mainly to match analyses that used stcox's Breslow default
allownonconverged	Off	Continue after a nonconverged weight model; use only as an explicit sensitivity/debugging choice
allowmissingweights	Off	Accept rows with no computed weight as a complete-case analysis; otherwise missing-weight rows are an error

truncate() is removed and errors. Choose truncvisit(), trunctreat(), or truncfinal() so the component being altered is explicit.

iivw_balance

Syntax:

iivw_balance [varlist] [if] [in], [options]

The optional numeric varlist adds covariates to the displayed table; the stored visit-model covariates remain the target of the balance verdict. The command applies to IIW and FIPTIW metadata, not IPTW-only weights.

Option	Default	Purpose
component(iiw|final)	iiw	Describe the visit component or the final analysis weight; the target verdict always uses the IIW component
cvcut(#)	0.10	CV below this threshold is classified as low leverage
essratiocut(#)	0.95	ESS/N above this threshold is classified as low leverage
balcut(#)	0.10	Maximum absolute target SMD allowed for within_rule
agrefit	Off	Display hazard ratios from the refitted visit-intensity model
level(#)	c(level)	Confidence level for refit hazard-ratio intervals
efron, breslow	Stored setting	Ignored as fit requests; the refit replays the tie method used to create the weights
nolog	Off	Suppress Cox iteration logs in the refit

For the workbook options shared by the reporting commands, see Excel reporting.

iivw_fit

Syntax:

iivw_fit depvar [indepvars] [if] [in], [options]

Option	Default	Purpose
unweighted	Off	Fit without applying stored weights
id(varname)	Stored metadata for weighted fits	Panel ID for an unweighted fit without package metadata
time(varname)	Stored metadata for weighted fits	Time variable for an unweighted fit without metadata when time is modeled
family(string)	gaussian	GLM family for the GEE fit
link(string)	Canonical link	GLM link override
timespec(string)	linear	linear, quadratic, cubic, ns(#), categorical, or none
interaction(varlist)	None	Create interactions between listed covariates and all generated time terms
categorical(varlist)	None	Expand integer-valued outcome predictors into labeled dummies
basecat(#)	Lowest observed level	Reference level for variables in categorical()
timebasecat(#)	Lowest observed time	Reference level for timespec(categorical)
cluster(varname)	Stored panel ID	Cluster variable for analytic robust SEs
vce(bootstrap, reps(#) [seed(#)] [fixedweights]|fixed|stacked)	Weight-type dependent	Refit bootstrap, fixed-weight bootstrap, fixed-weight analytic sandwich, or FIPTIW two-step stacked sandwich; details are below
citype(none|wald|percentile|basic|bca)	wald, except bare FIPTIW is none	Select point-only, normal/Wald, percentile, basic, or BCa endpoints
allowfailedreps	Off	Accept an incomplete bootstrap and record the failed-replicate counts
level(#)	c(level)	Confidence level
nolog	Off	Suppress the underlying estimator's iteration log
allownonconverged	Off	Continue after outcome-model nonconvergence
replace	Off	Overwrite generated time, categorical, and interaction variables
collect	Off	Use Stata's collect framework for non-bootstrap GEE fits
geeopts(string)	None	Pass additional options to glm, except options that take control of the package-owned variance

iivw_exogtest

Syntax:

iivw_exogtest varlist [if] [in], id(varname) time(varname) [options]

Option	Default	Purpose
id(varname)	Required	Subject identifier
time(varname)	Required	Numeric, nonnegative visit or measurement time
adjust(varlist)	None	Baseline or design covariates in the timing model
by(varname)	None	Fit separate diagnostics by a subject-constant group
bystart	Off	Permit a time-varying by() variable and classify intervals by the value at their start
entry(varname)	0	Subject-specific study entry time
censor(varname), maxfu(#), endatlastvisit	Exactly one required	End-of-follow-up contract; use the same choice as iivw_weight
generate(name)	_iivw_exog_	Prefix for generated one-visit lag variables
replace	Off	Overwrite owned lag variables and an existing export worksheet
efron	On	Efron ties in stcox
breslow	Off	Breslow ties for compatibility or sensitivity analysis
nolog	Off	Suppress Cox iteration logs
level(#)	c(level)	Confidence level for hazard-ratio intervals

For the workbook options, see Excel reporting; decimals() defaults to 3 for this command.

iivw_diagnose

Syntax:

iivw_diagnose coefficient, unweighted(estname) weighted(estname) adjusted(estname) [options]

Option	Default	Purpose
unweighted(estname)	Required	Stored unweighted model
weighted(estname)	Required	Stored IIW, IPTW, or FIPTIW-weighted model
adjusted(estname)	Required	Stored weighted model with a direct measurement-process adjustment
exogeneity(exogenous|endogenous|unknown)	unknown	State how the adjustment should be interpreted; this is not tested by the command
estimand(marginal|contrast)	marginal	Compute shares for a marginal/reference slope or movement only for a contrast
true(#)	None	Supply a known truth and return bias quantities
force	Off	Bypass the comparability check and label the result descriptive/non-decomposable
level(#)	c(level)	Confidence level for the three stored estimates

The three estimates must refer to the same outcome, coefficient, model scale, and clustering variable unless force is used. For workbook options, see Excel reporting; decimals() defaults to 4 for this command.

Key Options

Weight construction

iivw_weight auto-detects IIW versus FIPTIW from treat(); use wtype(iptw) for treatment weighting without a visit model. The treatment variable must be binary and time-invariant within subject, and the treatment model is fit cross-sectionally on one row per subject. Under FIPTIW, treatment is included in the visit-intensity model by construction.

The default baseline(entry) treats the first visit per subject as study entry and assigns it the normalized entry weight. baseline(event) models the first visit as a recurrent event and is retained for designs where that first visit is genuinely part of the monitoring process.

The default tie method is Efron in iivw_weight and iivw_exogtest. iivw_balance replays the stored method and may withhold its target-SMD verdict for tied Efron fits because the target is based on the Breslow score residual; leverage and effective-sample-size summaries remain available.

Inference

For IIW and IPTW GEE fits with no explicit variance request, iivw_fit uses vce(bootstrap, reps(999)) with subject-level nuisance-model refitting. The current build stamps this route uncleared-current-build until the long-run coverage gate is reproduced against its source manifest. vce(bootstrap, reps(#) fixedweights) resamples subjects while holding weights fixed, and vce(fixed) uses the analytic cluster-robust sandwich with weights treated as known. The latter two omit weight-estimation uncertainty and should be described as such.

For a bare weighted FIPTIW GEE fit, the default is point-only at every sample size: coefficients are reported without a covariance matrix or nominal interval. Explicit vce() or citype() requests nominal inference and records its status in e(iivw_inference_status); explicit vce(stacked) propagates the fitted-weight terms but remains uncleared-stacked-analytic because its larger-sample study was diagnostic rather than a release gate. Fewer than 999 bootstrap draws are allowed but are marked uncleared-low-reps; failed draws are an error unless allowfailedreps explicitly accepts them.

citype(wald) uses a normal/Wald transformation, citype(percentile) uses empirical bootstrap quantiles, citype(basic) reflects those quantiles around the observed estimate, and citype(bca) adds bias correction and delete-one-subject acceleration. The asymmetric choices require bootstrap draws.

Sharded bootstraps

A 999-draw refit bootstrap refits the visit-intensity model and the outcome model on every draw, and that work is the same on every draw and depends on no other draw. Profiling on a 1200-subject, 7100-row panel put 99.9% of the run in per-draw work, against a fixed overhead of about one draw, so the run is serial only because Stata's bootstrap prefix is.

iivw_fit's saving() writes the replicate draws to a dataset, and rngstream(#) draws them from an independent RNG substream of one shared seed(). Together they let K concurrent Stata processes produce the same 999 draws, which iivw_bspool pools into a single estimation result.

Pooling is delegated to Stata's own bstat, so the pooled variance and percentile limits are computed by the same code path the bootstrap prefix uses: splitting one run's replicate file into K pieces and pooling them back reproduces that run to the last bit. Nine shards of 111 drawn from independent substreams agree with a single 999-draw run within Monte Carlo tolerance, with a worst observed relative standard-error difference of 0.053 against a 3-MCSE band of 0.095.

iivw_fit stamps each replicate file with the identity of the fit that produced it, and iivw_bspool refuses to pool shards that disagree on the weight contract, the weight type, the specification, the coefficient set, the observed estimates or the build. It also refuses two shards drawn from the same RNG state, a truncated shard file, and any pooled result with failed replicates unless allowfailedreps is given. The uncleared-low-reps stamp is recomputed from the pooled total, so nine honestly-stamped 111-draw shards pool to a 999-draw result that no longer carries it; every other stamp is inherited, because pooling improves the replicate count and nothing else. BCa intervals are not poolable, because their acceleration term comes from a per-shard jackknife rather than from the draws.

Measured on a 1200-subject, 7100-row panel, 360 refit draws:

Run	Wall clock
One serial process, 360 draws	92.2 s
Four concurrent shards of 90	26.7 s
Pooling step	0.06 s

That is a 3.4x reduction on four shards, about 86% of linear. The pooled standard error on the treatment coefficient was 0.04797 against 0.04754 for the serial run, a 0.9% difference against a 3-MCSE band of 16% at 360 draws.

A worked K-process driver is in demo/shard_driver.do.

Excel reporting

iivw_balance, iivw_exogtest, and iivw_diagnose write direct styled .xlsx sheets when xlsx(filename) is supplied. sheet() defaults are Balance, Exogeneity, and Diagnostics, respectively. replace overwrites only the named sheet, open opens the workbook, and title()/footnote() add optional rows.

The reporting defaults are decimals(4) for iivw_balance and iivw_diagnose, decimals(3) for iivw_exogtest, and borderstyle(thin) with header shading and zebra rows off. font(), fontsize(), borderstyle(), headershade, headercolor(), zebracolor(), and zebra control the workbook presentation and require xlsx() when they affect an export.

Stored Results

The commands also leave dataset characteristics or estimation results so later steps can verify that the weights and model specification still match the data. The in-Stata help files list every stored name; the following are the main user-facing results.

iivw

iivw returns r(version), r(commands), and r(n_commands) for the package version and available public workflow commands.

iivw_weight

r(weighttype), r(weight_var), r(iw_var), r(tw_var), and r(ps_var) identify the weight type and generated variables. Weight summaries include r(mean_weight), r(min_weight), r(max_weight), r(p1_weight), r(median_weight), r(p99_weight), r(ess), and r(ess_ratio).

The command also returns analysis and model counts such as r(N), r(n_ids), missing-weight and treatment-arm loss counts, truncation cutpoints, propensity-score extrema, and visit-model event counts. r(visit_b) contains the visit-intensity coefficients, and the returned macros record the raw and expanded covariate lists, lag variables, follow-up contract, treatment covariates, and contract version.

iivw_balance

Key scalars are r(weight_cv), r(ess), r(ess_ratio), r(ess_cluster), r(ess_cluster_ratio), r(balance_max_shift), r(balance_max_tsmd), and r(extra_max_tsmd). The diagnostic labels are in r(leverage), r(balance_flag), r(extra_flag), r(target_status), and r(component). r(balance) contains the covariate table, r(target) the per-covariate target SMDs, and r(hr_unweighted) contains refit hazard ratios when available.

iivw_fit

iivw_fit is an eclass command. Standard results include e(cmd), e(b), e(V) when an interval-bearing fit is requested, and e(sample). Important package metadata include e(iivw_model), e(iivw_weighttype), e(iivw_weight_var), e(iivw_vce), e(iivw_underlying_vce), e(iivw_underlying_cmd), e(iivw_refitweights), e(iivw_ci_type), e(iivw_ci), and e(iivw_inference_status).

Bootstrap provenance is recorded in e(iivw_bs_reps_requested), e(iivw_bs_reps_completed), e(iivw_bs_reps_failed), e(iivw_resample_unit), e(iivw_vce_seed), e(iivw_rng), and e(iivw_rngstate_start). The fitted design is recorded in e(iivw_id), e(iivw_time), e(iivw_timespec), e(iivw_time_vars), e(iivw_interaction), and e(iivw_categorical).

iivw_exogtest

The result matrix r(results) contains model-by-term coefficients, standard errors, tests, hazard ratios, intervals, and sample counts. Useful scalars include r(N), r(n_ids), r(n_models), r(n_skipped), r(n_unknown), r(min_p), r(joint_min_p), r(holm_min_p), r(history_association_flag), r(tie_multiplicity), r(n_event_times), and r(n_modeled_events).

Macros record r(id), r(time), r(testvars), r(lagvars), r(adjust), r(by), indexed group/term labels, r(result_row_labels), r(result_columns), and r(conclusion). Export results are returned in r(xlsx), r(sheet), and r(decimals) when an export succeeds.

iivw_diagnose

Scalars include r(decomposable), r(sample_identical), r(n_sample_unweighted), r(n_sample_weighted), and r(n_sample_adjusted). Macros record the coefficient, three stored model names, r(exogeneity), r(estimand), r(depvar), confidence-interval distributions, r(noncollapsible), and r(conclusion).

r(estimates) contains the three model estimates and limits, r(decomp) contains the sampling/artifact/total gaps and shares when decomposable, and r(bias) is returned when true() is supplied. Workbook results use r(xlsx), r(sheet), and r(decimals).

Assumptions and Limits

Study design

These weights address measured visit timing and, when requested, measured treatment confounding; they do not make an observational design causal without credible exchangeability, positivity, and model specification. Unmeasured confounding, unmeasured visit drivers, and misspecified functional forms remain limitations.

The treatment implementation is for a binary, time-invariant treatment. Treatment switching requires a time-varying treatment marginal structural model, which this package does not implement. The package models when visits occur; it does not fit a dropout or censoring model. censor() and maxfu() define the at-risk window rather than estimating a censoring weight.

Current visit measurements must not be used as if they were known before the visit. Use lagvars() for previous-visit information, and remember that time-varying covariates are carried forward over the terminal at-risk interval.

Diagnostics and stability

A large composition shift is descriptive and can be evidence that the weights are doing work; the target SMD is the diagnostic with a reference distribution. Modeled covariates are close to balanced almost by construction, so list unmodeled covariates or transforms (for example z2 or an omitted u) and read their separate verdict r(extra_flag) to check the visit model. Read r(balance_flag) together with leverage and effective sample size, and treat unknown or not_assessed as absence of a supported verdict rather than as balance.

Extreme weights, low effective sample size, rare treatment patterns, few clusters, and nonconvergence can make estimates unstable. The default is to stop on nonconvergence or missing weights; allownonconverged, allowmissingweights, and allowfailedreps are explicit acknowledgments of weakened analyses, not repairs.

Compatibility notes

The end-of-follow-up contract is explicit, the first-visit default is baseline(entry), there is no single truncate() option, and stale or missing weight state fails closed. Re-run iivw_weight when a dataset's stored weighting contract lacks the replay information required by refit bootstrap or balance replay.

The outcome model is a GEE (independence working correlation, cluster-robust variance), which estimates the marginal effect the IIW theory identifies. Weighted random-effects models are not offered: a single observation-level probability weight does not consistently weight random-effects variance components.

References

• Buzkova P, Lumley T. Longitudinal data analysis for generalized linear models with follow-up dependent on outcome-related variables. Canadian Journal of Statistics. 2007;35(4):485–500. doi:10.1002/cjs.5550350402.
• Coulombe J, Moodie EEM, Platt RW. Weighted regression analysis to correct for informative monitoring times and confounders in longitudinal studies. Biometrics. 2021;77(1):162–174. doi:10.1111/biom.13285.
• Lin H, Scharfstein DO, Rosenheck RA. Analysis of longitudinal data with irregular, outcome-dependent follow-up. Journal of the Royal Statistical Society: Series B. 2004;66(3):791–813. doi:10.1111/j.1467-9868.2004.b5543.x.
• Rabe-Hesketh S, Skrondal A. Multilevel modelling of complex survey data. Journal of the Royal Statistical Society: Series A. 2006;169(4):805–827. doi:10.1111/j.1467-985X.2006.00426.x.
• Tompkins G, Dubin JA, Wallace M. On flexible inverse probability of treatment and intensity weighting: Informative censoring, variable selection, and weight trimming. Statistical Methods in Medical Research. 2025;34(5):915–937. doi:10.1177/09622802241313289.
• Hertz-Picciotto I, Rockhill B. Validity and efficiency of approximation methods for tied survival times in Cox regression. Biometrics. 1997;53(3):1151–1156.

QA

QA suites and how to run them are documented in qa/README.md.

Version History

• 1.0.0 (2026-10-15): Initial release.

Author

Timothy P Copeland, Karolinska Institutet

Questions and bug reports: GitHub issues. Please include the first line of which iivw and your Stata version.

To cite: Copeland TP. iivw: Inverse intensity weighting for irregular visit data. Stata-Tools, https://github.com/tpcopeland/Stata-Tools/tree/main/iivw. Please give the version you used (which iivw prints it).

License

MIT (see LICENSE)