msm

Version 2.0.0 | 2026-10-26

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

msm estimates marginal structural models for long person-period data with a time-varying treatment and time-varying confounding. It takes you from variable mapping through stabilized inverse-probability weights, longitudinal diagnostics, and a weighted outcome model to standardized risks under always-treated and never-treated strategies; companion packages make the tables and figures.

Quick Start

After installing, use the bundled person-period example to fit a pooled logistic MSM and obtain counterfactual cumulative-incidence predictions:

capture confirm file msm_example.dta
if _rc net get msm, from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/msm") replace
use msm_example.dta, clear

msm_prepare, id(id) period(period) treatment(treatment) outcome(outcome) ///
    covariates(biomarker comorbidity) baseline_covariates(age sex)
msm_weight, treat_d_cov(biomarker comorbidity age sex) ///
    treat_n_cov(age sex) truncate(1 99) nolog
msm_diagnose
msm_fit, model(logistic) outcome_cov(age sex) nolog
msm_predict, times(1 3 5 7 9) difference samples(500) seed(12345)

msm_predict compares standardized cumulative incidence under always-treated and never-treated strategies. The Demo section shows the figures and tables the demo script draws from the same data.

Features

• Standardized risks under static strategies. msm_predict returns cumulative incidence (or survival, type(survival)) under always-treated and never-treated strategies and, with difference, the always-minus-never contrast in r(rd), with Monte Carlo intervals drawn from the outcome model's coefficient covariance.
• Stabilized IPTW and IPCW. msm_weight fits the treatment and optional censoring models with the prior treatment added automatically, linear to natural-spline time bases, and percentile truncation; failed fits and probabilities at 0 or 1 are errors unless you name a fallback policy.
• Diagnostics built for longitudinal data. msm_diagnose reports treatment balance within period and prior-treatment history, censoring balance, and a positivity floor on the probability of the treatment actually observed; numerator covariates are flagged, not counted as imbalanced.
• Weighted pooled logistic, linear, or Cox models. msm_fit clusters sandwich standard errors on the subject identifier by default, adds treatment-history terms (lag1, cumulative, duration, interaction), and leaves the native glm, regress or stcox result in e().
• State that travels with the data. The mapping and a signed contract for each stage are stored in dataset characteristics; msm, status reports the current stage and the next command, and re-running msm_prepare clears what no longer applies.
• Graph-ready results. msm_predict, frame() writes one row per period and quantity in the schema eplot reads, so tables (regtab, effecttab), figures, and E-values (qba_confound) come from the companion packages.

Requirements

• Stata 16 or later
• The core commands need no other package.
• Optional companions for tables, diagnostics, figures and E-values: tabtools, psdash, eplot, swimlane and qba, all from Stata-Tools. psdash must be 1.8.0 or later to read data weighted by msm 2.0.0.

Installation

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

Uninstall first when upgrading from 1.4.x: net install, replace does not delete files that a new version no longer ships. If ado uninstall msm reports that more than one package matches, run ado dir and uninstall each msm entry by its number, highest first.

Download the example dataset into the working directory:

net get msm, from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/msm") replace

Install the companions you need the same way, for example:

foreach p in tabtools psdash eplot qba swimlane {
    net install `p', from("https://raw.githubusercontent.com/tpcopeland/Stata-Tools/main/`p'") replace
}

Commands

Command	Description
msm	Overview, companion map, and msm, status for the pipeline state
msm_prepare	Map ID, period, treatment, outcome, censoring and covariates; refuse invalid person-period data; report data checks
msm_weight	Stabilized treatment weights and optional censoring weights
msm_diagnose	Balance by period and prior treatment, censoring balance, and positivity
msm_fit	Weighted pooled logistic, linear, or Cox outcome model
msm_predict	Standardized risks under always and never treated, risk differences, and a graph-ready frame

msm_protocol, msm_validate, msm_plot, msm_report, msm_table, msm_diagtab and msm_sensitivity were removed in 2.0.0. They remain installed as stubs for all of 2.x; each prints its replacement and exits with return code 199.

How It Works

msm stores the variable mapping and a signed contract for each stage in dataset characteristics. msm_weight and msm_fit verify that the data still match the stage they build on before using it, and re-running msm_prepare clears every downstream weight and fit artifact. msm_diagnose and msm_predict only read: their results are in r() (and msm_predict, frame()), so nothing they produce can go stale.

msm_prepare → msm_weight → msm_diagnose → msm_fit → msm_predict
                  │              │            │           │
               psdash         psdash       regtab     effecttab · eplot · twoway
                                           qba_confound

msm_weight creates _msm_weight (treatment × censoring weight, the one msm_fit uses), _msm_tw_weight, optional _msm_cw_weight, the propensity score _msm_ps, the raw and used model probabilities, and _msm_decision_risk, which marks the rows that are real treatment decisions. Post-event and post-censoring rows carry the last weight forward but stay out of every fit and summary.

Run msm, status at any point to see the stage, the mapping, and the next step.

Choosing a Workflow

Goal	Workflow
Standardized risks and risk differences	msm_prepare, msm_weight, msm_diagnose, msm_fit, model(logistic), msm_predict
Weighted identity-scale effect	msm_fit, model(linear); the treatment coefficient is a period-specific probability difference, not a standardized risk difference
Weighted hazard ratio	msm_fit, model(cox); prediction is unavailable, report e()
Delayed or cumulative treatment effects	msm_fit, history(lag1 cumulative duration interaction); still prediction-compatible
Continuous treatment-history summary	msm_fit, exposure(varname) and, where appropriate, tvcov(varlist); estimation only

Replacements for the removed commands

Task (former command)	Replacement
Coefficient table (msm_table, msm_report)	collect: msm_fit ... then regtab (tabtools)
Risk-difference table	msm_predict, difference then matrix RD = r(rd) and effecttab, from(RD) (tabtools)
Raw matrix export	puttab using file.xlsx, matrix(name) (tabtools)
Weight distribution (msm_plot, msm_diagnose pooled block)	psdash weights treatment _msm_ps if _msm_decision_risk, wvar(_msm_weight)
Per-period weights (msm_diagnose, by_period)	psdash combined
Pooled balance and Love plot (msm_plot)	psdash balance ..., covariates(denominator covariates) wvar(_msm_weight) loveplot
Risk curve (msm_plot, type(survival))	msm_predict, frame() then twoway
Risk-difference forest plot	msm_predict, frame() then eplot if strategy == "rd", frame()
Treatment trajectories (msm_plot)	swimlane
E-value (msm_sensitivity)	qba_confound, from_model coef(treatment) evalue measure(HR) (qba)
Data checks (msm_validate)	Refusals and r() notes in msm_prepare and msm_weight; positivity in msm_diagnose
Protocol (msm_protocol)	None; record it in the manuscript or a protocol document
Cross-contrast export (msm_diagtab)	None; frame post the returns and puttab, frame()

Worked Examples

1. Prediction with censoring weights

use msm_example.dta, clear
msm_prepare, id(id) period(period) treatment(treatment) outcome(outcome) ///
    censor(censored) covariates(biomarker comorbidity) baseline_covariates(age sex)
msm_weight, treat_d_cov(biomarker comorbidity age sex) treat_n_cov(age sex) ///
    censor_d_cov(biomarker comorbidity age sex) censor_n_cov(age sex) ///
    truncate(1 99) nolog
msm_diagnose, threshold(0.1)
msm_fit, model(logistic) outcome_cov(age sex) period_spec(quadratic) nolog
msm_predict, times(1 3 5 7 9) difference samples(500) seed(12345) frame(pred, replace)
matrix list r(rd)

msm_diagnose reports balance within each period and prior-treatment stratum. Age and sex are in treat_n_cov(), so the stabilized weights keep their distribution on purpose: they appear as target=0 and are flagged (numerator) in the pooled table rather than counted as imbalanced.

2. Tables

collect clear
collect: msm_fit, model(logistic) outcome_cov(age sex) period_spec(quadratic) nolog
regtab, xlsx("msm_results.xlsx") sheet("Outcome model") coef(OR)

msm_predict, times(1 3 5 7 9) difference samples(500) seed(12345)
matrix RD = r(rd)
effecttab, from(RD) effect("Risk difference") xlsx("msm_results.xlsx") sheet("Risk differences")

r(rd) holds estimate, lower and upper Monte Carlo limits, and an empty p-value column, the layout effecttab, from() reads. Any other returned matrix goes to Excel with puttab using msm_results.xlsx, matrix(name) sheet(name).

3. Weight and balance diagnostics with psdash

psdash combined
psdash weights treatment _msm_ps if _msm_decision_risk, wvar(_msm_weight) detail graph
psdash balance treatment _msm_ps if _msm_decision_risk, ///
    covariates(biomarker comorbidity) wvar(_msm_weight) loveplot

psdash combined reads the msm contract, restricts to the decision risk set and diagnoses the analysis weight _msm_weight by default. Pass only denominator covariates to the Love plot; numerator covariates are not balance targets. psdash standardizes the weighted SMD by the unweighted pooled SD (the cobalt default) while msm_diagnose uses the weighted SD (Austin and Stuart 2015, section 4.1.1), so the two can differ slightly near a threshold.

4. Figures

msm_predict, times(1 3 5 7 9) difference samples(500) seed(12345) frame(pred, replace)
eplot if strategy == "rd", frame(pred) values xline(0)

frame pred: twoway ///
    (rarea ll ul horizon if strategy == "never", pstyle(p1) color(%25) lwidth(none)) ///
    (connected estimate horizon if strategy == "never", pstyle(p1)) ///
    (rarea ll ul horizon if strategy == "always", pstyle(p2) color(%25) lwidth(none)) ///
    (connected estimate horizon if strategy == "always", pstyle(p2)), ///
    legend(order(2 "Never treated" 4 "Always treated")) ///
    ytitle("Cumulative incidence") xtitle("Period")

The frame has one row per period and quantity (strategy is never, always or rd) with label, estimate, ll, ul and rowtype, eplot's frame schema. Treatment histories: generate pstop = period + 1, then swimlane, id(id) start(period) stop(pstop) state(treatment).

5. E-value for unmeasured confounding

msm_fit, model(logistic) outcome_cov(age sex) period_spec(quadratic) nolog
qba_confound, from_model coef(treatment) evalue measure(HR)

The pooled-logistic odds ratio is a ratio of per-period odds, close to a discrete-time hazard ratio when the per-period risk is small; hence measure(HR). Add commonoutcome when the cumulative incidence at the end of follow-up is 15% or more.

6. Cox estimation without prediction

use msm_example.dta, clear
msm_prepare, id(id) period(period) treatment(treatment) outcome(outcome) ///
    covariates(biomarker comorbidity) baseline_covariates(age sex)
msm_weight, treat_d_cov(biomarker comorbidity age sex) treat_n_cov(age sex) nolog
collect clear
collect: msm_fit, model(cox) outcome_cov(age sex) nolog
regtab, coef(HR)

Demo

demo/demo_msm.do regenerates the figures and the results workbook below from msm_example.dta. It is a checkout workflow, not part of net install.

Output	Command focus
	msm_predict, frame() and twoway
	eplot, frame()
	psdash weights
	psdash balance, loveplot
Results workbook	regtab and effecttab

The console transcripts of the demo are demo/console_pipeline.md (msm_prepare, msm_weight, psdash combined, msm_diagnose) and demo/console_results.md (msm_fit, regtab, msm_predict, effecttab, qba_confound).

Command Reference

Square brackets mark optional syntax; help command in Stata has the full narrative.

msm

Syntax: msm [, list detail status]

• list prints the five estimator commands; detail describes them and the companion map; status reports the verified stage, mapping and artifacts. At most one may be given. protocol was removed in 2.0.0.

msm_prepare

Syntax: msm_prepare, id(varname) period(varname) treatment(varname) outcome(varname) [censor(varname) covariates(varlist) baseline_covariates(varlist)]

• Refuses an empty dataset, missing id() or period(), duplicate id()-period() rows, non-integer periods, non-binary treatment, outcome or censoring, same-period event and censoring, and baseline covariates that vary within individual.
• Reports, without refusing, individuals with non-consecutive periods, mapped variables with missing values, covariates without variation, and the size of the smallest period (r(n_gap_ids), r(missing_vars), r(constant_covariates), r(min_period_n)).

msm_weight

Syntax: msm_weight [, treat_d_cov(varlist) treat_n_cov(varlist) censor_d_cov(varlist) censor_n_cov(varlist) period_d_spec(spec) period_n_spec(spec) truncate(numlist) fitfailure(policy) probpolicy(policy) clip(#) preview replace nolog]

• treat_d_cov() defaults to the prepared time-varying plus baseline covariates; treat_n_cov() defaults to none, leaving an intercept and the automatically added lagged treatment.
• censor_d_cov() enables IPCW and requires a mapped censor(); censor_n_cov() requires censor_d_cov().
• period_d_spec() defaults to linear and period_n_spec() to none; each accepts none, linear, quadratic, cubic, or ns(#).
• truncate() is off by default; truncate(1) is shorthand for truncate(1 99).
• fitfailure() and probpolicy() default to error; fitfailure(marginal) and probpolicy(clip) clip(#) are explicit sensitivity policies.
• preview resolves the specification without fitting; replace overwrites existing _msm_* weight variables.

msm_diagnose

Syntax: msm_diagnose [, balance_covariates(varlist) threshold(#) positivity(#)]

• balance_covariates() defaults to all covariates registered by msm_prepare; with none at all the command refuses.
• threshold() defaults to 0.1; positivity() defaults to 0.01 and is applied to the smallest probability of the treatment actually observed.

msm_fit

Syntax: msm_fit [, model(string) outcome_cov(varlist) exposure(varname) tvcov(varlist) history(string) period_spec(string) cluster(varname) vce(string) strata(varlist) bootstrap(#) level(#) nolog]

• model() defaults to logistic and accepts logistic, linear, or cox. The fit stays in e() as the native glm, regress or stcox result.
• outcome_cov() adds time-fixed outcome covariates; every numerator covariate must appear here or, for Cox, in strata().
• history() accepts lag1, cumulative, duration, interaction; period_spec() defaults to quadratic and accepts none, linear, quadratic, cubic, or ns(#).
• vce() defaults to sandwich SEs clustered at the mapped ID. bootstrap() is reserved.

msm_predict

Syntax: msm_predict, times(numlist) [strategy(string) type(string) samples(#) seed(#) level(#) difference extrapolate frame(name[, replace])]

• strategy() defaults to both; type() to cum_inc (or survival); samples() to 100 draws (at least 10; use 500 or more for published intervals).
• difference adds the always-minus-never contrast: r(rd) and r(rd_#), or r(sd) and r(sd_#) under type(survival). A negative period t is named m# (r(rd_m2), row tm2).
• frame() writes one graph-ready row per period and quantity; an existing frame needs replace.
• extrapolate allows times beyond the fitted risk-set support.

Options

The flagship msm takes the first four options (protocol is still parsed so that it can refuse with a pointer to the manuscript; it was removed in 2.0.0). The remaining rows list the options of the five estimator commands.

Option	Command	Meaning
list	msm	Print the five estimator commands
detail	msm	Describe the commands and the companion map
status	msm	Report the verified stage, mapping and artifacts
protocol	msm	Removed in 2.0.0; refuses with return code 198
id(varname)	msm_prepare	Individual identifier (required)
period(varname)	msm_prepare	Integer period variable (required)
treatment(varname)	msm_prepare	Binary treatment (required)
outcome(varname)	msm_prepare	Binary outcome (required)
censor(varname)	msm_prepare	Binary censoring indicator
covariates(varlist)	msm_prepare	Time-varying covariates
baseline_covariates(varlist)	msm_prepare	Time-fixed baseline covariates
treat_d_cov(varlist)	msm_weight	Treatment denominator covariates
treat_n_cov(varlist)	msm_weight	Treatment numerator covariates
censor_d_cov(varlist)	msm_weight	Censoring denominator covariates; enables IPCW
censor_n_cov(varlist)	msm_weight	Censoring numerator covariates
period_d_spec(spec)	msm_weight	Denominator time basis (default linear)
period_n_spec(spec)	msm_weight	Numerator time basis (default none)
truncate(numlist)	msm_weight	Percentile truncation; off by default
fitfailure(policy)	msm_weight	Failed-fit policy (default error)
probpolicy(policy)	msm_weight	Probability-support policy (default error)
clip(#)	msm_weight	Clip bound for probpolicy(clip)
preview	msm_weight	Resolve the specification without fitting
replace	msm_weight	Overwrite existing _msm_* weight variables
nolog	msm_weight, msm_fit	Suppress iteration logs
balance_covariates(varlist)	msm_diagnose	Covariates to balance
threshold(#)	msm_diagnose	SMD threshold (default 0.1)
positivity(#)	msm_diagnose	Positivity floor (default 0.01)
model(string)	msm_fit	logistic (default), linear or cox
outcome_cov(varlist)	msm_fit	Time-fixed outcome covariates
exposure(varname)	msm_fit	Continuous treatment-history summary
tvcov(varlist)	msm_fit	Time-varying companion terms
history(string)	msm_fit	lag1, cumulative, duration, interaction
period_spec(string)	msm_fit	Outcome time basis (default quadratic)
cluster(varname)	msm_fit	Cluster variable (default the mapped ID)
vce(string)	msm_fit	Variance estimator
strata(varlist)	msm_fit	Cox strata
bootstrap(#)	msm_fit	Reserved
level(#)	msm_fit, msm_predict	Confidence level
times(numlist)	msm_predict	Periods to predict (required)
strategy(string)	msm_predict	Strategy (default both)
type(string)	msm_predict	cum_inc (default) or survival
samples(#)	msm_predict	Monte Carlo draws (default 100)
seed(#)	msm_predict	Random-number seed
difference	msm_predict	Always-minus-never contrast
extrapolate	msm_predict	Allow times beyond the fitted support
frame(name[, replace])	msm_predict	Write a graph-ready frame

Stored Results

msm

r(version), r(commands), r(removed) and r(n_commands). With status: r(stage), r(next_step), r(model), the mapping macros r(id), r(period), r(treatment), r(outcome), r(censor), r(covariates), r(baseline_covariates), and r(prepared), r(weighted), r(fitted).

msm_prepare

Scalars r(N), r(n_ids), r(n_periods), r(period_span), r(n_events), r(n_treated), r(n_censored), r(n_gap_ids), r(min_period_n); macros r(id), r(period), r(treatment), r(outcome), r(censor), r(covariates), r(baseline_covariates), r(missing_vars), r(constant_covariates).

msm_weight

After fitting: r(mean_weight), r(sd_weight), r(min_weight), r(max_weight), r(p1_weight), r(median_weight), r(p99_weight), r(ess) (risk set), r(n_truncated), r(n_fitfail_fallback), r(fitfailure_fallback), r(n_probability_repairs), the matrix r(probability_repairs), and the resolved specification macros. preview returns the resolved specification only.

msm_diagnose

Scalars r(threshold), r(n_imbalanced), r(n_numerator), r(n_unavailable), r(positivity_threshold), r(n_positivity_violations), r(min_obs_probability), r(n_support_violations); macros r(covariates) and r(support_violation_periods); matrices r(treatment_balance), r(support), optional r(censor_balance), and the secondary pooled r(balance).

msm_fit

The standard results of glm, regress or stcox, plus e(msm_cmd), e(msm_model), e(msm_treatment), e(msm_exposure), e(msm_tvcov), e(msm_history_spec), e(msm_history_assumption), e(msm_period_spec), e(msm_vce), e(msm_cluster), e(msm_strata), e(msm_inf_dist), e(msm_inf_df), e(msm_n_clusters), e(msm_n_dropped) and the matrix e(effects).

msm_predict

Matrices r(predictions) and, with difference, r(rd) or r(sd); scalars r(n_times), r(n_ref), r(samples), r(level), r(min_support), r(max_support), r(extrapolated), and r(rd_#) or r(sd_#); macros r(seed), r(seed_source), r(seed_state), r(type), r(strategy), r(history_spec), r(diff_type), r(draw_method), r(frame).

Assumptions and Limits

• The causal interpretation requires consistency, sequential exchangeability given the measured history, positivity, and adequate treatment, censoring and marginal outcome models. The commands diagnose parts of these conditions but cannot establish them.
• Data must be long person-period data with unique id()-period() rows, integer periods, binary treatment and outcome, and a common baseline. Delayed entry is not supported.
• Treatment is meant to vary over time; for a single baseline treatment use teffects ipw.
• Prediction supports only static always-treated and never-treated strategies.
• The weight models add only the immediately prior treatment automatically. Build richer histories yourself and put them in treat_d_cov() and, where relevant, censor_d_cov().
• outcome_cov() variables must be time-fixed, and numerator covariates must enter the outcome model.
• exposure() and tvcov() are for treatment-history-derived terms and disable msm_predict.
• The censoring models condition on current treatment (within-period order L, A, C, Y), which differs from the printed sw-dagger formula in Hernán, Brumback and Robins (2000).
• msm_fit standard errors and msm_predict Monte Carlo intervals condition on the estimated weights; they do not propagate weight-model estimation.

Troubleshooting

Symptom	Response
A removed command prints "was removed in msm 2.0.0"	Use the replacement it names; see Companion Packages
msm_prepare notes gaps	msm_weight will refuse them; expand the grid or drop the incomplete individuals
msm_weight rejects delayed entry	Align the panel to a shared baseline
Weight means or ESS are poor	Inspect the treatment histories, psdash combined, and r(support) before a justified truncation or model revision
msm_weight reports probability support failure	Investigate missing predictors, separation and positivity first; probpolicy(clip) clip(#) only as a named sensitivity analysis
Balance remains poor	Inspect r(treatment_balance) by period and prior treatment; numerator covariates (target=0) are not expected to balance
msm_predict refuses	Confirm a logistic fit without exposure()/tvcov() and times() within the fitted support

References

• Robins JM, Hernan MA, Brumback B. Marginal structural models and causal inference in epidemiology. Epidemiology. 2000;11(5):550-560.
• Hernan MA, Brumback B, Robins JM. Marginal structural models to estimate the causal effect of zidovudine on the survival of HIV-positive men. Epidemiology. 2000;11(5):561-570.
• Cole SR, Hernan MA. Constructing inverse probability weights for marginal structural models. American Journal of Epidemiology. 2008;168(6):656-664.
• Austin PC, Stuart EA. Moving towards best practice when using inverse probability of treatment weighting using the propensity score to estimate causal treatment effects in observational studies. Statistics in Medicine. 2015;34(28):3661-3679. doi:10.1002/sim.6607.
• Adenyo D, Guertin JR, Candas B, Sirois C, Talbot D. Evaluation and comparison of covariate balance metrics in studies with time-dependent confounding. Statistics in Medicine. 2024. doi:10.1002/sim.10188.
• VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Annals of Internal Medicine. 2017;167(4):268-274.
• Hernan MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC, 2020.

QA

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

Version History

• 2.0.0 (2026-10-13): Breaking. msm is now the estimator core: msm_prepare, msm_weight, msm_diagnose (longitudinal balance and positivity), msm_fit and msm_predict. Removed msm_protocol, msm_validate, msm_plot, msm_report, msm_table, msm_diagtab and msm_sensitivity; running one prints its replacement (tabtools regtab/effecttab/puttab, psdash, eplot, swimlane, qba_confound) and exits with return code 199, and the stubs stay for all of 2.x. msm_diagnose no longer prints pooled weight summaries (use psdash) or accepts by_period or accumulate(); it adds r(covariates), flags numerator covariates in the pooled table instead of counting them as imbalanced, and flags periods with no treated or no untreated row. msm_prepare reports gaps, missing values, constant covariates and sparse periods in r(). msm_predict adds frame() and r(rd), writes nothing to the dataset, and returns r(rd_m#) for negative periods (it failed with r(101) before). msm_weight writes contract 1.1 (_msm_risk_var) so psdash restricts to the decision risk set, and prints truncation percentiles as 1st/99th; data weighted by 2.0.0 needs psdash 1.8.0 or later. With no covariates mapped in msm_prepare, msm_diagnose balances the treat_d_cov() covariates. Run ado uninstall msm before installing.
• 1.4.12 (2026-10-09): Help References now give a DOI for every entry.
• 1.4.11 (2026-09-30): Failed msm_prepare and msm_weight calls restore the caller's original sort metadata as well as its physical row order. Protocol descriptions retain literal quotes, macro-like text and Unicode in display, returned text and exported files. Report text and CSV numbers retain the requested decimal precision. Failed predictions preserve the caller's RNG state, returned results and prediction identifier counter. Weighting preserves active estimation results and restores caller returned results, RNG state and the identifier counter on failure. Guard and operational refusals in the dispatcher, preparation, validation, diagnosis, fitting, plotting, reporting, protocol and sensitivity commands preserve caller returned results. Validation retains its documented findings on a completed strict check. Commands that temporarily sort data restore the original sort keys and row order.

Earlier versions

• 1.4.10 (2026-09-27): msm_weight, truncate() counts and caps weights at the exact percentile cutoffs, and msm_diagnose counts extreme weights above the exact P99. The cutoffs used to pass through a macro, which could move them by one unit in the last place, so a weight equal to the cutoff was counted as truncated and rewritten slightly off it (off by one in r(n_truncated) on ordinary data). msm_diagnose no longer prints its imbalanced count when run quietly.
• 1.4.9 (2026-09-26): The effective sample size reported by msm_weight, msm_diagnose (overall, by treatment group, and in r(support), r(treatment_balance) and r(censor_balance)) and msm_report, and the weighted SMD of a continuous covariate, no longer become missing when the weights are extremely small or large (near 1e-200 or 1e200, where their squares underflow or overflow); both are computed on weights divided by a power of two. Ordinary weights give the same values to within the last digit or two.
• 1.4.8 (2026-08-30): Guarded auto-open paths against shell metacharacters, made accumulated diagnostics fail before state changes on incompatible frames, restored caller abbreviation settings on helper early returns, and corrected Viewer-width help tables.
• 1.4.7 (2026-08-28): Vectorized counterfactual prediction and Monte Carlo simulation in Mata while preserving seeded results, reducing realistic msm_predict runtimes by more than tenfold.
• 1.4.6 (2026-08-11): Bound downstream artifacts to their owning fit or weights, persisted prediction and balance matrices with saved datasets, corrected fitted-sample sensitivity incidence and RR-scale labels, and exported confounding-strength assumptions and bounds.
• 1.4.5 (2026-08-05): Made seeded Monte Carlo prediction intervals reproducible across numerically equivalent refits by using the symmetric positive-semidefinite covariance square root.
• 1.4.4 (2026-08-05): Refused rank-deficient fitted specifications before committing package state and corrected stored-result documentation.
• 1.4.3 (2026-07-29): Hardened empty-dataset handling in msm_prepare and aligned the flagship version readout with its shipped header.
• 1.4.2 (2026-07-28): Corrected weighted SMD definitions, risk-set plotting, signed-period survival plots, diagnostic reuse, and interpretation of fitted support and conditional Monte Carlo intervals.
• 1.4.1 (2026-07-27): Refined installed help and documentation contracts for the released user surface.
• 1.4.0 (2026-07-26): Added configurable denominator and numerator time bases, including natural cubic splines, with prior-weight-preserving defaults.
• 1.3.0 (2026-07-25): Corrected observed-treatment positivity diagnostics, risk-set pooling, observation-order restoration, and fitted-sample metadata.
• 1.2.4 (2026-07-23): Improved Excel export failure propagation and contextual interpretation of E-values and confounding bounds.
• 1.2.3 (2026-07-04): Improved the failure message for time-invariant treatment processes and documented the baseline-treatment alternative.
• 1.2.2 (2026-07-02): Corrected protective-effect sensitivity bounds, natural-spline edge cases, Excel title handling, weight-specification display, and missing SMD reporting.
• 1.2.1 (2026-06-25): Added a known-truth recovery workflow for the marginal structural log-odds.
• 1.2.0 (2026-06-17): Added continuous exposure and time-varying companion terms for estimation-only dose-duration analyses.
• 1.1.0 (2026-06-14): Added the retained per-period propensity variable and the optional psdash diagnostic contract.
• 1.0.4 (2026-05-29): Added accumulated cross-contrast weight diagnostics and the msm_diagtab Excel export.
• 1.0.3 (2026-05-06): Added explicit vce() control and Cox strata() support.
• 1.0.2 (2026-05-06): Hardened state invalidation, missing-weight handling, export restoration, and binary-outcome scope guidance.
• 1.0.1 (2026-04-30): Hardened validation edge cases, time-fixed outcome-covariate enforcement, Cox guidance, and protocol export escaping.
• 1.0.0 (2026-04-26): Initial Stata-Tools release of the MSM workflow suite.

Author

Timothy P Copeland, Karolinska Institutet

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

To cite: Copeland TP. msm: Marginal structural models with IPTW, end to end. Stata-Tools, https://github.com/tpcopeland/Stata-Tools/tree/main/msm. Please give the version you used (which msm prints it).

License

MIT (see LICENSE)