<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Lifeflexi on Medium]]></title>
        <description><![CDATA[Stories by Lifeflexi on Medium]]></description>
        <link>https://medium.com/@lifeflexi818?source=rss-04ceed8dafbf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*d6suFA2mU7LOiplf</url>
            <title>Stories by Lifeflexi on Medium</title>
            <link>https://medium.com/@lifeflexi818?source=rss-04ceed8dafbf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 23:06:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lifeflexi818/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Bland-Altman Agreement Analysis in Clinical Measurement]]></title>
            <link>https://medium.com/@lifeflexi818/bland-altman-agreement-analysis-in-clinical-measurement-8e4468d0dd8a?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e4468d0dd8a</guid>
            <category><![CDATA[clinicalmeasurement]]></category>
            <category><![CDATA[medical-research]]></category>
            <category><![CDATA[blandaltman]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[methodcomparison]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 04:10:42 GMT</pubDate>
            <atom:updated>2026-11-02T04:10:42.255Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*udGwWUCByZcXN6JH.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Abc_mgcQTeAu4bcq.png" /></figure><p>Biostatistics * Clinical Measurement * November 2, 2026</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NB_XTMHNnt1K01Eh.png" /></figure><p><strong>Bland-Altman analysis</strong> evaluates whether two quantitative measurement methods agree closely enough for a defined clinical use by estimating mean bias and limits of agreement; it does not replace clinical acceptance criteria and it is not a test of correlation.</p><p>Clinical researchers often compare a new instrument, laboratory assay, imaging algorithm, or bedside measurement with an established method. The tempting first step is to calculate Pearson’s correlation or fit a regression line. Those analyses describe association, not whether two methods produce values close enough to be used interchangeably. <strong>Bland-Altman agreement analysis</strong> focuses on paired differences, their average magnitude, and whether the observed disagreement is acceptable for the intended decision.</p><h3>Agreement is different from association</h3><p>Suppose method A and method B measure the same physiological quantity. If every value from method B is consistently higher than the corresponding value from method A, the two methods may be strongly correlated while still being clinically non-interchangeable. A correlation coefficient can be close to one because the measurements rise and fall together, even when their differences are large enough to alter treatment or eligibility decisions.</p><p>Regression has a related limitation. A slope and intercept can reveal proportional or constant differences, but a good fitted line does not directly quantify the range of disagreement expected for an individual measurement. Method comparison requires a question about individual-level differences: if the new method replaces the comparator, how far could its value reasonably depart from the comparator for a future patient?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*KSmooEytRvbc3ChC.png" /></figure><p>Analysis Main question Why it is insufficient alone for agreement</p><h3>How the Bland-Altman plot is constructed</h3><p>For each paired observation, calculate the difference between methods and the average of the two methods. If method A is the index method and method B is the comparator, define the difference as and the mean as . Plot each difference against its corresponding mean. The horizontal axis represents measurement magnitude; the vertical axis represents disagreement.</p><p>The <strong>mean difference</strong> estimates average bias. A positive bias indicates that method A tends to produce higher values than method B under the chosen direction; reversing the subtraction reverses the sign but not the underlying disagreement. The standard limits-of-agreement calculation uses the mean difference plus or minus approximately 1.96 standard deviations of the differences when the distributional assumptions are reasonable. These limits describe the interval expected to contain most individual differences, not a confidence interval for the mean bias.</p><p>Three visual features matter. First, a cloud displaced from zero suggests systematic bias. Second, a trend in the differences across the mean suggests proportional bias. Third, a fan-shaped spread suggests that variability changes with measurement magnitude. A single pair of constant-scale limits may be misleading when heteroscedasticity is substantial.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U9ACqX1UdcSg1CJB.png" /></figure><p><strong>Clinical interpretation boundary:</strong> Statistical limits of agreement are estimated from the sample. They become clinically meaningful only after comparison with an acceptance range defined from the measurement’s intended use, biological variation, decision thresholds, or a validated reference standard.</p><h3>Define the measurand and intended use before calculating limits</h3><p>A defensible method-comparison study starts with the quantity being measured and the decision that depends on it. “Agreement” is not a universal property of an instrument pair. A difference that is negligible for exploratory screening may be unacceptable when a value determines medication titration, transfusion, escalation of care, or trial eligibility.</p><p>Specify the target population, sampling conditions, measurement units, timing, and whether the comparator is a reference method or simply another imperfect method. Both methods should be applied to the same specimen, participant, or imaging episode under comparable conditions. If measurements are taken at different times, changes in the underlying state can be mistaken for method disagreement.</p><p>Prespecify an acceptable difference or an acceptance rule. The rule may be expressed in absolute units, a percentage, a clinically meaningful change, or a decision-specific threshold. If acceptance limits are unavailable, the study can still estimate bias and statistical limits, but it should not claim interchangeability merely because most points lie within the plotted band.</p><h3>Diagnose bias and non-constant variability</h3><p>Mean bias can conceal important patterns. A new assay may agree at low concentrations but diverge at high concentrations. An automated image measurement may have larger errors for larger lesions. A wearable sensor may show different agreement during motion than at rest. Plotting differences against means helps reveal these patterns, but formal modelling may be needed when the visual signal is complex or repeated observations are present.</p><p>When variability increases with magnitude, consider a scale transformation or a ratio or percentage-difference analysis if scientifically justified. Log transformation can make multiplicative error easier to interpret, but it changes the estimand and requires careful back-transformation. Reporting only transformed limits without explaining their clinical meaning can make the result difficult to use.</p><p>Outliers deserve investigation, not automatic deletion. A point may reflect a data-entry error, device failure, protocol violation, or a true clinically important disagreement. Report prespecified exclusion rules, the number of excluded pairs, and a sensitivity analysis when exclusions could change the conclusion. Bland-Altman plots are diagnostic tools; they do not turn an unexamined outlier into a statistical nuisance.</p><h3>Repeated measurements require dependence-aware analysis</h3><p>The simple formula assumes one independent pair per subject. Clinical studies frequently collect multiple measurements per participant, multiple raters, multiple body sites, or repeated device readings. Treating all pairs as independent can underestimate uncertainty and produce limits that do not reflect between-person and within-person variation.</p><p>For repeated measurements, consider mixed-effects or hierarchical agreement models, subject-level resampling, or repeated-measures extensions that estimate within-subject agreement. The analysis should distinguish repeatability of one method from agreement between two methods. A method can be highly repeatable yet systematically disagree with the comparator; two methods can have similar average values while showing poor repeatability in individuals.</p><p>The data structure should be reported explicitly: number of participants, number of pairs per participant, order of methods, blinding or randomization of measurement order, replicate design, and whether analysts knew prior measurements. A transparent flow of included pairs and exclusions is as important as the final plot.</p><h3>Confidence intervals and precision</h3><p>Point estimates of bias and limits of agreement are not enough. The mean bias has sampling uncertainty, and the upper and lower limits also have uncertainty. Confidence intervals can be obtained using analytical methods under appropriate assumptions or by bootstrap procedures that resample at the participant level when observations are clustered.</p><p>Sample size should be justified by the precision needed for bias and agreement limits, not by a generic correlation power calculation. A study may have enough observations to estimate a high correlation but not enough to characterize the tails of the difference distribution with useful precision. The intended acceptance rule, expected variability, number of repeated measures, and target confidence-interval width should inform planning.</p><h3>Seven-step Bland-Altman workflow</h3><h3>Evidence Summary Table</h3><h3>Common reporting errors</h3><ul><li>Calling two methods interchangeable because the correlation coefficient is high.</li><li>Reporting statistical limits of agreement without stating whether they are clinically acceptable.</li><li>Failing to define the subtraction direction, measurement units, or intended-use population.</li><li>Using ordinary limits-of-agreement formulas for repeated observations without accounting for clustering.</li><li>Ignoring proportional bias or increasing spread across the measurement range.</li><li>Removing outliers after seeing the plot without a prespecified rule or sensitivity analysis.</li><li>Reporting a plot without confidence intervals, sample flow, repeatability information, or measurement conditions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*I3AzuibOjQz0XNAx.png" /></figure><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI tools to connect agreement analysis with rigorous publication workflows:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: audit Bland-Altman calculations, repeated-measure assumptions, bias interpretation, and clinical acceptance claims.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: organize evidence on method comparison, repeatability, agreement limits, and measurement-error reporting with traceable citations.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with laboratory medicine, diagnostic methods, clinical measurement, and biostatistics.</li></ul><h3>Conclusion</h3><p>Bland-Altman analysis answers a practical method-comparison question: how large are the differences between two measurements, and are those differences acceptable for the intended clinical use? The analysis is strongest when the measurand, population, paired design, difference direction, acceptance rule, and uncertainty plan are specified before the plot is interpreted. Correlation and regression can describe association or systematic trends, but they do not replace agreement analysis. For repeated measurements or non-constant variability, the design requires extensions that reflect clustering and scale. Transparent reporting turns a familiar plot into a defensible measurement-evidence argument.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/bland-altman-agreement-clinical-measurement"><em>https://sci.lingcorehealth.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e4468d0dd8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Micro-Randomized Trials for JITAI Clinical Research]]></title>
            <link>https://medium.com/@lifeflexi818/micro-randomized-trials-for-jitai-clinical-research-6163c355f878?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/6163c355f878</guid>
            <category><![CDATA[clinical-trials]]></category>
            <category><![CDATA[microrandomizedtrials]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[digital-health]]></category>
            <category><![CDATA[health-research]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 01:20:32 GMT</pubDate>
            <atom:updated>2026-10-29T01:20:32.233Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JxiALd6ovgdaYL3y.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VqJa2GFaPZRLydc-.png" /></figure><p>Clinical Trial Design * Digital Health Methods * October 29, 2026</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g2g2x2oVJQrdPtwt.png" /></figure><p><strong>Micro-randomized trials (MRTs)</strong> repeatedly randomize an intervention option at eligible decision points, allowing researchers to estimate short-term proximal effects and evaluate when, where, and for whom a just-in-time intervention works.</p><p>Digital interventions are often expected to respond to changing context: a reminder when a patient is available, a coping exercise during a high-risk moment, or a prompt after a period of inactivity. A conventional parallel-group trial can estimate the average effect of an intervention package, but it is poorly suited to questions about timing, repeated delivery, component-level effects, and context-dependent benefit. A <strong>micro-randomized trial</strong> addresses this problem by embedding many small randomizations within each participant’s longitudinal follow-up.</p><h3>What an MRT is designed to answer</h3><p>An MRT is not simply a trial with a large number of observations. Its repeated randomizations answer a particular class of causal questions. At a decision point, the investigator defines whether a participant is <strong>available</strong> for an intervention component. If available, the participant is randomized to one option, such as a prompt versus no prompt or message A versus message B. A proximal outcome is then measured over a prespecified short interval. Repeating this process creates information about the average short-term effect of that component and how the effect changes with time-varying context.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Mt1-fnhuINKv3pcl.png" /></figure><p>Design element Question it supports What must be specified</p><h3>Decision points, availability, and randomization</h3><p>The decision point is the unit of opportunity. It may occur at fixed times, after a sensor-defined state, or after a participant reports a need. The choice should reflect the intervention mechanism rather than the convenience of data collection. If an intervention is intended to reduce sedentary time, randomizing a prompt only when the participant is awake, carrying the device, and not already active may be more informative than randomizing at every clock hour.</p><p><strong>Availability</strong> is not a minor eligibility variable. If a participant cannot receive the intervention because of sleep, driving, technical failure, or a competing intervention, randomizing the unavailable participant can dilute the estimand and create implementation problems. Availability must therefore be recorded at every decision point, with the reason for non-availability when feasible. The analysis population for a proximal effect is generally defined through the design’s availability rule, not by post hoc removal of inconvenient observations.</p><p>Randomization probabilities may be constant or may vary by decision point. Unequal probabilities can protect burden, reduce intervention fatigue, or allocate more opportunities to a clinically promising component. The probabilities must be known or reproducibly generated, because they are part of the design and analysis. If treatment delivery deviates from assignment, the primary intention-to-treat-like proximal effect remains tied to randomization; per-protocol interpretations require additional assumptions and should not be substituted silently.</p><h3>Proximal outcomes and causal estimands</h3><p>The proximal outcome is deliberately close to the intervention component in time. A message intended to reduce craving may use craving over the next hour; a prompt intended to increase activity may use steps during a subsequent window. The outcome must be clinically meaningful enough to support the intervention mechanism, but short enough that competing events and long chains of treatment exposure do not overwhelm the interpretation.</p><p>The principal estimand is often a <strong>causal excursion effect</strong>: the contrast in the expected proximal outcome if an eligible participant received one intervention option rather than another at a decision point, averaged over the observed distribution of history and context. This is not automatically the effect on a distal endpoint such as hospitalization, remission, or long-term mortality. A statistically precise proximal effect may fail to translate into durable clinical benefit if the component is burdensome, effects decay rapidly, or repeated exposure changes behavior.</p><p>Time-varying moderation asks whether the effect differs by variables measured before randomization at that decision point. Candidate moderators may include recent stress, location, time of day, prior activity, symptom burden, or treatment history. Moderation terms should be prespecified or clearly labeled exploratory. A large collection of interaction terms can produce unstable findings, especially when availability is rare or the effective number of randomized opportunities is smaller than the raw row count suggests.</p><h3>Analysis of the repeated person-time data</h3><p>MRT data usually contain multiple rows per participant and many decision points per day. The treatment indicator is randomized only at available opportunities, while the outcome may be missing because the participant does not respond or the sensor fails. A valid analysis must preserve the decision-point structure and account for within-person dependence.</p><p>Common analyses use a weighted and centered treatment indicator in a longitudinal regression model, with participant-level clustering or robust standard errors. The model may include time functions, baseline characteristics, time-varying moderators, and treatment-by-moderator interactions. The randomization probability and availability rule determine how treatment is centered and weighted. The exact estimator should be chosen from the estimand and design, not copied from a standard parallel-arm RCT.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YiV0AG6hKHFLeKYU.png" /></figure><p><strong>Interpretation boundary:</strong> A repeated randomization protects the treatment contrast at eligible decision points. It does not automatically protect analyses of treatment adherence, cumulative exposure, long-term outcomes, or post-randomization subgroups. Those questions may require additional causal methods and assumptions.</p><h3>Sample size and information in an MRT</h3><p>More decision points do not create unlimited independent information. The effective information depends on the number of participants, the number of available opportunities, treatment randomization probabilities, within-person correlation, outcome variance, and the strength of any time-varying moderation. A trial with many opportunities but little variation in availability or treatment uptake may have less information than its row count suggests.</p><p>Sample-size planning should therefore use design-based simulation whenever the decision schedule, missingness, autocorrelation, moderation, or treatment burden is complex. The simulation should reproduce the expected number of decision points, availability process, randomization probabilities, proximal outcome model, clustering, and the effect size that would be clinically meaningful. Reporting only a conventional two-arm sample-size calculation can misrepresent the precision of an MRT.</p><h3>Data quality and implementation risks</h3><p>Digital trials add measurement and delivery risks. Device non-wear, delayed notifications, battery failure, missing ecological momentary assessments, changes in operating-system permissions, and intervention fatigue can all affect availability and outcome measurement. These are not merely technical footnotes. If data capture or delivery changes over calendar time, the estimated effect may mix intervention efficacy with implementation drift.</p><p>Researchers should monitor randomization balance at eligible opportunities, treatment delivery, response latency, missing proximal outcomes, participant burden, and changes in the number of available decision points. Diagnostics should be defined before unblinding outcome patterns when possible. The analysis should distinguish a failure to deliver an assigned component from a null biological or behavioral effect.</p><h3>Five-step MRT workflow</h3><h3>Evidence Summary Table</h3><h3>Common reporting errors</h3><ul><li>Calling a high-frequency longitudinal study an MRT without describing the randomization at each decision point.</li><li>Failing to define availability and then interpreting unavailable opportunities as treatment failures.</li><li>Reporting proximal effects as evidence of long-term clinical benefit without a distal outcome analysis.</li><li>Ignoring within-person clustering or using a model that does not match the randomization probabilities.</li><li>Testing many moderators without prespecification, multiplicity control, or uncertainty reporting.</li><li>Reporting the number of rows instead of the number of participants, eligible opportunities, randomized opportunities, and observed proximal outcomes.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uSAyZ1X9gYMGQD7Q.png" /></figure><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI tools to connect MRT design, analysis, and publication decisions:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: audit MRT randomization, availability rules, proximal outcomes, and causal claims.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: organize JITAI and MRT evidence with traceable citations and design comparisons.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with digital health, clinical trials, biostatistics, and intensive longitudinal methods.</li></ul><h3>Conclusion</h3><p>Micro-randomized trials are designed for a narrow but important problem: learning how intervention components work at repeated, context-dependent decision points. Their strength comes from aligning randomization, availability, proximal outcomes, and time-varying moderators. Their limitations are equally important. Repeated observations do not eliminate within-person dependence, high-frequency opportunities do not guarantee adequate information, and a short-term causal effect does not establish long-term clinical benefit. A credible MRT report makes the decision schedule, randomization mechanism, availability rule, estimand, analysis model, diagnostics, and interpretation boundary visible.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/micro-randomized-trials-jitai-clinical-research"><em>https://sci.lingcorehealth.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6163c355f878" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Latent Class Trajectory Models in Clinical Research]]></title>
            <link>https://medium.com/@lifeflexi818/latent-class-trajectory-models-in-clinical-research-c1deeabca0be?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1deeabca0be</guid>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[latentclasstrajectory]]></category>
            <category><![CDATA[longitudinal-data]]></category>
            <category><![CDATA[trajectory-analysis]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 01:17:18 GMT</pubDate>
            <atom:updated>2026-10-28T01:17:18.036Z</atom:updated>
            <content:encoded><![CDATA[<p><a href="https://sci.lingcorehealth.com/blog.html">← Research Methodology Blog</a></p><p><strong>Latent class trajectory models</strong> identify unobserved subgroups with similar longitudinal patterns. They can summarize heterogeneity in symptoms, biomarkers, adherence, or utilization, but classes are <strong>probabilistic model-based summaries</strong>, not directly observed diseases or guaranteed clinical phenotypes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DrqMdamxY97DlSYq.png" /></figure><p>Clinical cohorts rarely follow one average path. Pain scores may improve quickly for one subgroup, remain stable for another, and worsen for a third. A biomarker can show early response, delayed response, or persistent elevation. Medication adherence can fluctuate rather than follow a single population mean. <strong>Latent class trajectory modeling</strong> provides a framework for describing such patterns by estimating a finite set of unobserved classes with distinct longitudinal trajectories.</p><p>The method is attractive because it turns repeated measurements into an individual-centered description of heterogeneity. It is also easy to overinterpret. The estimated classes depend on the selected outcome distribution, time scale, polynomial form, number of classes, missing-data assumptions, and sample size. A trajectory class is not automatically a biological subtype, a treatment target, or a causal mechanism.</p><h3>What question does a trajectory model answer?</h3><p>A trajectory analysis asks whether the observed longitudinal data are better described by several groups with different patterns over time than by a single population-average curve or a conventional random-effects model alone. The primary target is usually descriptive: identify clinically interpretable patterns and estimate the probability that participants belong to each pattern.</p><p>The question should be written before model fitting. Examples include:</p><ul><li>Do patients with chronic pain show distinct symptom courses during follow-up?</li><li>Are there subgroups with different medication-adherence patterns after treatment initiation?</li><li>Do biomarker trajectories identify groups with different subsequent risk?</li><li>Does a behavioral or functional measure follow different developmental or recovery paths?</li></ul><p>These are not interchangeable questions. A model describing latent patterns is different from a causal analysis of whether an exposure changes trajectory membership. If class membership is later used as a predictor of outcome, the analysis must account for classification uncertainty and avoid treating estimated labels as error-free observations.</p><h3>GBTM, growth mixture modeling, and related approaches</h3><p>Approach Core idea Typical use Important boundary</p><p>Terminology is inconsistent across applied literature. Some papers use “trajectory analysis,” “latent class growth analysis,” “group-based trajectory modeling,” and “growth mixture modeling” as if they were interchangeable. They are not fully interchangeable. The methods differ in whether within-class random effects are included, how the outcome distribution is specified, and how class membership uncertainty is handled.</p><h3>Data structure and time scale</h3><p>Trajectory models require repeated measurements aligned to a meaningful time scale. The time variable may be age, months since diagnosis, time since treatment initiation, days after discharge, or study visit. A clinically plausible time origin is more important than a convenient index. If participants enter at different disease stages, calendar time may obscure the process of interest.</p><p>At least several measurement occasions are usually needed to distinguish a trajectory from noise. Two observations per participant may describe change, but they provide limited information for estimating curvature or separating classes. More time points improve the opportunity to assess nonlinear patterns, although repeated measurements do not solve sparse outcome information or informative dropout.</p><p>The outcome distribution should follow the measurement scale. Continuous approximately Gaussian outcomes may use normal trajectories; counts may require Poisson or negative-binomial distributions; binary outcomes require a suitable link; bounded or censored outcomes require additional modeling decisions. A normal model for a highly skewed count or a bounded score can produce classes that reflect distributional misspecification rather than clinically meaningful patterns.</p><h3>Choosing the trajectory shape</h3><p>Within each class, the mean trajectory may be linear, quadratic, cubic, spline-based, or otherwise structured. Higher-order polynomials can fit complex curvature but may create implausible behavior at the boundaries and unstable extrapolation. A visually appealing curve is not sufficient evidence that the underlying clinical process has that shape.</p><p>Time should be centered and scaled when appropriate, especially when polynomial terms are included. Researchers should pre-specify a small set of plausible shapes, compare them using fit criteria and clinical diagnostics, and display the fitted curves over the observed data. The selected model should not be chosen solely by the lowest information criterion if it creates tiny classes, implausible turning points, or clinically uninterpretable patterns.</p><h3>Class enumeration: how many classes?</h3><p>Class enumeration is the process of deciding how many latent groups to retain. BIC, likelihood-based criteria, posterior probabilities, entropy, class size, stability across starting values, and clinical interpretability may all contribute. No single criterion determines the correct number of classes.</p><p>Adding classes almost always improves or preserves some measure of fit. That does not mean every additional class represents a real subgroup. Very small classes can be artifacts of outliers, non-normality, missingness, or an overly restrictive model. A class solution that changes substantially with starting values should be treated as unstable rather than presented as a settled discovery.</p><p>Researchers should report the candidate class counts considered, the information criteria, average posterior probabilities, class sizes, convergence checks, and the substantive reason for selecting the final solution. When the two-class and three-class solutions tell the same clinical story, the simpler model may be preferable. When a small additional class changes clinical interpretation, the decision should be justified and sensitivity analyses should be shown.</p><h3>Posterior probabilities and classification uncertainty</h3><p>Each participant receives posterior probabilities of membership across the estimated classes. The highest-probability class is often used for description, but it is not a directly observed label. A participant with probabilities of 0.52, 0.43, and 0.05 is not equivalent to a participant with probabilities of 0.98, 0.01, and 0.01.</p><p>Classification quality can be summarized by average posterior probabilities, classification error, and related diagnostics. If class membership is used in a subsequent regression, survival analysis, or causal model, a one-step “assign then analyze” approach can attenuate associations or create bias. Three-step approaches or probability-based procedures may better preserve uncertainty, depending on the software and estimand.</p><p>Class labels should also be named cautiously. “High-risk trajectory” describes the modeled pattern; it does not establish a biological disease subtype. “Non-responder” implies a treatment response definition and should not be used merely because a curve is relatively flat. Naming should follow the observed scale, time window, and model output.</p><h3>Missing data, dropout, and informative observation</h3><p>Longitudinal studies commonly have intermittent missing measurements and dropout. Full-information maximum likelihood or multiple imputation may be reasonable under assumptions about the missing-data mechanism, but the assumptions must be stated. If dropout depends on unobserved future outcome or latent trajectory membership, standard missing-at-random assumptions may be inadequate.</p><p>Trajectory classes can be especially sensitive to missingness because dropout may remove the participants most likely to worsen or improve. Compare observation patterns by baseline characteristics and early outcome values. Report the number and timing of measurements, reasons for dropout, follow-up completeness, and whether sensitivity analyses alter class definitions or downstream associations.</p><p>Informative observation schedules also matter. Patients with worsening symptoms may be measured more frequently, while stable patients may have fewer visits. A model that treats irregular observations as if they were scheduled identically can confuse healthcare utilization with disease trajectory. Joint models, inverse-intensity weighting, or other approaches may be relevant when observation processes are clearly informative.</p><h3>Predictors of trajectory membership and downstream outcomes</h3><p>After estimating trajectories, researchers often examine baseline predictors of class membership or compare later outcomes across classes. These analyses are useful but require careful ordering. Predictors should be defined before class membership when the goal is to explain who follows a trajectory. Variables measured after the trajectory begins may be mediators, consequences, or selection variables rather than baseline predictors.</p><p>For downstream outcome analyses, classification uncertainty should be carried forward. If the outcome is survival, a naive comparison of hard-assigned classes can exaggerate differences because the same data helped define the classes and the uncertainty is ignored. The analysis should state whether the trajectory model is exploratory, whether class membership is treated as exposure, and how uncertainty and confounding are addressed.</p><p>A trajectory association is not a causal effect. If treatment influences class membership and later outcome, comparing classes may condition on a post-treatment variable and create selection bias. Causal questions require a target estimand, a time-varying treatment framework, and assumptions appropriate to the data-generating process.</p><h3>Reporting requirements and reproducibility</h3><p>A reproducible report should provide the time origin, measurement schedule, outcome distribution, link function, missing-data handling, class-specific trajectory form, random-effects structure when applicable, estimation method, starting values, convergence information, class enumeration results, posterior probabilities, and sensitivity analyses.</p><p>Plots should show raw or summarized observations together with fitted class trajectories and uncertainty. A plot of only the smooth class means can make a weakly supported solution look definitive. Include class sizes, uncertainty, and the observed follow-up range. If curves are shown beyond the observed time range, label that region as extrapolation and avoid implying direct data support.</p><h3>Five-step latent trajectory workflow</h3><p>Step Action Output</p><p><strong>Step 1: Define the longitudinal question</strong></p><p>Specify the population, time origin, repeated outcome, follow-up window, and whether the aim is description, prediction, or causal analysis.</p><p>Pre-specified trajectory estimand and data map</p><p><strong>Step 2: Build the measurement model</strong></p><p>Choose the outcome distribution, link, time scale, plausible trajectory shapes, and missing-data strategy.</p><p>Candidate model set with clinical rationale</p><p><strong>Step 3: Enumerate classes transparently</strong></p><p>Fit multiple class counts and starting values; inspect BIC, posterior probabilities, class size, convergence, and clinical plausibility.</p><p>Class-enumeration table and diagnostics</p><p><strong>Step 4: Preserve uncertainty</strong></p><p>Report posterior membership probabilities and use an uncertainty-aware procedure for predictors or downstream outcomes.</p><p>Probabilistic class summaries and uncertainty intervals</p><p><strong>Step 5: Stress-test interpretation</strong></p><p>Vary time shape, missingness assumptions, class count, and model specification; separate descriptive patterns from causal claims.</p><p>Robustness analysis and bounded clinical interpretation</p><h3>Evidence Summary Table</h3><p>Source Evidence type Main contribution Reporting implication</p><h3>Researcher’s Toolkit</h3><h3>Move from longitudinal data to defensible interpretation</h3><p>Use Lingcore SCI tools to keep trajectory modeling connected to the broader research workflow:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: audit whether a trajectory study reports class enumeration, posterior uncertainty, missing-data assumptions, and downstream analysis boundaries.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: organize verified literature on GBTM, GMM, LTA, missing longitudinal data, and reporting standards.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Journal Matcher</strong></a>: compare journals suited to longitudinal methods, epidemiology, biostatistics, and clinical prediction research.</li></ul><h3>Conclusion</h3><p>Latent class trajectory models can make clinically important longitudinal heterogeneity visible. Their value depends on disciplined specification and restrained interpretation. Researchers should distinguish GBTM from GMM and related approaches, align the outcome distribution and time scale with the clinical process, evaluate several class solutions, report posterior uncertainty, address missingness and informative observation, and avoid treating latent classes as observed diseases. A trajectory model is strongest when it clarifies a pre-specified longitudinal question and remains explicit about what the data identify and what still depends on modeling assumptions.</p><p>© 2026 Lingcore Health SCI. Not affiliated with any medical institution. For research purposes only. This article is methodological information, not clinical advice or a treatment recommendation.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/latent-class-trajectory-models-clinical-research"><em>https://sci.lingcorehealth.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1deeabca0be" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Principal Stratification for Clinical Trial Estimands]]></title>
            <link>https://medium.com/@lifeflexi818/principal-stratification-for-clinical-trial-estimands-8eaa29b90df5?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/8eaa29b90df5</guid>
            <category><![CDATA[intercurrentevents]]></category>
            <category><![CDATA[principalstratification]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[clinical-trials]]></category>
            <category><![CDATA[causal-inference]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 00:00:59 GMT</pubDate>
            <atom:updated>2026-10-27T08:07:20.589Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Principal stratification</strong> defines latent subgroups by participants’ joint potential outcomes for a post-randomization variable. It can target effects among always-survivors, compliers, protected patients, or other principal strata, but those groups are usually unobserved. Estimates therefore depend on design, cross-world assumptions, monotonicity or exclusion restrictions, and sensitivity analyses; a principal-stratum effect is not automatically the intention-to-treat effect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1SfOoSosECM7hJJ9.png" /></figure><h3>Why post-randomization variables create a causal problem</h3><p>Randomization identifies the intention-to-treat effect when treatment assignment is implemented and outcomes are observed according to the trial protocol. Clinical trials also produce variables after randomization: treatment receipt, adherence, rescue medication, biomarker response, hospitalization, survival, withdrawal, and other intercurrent events. These variables may affect how an investigator wants to interpret the outcome, but conditioning on an observed post-treatment variable can destroy the protection provided by randomization.</p><p>Consider a trial in which treatment assignment affects whether a patient receives a rescue intervention, and rescue intervention is also related to prognosis. Comparing treated and control patients within the observed rescue groups can mix treatment effects with selection effects. The subgroup “patients who received rescue treatment” is not a baseline randomized subgroup. Its composition may differ across arms because treatment assignment helped determine membership.</p><p>Principal stratification addresses this problem by defining subgroups through potential values of the post-randomization variable under every treatment assignment. The strata are counterfactual: they classify a participant according to what would have happened under treatment and under control, even though only one of those potential values is observed for each participant.</p><h3>The potential-outcome definition</h3><p>Let <em>Z</em> denote randomized assignment, <em>S</em> an intermediate or intercurrent-event variable, and <em>Y</em> the final outcome. For a binary intermediate variable, each participant has two potential values:</p><p>S 1i = intermediate outcome under treatment; S 0i = intermediate outcome under control</p><p>The principal stratum for participant <em>i</em> is defined by the pair . The principal-stratum-specific causal effect compares the potential final outcomes under treatment and control within that latent subgroup:</p><p>The definition is conceptually simple but empirically difficult. A participant assigned to treatment reveals information about , not . A participant assigned to control reveals information about , not . The joint pair is therefore only partially observed.</p><h3>Common principal strata</h3><p>Stratum Potential intermediate outcomes Illustrative interpretation</p><p><strong>Always-event-free or always-survivor</strong></p><p>The participant would remain event-free or survive under either assignment by the specified time.</p><p><strong>Protected</strong></p><p>The intermediate event is avoided under treatment but occurs under control, depending on how the variable is coded.</p><p><strong>Harmed</strong></p><p>The intermediate event occurs under treatment but not control.</p><p><strong>Never-event-free or never-responder</strong></p><p>The participant has the intermediate event under both assignments or does not respond under either assignment.</p><p>The labels depend on coding and clinical context. “Protected” may refer to survival, absence of toxicity, treatment receipt, or biomarker response. Researchers should define the variable and direction before naming strata. A label is not a substitute for the counterfactual definition.</p><h3>Principal stratification is not ordinary subgroup analysis</h3><p>A baseline subgroup such as age below 65 years is observed before randomization. Randomization remains valid within that subgroup if the trial and analysis support the comparison. A principal stratum is different: membership is determined by potential post-randomization values and is generally unobserved.</p><p>This distinction matters for interpretation. An analysis that reports the treatment effect among “patients who adhered” may be describing a post-treatment selected subgroup. An analysis that targets the effect among “compliers” is a principal-stratification analysis only if compliance strata are defined through potential compliance under both assignments and the identifying assumptions are stated.</p><p>Principal stratification should also be distinguished from mediation. A mediator is an intermediate variable through which treatment may operate. A principal stratum is a latent subgroup defined by joint potential values of that variable. Conditioning on a principal stratum does not automatically identify a natural direct effect, and a principal-stratum-specific contrast should not be relabeled as mediation without a separate causal definition.</p><h3>Noncompliance and the complier average causal effect</h3><p>Noncompliance is one of the most familiar applications. Let and denote treatment receipt under assignment to treatment and control. Under a monotonicity assumption that rules out individuals who would take treatment only when assigned to control, the strata may include always-takers, compliers, and never-takers.</p><p>The complier average causal effect (CACE), also called the local average treatment effect in the instrumental-variable literature, targets the causal effect among participants whose treatment receipt is changed by assignment. This is not necessarily the effect among all randomized participants. It is a local effect for a latent population whose membership cannot usually be identified for an individual.</p><p>Monotonicity is substantive, not merely algebraic. It asserts that the instrument or assignment does not move some participants in opposite directions. If defiers are plausible, the standard CACE interpretation may fail or require alternative bounds and sensitivity analyses. Exclusion restrictions, treatment versions, stable unit treatment value, and sufficiently strong first-stage association also matter.</p><h3>Truncation by death and always-survivor effects</h3><p>When treatment affects survival, a quality-of-life or biomarker outcome measured after a fixed time is undefined for participants who have died before that time. Comparing observed outcomes among survivors can be misleading because survival itself is post-treatment and may be affected by assignment.</p><p>A principal-stratification estimand may target the effect among always-survivors: participants who would survive to the assessment time under both treatment and control. This can be clinically meaningful when the scientific question concerns outcome quality among patients who could be observed under either treatment. It does not answer the intention-to-treat effect on the combined outcome “survival and quality of life,” and it may not be identifiable without strong assumptions.</p><p>Alternatives include composite outcomes, while-alive estimands, survivors-average causal effects, semi-competing-risk models, or estimands that assign a value to death. The choice should be made before analysis and aligned with the trial’s clinical question. Principal stratification is one option, not a universal correction for missing post-treatment outcomes.</p><h3>Intercurrent events and ICH E9(R1)</h3><p>The ICH E9(R1) estimand framework treats intercurrent events as events that occur after treatment initiation and affect the interpretation or existence of the outcome. Principal stratification is one strategy for addressing such events. It asks about treatment effects within a subgroup defined by potential outcomes for the intercurrent event under each treatment condition.</p><p>Other strategies include treatment policy, hypothetical, composite, and while-on-treatment strategies. These strategies answer different questions. For example, a treatment-policy strategy may compare outcomes regardless of rescue medication, whereas a hypothetical strategy may ask what would have happened if rescue had not occurred. A principal-stratum strategy may ask about the effect among patients who would have the same or specified intermediate-event status under both assignments.</p><p>The estimand should therefore be written in words and notation before selecting an estimator. The analysis section should state the treatment conditions, population, endpoint, time frame, intercurrent-event strategy, summary measure, and assumptions. A statistical model cannot repair an estimand that was never defined.</p><h3>Identification: what must be assumed</h3><p>Because principal strata are latent, identification requires more than randomization. The exact assumptions depend on the endpoint and data structure, but common ingredients include:</p><ul><li><strong>Consistency:</strong> the observed intermediate and final outcomes under the assigned treatment correspond to the relevant potential outcomes.</li><li><strong>Random assignment:</strong> treatment assignment is independent of potential outcomes under the trial design, subject to implementation and analysis conditions.</li><li><strong>Positivity:</strong> treatment assignment probabilities and relevant observation processes support the comparisons required by the estimator.</li><li><strong>Monotonicity:</strong> in some noncompliance analyses, certain response types are ruled out, such as defiers.</li><li><strong>Exclusion restrictions:</strong> in instrumental-variable settings, the instrument affects the final outcome only through treatment receipt, if that restriction is used.</li><li><strong>Cross-world restrictions:</strong> assumptions connect potential outcomes under different treatment assignments, because both intermediate outcomes cannot be observed for one person.</li><li><strong>Measurement and missingness assumptions:</strong> intermediate events, final outcomes, censoring, and loss to follow-up are measured in a way compatible with the model.</li></ul><p>Some assumptions can be partially examined; many cannot be tested from observed trial data alone. A credible analysis labels assumptions by role, explains which are design-based and which are modeling choices, and shows how conclusions change when assumptions are relaxed.</p><h3>Estimation and sensitivity analysis</h3><p>Principal-stratification estimators may use likelihood-based latent-class models, Bayesian models, instrumental-variable methods, bounds, weighting, g-computation, or semiparametric approaches. The estimator should follow the estimand and data structure rather than being chosen because a software package produces a principal-stratum label.</p><p>Posterior probabilities can summarize uncertainty about stratum proportions, but a posterior classification is not an observed subgroup assignment. Reporting a table that assigns each patient to “complier” or “always-survivor” without acknowledging classification uncertainty can overstate what the data identify.</p><p>Sensitivity analysis is central. Vary cross-world association parameters, monotonicity, exclusion restrictions, outcome models, missing-data assumptions, and prior distributions when a Bayesian model is used. Report ranges or curves of principal-stratum effects over scientifically plausible values. If conclusions change across modest assumption ranges, the primary result should be described as assumption-sensitive rather than definitive.</p><h3>Five-step principal-stratification workflow</h3><p><strong>Step 1: Define the intercurrent event</strong></p><p>Name the post-randomization variable, coding direction, measurement time, and final outcome that may be undefined or selected.</p><p>Explicit estimand statement and potential-outcome notation</p><p><strong>Step 2: Define the principal strata</strong></p><p>List the joint potential values under treatment and control; do not rely on informal labels such as “adherent patients.”</p><p>Stratum table with clinical interpretation</p><p><strong>Step 3: State identification assumptions</strong></p><p>Specify randomization, consistency, positivity, monotonicity, exclusion restrictions, cross-world restrictions, and missingness assumptions.</p><p>Assumption matrix with sensitivity parameters</p><p><strong>Step 4: Estimate with uncertainty</strong></p><p>Use an estimator aligned with the estimand and report uncertainty in stratum proportions and stratum-specific effects.</p><p>Effect estimates, intervals, and latent-class uncertainty</p><p><strong>Step 5: Stress-test interpretation</strong></p><p>Vary assumptions, compare with intention-to-treat and alternative intercurrent-event strategies, and avoid individual-level stratum claims.</p><p>Robustness figure and bounded clinical conclusion</p><h3>Evidence Summary Table</h3><h3>Researcher’s Toolkit</h3><h3>Make intercurrent-event estimands explicit</h3><p>Principal stratification is useful only when the latent target and its assumptions are made visible. Lingcore SCI tools support the surrounding research workflow:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: Audit whether clinical-trial papers distinguish intention-to-treat, per-protocol, complier, survivor, and principal-stratum effects.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: Build an evidence-based review of intercurrent-event strategies with verified methodological and regulatory sources.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Journal Matcher</strong></a>: Identify journals suited to causal inference, clinical trials, estimands, and regulatory methodology manuscripts.</li></ul><h3>Conclusion</h3><p>Principal stratification provides a disciplined language for causal questions involving post-randomization variables. It can target effects among compliers, always-survivors, protected patients, or other latent response types without conditioning directly on an observed post-treatment variable. Its central limitation is equally important: principal strata are usually unobserved, and their effects depend on assumptions linking counterfactual outcomes. Clinical researchers should define the estimand before choosing the estimator, state identification assumptions in a structured way, report uncertainty in latent strata, compare with intention-to-treat and alternative strategies, and use sensitivity analyses to show how much of the conclusion is assumption-driven.</p><p>© 2026 Lingcore Health SCI. Not affiliated with any medical institution. For research purposes only. This article is methodological information, not clinical advice or a treatment recommendation.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/principal-stratification-intercurrent-events-clinical-trials"><em>https://sci.lingcorehealth.com</em></a><em> on October 27, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8eaa29b90df5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Flexible Parametric Survival Models: Royston-Parmar]]></title>
            <link>https://medium.com/@lifeflexi818/flexible-parametric-survival-models-royston-parmar-e35f8268aa36?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/e35f8268aa36</guid>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[clinicalprediction]]></category>
            <category><![CDATA[roystonparmar]]></category>
            <category><![CDATA[survival-analysis]]></category>
            <category><![CDATA[clinical-research]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 00:00:32 GMT</pubDate>
            <atom:updated>2026-10-26T03:13:14.114Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Flexible parametric survival models</strong>, especially the <strong>Royston-Parmar model</strong>, use restricted cubic splines to model the baseline survival process smoothly. They can produce absolute survival and risk estimates at arbitrary time points, represent time-varying effects, and support competing-risk or extrapolation analyses, but their credibility depends on knot choice, scale selection, validation, and transparent assumptions beyond observed follow-up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*roe1rW8A5KvPmTqo.png" /></figure><h3>Why move beyond a Cox baseline hazard?</h3><p>The Cox proportional hazards model is valuable because it estimates relative effects without specifying the baseline hazard. That convenience also limits some clinical outputs. A standard Cox analysis does not directly parameterize a smooth baseline survival function, and absolute risk estimates are often obtained afterward through stepwise estimators at observed event times.</p><p>Flexible parametric survival models take a different route. They model a transformed baseline survival quantity with a parametric function, commonly a restricted cubic spline of log time. The resulting curve is smooth over the modeled time range. Researchers can estimate survival probability, cumulative hazard, hazard, risk differences, restricted mean survival time, and other quantities directly from the fitted model.</p><p>The gain is not that splines are automatically more correct. The gain is that the model makes the time structure explicit and allows clinical questions to be expressed on an absolute-risk scale. The cost is a greater need to justify the transformation, degrees of freedom, knot placement, effect scale, and behavior beyond the data.</p><h3>What the Royston-Parmar model represents</h3><p>In the proportional-hazards formulation, the model represents the log cumulative hazard as a smooth function of log time plus covariate effects:</p><p>log H(t | X) = s(log t; γ, knots) + β TX</p><p>The function <em>s</em> is a restricted cubic spline. It is constructed from basis functions joined at knots, with linear behavior in the tails beyond the boundary knots. The spline can represent a baseline cumulative hazard that changes direction over time, unlike a simple Weibull model whose hazard shape is constrained by a small number of parameters.</p><p>The same family can be fitted on other scales. A proportional-odds formulation models a transformation related to survival odds, and a probit or normal-scale formulation provides another link between the linear predictor and the survival function. The selected scale is part of the model specification, not a cosmetic choice.</p><h3>Flexible parametric survival models versus common alternatives</h3><h3>Knot placement and degrees of freedom</h3><p>The number of interior knots controls how much the baseline function can bend. Too few degrees of freedom may underfit important hazard changes. Too many can follow random fluctuations, especially when the number of events is limited. Knot placement also matters because the data may contain dense event information in one time region and sparse information in another.</p><p>Clinical researchers should report the number of knots, the scale on which the spline was fitted, the rule used to place knots, boundary handling, and any sensitivity analyses. A model is not reproducible when the article says only “a flexible parametric model was fitted.” The spline basis and time origin are part of the estimand’s operational definition.</p><p>Visual inspection should be paired with quantitative checks. Compare observed Kaplan-Meier or cumulative-incidence estimates with model-based curves, inspect calibration at clinically relevant horizons, examine residual or pseudo-residual behavior when available, and assess whether the fitted hazard has implausible oscillations. Knot selection should not be based solely on the lowest in-sample information criterion.</p><h3>Time-varying effects within the model</h3><p>A major advantage of flexible parametric survival models is that covariate effects can vary smoothly with time. Instead of imposing a constant coefficient β, the model can include an interaction between a covariate and spline functions of log time:</p><p>log H(t | X) = s 0(log t) + βX + s X(log t) × X</p><p>This allows a treatment effect or prognostic association to change over follow-up. The resulting curve may show an early treatment difference that attenuates, a delayed separation, or a changing association in chronic disease. Such patterns should be interpreted on the scale of the model and translated into time-specific hazard ratios, survival differences, risk differences, or RMST contrasts.</p><p>A time-varying coefficient is not evidence that the treatment effect truly changes biologically. It may reflect treatment switching, informative follow-up, competing events, non-proportional measurement processes, model misspecification, or random variation. The analysis should prespecify how time variation will be evaluated and should show uncertainty bands around the estimated effect.</p><h3>Absolute risk and clinically interpretable estimands</h3><p>Hazard ratios are relative measures. Clinical decisions often require absolute quantities: the probability of surviving to 12 months, the risk difference by 24 months, the number of events prevented per 1,000 patients, or the mean event-free time over a restricted horizon. Flexible parametric models are useful because they provide a smooth estimated survival function from which these quantities can be calculated at arbitrary times within the supported range.</p><p>For example, a treatment comparison can be expressed as survival probability under each treatment at 6, 12, and 24 months, an absolute risk difference at each horizon, or a restricted mean survival-time difference through 24 months. These summaries can be more interpretable than a single hazard ratio when hazards are non-proportional or when the clinical decision concerns a fixed time point.</p><p>Absolute predictions must be validated. A smooth curve can look persuasive while being miscalibrated. Report calibration-in-the-large, calibration slope or flexible calibration plots, prediction error, and uncertainty at prespecified horizons. If the model is intended for individual risk prediction, also assess transportability across sites, calendar periods, disease severity, and clinically important subgroups.</p><h3>Competing risks and multiple event processes</h3><p>When a competing event prevents the event of interest from occurring, standard survival probabilities require careful interpretation. Flexible parametric frameworks can be extended to model cause-specific hazards, cumulative incidence functions, or other competing-risk estimands. These are not interchangeable.</p><p>A cause-specific hazard describes the instantaneous event rate among individuals who are still free of any event. A cumulative incidence function describes the probability of experiencing a particular event by a time point in the presence of competing events. A model that produces a smooth cause-specific hazard does not automatically provide a direct model of the clinical probability of the event of interest. The article should name the estimand before describing the model.</p><p>For competing-risk analyses, report how event types were coded, whether covariates were allowed to have cause-specific time-varying effects, how cumulative incidence was derived, and how uncertainty was calculated. If the clinical question concerns treatment policy or resource planning, cumulative incidence and absolute risk may be more relevant than a cause-specific hazard ratio.</p><h3>Extrapolation beyond observed follow-up</h3><p>Flexible parametric models are frequently used to estimate survival beyond the end of a trial or registry. The smooth mathematical form makes extrapolation possible, but it does not make the extrapolated tail identifiable from the observed data. Once the curve extends beyond the region with reliable follow-up, the result is driven increasingly by the chosen functional form and external assumptions.</p><p>A defensible extrapolation analysis should show the observed follow-up boundary, identify where censoring becomes substantial, compare multiple plausible models, and perform sensitivity analyses using alternative tail behaviors. External mortality or registry information may help in relative-survival or health-economic settings, but such information must be aligned with population, calendar period, age, and outcome definitions.</p><p>Do not present a single extrapolated mean survival time as if it were observed. Display prediction intervals or confidence intervals, distinguish parameter uncertainty from structural uncertainty, and state which parts of the result depend on assumptions that cannot be checked using the trial data alone.</p><h3>Five-step flexible survival modeling workflow</h3><p><strong>Step 1: Define the estimand</strong></p><p>Specify the event, competing events, time origin, follow-up, target population, scale, and whether the primary target is hazard, survival, risk, RMST, or cumulative incidence.</p><p>Versioned estimand and event-coding plan</p><p><strong>Step 2: Diagnose the time structure</strong></p><p>Inspect Kaplan-Meier or cumulative-incidence curves, event density, censoring, and plausible non-proportionality before choosing the baseline form.</p><p>Pre-model diagnostic plots and rationale</p><p><strong>Step 3: Fit and compare candidate forms</strong></p><p>Vary degrees of freedom, knot placement, link scale, and time-varying-effect terms; compare fit while prioritizing clinical plausibility.</p><p>Model specification and sensitivity table</p><p><strong>Step 4: Validate absolute predictions</strong></p><p>Assess calibration, discrimination, prediction error, subgroup performance, and uncertainty at prespecified horizons.</p><p>Validation plots and horizon-specific estimates</p><p><strong>Step 5: Translate and stress-test</strong></p><p>Report absolute risks, survival differences, RMST, or cumulative incidence; mark extrapolation and test alternative assumptions.</p><p>Clinically interpretable estimates with structural uncertainty</p><h3>Evidence Summary Table</h3><p>Source Evidence type Methodological contribution Reporting implication</p><h3>Researcher’s Toolkit</h3><h3>Strengthen your flexible survival analysis workflow</h3><p>Royston-Parmar models require careful estimand definition, spline specification, validation, and transparent extrapolation. Lingcore SCI tools support the full research workflow:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: Audit survival-model papers for proportionality, spline specification, competing-risk estimands, calibration, and extrapolation claims.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: Build an evidence-based review of flexible parametric survival models with verified methodological sources.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Journal Matcher</strong></a>: Identify journals suited to survival analysis, clinical prediction, health economics, and clinical epidemiology manuscripts.</li></ul><h3>Conclusion</h3><p>Flexible parametric survival models make the time structure of a survival analysis explicit. The Royston-Parmar approach can model complex baseline hazard shapes, estimate smooth absolute risks, represent time-varying effects, and support clinically useful summaries such as survival differences and RMST. These advantages depend on disciplined specification. Researchers should report knots, degrees of freedom, modeled scale, time origin, competing-risk estimand, validation results, and extrapolation assumptions. A smooth curve is a modeling choice, not a guarantee of truth beyond the observed data.</p><p>© 2026 Lingcore Health SCI. Not affiliated with any medical institution. For research purposes only. This article is methodological information, not clinical advice or a treatment recommendation.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/flexible-parametric-survival-royston-parmar"><em>https://sci.lingcorehealth.com</em></a><em> on October 26, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e35f8268aa36" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bayesian Additive Regression Trees in Clinical Research]]></title>
            <link>https://medium.com/@lifeflexi818/bayesian-additive-regression-trees-in-clinical-research-f04ee3a3b695?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/f04ee3a3b695</guid>
            <category><![CDATA[clinicalepidemiology]]></category>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[medical-paper]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[medical-research]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 00:00:24 GMT</pubDate>
            <atom:updated>2026-10-20T10:58:49.049Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Bayesian Additive Regression Trees (BART)</strong> combine many regularized shallow trees to estimate nonlinear relationships and interactions while retaining a posterior distribution for prediction uncertainty. In clinical research, BART can support risk prediction and treatment-effect heterogeneity, but it still requires prespecified targets, leakage-safe validation, calibration assessment, and careful causal assumptions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AnovouO2MwqGUeZw.png" /></figure><h3>What BART estimates</h3><p>BART is a Bayesian sum-of-trees model. Instead of asking one large decision tree to represent the entire relationship between predictors and an outcome, it uses many small trees. Each tree is intentionally constrained so that it acts as a weak learner. The ensemble then adds their contributions to estimate a flexible function of the covariates.</p><p>Here, describes the structure of tree <em>j</em>, contains its terminal-node parameters, and <em>g</em> maps a patient’s covariates to that tree’s contribution. The regularizing prior favors shallow trees and small terminal-node effects. This prevents any single tree from becoming an unstable explanatory model and spreads the signal across the ensemble.</p><p>The Bayesian component is not an ornamental label. BART produces posterior draws of the fitted function, so a prediction can be summarized by a posterior mean or median together with a credible interval. That uncertainty reflects the model and data under the specified prior and likelihood; it is not automatically a complete interval for all clinical uncertainty.</p><h3>Why clinical researchers consider BART</h3><p>Clinical data frequently contain threshold effects, nonlinear associations, and interactions that are difficult to specify in advance. Age may have a different association with risk in patients with severe renal impairment than in patients with preserved function. A treatment may have different estimated effects across baseline risk, disease severity, or biomarker profiles. A conventional linear model can represent these patterns only when the analyst specifies transformations and interaction terms before fitting.</p><p>BART can learn a broad class of nonlinear surfaces without requiring the analyst to enumerate every possible interaction. This makes it useful as a flexible prediction model, a benchmark for a prespecified regression model, or one component of a causal analysis. Flexibility, however, does not turn observational associations into causal effects. The estimand, treatment assignment mechanism, confounding control, positivity, and outcome definition remain central.</p><h3>Prediction versus causal effect heterogeneity</h3><h3>How the ensemble avoids a single unstable tree</h3><p>A deep tree can fit highly specific combinations of predictors and become sensitive to small changes in the sample. BART takes a different route. The prior discourages deep, complex trees, while the sum of many weak trees allows the ensemble to approximate a complicated response surface. The individual trees are not intended to be interpreted as separate clinical pathways. Their joint posterior function is the fitted object.</p><p>This distinction matters when communicating results. A variable that appears often in tree splits is not automatically a causal determinant, and variable inclusion frequency is not a substitute for a prespecified importance analysis. If the scientific goal is explanation, report partial dependence or accumulated local effects only with appropriate caveats, and assess whether the summaries are stable across posterior draws, resamples, and clinically plausible covariate ranges.</p><h3>Posterior uncertainty is part of the result</h3><p>For each patient or covariate pattern, BART can provide a posterior distribution of predicted risk. The distribution may be summarized using a posterior mean and a credible interval. For treatment-effect analysis, the same logic can yield a posterior distribution for a conditional effect or a contrast averaged over a target population.</p><p>Intervals should not be presented as if they solve external validity. A narrow posterior interval can occur when a model is confidently wrong because the target population differs from the development data, because a key predictor was measured differently, or because treatment assignment created unsupported extrapolation. Report the covariate support, missingness, overlap, and validation design alongside the interval.</p><h3>Clinical prediction workflow</h3><p>Step Analysis action Minimum reporting output</p><p><strong>Step 1: Define the target</strong></p><p>Specify population, index time, outcome, prediction horizon, and intended decision before fitting BART.</p><p>Protocol-level estimand and outcome definition</p><p><strong>Step 2: Lock the data pipeline</strong></p><p>Separate development and validation data; perform preprocessing, feature selection, and imputation without using validation outcomes.</p><p>Leakage-safe analysis diagram and missing-data plan</p><p><strong>Step 3: Fit and tune transparently</strong></p><p>Document prior settings, number of trees, burn-in, posterior draws, convergence checks, and computational seeds.</p><p>Reproducible model specification</p><p><strong>Step 4: Validate performance</strong></p><p>Assess discrimination, calibration, prediction error, subgroup performance, and uncertainty in an independent or internal-external validation design.</p><p>Performance estimates with uncertainty intervals</p><p><strong>Step 5: Test clinical relevance</strong></p><p>Compare against a prespecified benchmark and evaluate whether predictions change decisions at plausible thresholds.</p><p>Decision-analytic interpretation, not only an AUC comparison</p><h3>BART for heterogeneous treatment effects</h3><p>In causal research, BART is often used to estimate flexible outcome regressions, propensity models, or treatment-effect surfaces. The method can be attractive when effect modification is likely to be nonlinear or involves interactions that are difficult to prespecify. A causal BART analysis may estimate potential outcomes under each treatment and then contrast them for a target population or covariate pattern.</p><p>The causal interpretation does not come from BART alone. Researchers must define the treatment strategies, time zero, outcome, follow-up, and estimand; identify a sufficient adjustment set; assess whether treated and untreated patients have comparable covariate support; and explain how interference, treatment adherence, missing outcomes, and competing events are handled. Posterior uncertainty quantifies uncertainty conditional on the model and assumptions. It does not measure the bias caused by an unmeasured confounder.</p><p>For subgroup discovery, avoid reporting every apparent interaction. A data-driven subgroup with a favorable posterior treatment effect can be a hypothesis for validation rather than a ready-to-use treatment rule. Report subgroup size, covariate support, multiplicity or selection procedures, calibration of the estimated effects, and performance in held-out patients.</p><h3>Validation and comparison</h3><p>BART should be compared with clinically credible alternatives, not only with weak baselines. Depending on the outcome, benchmarks may include a prespecified logistic or Cox model, penalized regression, generalized additive model, random forest, gradient boosting, or another model already used in the field. The comparison should use the same outcome definition, prediction horizon, missing-data strategy, and validation splits.</p><p>For binary outcomes, report calibration-in-the-large, calibration slope or flexible calibration plots, discrimination, and a proper scoring rule such as the Brier score. For survival outcomes, use time-dependent calibration and prediction-error measures that account for censoring. For treatment-effect models, evaluate both effect-estimation accuracy in simulation or known-effect settings and clinical performance of any treatment rule. A higher ranking metric alone does not establish clinical usefulness.</p><h3>Common reporting errors</h3><ul><li><strong>Calling BART explainable because it uses trees:</strong> the fitted ensemble is not a single decision tree or a set of independent clinical rules.</li><li><strong>Using posterior intervals as causal guarantees:</strong> posterior uncertainty is conditional on the likelihood, prior, data, and causal assumptions.</li><li><strong>Tuning before the validation split:</strong> feature processing, hyperparameter selection, and outcome-informed transformations can leak information.</li><li><strong>Reporting variable importance as mechanism:</strong> predictive contribution does not establish biological or causal importance.</li><li><strong>Ignoring support:</strong> individualized effects are especially fragile where one treatment group has little comparable information.</li><li><strong>Comparing only AUC:</strong> calibration, absolute risk error, uncertainty, decision consequences, and transportability also matter.</li></ul><h3>Evidence Summary Table</h3><p>Source Evidence type Methodological contribution Use in clinical reporting</p><h3>Researcher’s Toolkit</h3><h3>Strengthen your BART and clinical prediction workflow</h3><p>BART analyses require careful target definition, leakage-safe validation, causal assumption review, and manuscript-ready interpretation. Lingcore SCI tools support the full research workflow:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: Audit a BART or machine-learning paper for outcome definition, validation, calibration, uncertainty, and claim-evidence alignment.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: Build an evidence-based review of BART, precision medicine, heterogeneous treatment effects, and clinical prediction with verified citations.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Journal Matcher</strong></a>: Identify journals suited to clinical prediction, biostatistics, causal inference, and precision-medicine manuscripts.</li></ul><h3>Conclusion</h3><p>BART is a flexible Bayesian ensemble for nonlinear prediction and complex effect surfaces. Its value in clinical research depends less on the novelty of using trees than on whether the analysis defines a credible target, protects validation data, quantifies uncertainty, checks calibration and support, and reports causal assumptions explicitly. Treat BART as a statistical tool within a research design, not as a replacement for that design. When the model improves prediction or clarifies clinically plausible treatment-effect heterogeneity in independent data, it can add useful information without overstating what the data can identify.</p><p>© 2026 Lingcore Health SCI. Not affiliated with any medical institution. For research purposes only. This article is methodological information, not clinical advice or a treatment recommendation.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/bayesian-additive-regression-trees-clinical-prediction"><em>https://sci.lingcorehealth.com</em></a><em> on October 20, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f04ee3a3b695" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Zero-Inflated and Hurdle Models for Clinical Counts]]></title>
            <link>https://medium.com/@lifeflexi818/zero-inflated-and-hurdle-models-for-clinical-counts-848b14bb327c?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/848b14bb327c</guid>
            <category><![CDATA[zeroinflatedmodels]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[hurdlemodels]]></category>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[clinicalepidemiology]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 00:00:21 GMT</pubDate>
            <atom:updated>2026-10-19T00:48:18.038Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_BvspDoTgwJuLKDa.png" /></figure><p><strong>Zero-inflated models</strong> treat observed zeros as arising from a mixture of structural and count-generating processes, whereas <strong>hurdle models</strong> use one model for whether any event occurs and a separate truncated-count model for positive values. The correct choice depends on the clinical data-generating process, not on the percentage of zeros alone.</p><p>Clinical studies frequently record outcomes as counts: hospital visits, readmissions, symptom episodes, adverse events, rescue medication use, or healthcare contacts during a defined period. These outcomes are non-negative integers and are often right-skewed. A substantial group may have a count of zero, while a smaller group has a long right tail. That combination can make ordinary linear regression a poor description of the outcome distribution.</p><p>Poisson regression is a natural starting point for counts, but it assumes that the conditional mean and variance are equal. Negative binomial regression relaxes that equality and can accommodate overdispersion. When zeros occur more often than the selected count distribution predicts, analysts may consider zero-inflated or hurdle models. These models are not interchangeable. They encode different assumptions about where zeros come from and how the positive counts are generated.</p><h3>Start with the clinical data-generating process</h3><p>Before choosing a model, define what a zero means. A patient may have no hospital visit because the patient was not susceptible to the event during follow-up, because the event did not occur by chance, because care was received elsewhere, or because follow-up was incomplete. Those explanations are not statistically equivalent.</p><p>A structural zero is a zero that arises from a process that makes a positive count impossible or outside the relevant risk set. A sampling zero is a zero that can occur from the ordinary count-generating distribution even when a positive count is possible. These labels are conceptual assumptions, not facts that can be observed directly from the final outcome column.</p><p>For example, in a study of symptom episodes, some participants might be in a state where the symptom cannot occur under the defined measurement window, while other participants could experience the symptom but happen to report none. In a hospital-use study, however, a recorded zero may reflect no observed contact rather than true absence of need. The analyst must examine eligibility, observation time, ascertainment, access to care, and competing data sources before interpreting zero inflation clinically.</p><p><strong>Key boundary:</strong> a large zero fraction is evidence that the ordinary model may fit poorly; it is not proof that the population contains a separate structural-zero class. A hurdle or zero-inflated model should be justified by the outcome mechanism and evaluated against plausible alternatives.</p><h3>What Poisson and negative binomial models already handle</h3><p>Poisson regression models the expected count as a function of predictors, usually through a log link. Its conditional variance equals its conditional mean. When the observed data show more variability than this assumption allows, standard errors and fitted probabilities may be poorly calibrated if the overdispersion is ignored.</p><p>Negative binomial regression introduces an additional dispersion parameter. It is often useful when the count distribution has a longer right tail than the Poisson distribution. Importantly, negative binomial regression can already generate zeros through its ordinary count process. A high number of zeros does not automatically require a mixture or two-part model if a negative binomial distribution describes the data adequately and the clinical question does not require a separate zero process.</p><p>Model selection should therefore proceed from the simplest clinically defensible model. Compare the observed mean, variance, zero frequency, upper tail, exposure time, repeated-measure structure, and covariate relationships. Do not use a more complex model merely because it produces a lower information criterion in one sample without checking interpretation, prediction, and stability.</p><h3>Zero-inflated models: two sources of zero</h3><p>A zero-inflated model combines a point mass at zero with a count distribution such as Poisson or negative binomial. The mixture component represents an additional zero-generating process, while the count component can produce both zero and positive counts. Thus, a zero-inflated model can contain two types of zero: a mixture or structural zero and a sampling zero from the count distribution.</p><p>P(Y = 0) = π + (1 − π)P count(Y = 0)</p><p>Here, <em>π</em> represents the probability of belonging to the extra-zero component, while the count component describes the remaining distribution. Predictors can be included in the zero component, the count component, or both. The two parts answer different questions. A predictor associated with the zero component changes the probability of being in the extra-zero class; a predictor in the count component changes the expected count among the count-generating group.</p><p>This separation can be clinically useful when the population plausibly contains individuals who cannot experience the event under the study definition. It can also be difficult to explain when class membership is not directly observed. A coefficient from the zero component should not be described as if it were an incidence-rate ratio for the overall population. Likewise, an incidence-rate ratio from the count component is conditional on the mixture structure rather than automatically an effect on the marginal mean count.</p><p>Zero-inflated negative binomial models add an overdispersion parameter to the count component. They may be considered when the data have both extra zeros and a long, variable positive tail. Their additional flexibility increases the need for transparent diagnostics, because a complex mixture can fit the observed distribution while producing unstable or weakly identified component-specific estimates.</p><h3>Hurdle models: one decision to cross, then positive counts</h3><p>A hurdle model uses a two-part process. The first part models whether the outcome is zero or positive. The second part models the positive counts only, using a zero-truncated Poisson or zero-truncated negative binomial distribution. In this formulation, every zero is generated by the first part, and the positive-count model begins after the hurdle has been crossed.</p><p>P(Y &gt; 0) = 1 − P(Y = 0); positive counts follow a zero-truncated count distribution</p><p>The first component can be estimated with logistic regression or another binary-response model. Its result describes the probability of any event during the defined period. The second component describes the count intensity among observations with at least one event. This can be a natural representation when the clinical pathway has a clear initiation step: whether a patient is readmitted at all, whether a patient reports any episode, or whether a patient uses any rescue medication before the number of episodes is considered.</p><p>The hurdle interpretation is not that every zero is biologically structural. It is that the analyst has chosen to separate the probability of crossing from the positive-count process. That distinction can be useful even when the first process is behavioral, clinical, or operational rather than an immutable susceptibility state.</p><p>As with zero-inflated models, the two components require separate interpretation. A treatment can increase the probability of any event but reduce the number of events among those who experience at least one. Reporting only one coefficient or one p-value can conceal this pattern.</p><h3>Zero-inflated versus hurdle: the central difference</h3><p>How are zeros represented?</p><p>Zeros can arise from an extra-zero mixture and from the ordinary count distribution.</p><p>All zeros arise from the binary zero-versus-positive part.</p><p>What does the positive-count component include?</p><p>Zero and positive values from the count-generating component.</p><p>Positive values only, using a truncated count distribution.</p><p>When is the structure plausible?</p><p>When a separate class or process can produce extra zeros while the count process can still produce zeros.</p><p>When the clinical process naturally separates event initiation from the number of events after initiation.</p><p>How should coefficients be read?</p><p>Separately for extra-zero membership and the count process; component-specific effects are not automatically marginal effects.</p><p>Separately for any-event probability and positive-count intensity.</p><p>Main communication risk</p><p>Calling the latent mixture class a verified biological subgroup.</p><p>Calling the binary part a structural susceptibility mechanism when it is only a modeling decomposition.</p><p>These models may yield different estimates because they impose different data-generating assumptions. The choice should be explained in the statistical analysis plan and linked to the clinical outcome process. A model comparison table can report fit, prediction, residual behavior, component stability, and interpretability rather than relying on a single selection statistic.</p><h3>Exposure time, follow-up, and repeated observations</h3><p>Counts are often observed over unequal periods. A patient followed for 12 months has more opportunity to experience a hospital visit than a patient followed for 3 months. If exposure time is part of the design, include a log exposure offset when the estimand is a rate rather than a raw count. State whether the outcome is a fixed-window count, a rate, or a recurrent-event process.</p><p>Repeated counts from the same patient are correlated. A standard zero-inflated or hurdle model that treats every row as independent may underestimate uncertainty. Mixed-effects, generalized estimating equation, cluster-robust, or other dependence-aware approaches may be appropriate depending on the target estimand and study design. The zero process may also be repeated or patient-specific, so it should not be assumed that only the positive-count component needs a random effect.</p><p>Incomplete follow-up can generate apparent zeros when an event was not observable for the full window. Distinguish a true zero from an unobserved outcome. Administrative censoring, loss to follow-up, missed visits, and differential healthcare access can affect both the zero-versus-positive process and the positive-count distribution.</p><h3>Model diagnostics and comparison</h3><p>Model comparison should include the observed and fitted distribution of counts, the zero frequency, the right tail, and covariate-specific predictions. A global information criterion can help compare non-identical models, but it does not establish that the latent zero mechanism is scientifically correct.</p><p>Useful checks include:</p><ul><li>Compare observed and predicted proportions of zero, one, two, and higher counts.</li><li>Examine fitted mean counts and event probabilities across clinically meaningful covariate profiles.</li><li>Assess residual behavior and overdispersion in the positive-count region.</li><li>Check whether component-specific estimates are stable under alternative starting values or reasonable model specifications.</li><li>Use bootstrap or resampling methods when uncertainty for derived marginal means or contrasts is not adequately represented by default output.</li><li>Compare predicted overall means with observed means rather than interpreting only the zero component or positive-count component.</li><li>Evaluate out-of-sample prediction when the stated goal is risk prediction rather than explanation.</li></ul><p>For nested models, likelihood-based comparisons may be available, but the assumptions and parameter boundaries should be considered. For non-nested models, information criteria and predictive checks can be informative. A test designed for a specific comparison should not be treated as a universal decision rule.</p><h3>Interpretation on the clinical scale</h3><p>Component-specific log coefficients are rarely the clearest final result for clinical readers. Present predicted probability of any event, predicted positive-count mean or median where appropriate, and predicted overall mean count for representative patient profiles. If treatment groups are compared, report absolute differences and uncertainty for the two parts and for the overall outcome.</p><p>Consider a treatment that lowers the probability of any readmission but, among patients who are readmitted, is associated with more repeated admissions. A single overall rate ratio may mask the pathway. A hurdle presentation can show both stages. A zero-inflated presentation can show how the treatment relates to the extra-zero component and the count-generating component, while clearly stating that the mixture class is latent.</p><p>Do not translate a zero-inflated component into a causal claim without a causal design and assumptions. These models describe conditional outcome distributions; they do not by themselves solve confounding, informative follow-up, measurement error, or selection bias. If the study is observational, connect the count model to a prespecified confounding strategy and target estimand. The site’s guides to <a href="https://sci.lingcorehealth.com/posts/causal-inference-observational-medical-research">causal inference</a> and <a href="https://sci.lingcorehealth.com/posts/quantitative-bias-analysis-clinical-epidemiology">quantitative bias analysis</a> provide related context.</p><h3>Five-step workflow for excess-zero clinical counts</h3><p><strong>1. Define the count outcome</strong></p><p>Specify the event, observation window, exposure time, ascertainment method, and meaning of a zero.</p><p>A zero is not silently conflated with missing follow-up or no opportunity for the event.</p><p><strong>2. Inspect the distribution</strong></p><p>Compare mean, variance, zero frequency, upper tail, and patient-level repetition.</p><p>Poisson, negative binomial, zero-inflated, and hurdle candidates are considered from the observed structure.</p><p><strong>3. State the two-part assumption</strong></p><p>Decide whether the analysis requires a latent extra-zero process or a binary any-event process followed by positive counts.</p><p>The model choice is linked to a clinical data-generating rationale, not only a fit statistic.</p><p><strong>4. Fit and diagnose</strong></p><p>Use offsets, dependence structures, and censoring or missing-data methods appropriate to the design; compare fitted and observed distributions.</p><p>Component-specific estimates and overall predictions are stable and interpretable.</p><p><strong>5. Report marginal clinical contrasts</strong></p><p>Present event probability, positive-count intensity, overall mean, uncertainty, and sensitivity analyses.</p><p>Readers can distinguish the two processes without mistaking them for separate observed subgroups.</p><h3>Common reporting errors</h3><ul><li><strong>Choosing a zero-inflated model because zeros are common:</strong> common zeros can arise from an ordinary negative binomial process or from a hurdle mechanism.</li><li><strong>Calling all extra zeros structural:</strong> structural membership is usually latent and must be justified, not read directly from the data.</li><li><strong>Ignoring exposure time:</strong> unequal observation windows can distort raw count comparisons.</li><li><strong>Assuming independent repeated rows:</strong> patient-level clustering and repeated observations require a dependence-aware plan.</li><li><strong>Reporting only one component:</strong> a treatment may affect event initiation and positive counts in different directions.</li><li><strong>Interpreting component coefficients as marginal effects:</strong> transform results to the overall clinical scale.</li><li><strong>Using fit statistics as proof of mechanism:</strong> a better fit does not establish the assumed data-generating process.</li><li><strong>Claiming causality from a two-part count model:</strong> outcome modeling does not replace confounding control or a causal estimand.</li></ul><h3>Evidence Summary Table</h3><p><strong>Hu, Pavlicova, and Nunes, 2011</strong><br><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC3238139/">Open-access article</a></p><p>Explains Poisson, negative binomial, zero-inflated, and hurdle models using a clinical-trial count outcome with extra zeros.</p><p>Illustrative applied methodology; model interpretation remains dependent on the outcome process and study design.</p><p><strong>Atkins and Gallop, 2007</strong><br><a href="https://pubmed.ncbi.nlm.nih.gov/16892908/">PubMed record</a></p><p>Discusses zero-inflated and hurdle models for count outcomes and when they may be advantageous over standard count models.</p><p>Model suitability is conditional on the scientific question, distribution, and assumptions.</p><p><strong>Loeys et al., 2012</strong><br><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC8570364/">Comparison review</a> · <a href="https://pubmed.ncbi.nlm.nih.gov/34760432/">PubMed record</a></p><p>Reviews differences in the data-generating processes of zero-inflated and hurdle models.</p><p>Conceptual distinctions should be translated to the clinical outcome rather than treated as automatic software choices.</p><p>Discusses zero-inflated models for epidemiologic count data with varying exposures and two-part interpretations.</p><p>Exposure adjustment and interpretation require explicit attention to the target population and observation process.</p><h3>How to write the results section</h3><p>Begin by stating the count outcome, observation period, opportunity for events, and handling of unequal follow-up. Report the observed mean, variance, zero proportion, and upper-tail behavior. Explain why Poisson, negative binomial, zero-inflated, or hurdle models were considered.</p><p>If a two-part model is used, describe each component in plain language. For a zero-inflated model, state that the extra-zero class is latent and that the count component may generate both zero and positive observations. For a hurdle model, state that the first component models any event and the second models positive counts using a truncated distribution.</p><p>Present model estimates with transformed clinical contrasts. Include predicted probability of any event, expected positive count, and overall expected count when these quantities answer the clinical question. Report uncertainty and sensitivity analyses for the zero-process specification, dispersion, missing follow-up, exposure offset, and correlation structure.</p><p>A defensible report does not ask readers to infer the meaning of zero from a regression table. It makes the measurement process, opportunity for events, model structure, component interpretation, and limitations explicit.</p><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI to audit and structure count-outcome research:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: extract outcome definitions, exposure windows, zero-process assumptions, offsets, dependence structures, and model diagnostics from clinical papers.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: synthesize evidence on Poisson, negative binomial, zero-inflated, and hurdle models with source-linked methodological boundaries.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with clinical biostatistics, epidemiology, health services research, and medical outcomes modeling.</li></ul><p><strong>Medical Disclaimer:</strong> This article is for research and educational purposes only. It does not provide medical advice, treatment recommendations, or a substitute for protocol-specific epidemiologic or statistical review. Researchers must verify methods against the original sources, data structure, outcome definition, and clinical context before using them in a study or manuscript.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/zero-inflated-hurdle-models-clinical-count-outcomes"><em>https://sci.lingcorehealth.com</em></a><em> on October 19, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=848b14bb327c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Brier Score and Prediction Error Curves in Clinical Models]]></title>
            <link>https://medium.com/@lifeflexi818/brier-score-and-prediction-error-curves-in-clinical-models-d184f5bb71ea?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/d184f5bb71ea</guid>
            <category><![CDATA[calibration]]></category>
            <category><![CDATA[model-validation]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[survival-analysis]]></category>
            <category><![CDATA[clinicalprediction]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 00:00:20 GMT</pubDate>
            <atom:updated>2026-10-15T11:09:17.221Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wUk7fos4aq0vshJt.png" /></figure><p><strong>The Brier score</strong> is a proper scoring rule that summarizes the squared error of probabilistic predictions. It reflects both calibration and discrimination, so it should not be presented as a pure calibration statistic or as evidence of clinical utility. For survival outcomes, a time-dependent Brier score uses censoring-aware weighting, while a prediction-error curve shows when model performance changes across the follow-up horizon.</p><p>Clinical prediction models produce probabilities, not only rankings. A model may assign a higher risk to patients who experience an event and therefore discriminate reasonably well, yet systematically overpredict or underpredict absolute risk. Conversely, a model can be well calibrated on average while failing to separate clinically distinct risk profiles. Evaluation must therefore match the intended use: screening, prognosis, treatment selection, follow-up planning, or research stratification.</p><p>The Brier score is useful because it evaluates probabilistic accuracy on the same scale as the outcome indicator. Its usefulness is also limited because it combines several performance properties and depends on the outcome prevalence, prediction horizon, validation population, and censoring mechanism. A single Brier score cannot explain why a model performs poorly, when it fails, or whether using it improves decisions.</p><h3>What the Brier score measures</h3><p>For a binary outcome <em>Y</em> and a predicted risk <em>p</em>, the individual squared prediction error is the squared difference between the observed outcome and the predicted probability. Averaging this quantity across the validation population gives the binary Brier score:</p><p>Brier score = mean of (observed outcome − predicted risk) 2</p><p>A perfect probabilistic prediction has a score of zero. Larger values indicate greater squared prediction error. The score is a <strong>proper scoring rule</strong>: in expectation, a forecaster is encouraged to report probabilities that correspond to the true event risk. This property does not mean that a low score proves clinical usefulness, transportability, or causal validity.</p><p>The scale requires context. For a binary outcome, the score depends partly on event prevalence. A model that predicts the same prevalence for everyone can perform surprisingly well in an imbalanced population even though it has little discrimination. Comparisons should therefore use the same target population, outcome definition, prediction horizon, and evaluation sample. Comparing a Brier score from one cohort with a score from another cohort without accounting for prevalence and case mix can be misleading.</p><h3>Why Brier score is not a calibration-only measure</h3><p>Calibration asks whether predicted probabilities agree with observed risks. If a group of patients receives a predicted risk of 20%, approximately 20% should experience the event under the specified outcome definition and time horizon. Discrimination asks whether patients with higher observed risk tend to receive higher predictions than patients with lower observed risk.</p><p>The Brier score is affected by both properties. A model can have good discrimination but poor absolute calibration if predictions are consistently too high. It can have acceptable calibration-in-the-large but weak discrimination if every patient receives a similar prediction. The overall Brier score compresses these patterns into one number, which is why it should be paired with calibration plots, calibration-in-the-large, calibration slope, and a discrimination measure such as the C-statistic or time-dependent AUC when appropriate.</p><p>Question Metric family What the result can support</p><p>Are predicted risks numerically close to observed risks?</p><p>Calibration plot, calibration-in-the-large, calibration slope, observed-to-expected summaries</p><p>Assessment of absolute-risk alignment at a defined horizon</p><p>Does the model rank patients by risk?</p><p>C-statistic, time-dependent AUC, concordance measures</p><p>Assessment of relative ordering, not absolute probability accuracy</p><p>How large is the squared probabilistic error overall?</p><p>Brier score or time-dependent Brier score</p><p>Summary of prediction accuracy combining calibration and discrimination</p><p>Does model use improve a clinical decision?</p><p>Decision Curve Analysis, impact evaluation, implementation study</p><p>Clinical utility under a decision threshold and care pathway</p><p>A lower Brier score can be useful when models are evaluated on the same population and outcome definition. It is not a substitute for examining the calibration curve. It also should not be used to claim that one model is clinically preferable without specifying the action, threshold, harms, benefits, and implementation context.</p><h3>From binary outcomes to time-to-event predictions</h3><p>Survival models usually predict a probability at a time horizon. For a patient <em>i</em> and horizon <em>t</em>, the model may provide a predicted survival probability or a predicted event risk. The time-dependent Brier score evaluates the discrepancy between that prediction and the patient’s event status by time <em>t</em>.</p><p>Censoring creates a problem: for a patient censored before <em>t</em>, the event status at <em>t</em> is not fully observed. Treating that patient as event-free would be incorrect. Treating the patient as an event would also be incorrect. Time-dependent Brier-score estimators therefore use inverse probability of censoring weighting or related censoring-aware approaches to construct an estimate from the observed data.</p><p>Conceptually, the estimator gives more influence to observations whose event status at the horizon is observed and weights them according to the estimated probability of remaining uncensored. The exact estimator depends on the target quantity, censoring assumptions, competing-event structure, and whether censoring weights are estimated marginally or conditional on covariates.</p><p><strong>Interpretation boundary:</strong> censoring weights do not make censoring non-informative by magic. The validity of a time-dependent Brier score depends on the assumptions and model used for the censoring distribution, adequate follow-up support, and a validation design that matches the prediction question.</p><h3>Prediction-error curves show where performance changes</h3><p>A prediction-error curve plots the time-dependent Brier score over a sequence of follow-up horizons. The curve can show early performance, late degradation, periods with sparse information, and differences between candidate models that are hidden by a single integrated summary.</p><p>When two survival models cross, one model may have lower error early and higher error later. That pattern can matter clinically: a model used to support discharge decisions may need strong short-term prediction, whereas a survivorship model may be judged over a longer horizon. An integrated Brier score summarizes performance over a prespecified time interval, but it can conceal these local differences.</p><p>The time range should be chosen from the clinical question and data support. Do not extend a curve into a period with very few individuals at risk, unstable censoring estimates, or no plausible use case merely to make the plot look complete. Report the evaluation interval, number of events, censoring pattern, and any truncation of the time horizon.</p><h3>Integrated Brier score and model comparison</h3><p>The integrated Brier score is an average or weighted integral of the time-dependent Brier score over a specified interval. It can be useful for comparing overall prediction error across a clinically meaningful period. The integration rule, time interval, weighting, and treatment of censoring should be prespecified.</p><p>An integrated number should always be accompanied by the underlying prediction-error curve. Otherwise, the reader cannot determine whether the difference is driven by a brief early period, a long late period, or a region with sparse information. Confidence intervals or bootstrap intervals should also be considered, especially when models are compared directly.</p><p>For nested models, report the direction and uncertainty of the difference in prediction error rather than describing a small numerical change as automatically important. A statistically distinguishable improvement may be too small to alter risk communication or care. Conversely, a modest average improvement may matter if it occurs at the decision horizon that clinicians actually use.</p><h3>Competing risks and event-specific prediction</h3><p>When a competing event prevents the event of interest from occurring, the prediction target must be stated precisely. A model may predict all-cause event-free survival, the cumulative incidence of a specific event in the presence of competing events, or a cause-specific quantity under a different estimand. The observed status and censoring weights must correspond to that target.</p><p>A Brier score for one event type should not be interpreted as if competing events were ordinary censoring. Treating a competing event as independent censoring can target a different quantity and may produce an inappropriate performance assessment. The analyst should state whether the prediction is for a cumulative incidence function, a cause-specific event probability, or another clinically defined outcome.</p><p>Model comparisons should use the same event definition, competing-risk convention, horizon grid, and validation population. If a model predicts several event types, report performance for each clinically relevant target rather than allowing a favorable all-cause score to obscure poor performance for a specific outcome.</p><h3>Internal validation, external validation, and optimism</h3><p>Performance measured on the development sample is usually optimistic because the model has been fitted to those observations. A low development-set Brier score is therefore not evidence that the model will perform similarly in a new hospital, time period, or care pathway.</p><p>Internal validation methods such as bootstrap optimism correction or cross-validation estimate how performance may change under repeated sampling from a similar development setting. External validation evaluates locked predictions in new data. The validation dataset should be separated from model updating until the original performance has been reported. For a practical transportability workflow, see the site’s article on <a href="https://sci.lingcorehealth.com/posts/external-validation-clinical-prediction-models">external validation of clinical prediction models</a>.</p><p>For clustered data, internal-external cross-validation can display how prediction error varies across hospitals, regions, or time periods. A pooled Brier score may hide site-specific miscalibration. Report cluster-level sample sizes, events, censoring, prediction horizons, and uncertainty when the intended implementation spans multiple settings.</p><h3>Five-step workflow for reporting Brier-based evaluation</h3><p><strong>1. Define the prediction target</strong></p><p>Specify the outcome, prediction time, horizon, event type, competing-event convention, and target population.</p><p>The Brier score corresponds to a named clinical prediction question.</p><p><strong>2. Lock the validation design</strong></p><p>Separate development, internal validation, external validation, or temporal validation data before comparing performance.</p><p>Predictions are generated without refitting on the evaluation observations.</p><p><strong>3. Estimate censoring-aware error</strong></p><p>Use a suitable time-dependent estimator and document the censoring-weight model, support, and assumptions.</p><p>Censored observations and competing events are not silently treated as ordinary negative outcomes.</p><p><strong>4. Display components and time</strong></p><p>Report Brier score with calibration, discrimination, and a prediction-error curve across the prespecified horizon.</p><p>The reader can see when and why models differ.</p><p><strong>5. Interpret clinical relevance</strong></p><p>Relate performance to decision thresholds, action timing, harms, benefits, and implementation feasibility.</p><p>No claim of clinical utility is made from Brier score alone.</p><h3>Common reporting errors</h3><ul><li><strong>Reporting only the C-index:</strong> ranking performance does not establish absolute-risk accuracy.</li><li><strong>Calling Brier score a calibration test:</strong> it is an overall prediction-error measure influenced by calibration and discrimination.</li><li><strong>Using a naive binary score after censoring:</strong> unknown event status at the horizon requires an appropriate censoring-aware estimator.</li><li><strong>Ignoring the null model:</strong> a score should be compared with a relevant baseline prediction strategy when model value is being assessed.</li><li><strong>Integrating over an arbitrary period:</strong> an integrated score is only interpretable with its time interval and weighting rule.</li><li><strong>Comparing scores across different populations:</strong> event prevalence, case mix, follow-up, and censoring can change the score.</li><li><strong>Calling a lower error clinically useful:</strong> utility requires a decision, threshold, outcome consequence, and implementation evaluation.</li><li><strong>Using development performance as transportability evidence:</strong> optimism correction and external validation are separate requirements.</li></ul><h3>Evidence Summary Table</h3><p><strong>Steyerberg et al., 2010</strong><br><a href="https://pubmed.ncbi.nlm.nih.gov/20010215/">PubMed record</a> · <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC3575184/">Open-access article</a></p><p>Provides a framework for discrimination, calibration, Brier score, and decision-analytic measures in prediction-model evaluation.</p><p>Methodological framework; metric choice must follow the intended prediction and decision context.</p><h3>How to write the results section</h3><p>Begin by naming the population, outcome, prediction horizon, and validation design. State whether the reported predictions came from development data, resampling, a temporal split, or an external dataset. Then report the Brier score with its uncertainty interval and the corresponding calibration and discrimination summaries.</p><p>For survival outcomes, describe the censoring estimator and the time interval used for the prediction-error curve. If an integrated Brier score is reported, provide the integration limits and weighting rule. Include the curve rather than presenting only the integrated number. When models cross, describe the horizon-specific pattern instead of selecting a single period that favors one model.</p><p>Use cautious language. “The model had lower prediction error over the prespecified 1-year horizon” is different from “the model improved clinical care.” The latter requires evidence about clinician behavior, patient outcomes, and implementation. A Brier score is a model-performance result, not an impact study.</p><p>For studies that use risk predictions to guide treatment or referral, pair the performance analysis with <a href="https://sci.lingcorehealth.com/posts/decision-curve-analysis-clinical-utility">Decision Curve Analysis</a> when the decision threshold and action are clinically specified. For longitudinal or joint prediction settings, the site’s guide to <a href="https://sci.lingcorehealth.com/posts/joint-models-longitudinal-time-to-event-clinical-research">joint models</a> provides additional context on dynamic prediction and time-to-event outcomes.</p><p>A rigorous Brier-based evaluation makes the prediction target, validation population, censoring assumptions, time horizon, error pattern, and clinical interpretation visible. That transparency is more informative than a single favorable score.</p><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI to audit prediction-model evidence and reporting:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: extract prediction targets, validation design, calibration measures, censoring assumptions, and performance claims from clinical-model papers.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: organize evidence on Brier scores, prediction-error curves, calibration, discrimination, and clinical utility with source-linked limitations.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with clinical prediction, biostatistics, medical informatics, and prognostic modeling.</li></ul><p><strong>Medical Disclaimer:</strong> This article is for research and educational purposes only. It does not provide medical advice, treatment recommendations, or a substitute for protocol-specific epidemiologic or statistical review. Researchers must verify methods against the original sources, data structure, prediction target, and clinical context before using them in a study or manuscript.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/brier-score-prediction-error-curves-clinical-models"><em>https://sci.lingcorehealth.com</em></a><em> on October 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d184f5bb71ea" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Collider Bias in Clinical Epidemiology: DAGs and Selection]]></title>
            <link>https://medium.com/@lifeflexi818/collider-bias-in-clinical-epidemiology-dags-and-selection-9c33a8102ef4?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c33a8102ef4</guid>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[colliderbias]]></category>
            <category><![CDATA[clinicalepidemiology]]></category>
            <category><![CDATA[directed-acyclic-graph]]></category>
            <category><![CDATA[observationalresearch]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Tue, 13 Oct 2026 00:00:41 GMT</pubDate>
            <atom:updated>2026-10-13T11:10:03.156Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IKDjcV4a_5p0MC7e.png" /></figure><p><strong>Collider bias</strong> arises when an analysis conditions on a variable influenced by both the exposure and the outcome, or by their causes. Conditioning can open a non-causal pathway, creating, reversing, or exaggerating an association. In clinical epidemiology, the collider may be explicit-such as hospital admission-or implicit, such as healthcare utilization, complete-record availability, or selection into an analytic cohort.</p><p>Researchers often learn that confounders should be controlled. That principle is important, but it is incomplete. A variable can be associated with both the exposure and the outcome without being a confounder, and adjusting for it can make an analysis worse. The central distinction is causal structure: a common cause creates an open back-door path that may need to be blocked, whereas a common effect creates a closed path that may become open after conditioning.</p><p>This distinction matters in retrospective clinical studies because inclusion criteria, referral patterns, diagnostic testing, hospital admission, and data completeness are rarely random. A dataset can therefore look large and clinically detailed while representing a selected population whose membership depends on causes of both the exposure and outcome. The resulting association may be precise but not causally interpretable.</p><p>A <strong>directed acyclic graph (DAG)</strong> makes the problem explicit. It does not prove that the assumed causal relationships are correct. Instead, it records the investigator’s causal knowledge and shows which paths are blocked or opened by selection and adjustment. The appropriate workflow is to define the target population and estimand, draw a prespecified DAG, identify the selection mechanism, and only then decide which variables belong in the analysis.</p><h3>What makes a variable a collider?</h3><p>Suppose an exposure <em>A</em> and an outcome <em>Y</em> both cause a third variable <em>C</em>. The structure is <em>A → C ← Y</em>. The two arrowheads meet at <em>C</em>, so <em>C</em> is a collider on that path. Without conditioning on <em>C</em> or its descendant, the path between <em>A</em> and <em>Y</em> is closed through the collider. If the analysis restricts the sample to a level of <em>C</em>, adjusts for <em>C</em>, or conditions on a downstream variable, the path can open.</p><p>Once opened, information about <em>A</em> can become statistically related to information about <em>Y</em> through the selection mechanism, even when there is no causal effect of <em>A</em> on <em>Y</em>. The induced association can have either direction. It may look protective, harmful, or null depending on the causal strengths, prevalence, measurement, and selection rule. Collider bias is therefore not identified by the sign of an estimate; it is identified by the mechanism that created the analytic sample or adjustment set.</p><p>A collider can also be a descendant of a collider. For example, a latent disease severity variable may influence both treatment and admission, while admission influences whether a patient enters an EHR cohort. Conditioning on admission can transmit the same selection problem even if the original common effect is not directly included in the regression model.</p><h3>Selection, adjustment, and conditioning are the same causal concern</h3><p>Selection bias is often discussed as a sampling problem, while collider bias is discussed as a regression problem. In practice, both can arise from conditioning on a common effect. Selecting only hospitalized patients, adjusting for hospital attendance, requiring a specialist visit, and including only patients with complete measurements may all impose a condition on the data-generating process.</p><p>Not every selection variable is a collider, and not every selection problem has the same mathematical form. The practical point is not to label all selection as collider bias. The point is to make the selection mechanism explicit and use the DAG to determine whether conditioning closes or opens relevant paths.</p><h3>Why EHR reuse is especially vulnerable</h3><p>Electronic health records are generated through clinical care, not through a probability sample of the community. The probability of having a diagnosis code, laboratory result, imaging record, medication record, or repeated visit can depend on symptoms, disease severity, clinician suspicion, insurance, transportation, local services, and prior healthcare use. These causes may also be related to the exposure and outcome under study.</p><p>The article by Weiskopf and colleagues describes healthcare utilization as a collider in EHR data reuse. A patient can enter an analytic dataset because an exposure-related condition increases care-seeking or because an outcome-related condition increases referral. Restricting the analysis to patients who have used a particular service may then create an association that would not appear in the target population.</p><p>This problem also affects generalizability. A selected EHR cohort may not represent the target population even if collider bias does not create a strong internal association. External validity and internal validity should therefore be evaluated separately. A study can be internally biased, externally non-transportable, or both.</p><p>“More complete data” is not automatically “less biased data.” Requiring a minimum number of visits or laboratory measurements may improve measurement availability but also select patients with more severe disease or more stable access to care. The inclusion rule should be treated as part of the causal design, not as a neutral data-cleaning step.</p><h3>Collider bias is not confounding</h3><p>Consider three common structures:</p><ul><li><strong>Confounding:</strong> <em>A ← U → Y</em>. The common cause <em>U</em> creates a non-causal path between exposure and outcome. If <em>U</em> is measured and correctly specified, adjustment can block that path.</li><li><strong>Mediation:</strong> <em>A → M → Y</em>. The mediator lies on part of the causal effect. Adjusting for it changes the estimand from a total effect toward a direct-effect question.</li><li><strong>Collider structure:</strong> <em>A → C ← Y</em>. The common effect closes the path by default. Conditioning on <em>C</em> can open a non-causal pathway.</li></ul><p>A variable’s clinical importance does not determine which structure it occupies. A measurement can be a confounder in one study, a mediator in another, and a collider in a third because the exposure, outcome, timing, and target estimand differ. Covariate selection should therefore not be justified only by statistical association, availability, or a rule that every baseline variable must enter the model.</p><p>This is also why automated variable selection is not a causal adjustment strategy. A variable selected by a prediction algorithm may improve discrimination while opening a causal path or changing the population represented by the analysis. Prediction and causal effect estimation require different variable-selection logic.</p><h3>How to use a DAG before analysis</h3><p>A DAG should be drafted before outcome-driven model searching. Start with the target population, exposure, outcome, time zero, and estimand. Add variables that plausibly cause the exposure, outcome, selection process, or measurement process. Mark variables that define inclusion, follow-up, testing, or data availability. Then inspect whether conditioning on any of these variables opens a path between exposure and outcome.</p><p>DAGitty can help researchers represent causal assumptions and identify adjustment sets, but software cannot supply the underlying clinical knowledge. The graph should be informed by subject-matter expertise, temporal ordering, prior evidence, and the protocol. A graph that is technically valid but clinically implausible is not rescued by a polished visualization.</p><h3>Five-step workflow for a defensible clinical analysis</h3><h3>What can and cannot fix collider bias?</h3><p>The first remedy is design: avoid unnecessary conditioning on a common effect. If a hospital-based sample is not required by the research question, consider a sampling frame that better matches the target population. If testing is part of the scientific question, define the estimand conditional on testing rather than presenting the result as a population-wide effect.</p><p>When selection is unavoidable, researchers may use methods that explicitly model sampling or selection, such as inverse-probability-of-selection weighting, transportability methods, two-phase designs, or quantitative selection-bias analyses. These methods require measured variables and defensible models for selection. Weighting does not automatically repair unmeasured causes of selection or violations of positivity.</p><p>Negative control analyses can sometimes help detect residual systematic error, but they do not make a collider safe to adjust for. Likewise, propensity-score methods address a specified treatment-assignment structure; they do not guarantee protection from selection induced by conditioning on healthcare utilization or data completeness. Methods should be matched to the causal problem rather than used as generic bias-removal labels.</p><p>For a broader framework on unmeasured confounding, misclassification, and selection assumptions, see the site’s guide to <a href="https://sci.lingcorehealth.com/posts/quantitative-bias-analysis-clinical-epidemiology">quantitative bias analysis</a>. For a general causal design perspective, see <a href="https://sci.lingcorehealth.com/posts/causal-inference-observational-medical-research">causal inference in observational medical research</a>. The distinction between a collider problem and a measurement-error problem should remain explicit in the analysis plan.</p><h3>Evidence Summary Table</h3><h3>How to report collider risk in a manuscript</h3><p>Describe the source population and every analytic restriction in operational terms. State whether inclusion required admission, referral, a test, a minimum number of visits, complete covariates, or follow-up data. Explain why the selected population corresponds to the estimand. If the analysis is conditional on a selected population, do not silently generalize the estimate to all eligible patients.</p><p>Include a DAG or a clear causal rationale for covariate selection. Distinguish variables adjusted for confounding from variables included for prediction, precision, missing-data handling, or transportability. If a variable is post-treatment or downstream of selection, state its role and explain why it is or is not included.</p><p>Report sensitivity analyses that change the selection definition or plausible selection parameters when justified. Present how the estimate changes, not only whether a p-value crosses a threshold. A stable result across several defensible selection specifications is informative, but it is not proof that the underlying assumptions are true.</p><p>Collider bias is best treated as a design problem rather than a late-stage regression diagnosis. The strongest manuscript makes the target population, selection mechanism, DAG assumptions, adjustment strategy, and interpretation boundary visible before presenting the effect estimate.</p><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI to audit causal structure and reporting decisions:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: extract study populations, selection criteria, conditioning variables, DAG claims, and causal interpretation from published papers.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: organize evidence on collider bias, EHR selection, negative controls, weighting, and quantitative sensitivity analysis with source-linked caveats.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with clinical epidemiology, causal inference, medical informatics, and real-world evidence.</li></ul><p><strong>Medical Disclaimer:</strong> This article is for research and educational purposes only. It does not provide medical advice, treatment recommendations, or a substitute for protocol-specific epidemiologic or statistical review. Researchers must verify methods against the original sources, data structure, estimand, and clinical context before using them in a study or manuscript.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/collider-bias-clinical-epidemiology"><em>https://sci.lingcorehealth.com</em></a><em> on October 13, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c33a8102ef4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pseudo-Observations for Survival and Competing-Risk Regression]]></title>
            <link>https://medium.com/@lifeflexi818/pseudo-observations-for-survival-and-competing-risk-regression-0b7ee6f74a1b?source=rss-04ceed8dafbf------2</link>
            <guid isPermaLink="false">https://medium.com/p/0b7ee6f74a1b</guid>
            <category><![CDATA[eventhistoryanalysis]]></category>
            <category><![CDATA[biostatistics]]></category>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[pseudoobservations]]></category>
            <category><![CDATA[survival-analysis]]></category>
            <dc:creator><![CDATA[Lifeflexi]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 00:00:32 GMT</pubDate>
            <atom:updated>2026-10-12T12:15:48.832Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pV7560hRPTkO-pgq.png" /></figure><p><strong>Pseudo-observations</strong> convert a censoring-adjusted sample-level survival quantity into subject-indexed quantities that can be modeled with generalized linear or generalized estimating equation methods. They support direct regression of survival probability, restricted mean survival time, cumulative incidence, and selected multi-state probabilities, but their validity still depends on the event estimand, censoring assumptions, time points, and variance estimation.</p><p>Clinical researchers often want a quantity that is easy to communicate: survival probability at 12 months, cumulative incidence of relapse by day 180, restricted mean survival time through a prespecified horizon, or the probability of occupying a disease state at a given time. The familiar Cox model does not directly model each of these quantities. It models a relative hazard under a proportional-hazards structure, and the relationship between a hazard ratio and an absolute risk can be difficult to explain when hazards are non-proportional or competing events are common.</p><p>Pseudo-observation regression provides another route. A jackknife construction compares the full-sample estimator with leave-one-out estimators. The resulting pseudo-values are indexed by patient and target time, then used as responses in a regression or GEE framework. This lets investigators relate a directly interpretable survival quantity to baseline covariates without pretending that the pseudo-value is a newly observed clinical outcome.</p><p>The method is useful because it separates two tasks. First, estimate a censoring-adjusted population quantity such as a Kaplan-Meier survival probability or cumulative incidence. Second, model how that quantity varies with covariates. The separation is also a warning: the regression remains tied to the estimator used to construct the pseudo-values and to the assumptions supporting that estimator.</p><h3>How pseudo-values are constructed</h3><p>Let the target parameter be an estimator of a survival-related quantity at time <em>t</em>. The full-sample estimate is obtained from all participants. For participant <em>i</em>, a leave-one-out estimate is obtained after removing that participant. A commonly used jackknife pseudo-value has the form <em>n</em> times the full-sample estimate minus <em>n − 1</em> times the leave-one-out estimate. The exact notation changes with the target quantity, but the principle is the same: each participant receives a constructed value whose collection preserves the target estimator in an estimating-equation sense.</p><p>Pseudo-values can be generated for multiple time points. A long-format data set then contains repeated pseudo-value records for each participant, with a time-point indicator and covariates. GEE or another method accounting for within-participant correlation can estimate covariate effects across the selected time points. The working correlation is a modeling choice, while the sandwich variance is commonly used to account for the dependence created by repeated pseudo-values.</p><p>These values should not be described as observed survival times, individual probabilities, or guaranteed unbiased subject-level outcomes. They are estimating constructs. Some may fall outside the usual 0-to-1 range, particularly in finite samples or near the boundary of the estimated survival curve. That behavior is not automatically an error; it reflects the jackknife construction and reinforces the need to interpret model-based predictions rather than isolated pseudo-values.</p><h3>What can be modeled directly</h3><p>Target quantity Typical estimator Clinical interpretation</p><p>The target must be named before the regression model is selected. For example, a covariate effect on cumulative incidence is not the same as a covariate effect on a cause-specific hazard. Similarly, an effect on RMST is not a hazard ratio. Pseudo-observations make direct modeling possible; they do not make different estimands interchangeable.</p><h3>Pseudo-observations and competing risks</h3><p>When death or another event prevents the outcome of interest, cumulative incidence is often the clinically relevant target. Pseudo-values can be constructed for the cumulative incidence function and then modeled with covariates. This supports direct statements such as the adjusted difference in the probability of relapse by day 180, provided the competing-event definition and censoring assumptions are clear.</p><p>The competing-risk estimand must remain visible in the manuscript. A model for the cumulative incidence of relapse is not automatically a Fine-Gray model, and a pseudo-value regression coefficient should not be labeled a subdistribution hazard ratio unless the fitted model actually estimates that quantity. Investigators should report the event of interest, competing events, administrative censoring, and the time horizon together.</p><p>Researchers comparing approaches can consult the site’s guides to <a href="https://sci.lingcorehealth.com/posts/mastering-competing-risks-survival-analysis">competing-risks analysis</a>, <a href="https://sci.lingcorehealth.com/posts/restricted-mean-survival-time-rmst-oncology">restricted mean survival time</a>, and <a href="https://sci.lingcorehealth.com/posts/joint-frailty-models-recurrent-terminal-events">joint frailty models</a>. These methods answer different questions. A pseudo-value model can be useful for direct risk or RMST regression, while a joint frailty model addresses dependence between recurrent and terminal processes.</p><h3>Choosing the regression layer</h3><p>Once pseudo-values have been computed, researchers often use a generalized linear model or GEE with a link function chosen for the target quantity. The link should respect the scientific scale and the behavior of the fitted predictions. A complementary-log-log link may be useful for certain survival-probability representations, while identity or other links may be considered for RMST differences. The choice is not a cosmetic software setting; it changes how coefficients map to the target parameter.</p><p>For several time points, include time indicators or a prespecified time structure. An interaction between a covariate and time can examine whether an effect varies across the follow-up grid. Do not report a single time-averaged effect if the model allows strong time variation without showing the time-specific results. Conversely, do not test many time points and highlight only the most favorable one without addressing multiplicity and selective reporting.</p><p>For non-randomized data, pseudo-observations do not replace confounding control. Covariate adjustment, propensity-score weighting, overlap weighting, or <a href="https://sci.lingcorehealth.com/posts/augmented-inverse-probability-weighting-doubly-robust-clinical-causal-inference">AIPW</a> may be incorporated depending on the causal estimand and design. The pseudo-value layer concerns the outcome quantity; it does not by itself identify a causal effect.</p><h3>Assumptions and diagnostics</h3><p>The first assumption is tied to the estimator used to create the pseudo-values. Kaplan-Meier-based quantities require a defensible independent-censoring structure conditional on the variables included in the analysis. If censoring depends on unmeasured prognosis, the pseudo-values do not repair that problem. Left truncation, delayed entry, time-dependent covariates, informative loss to follow-up, and competing events require an estimator and data structure that explicitly address them.</p><p>Second, the regression model must be adequate for the target quantity. Inspect fitted survival or cumulative-incidence curves against nonparametric estimates where possible. Examine predictions at clinically relevant time points and check whether they remain in an interpretable range. Report how standard errors were calculated and whether repeated pseudo-values from the same participant were handled in the variance estimator.</p><p>Third, time-point selection matters. A dense grid can create highly correlated outcomes and unstable interactions; a sparse grid can miss clinically important changes. Choose time points from the clinical question, event density, and prespecified horizon. If the analysis uses a data-driven grid, explain that choice and assess whether conclusions are sensitive to reasonable alternatives.</p><h3>Five steps for a defensible analysis</h3><h3>Evidence Summary Table</h3><h3>How to report the method</h3><p>A manuscript should name the target parameter, estimator, time points, censoring assumptions, and regression link. State whether pseudo-values were generated for survival, RMST, cumulative incidence, or a multi-state probability. Explain the leave-one-out construction at a level that permits reproduction, and describe how repeated pseudo-values were handled for standard errors.</p><p>Results should include the number of participants, events, competing events, and censored observations. Present estimates on the clinical scale whenever possible: survival probabilities, absolute risk differences, RMST differences, or state-occupation probabilities. If a regression coefficient is the primary result, show how it translates into an interpretable contrast at clinically relevant covariate patterns and time points.</p><p>Finally, avoid claiming that pseudo-observations make a survival analysis “non-parametric” in every meaningful sense or that they automatically solve censoring, confounding, or competing-risk bias. They are a flexible bridge between censoring-adjusted estimators and regression models. Their credibility depends on a coherent estimand, an appropriate estimator, valid variance handling, and transparent reporting.</p><h3>Researcher’s Toolkit</h3><p>Use Lingcore SCI to structure and audit time-to-event evidence:</p><ul><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Paper Analyzer</strong></a>: extract event definitions, censoring rules, time horizons, pseudo-value construction, and reported uncertainty from published studies.</li><li><a href="https://sci.lingcorehealth.com/portal.html"><strong>Review Builder</strong></a>: compare pseudo-observation, Cox, RMST, competing-risk, multi-state, and joint-frailty evidence with source-linked boundaries.</li><li><a href="https://sci.lingcorehealth.com/pricing.html"><strong>Journal Matcher</strong></a>: identify journals aligned with survival analysis, clinical trials, outcomes research, and medical biostatistics.</li></ul><p><strong>Medical Disclaimer:</strong> This article is for research and educational purposes only. It does not provide medical advice, treatment recommendations, or a substitute for protocol-specific statistical review. Researchers must verify methods against the original sources, data structure, estimand, and clinical context before using them in a study or manuscript.</p><p><em>Originally published at </em><a href="https://sci.lingcorehealth.com/posts/pseudo-observations-survival-competing-risk-regression"><em>https://sci.lingcorehealth.com</em></a><em> on October 12, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0b7ee6f74a1b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>