UH3 Pragmatic Suicide Prevention Trial: Complete Statistical Analysis

1. Trial at a Glance

This page covers the registry results for UH3 Pragmatic Suicide Prevention Trial (NCT02326883), a randomized na study. The primary outcome is Time to Fatal or Non-fatal Suicide Attempt or Self-Harm.

2. Clinical Question

3. Trial Design

This was a completed, randomized, single-masked, parallel-group pragmatic trial sponsored by Kaiser Permanente. The registry lists enrollment of 18,882 and three behavioral intervention groups: Care Management, Skills Training, and Usual Care. The study was conducted at Kaiser Permanente Washington as the lead site, as well as HealthPartners, Kaiser Permanente Colorado, and Kaiser Permanente Northwest. It began in 2015-03, with primary completion in 2020-04.

Electronic health records were used to identify outpatients aged 18 or older who met the questionnaire-based criterion for thoughts of death or self-harm. Eligible patients were automatically assigned to one of the three groups. The brief description states that the intended allocation was one-third to each group. The posted group sizes were 6,230 for Care Management, 6,227 for Skills Training, and 6,187 for Usual Care.

Both intervention programs used outreach through Epic secure messaging, with optional telephone outreach for people who did not read messages. Care Management involved ongoing risk assessment, care coordination, and support for outpatient follow-up. Skills Training used an online program and coaching messages. Usual Care participants were not contacted by study staff. The registry does not report who was masked, so the meaning of “single” masking cannot be specified further from the record.

The primary endpoint was time to fatal or non-fatal suicide attempt or self-harm during 18 months after randomization. Analyses followed initial assignment, regardless of intervention participation. Follow-up was censored at disenrollment from the participating health system or death due to a cause other than self-harm. The trial reports two comparisons against usual care, using log-rank methods and Cox proportional-hazards estimates.

4. Participant Flow and Accounting

The registry reports how participants moved through each stage of the study. All counts below are exactly as posted; missing counts are not inferred from other stages.

Overall Study

Reasons for not completing this period:

Analysis population for the primary outcome, as defined in the registry: Participants were analyzed according to initial assignment, regardless of intervention participation. Follow-up time was censored at time of disenrollment from the participating health system or death due to cause other than self-harm.

5. Baseline Characteristics

The table reproduces the baseline characteristics posted in the registry, which describe the participants analysed at baseline and help show whether the randomized groups were comparable.

Population: Analysis population excludes randomized participants who died or disenrolled prior to first follow-up day

6. Primary Endpoint

The registry lists the primary outcome as Time to Fatal or Non-fatal Suicide Attempt or Self-Harm, measured over 18 Months after randomization.

The primary endpoint measured time from randomization to a fatal or non-fatal suicide attempt or self-harm during the 18 months after randomization. Electronic health records and insurance claims were searched for non-fatal attempts or self-harm, while state mortality data were used to identify fatal suicide attempts. The outcome therefore combines events identified through several health and mortality data sources.

Time-to-event outcomes retain information about when an event occurs, rather than only whether an event occurred by a single fixed date. They can also account for participants whose event status is not observed throughout the entire follow-up period. In this trial, follow-up was censored when a participant disenrolled from a participating health system or died from a cause other than self-harm. Censoring means that the participant's later event experience was not included after that point; it does not mean that an event occurred.

The registry also reports median time-to-event values of 548 days in all three groups. The reported interquartile-range bounds are 404 to 548 days for Care Management, 398 to 548 days for Skills Training, and 409 to 548 days for Usual Care. The reported comparative analyses are hazard ratios, which compare event hazards over follow-up rather than comparing these medians alone. The record does not provide a separate count of primary endpoint events by group.

7. Results Posted in the Registry

• Care Management vs Usual Care: hazard ratio 1.07 (95% CI 0.86–1.33), reported p = 0.610 (as reported in the registry).
• Skills Training vs Usual Care: hazard ratio 1.29 (95% CI 1.05–1.59), reported p = 0.020 (as reported in the registry).

8. Statistical Interpretation

9. Statistical Methods Explained

The main comparative method was a log-rank analysis with a Cox proportional-hazards estimate for each intervention versus Usual Care. The reported estimates are hazard ratios: 1.07 for Care Management versus Usual Care and 1.29 for Skills Training versus Usual Care. Each is accompanied by a 95% confidence interval and a p-value. The record reports two such analyses, both labelled as superiority analyses.

What does a hazard ratio compare?

In general, a hazard is the instantaneous rate at which an event occurs among participants who have not yet had the event and remain under observation. A hazard ratio compares that rate between groups over follow-up. A value of 1 represents equal hazards under the model; a value above 1 represents a higher estimated hazard in the first group named in the comparison, and a value below 1 represents a lower estimated hazard. This is a relative time-to-event measure. It is not an absolute risk, a difference in the number of events, or a direct comparison of median survival time.

The Cox proportional-hazards model estimates a hazard ratio while accounting for event timing and censoring. Its usual interpretation relies on a proportional-hazards assumption: broadly, the relative hazard between groups is treated as stable over follow-up. Whether this assumption was checked, and what diagnostic results were obtained, is not reported in the registry record. No adjustment variables are listed for these comparisons, so the registry record does not establish whether the Cox estimates were adjusted beyond the stated group comparison.

What role does the log-rank method play?

In general, the log-rank method compares event experience across groups over follow-up while incorporating the timing of events and the time participants remain under observation. It is commonly used for time-to-event comparisons. The registry identifies the method as Log Rank and gives the comparative parameter as Cox Proportional Hazard. The detailed implementation, including any treatment of ties or additional model specifications, is not reported in the registry record.

How should censoring be understood?

Participants were censored at disenrollment from a participating health system or death from a cause other than self-harm. In general, censoring ends the period during which an outcome can be observed in the analysis; it does not count as an endpoint event. Time-to-event methods use the observed follow-up up to that point. Their interpretation depends in part on whether censoring is informative about the future event process. The registry does not report an assessment of that assumption or additional analyses addressing it.

Why can over-dispersion matter?

In general, over-dispersion occurs when outcome variability is greater than a model assumes. It can make uncertainty estimates too small if the model does not accommodate that extra variability. The primary comparison in this trial is described as a time-to-event analysis with hazard ratios, not as a count model, and the registry does not report an over-dispersion assessment. This general concept should not be mistaken for a reported problem in these analyses.

What does the median add?

The registry reports a median of 548 days in each group, together with interquartile-range bounds. A median describes the midpoint of a time distribution; an interquartile range describes the span containing the middle half of the distribution. In a censored time-to-event setting, the interpretation depends on how the median and distribution were estimated. That estimation detail is not reported in the registry record. The median alone does not replace the hazard-ratio comparison, which uses event timing across follow-up.

The trial's analyses follow initial assignment regardless of whether participants took part in their assigned intervention. This is an assignment-based comparison, preserving the randomized groups as defined by initial allocation. It answers a different question from an analysis limited to intervention participants, which is not reported in the registry record.

The registry does not report the full statistical analysis plan, covariate specifications, proportional-hazards diagnostics, handling of competing risks beyond the stated censoring rule, or a missing-data analysis. Those details should not be inferred from the use of Cox or log-rank methods. The sponsor's reported hazard ratios and associated uncertainty are the basis for describing the posted comparative results.

10. Confidence Intervals and Precision

The 95% confidence interval gives a range of effect estimates compatible with the data and the analysis under the method's assumptions. It reflects statistical uncertainty; it is not a guarantee that the true effect lies in the interval, nor does it describe the range of individual outcomes.

For Care Management versus Usual Care, the hazard ratio was 1.07 and the 95% confidence interval was 0.86 to 1.33. This interval includes 1, the value corresponding to equal hazards. It also spans values below and above 1, so the estimate is not precise enough to rule out either direction of difference based on this interval.

For Skills Training versus Usual Care, the hazard ratio was 1.29 and the 95% confidence interval was 1.05 to 1.59. This interval lies above 1, indicating that the estimated hazard was higher in Skills Training under the reported comparison and model. The interval still allows a range of relative effects; it does not mean the effect is exactly the point estimate for all follow-up times or participants.

Precision depends on the amount and distribution of event information, as well as the analysis assumptions. The registry reports enrollment and group sizes but does not report the primary endpoint event count by group. Consequently, the confidence intervals are the most direct posted summary of uncertainty for the comparative hazard ratios. They should be considered together with the endpoint definition, follow-up and censoring rules, rather than read as measures of clinical importance on their own.

11. Planning a Similar Trial: Sample Size

This is a planning illustration, not a re-evaluation of the original trial. It asks: if a new study wanted to detect a difference like the one observed, or a smaller one, how large would it need to be? It is deliberately not “observed power” of the finished trial, which is not a meaningful quantity.

The sample-size figures are planning illustrations based on the observed effect, not the trial's original design assumptions. For these time-to-event comparisons, the figures are the total number of events required across the two comparison groups, not participants per arm. The planning assumptions are a two-sided significance level of 0.05, equal allocation between the two arms, a normal approximation, and an effect equal to the stated fraction of the observed effect. The trial's original sample-size justification and design assumptions are not reported in the registry record.

The 80% power column gives the total event count that the planning calculation associates with an 80% chance of detecting the specified effect under those assumptions, if that effect were the true one. The 90% power column gives the corresponding event count for a 90% chance. These are prospective planning concepts: they describe what an assumed design might be expected to achieve, not the probability that the trial's result is correct.

The rows for planning for 75% or 50% of the observed effect use progressively smaller assumed effects than planning for 100% of the observed effect. Smaller effects require more information to detect, so the required total event counts increase. For example, for the Care Management comparison, the 80% power illustration ranges from 6,859 total events when planning for 100% of the observed effect to 27,434 when planning for 50%. For Skills Training, the corresponding figures are 485 and 1,937 total events.

These figures are not observed power. They do not replace the reported hazard ratios, confidence intervals, or p-values, and they should not be interpreted as the trial's actual event count or as a retrospective measure of its success.

12. How the Sample Size Is Derived

Power is the probability that a study of a given size rejects the null hypothesis when a specific true effect exists. It depends on four things: the effect to detect, the variability of the outcome, the significance level (here a two-sided α of 0.05, so z_1−α/2 = 1.960) and the desired power (z_1−β = 0.842 for 80% and 1.282 for 90%).

Formula

d is the total number of events needed (Schoenfeld's formula) with 1:1 allocation and proportional hazards.

Worked calculation for this trial

1. Hazard ratio HR = 1.070, so ln(HR) = 0.0677.
2. d = 4 × (1.96 + 0.842)² / [ln(1.070)]² = 6,858.4, rounded up to 6,859 events in total for 80% power.

The smaller-effect rows in the table above repeat the same calculation with the effect multiplied by 0.75 and 0.5 (for ratios, applied to the logarithm of the ratio).

R code

alpha <- 0.05
z_a   <- qnorm(1 - alpha / 2)          # two-sided
hr <- 1.07                              # observed hazard ratio
events_needed <- function(power, frac = 1) ceiling(round(4 * (z_a + qnorm(power))^2 / log(hr^frac)^2, 6))   # Schoenfeld, 1:1 allocation
for (f in c(1, 0.75, 0.5))
  cat(sprintf("%3.0f%% of observed effect: events = %s (80%% power), %s (90%% power)\n", 100 * f,
              format(events_needed(0.80, f), big.mark = ","), format(events_needed(0.90, f), big.mark = ",")))

# Simulation: given d total events under proportional hazards, events in the treatment arm are Binomial(d, hr / (1 + hr)).
# The score (log-rank type) statistic is (D1 - d / 2) / sqrt(d / 4).
set.seed(2326883)
d    <- events_needed(0.80)
reps <- 20000
D1 <- rbinom(reps, d, hr / (1 + hr))
z  <- (D1 - d / 2) / sqrt(d / 4)
cat(sprintf("Simulated power with %s events: %.3f (%d replicates)\n", format(d, big.mark = ","), mean(abs(z) > z_a), reps))

Output from running the code

100% of observed effect: events = 6,859 (80% power), 9,182 (90% power)
 75% of observed effect: events = 12,193 (80% power), 16,323 (90% power)
 50% of observed effect: events = 27,434 (80% power), 36,726 (90% power)
Simulated power with 6,859 events: 0.802 (20000 replicates)

The simulation uses the fact that, with d events in total and proportional hazards, the number of events in the treatment arm is binomial. It is an approximation that conditions on the total number of events, and it should land close to the target power.

13. Analyses Reported by the Sponsor

The comparative results are the sponsor's reported analyses: log-rank comparisons summarized with Cox proportional-hazards estimates. For Care Management versus Usual Care, the reported hazard ratio is 1.07, with a 95% confidence interval of 0.86 to 1.33 and p = 0.61. For Skills Training versus Usual Care, the reported estimate is 1.29, with a 95% confidence interval of 1.05 to 1.59 and p = 0.02.

No discrepancy is flagged between the posted comparison and a recalculated result. In any case, the sponsor's own analysis takes precedence when interpreting the trial's reported findings. The planning illustrations are separate from those comparative analyses: they use the observed effects to show event-count requirements under specified planning assumptions. They do not alter or supersede the sponsor's estimates.

The registry does not report sufficient detail to reconstruct every element of the sponsor's analysis, such as model diagnostics or any additional specification choices. Accordingly, the interpretation should remain tied to the stated methods and the results as reported, rather than implying an independent reanalysis of participant-level data.

14. Multiplicity, Missing Data and Analysis Populations

The registry lists two primary comparisons: Care Management versus Usual Care and Skills Training versus Usual Care. Each is reported as a superiority analysis, with its own hazard ratio, 95% confidence interval, and p-value. The record does not report whether a multiplicity adjustment was planned or applied, whether one comparison was designated as primary among the two, or how the two tests were interpreted together. The reported p-values should therefore be read as posted for their respective comparisons, without assuming a family-wise error adjustment.

Multiplicity matters in general when several hypotheses are tested: the chance of at least one result appearing statistically notable can change as the number of tests increases. That general consideration does not establish that the reported analyses are invalid; it identifies an aspect of the analysis plan that cannot be determined from the registry record.

The analysis population is described as participants analyzed according to initial assignment, regardless of intervention participation. This supports an assignment-based interpretation of the randomized comparisons. The record does not provide further details about exclusions after randomization, the number of participants with incomplete outcome ascertainment, or the handling of missing covariates, if any.

Outcome ascertainment relied on electronic health records and insurance claims for non-fatal events and state mortality data for fatal events. The registry states that follow-up was censored at disenrollment from a participating health system or death due to a cause other than self-harm. It does not report the number censored in each arm, the extent of missing outcome information, or sensitivity analyses for alternative censoring assumptions.

The registry does not report an interim-analysis plan, stopping rules, the full statistical analysis plan, or a detailed account of missing-data procedures. These omissions do not show that such procedures were absent; they mean that their details are not reported in the registry record.

15. Safety Results

The registry reports serious adverse events as affected participants divided by participants at risk.

Deaths

Other (non-serious) adverse events

Participants with at least one non-serious event at or above the registry's 0% reporting threshold.

Most frequent serious adverse event terms

Serious events with the highest arm-level rate: self-harm leading to death or hospitalization (1.5%).

The serious adverse event figures are counts recorded in the three groups: 71 in Care Management, 92 in Skills Training, and 69 in Usual Care. The numbers at risk shown for those serious adverse events are 6,230, 6,227, and 6,187, respectively. These are posted counts, not a comparison adjusted for follow-up time or other factors. No new percentages or comparative rates should be inferred from the counts alone.

The registry lists “Self-harm leading to death or hospitalization” among the serious-event terms, at 1.5%. A group-specific breakdown for this term is not reported in the registry record. The record also lists zero at risk and zero affected for other adverse events in each group; this describes the posted fields and does not establish that no other events occurred outside the registry's reported categories.

Deaths are separately listed as 133 for Care Management, 132 for Skills Training, and 118 for Usual Care. These figures are not identified as deaths caused by the intervention, and the endpoint censoring rule distinguishes deaths due to causes other than self-harm. The registry does not report further detail that would support attributing these deaths to an intervention or comparing causes.

16. What the Result Does, and Does Not, Mean

The trial provides randomized comparisons of two outreach programs with Usual Care among adult outpatients identified by a routine questionnaire in four participating health systems. The primary outcome combined fatal and non-fatal suicide attempts or self-harm, measured over 18 months after randomization using health records, claims, and mortality data.

The reported Care Management comparison has a hazard ratio of 1.07, with a 95% confidence interval of 0.86 to 1.33 and p = 0.61. The Skills Training comparison has a hazard ratio of 1.29, with a 95% confidence interval of 1.05 to 1.59 and p = 0.02. These estimates describe the relative hazard of the combined endpoint under the reported Cox analyses. They do not provide absolute risks, numbers needed to treat, or separate effects on fatal and non-fatal events.

The results do not determine what would happen in populations outside the participating health systems, among people not meeting the questionnaire criterion, or under different intervention delivery. They also do not establish how much participants engaged with each program. The assignment-based analysis includes participants regardless of intervention participation, which is relevant to interpreting the comparison as an effect of assignment in this trial setting.

A confidence interval and p-value are statistical summaries, not measures of clinical importance or individual certainty. The reported median time was 548 days in each group, but that summary alone cannot represent all features of event timing and censoring. Nor does a statistically notable comparison on the combined endpoint establish a difference in each outcome component or explain its cause.

These findings are not medical advice and do not establish that a program should or should not be used. The registry record supports a bounded statistical interpretation of the posted comparisons; important details about multiplicity, model diagnostics, and missing-data analyses are not reported in the registry record.

17. Why This Trial Matters Statistically

This trial illustrates several important features of pragmatic clinical research. Participants were identified through routine care data, and the interventions were delivered through health-system messaging and follow-up processes. The outcome was also assembled from routinely collected records, claims, and mortality data. Such a design can evaluate programs in the context of participating health systems, while making the definition and coverage of routine data central to interpretation.

The three-arm structure permits two separate randomized comparisons with a shared usual-care group. That structure makes it important to report each intervention contrast clearly and to understand how multiple comparisons were handled. The registry reports two hazard ratios and their confidence intervals, but does not report a multiplicity plan. This is a reminder that estimates, uncertainty intervals, and the testing strategy all contribute to interpretation.

The time-to-event endpoint also emphasizes that outcome timing and censoring matter. Participants were followed until an event or a stated censoring point, rather than simply classified by an endpoint measured at one visit. The hazard ratios summarize relative event hazards, while the medians offer a different view of the observed time distribution. Neither summary alone conveys every aspect of the data.

Finally, the planning illustrations show why assumptions about the effect size and desired power affect required event counts. They are not a substitute for a trial's original design justification or for the observed comparative results. Keeping planning calculations distinct from reported analyses helps prevent retrospective calculations from being mistaken for evidence about the trial's design or certainty.

18. Limitations

• The posted summary statistics are the only data these results rest on; patient-level data, full protocols and statistical analysis plans are not available.
• Recalculated results are simple, unadjusted calculations from posted summaries. They do not replace the sponsor's prespecified analysis.

The trial's population was limited to outpatients aged 18 or older in four participating health systems who met a specific questionnaire response criterion. This limits direct generalization to people identified in other ways, people receiving care elsewhere, and populations outside the trial's eligibility definition. The registry does not report demographic or clinical characteristics needed to assess how representative the enrolled groups were.

The primary endpoint combined fatal and non-fatal suicide attempts or self-harm, identified using different data sources. The registry does not report separate group-specific results for the fatal and non-fatal components, nor the count of primary events by group. The extent to which events outside the participating systems were captured is not reported in the registry record.

Follow-up was censored at disenrollment or death from a cause other than self-harm. Group-specific censoring counts, reasons for disenrollment, and analyses of the censoring assumptions are not reported. The registry also does not report checks of the proportional-hazards assumption, model diagnostics, or the details of any covariate adjustment.

Two comparisons were reported, but multiplicity handling is not reported. The registry record also does not report the full statistical analysis plan, interim analyses, sample-size justification, missing-data procedures, or who was masked. These are limits on what can be established from the record, not evidence that the study necessarily lacked such procedures.

The descriptions of intervention delivery are brief. The detailed protocol, levels of intervention uptake, and adherence information are not reported in the registry record. As a result, the comparative estimates describe assignment to the programs in this trial setting and should not be interpreted as estimates of the effect among only those who fully engaged with an intervention.

19. Related Tutorials

20. Related Statistical Calculators

21. Sources

• ClinicalTrials.gov: UH3 Pragmatic Suicide Prevention Trial — NCT02326883.
• Registry record last updated 2021-10-19; status last verified 2021-09.
• Results first posted 2021-10-19.
• Data retrieved from ClinicalTrials.gov on 2026-11-09.
• Group sizes, means, event counts and sponsor-reported p-values are reproduced as posted. Effect estimates, confidence intervals and p-values not attributed to the sponsor are independently calculated from the posted summary statistics.

22. Record Summary

NCT02326883 is a completed randomized, single-masked, parallel-group pragmatic trial sponsored by Kaiser Permanente, with enrollment of 18,882. It compared Care Management and Skills Training with Usual Care among adult outpatients identified as at risk through a routine depression questionnaire in four participating health systems.

The primary endpoint was time to fatal or non-fatal suicide attempt or self-harm during 18 months after randomization. The sponsor reported a hazard ratio of 1.07 for Care Management versus Usual Care (95% confidence interval 0.86 to 1.33; p = 0.61) and 1.29 for Skills Training versus Usual Care (95% confidence interval 1.05 to 1.59; p = 0.02). Participants were analyzed according to initial assignment, with follow-up censored at disenrollment or death from a cause other than self-harm.