Propensity Score-based Approach Research Articles

Treatment options for patients with pilonidal sinus disease: PITSTOP, a mixed-methods evaluation.

There is no consensus on optimal management of pilonidal disease. Surgical practice is varied, and existing literature is mainly single-centre cohort studies of varied disease severity, interventions and outcome assessments. A prospective cohort study to determine: • disease severity and intervention relationship • most valued outcomes and treatment preference by patients • recommendations for policy and future research. Observational cohort study with nested mixed-methods case study. Discrete choice experiment. Clinician survey. Three-stage Delphi survey for patients and clinicians. Inter-rater reliability of classification system. Thirty-one National Health Service trusts. Patients aged > 16 years referred for elective surgical treatment of pilonidal disease. Surgery. Pain postoperative days 1 and 7, time to healing and return to normal activities, complications, recurrence. Outcomes compared between major and minor procedures using regression modelling, propensity score-based approaches and augmented inverse probability weighting to account for measured potential confounding features. Clinician survey: There was significant heterogeneity in surgeon practice preference. Limited training opportunities may impede efforts to improve practice. Cohort study: Over half of patients (60%; N = 667) had a major procedure. For these procedures, pain was greater on day 1 and day 7 (mean difference day 1 pain 1.58 points, 95% confidence interval 1.14 to 2.01 points, n = 536; mean difference day 7 pain 1.53 points, 95% confidence interval 1.12 to 1.95 points, n = 512). There were higher complication rates (adjusted risk difference 17.5%, 95% confidence interval 9.1 to 25.9%, n = 579), lower recurrence (adjusted risk difference -10.1%, 95% confidence interval -18.1 to -2.1%, n = 575), and longer time to healing (>34 days estimated difference) and time to return to normal activities (difference 25.9 days, 95% confidence interval 18.4 to 33.4 days). Mixed-methods analysis: Patient decision-making was influenced by prior experience of disease and anticipated recovery time. The burden involved in wound care and the gap between expected and actual time for recovery were the principal reasons given for decision regret. Discrete choice experiment: The strongest predictors of patient treatment choice were risk of infection/persistence (attribute importance 70%), and shorter recovery time (attribute importance 30%). Patients were willing to trade off these attributes. Those aged over 30 years had a higher risk tolerance (22.35-34.67%) for treatment failure if they could experience rapid recovery. There was no strong evidence that younger patients were willing to accept higher risk of treatment failure in exchange for a faster recovery. Patients were uniform in rejecting excision-and-leave-open because of the protracted nursing care it entailed. Wysocki classification analysis: There was acceptable inter-rater agreement (κ = 0.52, 95% confidence interval 0.42 to 0.61). Consensus exercise: Five research and practice priorities were identified. The top research priority was that a comparative trial should broadly group interventions. The top practice priority was that any interventions should be less disruptive than the disease itself. Incomplete recruitment and follow-up data were an issue, particularly given the multiple interventions. Assumptions were made regarding risk adjustment. Results suggest the burden of pilonidal surgery is greater than reported previously. This can be mitigated with better selection of intervention according to disease type and patient desired goals. Results indicate a framework for future higher-quality trials that stratify disease and utilise broad groupings of common interventions with development of a patient-centred core outcome set. This trial is registered as ISRCTN95551898. This award was funded by the National Institute for Health and Care Research (NIHR) Health Technology Assessment programme (NIHR award ref: 17/17/02) and is published in full in Health Technology Assessment; Vol. 28, No. 33. See the NIHR Funding and Awards website for further award information.

647Use of menopausal hormone therapy before and after ovarian cancer diagnosis and ovarian cancer survival

Abstract Background Menopausal hormone therapy (MHT) use before ovarian cancer (OvCa) diagnosis has been suggested to improve survival but data on type, duration and use after treatment for OvCa are scarce. Methods We investigated MHT use and OvCa survival among participants with newly diagnosed OvCA in the Ovarian cancer Prognosis And Lifestyle (OPAL) Study. Analysis of pre-diagnosis use was restricted to 661 post-menopausal women and analysis of post-diagnosis use included 254 women aged ≤55-years. We used multivariable Cox proportional hazards regression to estimate hazard ratios (HR) and 95% confidence intervals (CI) for the association between MHT and OvCa-specific survival. We used propensity score-based approaches to account for potential bias due to confounding by indication. Results Approximately 14% of post-menopausal women were current/recent users of oestrogen-only (7%) or oestrogen-progestin/unknown MHT (E-P=7%) at the time of diagnosis. In the pre-diagnosis analysis, E-P use was associated with better survival (HR = 0.60, 95%CI=0.37-0.98; HR = 0.93, 95%CI=0.79-1.09 per 5-years/use). The association between oestrogen-only MHT and survival was weaker and non-significant (HR = 0.74, 95%CI=0.47-1.16). Among women ≤55-years at diagnosis, the HR was 0.91 (95%CI=0.50-1.67) for new use after diagnosis regardless of type; and 0.89 (95%CI 0.51-1.54) for any use post-diagnosis compared to never users. Propensity-score-based methods showed similar estimates. Conclusions Pre-diagnosis MHT use is associated with better ovarian cancer survival. Post-diagnosis MHT use might also improve survival for women younger than 55-years, even after accounting for bias due to confounding by indication. Key messages Menopausal hormone therapy may be considered to manage menopausal symptoms in women with ovarian cancer.

Contextualizing a single-arm trial of ceralasertib (cer) plus paclitaxel with real-world data (RWD) in patients (pts) with advanced melanoma previously treated with anti-PD-(L)1(PDx) therapies.

e21542 Background: Metastatic melanoma, especially in pts who do not respond to PDx therapy, is an area of high unmet need. Chemotherapy may be used but clinical efficacy is limited. A recent Phase I study (NCT02630199) showed that cer + paclitaxel had promising anti-tumor activity with durable responses in 34 melanoma pts after failing PDx therapy. Given the lack of published prospective clinical trial data in the post-PDx setting, RWD was used to contextualize the results of the trial. Methods: Adult pts (≥18 years) diagnosed with advanced melanoma between Jan2011-Jun2020, who received PDx therapy alone or in combination with anti-CTLA4, and also received chemotherapy immediately after the PDx line (chemo cohort) were selected from the nationwide US-based Flatiron Health EHR-derived de-identified database. A subgroup of pts who received taxane was also selected. Overall survival (OS) was assessed using Kaplan-Meier methods. Key covariates such as age, gender, LDH level, number of prior lines of therapy, presence of brain and liver metastases, and prior BRAF/MEK treatment, were adjusted using propensity score-based weighting approach. Cox regression models were used to obtain the effect estimates for OS comparing the trial arm against the RWD cohorts. Results: A cohort of 114 pts met the inclusion criteria, of which 69 received taxane. Compared with the trial arm, the RWD cohorts were older, contained more males, with higher proportions receiving prior BRAF/MEK inhibitors, with brain metastasis, and LDH≤ULN. Median OS was 7.4 months (mo) in the trial arm, compared to 5.6 (unadjusted) and 3.5 mo (adjusted) in the chemo cohort. Hazard ratio of 0.49 (95% CI: 0.29-0.83, trial vs. chemo) indicated significant benefits of the trial arm over the chemo cohort. Analyses using taxane subgroup showed similar findings. Conclusions: The RWD cohorts demonstrated poor outcomes for the pts receiving chemotherapy, in particular taxanes, confirming this is a population of unmet need. The results show the potential for improved outcomes using cer + paclitaxel in pts previously treated with PDx.[Table: see text]

Association of intranasal desmopressin therapy with overcorrection of severe hyponatremia: A retrospective, propensity score-based, single-center cohort study.

Severe hyponatremia, defined as serum sodium concentration ([sNa]) ≤ 120 mEq/L, requires aggressive treatment to prevent potentially fatal cerebral edema, seizures, and other sequelae, but overcorrection can also result in life-threatening cerebral hemorrhage and demyelination. We compared the safety and efficacy of nasal desmopressin to conventional management for the prevention of [sNa] overcorrection. This retrospective analysis compared 47 patients treated with desmopressin to 17 patients treated conventionally at a university hospital ICU in Japan between 2013 and 2018 using propensity score-based approaches. The primary outcome was safe [sNa] correction, defined as a ≤ 8 mEq/L difference between baseline and follow-up [sNa] at any time within 24h of diagnosis. The 24-h safe correction rate was significantly greater in the desmopressin group than the conventional treatment group (68% [32/47] vs. 41% [7/17], P = 0.039), and dose-response analysis indicated a positive association between cumulative 24-h desmopressin dose and safe correction at 24 h (P = 0.003). Few overcorrections precluded reliable assessment at 48 h. Exacerbation of hyponatremia was comparable in the two treatment groups. Intranasal desmopressin therapy increased the safe correction of severe hyponatremia. Large prospective trials are warranted to confirm this result.

Trends in Use of Low-Value Care in Traditional Fee-for-Service Medicare and Medicare Advantage

Decreasing use of low-value care is a major goal for Medicare given the potential to decrease costs and harms. Compared with traditional fee-for-service Medicare (TM), Medicare Advantage (MA) is more strongly financially incentivized to decrease use of low-value care. To compare use of low-value care among individuals enrolled in TM and those enrolled in MA overall and to examine trends in use of low-value care in both programs from 2006 to 2015. This cross-sectional study analyzed individuals enrolled in TM and MA using data from the 2006 to 2015 Medical Expenditure Panel Survey. To account for differences in characteristics between individuals enrolled in TM and those enrolled in MA, a propensity score-based approach was used. Data were analyzed from August 2020 through January 2021. Being enrolled in MA or TM. Binary measures of use were collected for 13 low-value services in 4 categories (ie, [1] cancer screening: cervical, colorectal, and prostate cancer screening in older adults; [2] antibiotic use: antibiotic for acute upper respiratory infection and antibiotic for influenza; [3] medication: anxiolytic, sedative, or hypnotic in an adult older than 65 years; benzodiazepine for depression; opioid for headache; opioid for back pain; and nonsteroidal anti-inflammatory drug [NSAID] for hypertension, heart failure, or chronic kidney disease; and [4] imaging: magnetic resonance imaging [MRI] or computed tomography [CT] for back pain, radiograph for back pain, and MRI or CT for headache) and 4 low-value composites corresponding to the categories (ie, cancer screening composite, antibiotic use composite, medication composite, and imaging composite). Among 11 677 individuals enrolled in TM and 5164 individuals enrolled in MA, 9429 (56.0%) were women and the mean (SD) age was 74.5 (6.3) years. Of 13 low-value services and 4 low-value composites, statistically significant differences were found in 2 measures. For the low-value medication composite, 2054 of 11 636 eligible individuals enrolled in TM (adjusted mean, 17.6%; 95% CI, 16.8%-18.3%) received the care, and 981 of 5141 eligible individuals enrolled in MA (adjusted mean, 19.7%; 95% CI, 18.3%-21.2%) received the care, for a rate of use that was significantly higher among individuals enrolled in MA, by 2.2 percentage points (95% CI, 0.5-3.8 percentage points; P = .02). For the NSAID use for hypertension, heart failure, or kidney disease metric, 807 of 7832 individuals enrolled in TM (adjusted mean, 10.0%; 95% CI, 9.2%-10.8%) received the care, and 447 of 3566 individuals enrolled in MA (adjusted mean, 12.9%; 95% CI, 19.7%-27.1%) received the care, for a rate of use that was significantly higher among individuals enrolled in MA, by 2.9 percentage points (95% CI, 1.3-4.6 percentage points; P = .001). Overall, there were no decreases in use of low-value care in TM or MA over time. This cross-sectional study found that use of low-value care was similarly prevalent in MA and TM, suggesting that MA enrollment was not associated with decreased provision of low-value care compared with TM.

Do data security measures, privacy regulations, and communication standards impact the interoperability of patient health information? A cross-country investigation

BackgroundThe lack of interoperability is one of the biggest obstacles to the complete digitalization of patient health information in electronic medical records (EMR). The high volume of data breaches has put pressure on care providers to adopt data protection measures to remain compliant with legal requirements. Extreme data protection measures can impede information flow, but they also instill confidence in secure information sharing. This study investigates how the adoption of security measures, privacy regulations, and communication standards has impacted patient health information interoperability at technical (TI), semantic (SI), and organizational (OI) levels within the hospitals. MethodsThe study utilizes a quasi-experimental research design to probe the relationships of interest. Secondary data from a survey of randomly selected 773 hospitals conducted by the European Commission in over 30 countries in Europe is used to understand the relationships. The study counters selection bias and accounts for systematic differences in adopting treatments of interest in the hospitals using the propensity score-based approaches for the observational data. ResultsThe empirical models that account for selection bias explain more observational data variations than those that did not. Access control measures on workstations are linked to 44 % lesser odds of experiencing TI problems. However, hospitals with regional and organizational level privacy regulations have 85 % and 76 % higher odds of experiencing SI and OI problems, respectively. On the other hand, hospitals with a single hospital-wide EMR are 53 % and 43 % less likely to experience TI and SI problems, respectively, in comparison to those with multiple EMR systems. ConclusionThe study highlights the differential impacts of data protection measures on the hospitals' three key types of interoperability problems (i.e., TI, SI, and OI). Homogenous EMR systems type and substantial investment in technology are critical to supporting health information interoperability within the hospitals. The study findings inform policy considerations for improving specific aspects of health information's interoperability while preserving patient data privacy and security.

Clinical Trial Emulation by Matching Time-dependent Propensity Scores: The Example of Estimating Impact of Kidney Transplantation.

No study to our knowledge has examined the use of observational data to emulate a clinical trial whereby patients at the time of kidney transplant proposal are randomly assigned to an awaiting transplantation or transplantation group. The main methodologic issue is definition of the baseline for dialyzed patients assigned to awaiting transplantation, resulting in the inability to use common propensity score-based approaches. We aimed to use time-dependent propensity score to better appraise the benefit of transplantation. We studied 23,231 patients included in the French registry and on a transplant waiting list from 2005 to 2016. The main outcome was time to death. By matching on time-dependent propensity score, we obtained 10,646 pairs of recipients (transplantation group) versus comparable patients remaining on dialysis (awaiting transplantation group). The baseline exposure, that is, pseudo-randomization, was matching time. Median follow-up time was 3.5 years. At 10 years' follow-up, the restricted mean survival time was 8.8 years [95% confidence interval (CI) = 8.7, 8.9] in the transplantation group versus 8.2 years (95% CI = 8.1, 8.3) in the awaiting transplantation group. The corresponding life expectancy gain was 6.8 months (95% CI = 5.5, 8.2), and this corresponded to one prevented death at 10 years for 13 transplantations. Our study has estimated the life expectancy gain due to kidney transplantation. It confirms transplantation as the best treatment for end-stage renal disease. Furthermore, we demonstrated that this simple method should also be considered for estimating marginal effects of time-dependent exposures.

A medicare-based comparative mortality analysis of active surveillance in older women with DCIS

Over 97% of individuals diagnosed with ductal carcinoma in situ (DCIS) will choose to receive guideline concordant care (GCC), which was originally designed to treat invasive cancers and is associated with treatment related morbidity. An alternative to GCC is active surveillance (AS) where therapy is delayed until medically necessary. Differences in mortality risk between the two approaches in women age 65+ are analyzed in this study. SEER and Medicare information on treatment during the first year after diagnosis was used to identify three cohorts based on treatment type and timing: GCC (N = 21,772; immediate consent for treatment), AS1 (N = 431; delayed treatment within 365 days), and AS2 (N = 205; no treatment/ongoing AS). A propensity score-based approach provided pseudorandomization between GCC and AS groups and survival was then compared. Strong influence of comorbidities on the treatment received was observed for all age-groups, with the greatest burden observed in the AS2 group. All-cause and breast-cancer-specific mortality hazard ratios (HR) for AS1 were not statistically different from the GCC group; AS2 was associated with notably higher risk for both all-cause (HR:3.54; CI:3.29, 3.82) and breast-cancer-specific (HR:10.73; CI:8.63,13.35) mortality. Cumulative mortality was substantially higher from other causes than from breast cancer, regardless of treatment group. Women managed with AS for DCIS had higher all-cause and breast-cancer-specific mortality. This effect declined after accounting for baseline comorbidities. Delays of up to 12 months in initiation of GCC did not underperform immediate surgery.

Effect of the Extent of Resection on Survival Outcome in Glioblastoma: Propensity Score Approach

Abstract Objective To evaluate the effectiveness of the extent of resection (EOR) on survival outcome using propensity score-based approaches. Materials and Methods A retrospective cohort study was performed in patients with newly diagnosed glioblastoma. Propensity score matching (PSM) and propensity score regression adjustment were used in the matched and unmatched dataset, respectively. Therefore, the Kaplan-Meier survival curve and Cox's regression analyses were performed to determine the effect of the EOR on survival outcomes. Results One hundred and sixty-eight patients were included for analyzes. The total tumor resection in the unmatched dataset was 22.6% of all cases. Using PSM, incomplete tumor resection had an unfavorable survival outcome when compared with total tumor resection (hazard ratio (HR) 2.92, 95% confidence interval [CI] 1.72–4.94). Additionally, biopsy and partial tumor resection were significantly associated with poor prognosis when compared with total tumor resection using propensity score regression adjustment (HR of biopsy 1.89, 95%CI 1.13–3.16 and HR of partial resection 1.89, 95%CI 1.28–2.80). Conclusions Patients with total tumor resection tend to have a more favorable prognosis than patients with partial tumor resection. The propensity score-based analysis is an alternative approach to evaluate the effect of an intervention that has limitations to perform a randomized controlled trial.

Novel statistical approaches and applications in leveraging real-world data in regulatory clinical studies

In medical product development, there has been a growing interest in utilizing real-world data which have become abundant owing to advances in biomedical science, information technology and engineering. High-quality real-world data may be utilized to generate real-world evidence for regulatory or healthcare decision-making. We discuss propensity score-based approaches for leveraging patients from a real-world data source to construct a control group for a non-randomized comparative study or to augment a single-arm or randomized prospective investigational clinical study. The proposed propensity score-based approaches leverage real-world patients that are similar to those prospectively enrolled into the investigational clinical study in terms of baseline characteristics. Either frequentist or Bayesian inference can then be applied for outcome data analysis, with the option of down-weighting information from the real-world data source. Examples based on pre-market regulatory review experience are provided to illustrate the implementation of the proposed approaches.

Non-small cell lung cancer (NSCLC) case study examining whether results in a randomized control arm are replicated by a synthetic control arm (SCA).

9108 Background: The FDA’s accelerated approval (AA) pathway provides conditional approval for an investigational product (IP) after positive effect on a surrogate endpoint has been provided, allowing patients earlier access to the therapy. Confirmation of a positive effect on the clinical endpoint after conditional approval is required and usually includes a randomized trial. However, such a trial is challenged by availability of the IP outside the trial. Recruitment becomes more difficult, and patients assigned to control are more likely to drop-out and use the non-assigned IP, which may bias the observed treatment effect. In AA settings we propose a SCA composed of patient level data from previous clinical trials to augment or replace the randomized control. Validity of this approach in one case study is assessed by examining if a SCA can replicate the outcomes of a target randomized control (TRC) from a recent NSCLC trial. Methods: The patients for the NSCLC SCA were required to have satisfied the key eligibility criteria of the target trial and were further selected using a propensity score-based approach to balance the baseline characteristics in the SCA and TRC. All patient selections were made without knowledge of patient outcomes. Results: The results show comparable balance in observed baseline characteristics of the SCA and TRC was achieved. Overall survival (OS) in TRC was replicated by SCA. The Kaplan Meier curves for OS in the SCA and TRC visually overlap. In addition, the log rank test (p = 0.65) and hazard ratio of 1.04 (95% CI: (0.88, 1.23)) were not statistically significant. Conclusions: If the SCA had been in place of the randomized control in this study, conclusions about the treatment effect would have been the same. While this may not hold when it is not possible to balance the groups on all confounders, this suggests that in some settings, SCA could augment or replace the randomized control in future trials easing recruitment, retention, and crossover challenges without compromising the understanding of the treatment effect. Future work should examine in what settings SCA is appropriate and consider the implications of potential unobserved confounders.

3371 The Devil is in the Details: Unbalanced Gains in Healthcare Access and Affordability in the Health Insurance Exchanges

OBJECTIVES/SPECIFIC AIMS: Evaluate how access and affordability has changed before and after the implementation HIEs in three subpopulations. The subpopulations are individuals who are currently insured through the HIE but were previously: 1. Insured through Employment-based insurance (PEBI subpopulation) 2. Insured through Private Insurance (PPI subpopulation) and 3. Uninsured (PU subpopulation). The three access and affordability measures are: Outcome measure 1. Did not fill a prescription in the past year due to cost Outcome measure 2. Could not get needed medical exam in the past year due to cost and Outcome measure 3. Had problems paying medical bills in the past year. METHODS/STUDY POPULATION: We analyzed the de-identified public use data from the 2012 and 2015 Ohio Medicaid Assessment Survey (OMAS). Sponsored by the Ohio Department of Medicaid, Ohio Department of Health, and the Ohio State University, the OMAS is a representative cross-sectional survey of non-institutionalized Ohio residents, regardless of their Medicaid status. In order to “longitudinalize” the 2012 and 2015 cross-sectional data of the OMAS, we employed a propensity score-based approach. We started with the 2015 OMAS, and carefully characterized each of the PEBI, PPI, and PU subpopulations along 17 demographic, health utilization, health behavior, and health status covariates using a propensity score model. Then, we identified controls for the three subpopulations within the 2012 OMAS data using the propensity scores. Finally, we estimated the odds ratios for the three outcome measures between 2012 and 2015. RESULTS/ANTICIPATED RESULTS: In 2015 there were approximately 201,381 adults (unweighted count = 996) who were insured through the HIE in Ohio. Of those individuals, 17.7% fell into the PEBI subpopulation, 17.6% fell into the PPI subpopulation, and 53.3% fell into the PU subpopulation; the balance of the respondents (11.4%) reported previously having Medicaid, or “Other” insurance. There are several key differences in the covariates at baseline between the three subpopulations. In general, the PU subpopulation tended to younger, more minority, more socioeconomically disadvantaged, and more likely to not have a primary care provider compared to both the PEBI and PPI subpopulations. In the 2012 data, we were able to identify 170 controls for the PEBI subpopulation, 167 controls for the PPI subpopulation, and 516 controls for the PU subpopulation. Compared to 2012, in 2016 (after the implementation of the HIEs):. Outcome measure 1: The PEBI subpopulation was more likely to report not filling a prescription in the past year due to cost (there were no significant changes in the PPI or PU subpopulations). Outcome measure 2: The PEBI subpopulation was more likely to report not getting a needed medical exam or medical supplies in the past year due to cost. The PPI subpopulation was less likely to report not getting a needed medical exam or medical supplies in the past year due to cost. There were no significant changes for the PU subpopulation for this outcome measure. Outcome measure 3: There were no changes in the “had problems paying medical bill in the past year” outcome across the three subpopulations. DISCUSSION/SIGNIFICANCE OF IMPACT: This is among the most detailed studies of health insurance exchanges known to the investigators. Analyzing outcomes at the subpopulation level illustrates that there have been unbalanced gains in access and affordability as a result of the HIEs. In general, those who were previously insured through employer-based insurance saw their access and affordability decrease; those previously insured through private insurance saw modest increases to access and affordability; and perhaps most surprising, those that were previously uninsured saw no changes to their access and affordability. Future studies will incorporate 2017 OMAS data (when it becomes available) to see if these trends persist over time. During this time of rapid health systems and health policy change, our study adds an important contribution to the discussion surrounding how to best deliver highly effective and efficient health care.

Propensity Score-Based Approaches in High Dimension for Pharmacovigilance Signal Detection: an Empirical Comparison on the French Spontaneous Reporting Database.

Classical methods used for signal detection in pharmacovigilance rely on disproportionality analysis of counts aggregating spontaneous reports of a given adverse drug reaction. In recent years, alternative methods have been proposed to analyze individual spontaneous reports such as penalized multiple logistic regression approaches. These approaches address some well-known biases resulting from disproportionality methods. However, while penalization accounts for computational constraints due to high-dimensional data, it raises the issue of determining the regularization parameter and eventually that of an error-controlling decision rule. We present a new automated signal detection strategy for pharmacovigilance systems, based on propensity scores (PS) in high dimension. PSs are increasingly used to assess a given association with high-dimensional observational healthcare databases in accounting for confusion bias. Our main aim was to develop a method having the same advantages as multiple regression approaches in dealing with bias, while relying on the statistical multiple comparison framework as regards decision thresholds, by considering false discovery rate (FDR)-based decision rules. We investigate four PS estimation methods in high dimension: a gradient tree boosting (GTB) algorithm from machine-learning and three variable selection algorithms. For each (drug, adverse event) pair, the PS is then applied as adjustment covariate or by using two kinds of weighting: inverse proportional treatment weighting and matching weights. The different versions of the new approach were compared to a univariate approach, which is a disproportionality method, and to two penalized multiple logistic regression approaches, directly applied on spontaneous reporting data. Performance was assessed through an empirical comparative study conducted on a reference signal set in the French national pharmacovigilance database (2000–2016) that was recently proposed for drug-induced liver injury. Multiple regression approaches performed better in detecting true positives and false positives. Nonetheless, the performances of the PS-based methods using matching weights was very similar to that of multiple regression and better than with the univariate approach. In addition to being able to control FDR statistical errors, the proposed PS-based strategy is an interesting alternative to multiple regression approaches.

Use of phosphate-binders and risk of infection-related and all-cause mortality in patients undergoing hemodialysis: The Q-Cohort Study

The use of phosphate (P)-binders allows hemodialysis patients to take in more protein and thus may maintain a good nutritional status. Protein-energy-malnutrition increases the risk of infection-related death. The association between use of P-binders and the relative risks of infection-related death remains unknown in hemodialysis patients. A total of 2926 hemodialysis patients registered to the Q-Cohort Study was followed up for 4-years. The association between use of P-binders and the risks for infection-related and all-cause mortality were estimated by Cox proportional hazards risk model with multiple adjustments by conventional and propensity-score based approaches. During the follow-up period, 106 patients and 492 patients died of infection and any cause, respectively. Cox proportional hazards models with multivariable adjustments including nutritional confounders showed that the incidence of infection-related death was significantly lower in patients with P-binders use compared with those without (hazard ratio [95% confidence interval] for infection-related mortality 0.63 [0.40–0.99]). The results remained significant even after applying four different propensity score-based analyses. Notably, use of P-binders was associated with a lower risk of all-cause mortality. Further studies including randomized controlled clinical trials and observational studies analyzed by an instrumental variable model will provide more robust evidences for the associations observed in our study.

Study on Dialysis Session Length and Mortality in Maintenance Hemodialysis Patients: The Q-Cohort Study

Objectives: Hemodialysis (HD) time has been recognized as an important factor in dialysis adequacy. However, few studies have reported on associations between HD time and prognosis among maintenance HD patients. We present some findings from a prospective cohort study, the ­Q-Cohort Study, which was set up to explore risk factors for mortality in Japanese HD patients. We hypothesized that HD ≥5 h was associated with a significant survival advantage compared with HD < 5 h. The present study examined association between HD time and mortality in Japanese HD patients. Methods: The prospective multicenter Q-Cohort Study was conducted between December 2006 and December 2010, following 3,456 Japanese HD patients for 4 years. We examined the association between HD time and prognosis using Cox proportional hazards modeling. Propensity scores were calculated using logistic regression. Results: During follow-up, 566 patients died from any cause. Patients with HD ≥5 h (n = 2,141) showed ­significantly lower risk of all-cause death (hazards ratio = 0.82; 95% CI 0.68–0.99) than those with HD < 5 h (n = 1,315), after adjusting for confounding risk factors. This ­association remained significant using a propensity score-based approach. After stratifying the analysis by patient age in 10-year increments, this finding remained ­significant only in patients who were ≥80 years of age. Conclusion: Our results suggest that HD ≥5 h has a more favorable effect on mortality than HD < 5 h.

Propensity score analysis with partially observed covariates: How should multiple imputation be used?

Inverse probability of treatment weighting is a popular propensity score-based approach to estimate marginal treatment effects in observational studies at risk of confounding bias. A major issue when estimating the propensity score is the presence of partially observed covariates. Multiple imputation is a natural approach to handle missing data on covariates: covariates are imputed and a propensity score analysis is performed in each imputed dataset to estimate the treatment effect. The treatment effect estimates from each imputed dataset are then combined to obtain an overall estimate. We call this method MIte. However, an alternative approach has been proposed, in which the propensity scores are combined across the imputed datasets (MIps). Therefore, there are remaining uncertainties about how to implement multiple imputation for propensity score analysis: (a) should we apply Rubin’s rules to the inverse probability of treatment weighting treatment effect estimates or to the propensity score estimates themselves? (b) does the outcome have to be included in the imputation model? (c) how should we estimate the variance of the inverse probability of treatment weighting estimator after multiple imputation? We studied the consistency and balancing properties of the MIte and MIps estimators and performed a simulation study to empirically assess their performance for the analysis of a binary outcome. We also compared the performance of these methods to complete case analysis and the missingness pattern approach, which uses a different propensity score model for each pattern of missingness, and a third multiple imputation approach in which the propensity score parameters are combined rather than the propensity scores themselves (MIpar). Under a missing at random mechanism, complete case and missingness pattern analyses were biased in most cases for estimating the marginal treatment effect, whereas multiple imputation approaches were approximately unbiased as long as the outcome was included in the imputation model. Only MIte was unbiased in all the studied scenarios and Rubin’s rules provided good variance estimates for MIte. The propensity score estimated in the MIte approach showed good balancing properties. In conclusion, when using multiple imputation in the inverse probability of treatment weighting context, MIte with the outcome included in the imputation model is the preferred approach.

High-Flow Nasal Cannula Oxygenation in Immunocompromised Patients With Acute Hypoxemic Respiratory Failure: A Groupe de Recherche Respiratoire en Réanimation Onco-Hématologique Study.

In immunocompromised patients with acute respiratory failure, invasive mechanical ventilation remains associated with high mortality. Choosing the adequate oxygenation strategy is of the utmost importance in that setting. High-flow nasal oxygen has recently shown survival benefits in unselected patients with acute respiratory failure. The objective was to assess outcomes of immunocompromised patients with hypoxemic acute respiratory failure treated with high-flow nasal oxygen. We performed a post hoc analysis of a randomized controlled trial of noninvasive ventilation in critically ill immunocompromised patients with hypoxemic acute respiratory failure. Twenty-nine ICUs in France and Belgium. Critically ill immunocompromised patients with hypoxemic acute respiratory failure. A propensity score-based approach was used to assess the impact of high-flow nasal oxygen compared with standard oxygen on day 28 mortality. Among 374 patients included in the study, 353 met inclusion criteria. Underlying disease included mostly malignancies (n = 296; 84%). Acute respiratory failure etiologies were mostly pneumonia (n = 157; 44.4%) or opportunistic infection (n = 76; 21.5%). Noninvasive ventilation was administered to 180 patients (51%). Invasive mechanical ventilation was ultimately needed in 142 patients (40.2%). Day 28 mortality was 22.6% (80 deaths). Throughout the ICU stay, 127 patients (36%) received high-flow nasal oxygen whereas 226 patients received standard oxygen. Ninety patients in each group (high-flow nasal oxygen or standard oxygen) were matched according to the propensity score, including 91 of 180 (51%) who received noninvasive ventilation. High-flow nasal oxygen was neither associated with a lower intubation rate (hazard ratio, 0.42; 95% CI, 0.11-1.61; p = 0.2) nor day 28 mortality (hazard ratio, 0.80; 95% CI, 0.45-1.42; p = 0.45). In immunocompromised patients with hypoxemic acute respiratory failure, high-flow nasal oxygen when compared with standard oxygen did not reduce intubation or survival rates. However, these results could be due to low statistical power or unknown confounders associated with the subgroup analysis. A randomized trial is needed.

Conducting Privacy-Preserving Multivariable Propensity Score Analysis When Patient Covariate Information Is Stored in Separate Locations.

Distributed networks of health-care data sources are increasingly being utilized to conduct pharmacoepidemiologic database studies. Such networks may contain data that are not physically pooled but instead are distributed horizontally (separate patients within each data source) or vertically (separate measures within each data source) in order to preserve patient privacy. While multivariable methods for the analysis of horizontally distributed data are frequently employed, few practical approaches have been put forth to deal with vertically distributed health-care databases. In this paper, we propose 2 propensity score-based approaches to vertically distributed data analysis and test their performance using 5 example studies. We found that these approaches produced point estimates close to what could be achieved without partitioning. We further found a performance benefit (i.e., lower mean squared error) for sequentially passing a propensity score through each data domain (called the "sequential approach") as compared with fitting separate domain-specific propensity scores (called the "parallel approach"). These results were validated in a small simulation study. This proof-of-concept study suggests a new multivariable analysis approach to vertically distributed health-care databases that is practical, preserves patient privacy, and warrants further investigation for use in clinical research applications that rely on health-care databases.

On estimating regression-based causal effects using sufficient dimension reduction

SUMMARY In many causal inference problems the parameter of interest is the regression causal effect, defined as the conditional mean difference in the potential outcomes given covariates. In this paper we discuss how sufficient dimension reduction can be used to aid causal inference, and we propose a new estimator of the regression causal effect inspired by minimum average variance estimation. The estimator requires a weaker common support condition than propensity score-based approaches, and can be used to estimate the average causal effect, for which it is shown to be asymptotically super-efficient. Its finite-sample properties are illustrated by simulation.

Causal effect estimation strategies in a longitudinal study with complex time-varying confounders: A tutorial.

The Dutch Sciatica Trial represents a longitudinal study with complex time-varying confounders as patients with poorer health conditions (e.g. more severe pain) are more likely to opt for surgery, which, in turn, may affect future outcomes (pain severity). A straightforward classical as-treated comparison at the end point would lead to biased estimation of the surgery effect. We present several strategies of causal treatment effect estimation that might be applicable for analyzing such data. These include an inverse probability of treatment weighted regression analysis, a marginal weighted analysis, an unweighted regression analysis, and several propensity score-based approaches. In addition, we demonstrate how to evaluate these approaches in a thorough simulation study where we generate various realistic complex confounding patterns akin to the sciatica study.