Bayesian Confidence Intervals for Multiplexed Proteomics Integrate Ion-statistics with Peptide Quantification Concordance

Multiplexed proteomics has emerged as a powerful tool to measure relative protein expression levels across multiple conditions. The relative protein abundances are inferred by comparing the signals generated by isobaric tags, which encode the samples' origins. Intuitively, the trust associated with a protein measurement depends on the similarity of ratios from the protein's peptides and the signal-strength of these measurements. However, typically the average peptide ratio is reported as the estimate of relative protein abundance, which is only the most likely ratio with a very naive model. Moreover, there is no sense on the confidence in these measurements. Here, we present a mathematically rigorous approach that integrates peptide signal strengths and peptide-measurement agreement into an estimation of the true protein ratio and the associated confidence (BACIQ). The main advantages of BACIQ are: (1) It removes the need to threshold reported peptide signal based on an arbitrary cut-off, thereby reporting more measurements from a given experiment; (2) Confidence can be assigned without replicates; (3) For repeated experiments BACIQ provides confidence intervals for the union, not the intersection, of quantified proteins; (4) For repeated experiments, BACIQ confidence intervals are more predictive than confidence intervals based on protein measurement agreement. To demonstrate the power of BACIQ we reanalyzed previously published data on subcellular protein movement on treatment with an Exportin-1 inhibiting drug. We detect ∼2× more highly significant movers, down to subcellular localization changes of ∼1%. Thus, our method drastically increases the value obtainable from quantitative proteomics experiments, helping researchers to interpret their data and prioritize resources. To make our approach easily accessible we distribute it via a Python/Stan package.

Prognostic value of transcriptomic determination of tumour-infiltrating lymphocytes in localised breast cancer

PURPOSE

Tumour-infiltrating lymphocyte (TIL) detection by histology is associated with outcomes in breast cancer; nevertheless, analysis standardisation is difficult. We determined whether transcriptomic data could generate a genomic signature that estimated TIL infiltrates and determined patient prognosis in localised breast cancer.

EXPERIMENTAL DESIGN

Using 1928 transcriptomic profiles of pure cells, we generated a genetic signature specific to lymphocyte, myeloid, stromal and cancer cells. We then computed a score based on this signature and tested the association between the score and the TILs estimated for patients in an adjuvant setting from public and private data sets. We tested the capacity of the transcriptomic RNA TIL score to predict disease-free survival (DFS) or overall survival (OS) through multivariate Cox models adjusted for classical clinical variables and PAM50 molecular classification in two public data sets (Carte d'Identité des Tumeurs [CIT], n = 530; Metabric, n = 1832).

RESULTS

A high RNA TIL score was significantly associated with the presence of a high level of TILs as assessed by histology. The score was also associated with DFS and OS in multivariate Cox models adjusted for molecular and clinical variables (CIT: OS hazard ratio [HR] = 0.15 [0.04, 0.61], p-value = 0.007; DFS: 0.27 [0.08, 0.8] p-value = 0.02; Metabric: OS HR = 0.87 [0.77, 0.97], p-value = 0.01). The association between the RNA TIL score and survival was tested by univariate analysis in each molecular subgroup; the RNA TIL score was associated with survival only in basal-like tumours.

CONCLUSIONS

Determination of the TIL rate using a transcriptomic signature is feasible and has a high prognostic value in patients with basal-like tumours in an adjuvant setting.

Comparative view for the impact of five eco factors on species distribution and weed community structure in Isthmus of Suez and adjoining farmland east Nile delta, Egypt

The study of the extent of variability and the order of importance for the impact of environmental factors on species distribution and weed community structure from one region to the other is an interesting research subject. The present study aimed to discuss the issue to give a comparative view with the author's findings about the impact of the prevailing climate, soil type, crop type, crop sustainability and urbanization on species distribution and weed community structure in the coastal farmland and adjacent territories in northwest delta region (Mahgoub, 2019). A new sample area selected and comprised the reclaimed land of El Ballah region in Isthmus of Suez and adjoining farmland east Nile delta. A total of 245 species were recorded. Agglomerative Hierarchical Clustering (AHC) identified four vegetative sociation groups (VSG). The diversity of the four identified VSG or weed communities was evaluated at different levels. Parincipal Component Analysis (PCA) indicated the influence of the five eco factors on species distribution and variability of weed community structure, summarized the relationships among variables and investigated the proximity among samples and how they related to variables. ANOVA followed by Tukey's test were applied twice for the resulted VSG, one depending on soil variables as explanatory variables and the other on sampling site's indicative scores for the five eco factors. The results of ANOVA (R², F, P), sample variance (S²) and other multivariate analyses indicated a different order of importance for the impact of the five eco factors in comparable to the former study denoted above. The soil type was the most impacting factor on species distribution and weed community structure, followed by crop type, crop sustainability, prevailing climate and urbanization, respectively.

The Hermans-Rasson test as a powerful alternative to the Rayleigh test for circular statistics in biology

BACKGROUND

Circular data are gathered in diverse fields of science where measured traits are cyclical in nature: such as compass directions or times of day. The most common statistical question asked of a sample of circular data is whether the data seems to be drawn from a uniform distribution or one that is concentrated around one or more preferred directions. The overwhelmingly most-popular test of the null hypothesis of uniformity is the Rayleigh test, even though this test is known to have very low power in some circumstances. Here we present simulation studies evaluating the performance of tests developed as alternatives to the Rayleigh test.

RESULTS

The results of our simulations demonstrate that a single test, the Hermans and Rasson test is almost as powerful as the Rayleigh test in unimodal situations (when the Rayleigh test does well) but substantially outperforms the Rayleigh test in multimodal situations.

CONCLUSION

We recommend researchers switch to routine use of the new Hermans and Rasson test. We also demonstrate that all available tests have low power to detect departures from uniformity involving more than two concentrated regions: we recommend that where researchers suspect such complex departures that they collect substantially-sized samples and apply another recent test due to Pycke that was designed specifically for such complex cases. We provide clear textual descriptions of how to implement each of these recommended tests and encode them in R functions that we provide.

HIV-1 mother-to-child transmission in Brazil (1994-2016): a time series modeling

HIV-1 mother-to-child transmission (HIV-1 MTCT), is an important cause of children mortality worldwide. Brazil has been traditionally praised by its HIV/Aids program, which provides free-of-charge care for people living with HIV-1. Using public epidemiology and demographic databases, we aimed at modeling HIV-1 MTCT prevalence in Brazil through the years (1994-2016) and elaborate a statistical model for forecasting, contributing to HIV-1 epidemiologic surveillance and healthcare decision-making. We downloaded sets of live births and mothers' data alongside HIV-1 cases notification in children one year old or less. Through time series modeling, we estimated prevalence along the years in Brazil, and observed a remarkable decrease of HIV-1 MTCT between 1994 (10 cases per 100,000 live births) and 2016 (five cases per 100,000 live births), a reduction of 50%. Using our model, we elaborated a prognosis for each Brazilian state to help HIV-1 surveillance decision making, indicating which states are in theory in risk of experiencing a rise in HIV-1 MTCT prevalence. Ten states had good (37%), nine had mild (33%), and eight had poor prognostics (30%). Stratifying the prognostics by Brazilian region, we observed that the Northeast region had more states with poor prognosis, followed by North and Midwest, Southeast and South with one state of poor prognosis each. Brazil undoubtedly advanced in the fight against HIV-1 MTCT in the past two decades. We hope our model will help indicating where HIV-1 MTCT prevalence may rise in the future and support government decision makers regarding HIV-1 surveillance and prevention.

Multicollinearity and misleading statistical results

Multicollinearity represents a high degree of linear intercorrelation between explanatory variables in a multiple regression model and leads to incorrect results of regression analyses. Diagnostic tools of multicollinearity include the variance inflation factor (VIF), condition index and condition number, and variance decomposition proportion (VDP). The multicollinearity can be expressed by the coefficient of determination (R_h²) of a multiple regression model with one explanatory variable (X_h) as the model’s response variable and the others (X_i [i≠h] as its explanatory variables. The variance (σ_h²) of the regression coefficients constituting the final regression model are proportional to the VIF (11-Rh2). Hence, an increase in R_h² (strong multicollinearity) increases σ_h². The larger σ_h² produces unreliable probability values and confidence intervals of the regression coefficients. The square root of the ratio of the maximum eigenvalue to each eigenvalue from the correlation matrix of standardized explanatory variables is referred to as the condition index. The condition number is the maximum condition index. Multicollinearity is present when the VIF is higher than 5 to 10 or the condition indices are higher than 10 to 30. However, they cannot indicate multicollinear explanatory variables. VDPs obtained from the eigenvectors can identify the multicollinear variables by showing the extent of the inflation of σ_h² according to each condition index. When two or more VDPs, which correspond to a common condition index higher than 10 to 30, are higher than 0.8 to 0.9, their associated explanatory variables are multicollinear. Excluding multicollinear explanatory variables leads to statistically stable multiple regression models.

A biostatistical support system in health sciences: is this sustainable in a resource-restricted environment?

Background

Training in biostatistics is important for strengthening capacity in health research. This is particularly true for Africa, where research output in the health sciences has been low. Training initiatives for the continent are therefore essential. The aim of the present study was to analyse the quality and financial sustainability of the expanded biostatistical support system at a South African health sciences institution between 2013 and 2017.

Methods

A cross-sectional investigation of the initiatives created between the years 2013 and 2017 in the University of the Witwatersrand, Faculty of Health Sciences Research Office was undertaken. An assessment of the one-on-one consultations carried out by postgraduate students and staff, financial costs of the support system and the number of postgraduate student graduations were analysed.

Results

The number of statistical consultations increased over the period examined. The consultations were highly recommended by the postgraduate students and staff (consulters). A clear rise in the number of Masters and PhD student graduates and an increase in research units were observed from 2013 to 2017, although these cannot be solely associated with the biostatistical support system. The finances for maintaining the support system are cost effective as the number of graduates increases. The total cost to the Research Office is US$ 225 per graduate per annum.

Conclusions

The expansion of the biostatistical support system has indirectly contributed to an increased number of graduates and research publication units in the institution. While the current finances support the system, any increases in enrolments or growth in diversification of biostatistical requirements may place a strain on the financial sustainability. This service is of value to developed and developing countries.

SSDs revisited: part II-practical considerations in the development and use of application factors applied to species sensitivity distributions

Application factors are routinely applied in the extrapolation of laboratory aquatic toxicity data to ensure protection from exposure to chemicals in the natural environment. The magnitude of the application factor is both a scientific and a policy decision, but in any case, it should be rooted in scientific knowledge so as to not be arbitrary. Information-rich chemicals are often subjected to species sensitivity distribution (SSD) analysis to transparently describe certain aspects of assessment uncertainty and are normally subjected to much smaller application factors than screening information data sets. We describe a new set of tools useful to assess the quality of SSDs. Twenty-two data sets and 19 chemicals representing agrochemicals, biocides, surfactants, metals, and common wastewater contaminants were compiled to demonstrate how the tools can be used. "Add-one-in" and "leave-one-out" simulations were used to investigate SSD robustness and develop quantitative evidence for the use of application factors. Theoretical new toxicity data were identified for add-one-in simulations based on the expected probabilities necessary to lower the hazardous concentration to 5% of a species (HC5) by a factor of 2, 3, 5, or 10. Simulations demonstrate the basis for application factors in the range of 1 to 5 for well-studied chemicals with high-quality SSDs. Leave-one-out simulations identify the fact that the most influential values in the SSD come from the extremes of the sensitive and tolerant toxicity values. Mesocosm and field data consistently demonstrate that HC5s are conservative, further justifying the use of small application factors for high-quality SSDs. Environ Toxicol Chem 2019;38:1526-1541. © 2019 SETAC.

General health measures in shoulder surgery: are we powered for success?

BACKGROUND

Surgeons, policymakers, and payers increasingly use changes in general health to guide decision-making. It is unknown how such measures are incorporated into shoulder surgery research, how strongly they are associated with changes in shoulder-specific outcomes, and whether they are appropriately powered.

METHODS

PubMed was searched for articles reporting shoulder-specific and general health measures after rotator cuff repair and total shoulder arthroplasty. Study characteristics, results at the study group level, reporting of power calculations, and statistical significance were recorded. Meta-regression was employed to describe the association of changes between shoulder-specific and general health measures.

RESULTS

Of 360 identified abstracts, 21 articles with 28 patient groups were included. Only 1 article was published before 2000. There was a strong association of changes between shoulder-specific and general health measures (r = 0.66; P < .001). Power calculations were mentioned in 33% of studies and based on shoulder-specific measures. Of 20 studies conducting hypothesis tests, 75% reported agreement regarding the statistical significance of shoulder-specific and general health tests. Of 5 discordant studies, 4 found the shoulder-specific measure statistically significant and not the general health measure.

CONCLUSION

Shoulder surgery research increasingly reports changes in general health measures that are associated with changes in shoulder-specific measures, suggesting that improvements in shoulder symptoms increase quality of life. When disagreement exists, it usually results from the general health measure's not meeting statistical significance, which may simply reflect type II error. Research reporting general health measures should carefully report power considerations to avoid misinterpretation of findings failing to reach statistical significance.

SSDs Revisited: Part I-A Framework for Sample Size Guidance on Species Sensitivity Distribution Analysis

We propose a framework on sample size for species sensitivity distribution (SSD) analyses, with perspectives on Bayesian, frequentist, and even nonparametric approaches to estimation. The intent of a statistical sample size analysis is to ensure that the implementation of a statistical model will satisfy a minimum performance standard when relevant conditions are met. It requires that a statistical model be fully specified and that the means of measuring its performance as a function of sample size be detailed. Defining the model conditions under which sample size is calculated is often the most difficult, and important, aspect of sample size analysis because if the model is not representative, then the sample size analysis will provide incorrect guidance. Definitive guidance on sample size requires general agreement on representative models and their performance from stakeholders in important domains such as chemical safety assessments involving government regulators and industry; the present study provides an initial framework that could be used to this end in the future. In addition, our analysis provides immediate value for understanding how well current SSD analyses perform under a few basic models, sample sizes, and quantitative performance criteria. The results confirm that many analyses are adequately sized to estimate hazardous concentration percentile values (typically the 5th percentile for chemical hazard assessments). However, on the low end of sizes seen in common practice, hazardous concentration estimates can be more than 1 order of magnitude greater than the model-defined value. Environ Toxicol Chem 2019;38:1514-1525. © 2019 SETAC.

Construction and preliminary evaluation of the inpatient glycemic control questionnaire (IGCQ): a survey tool assessing perceptions and knowledge of resident physicians

BACKGROUND

Uncontrolled hyperglycemia in hospitalized patients, with or without diabetes mellitus, is associated with many adverse outcomes. Resident physicians are the primary managers of inpatient glycemic control (IGC) in many academic and community medical centers; however, no validated survey tools related to their perceptions and knowledge of IGC are currently available. As identification of common barriers to successful IGC amongst resident physicians may help foster better educational interventions (ultimately leading to improvements in IGC and patient care), we sought to construct and preliminarily evaluate such a survey tool.

METHODS

We developed the IGC questionnaire (IGCQ) by using previously published but unvalidated survey tools related to physician perspectives on inpatient glycemic control as a framework. We administered the IGCQ to a cohort of resident physicians from the University of Mississippi Medical Center, University of Louisville, Emory University, and the University of Virginia. We then used classical test theory and Rasch Partial Credit Model analyses to preliminarily evaluate and revise the IGCQ. The final survey tool contains 16 total items and three answer-choice categories for most items.

RESULTS

Two hundred forty-six of 438 (56.2%) eligible resident physicians completed the IGCQ during various phases of development.

CONCLUSIONS

We constructed and preliminarily evaluated the IGCQ, a survey tool that may be useful for future research into resident physician perceptions and knowledge of IGC. Future studies could seek to externally validate the IGCQ and then utilize the survey tool in pre- and post-intervention assessments.

An Introduction to Statistics - Data Types, Distributions and Summarizing Data

The purpose of research is to gather data, which can then be used to inform decision-making. Data can be of various types and an understanding of this is crucial for its proper analysis and interpretation. In this article, we look at various types and distributions of data, and methods to summarize this data.

How to cite this article

Ranganathan P, Gogtay NJ, An Introduction to Statistics - Data Types, Distributions and Summarizing Data. Indian J Crit Care Med 2019;23(Suppl 2):S169-S170.

WASP (Write a Scientific Paper): Qualitative versus quantitative analyses

Quantitative and qualitative analyses and research are conventionally regarded as fundamentally different techniques. However, their objectives and applications overlap, and the decision of which of the two methods to apply depends not only on the nature and volume of data that is to be collected, but also the actual type of data and the resources at hand. Indeed, the outright combination of the two methods is becoming increasingly common. This collection will cover some aspects of qualitative analyses techniques.

[The role of biostatistics in institutional review boards]

Research in humans is associated with risks. These risks are only justifiable if an independent institutional review board (IRB) has evaluated the planned research in terms of scientific integrity. Only scientifically sound research can be considered ethical. A biostatistician should be a member of the IRB to assure adequate evaluation of fundamental topics like design, sample size estimation, and statistical analysis of the study.This paper presents core biostatistical concepts following the current guidelines of the International Council of Harmonization (ICH E6 and ICH E9). We discuss important pitfalls based on examples from published clinical trials. Furthermore, we discuss new concepts like estimands and their relevance for biostatisticians working in IRBs. Finally, we discuss the role of biostatisticians in IRBs and present thoughts on the way they should be trained.

Dr. Hossein Malekafzali: A Pioneer in Iran's Health Evolution and Development

Dr Malekafzali, an elite biostatistics professor at Tehran University of Medical Sciences, in his more than 50 years of glorious service, has played a crucial role in creation of fundamental evolution in public health, reproductive health and development of applied research in Iran. He has left lasting activities in administrative positions such as health and research deputies, health minister consultant, director of health faculty and director of health research institutes. He published several books and articles on statistics, epidemiology and public health. This article is a review of his worthy and interesting activities in the country's health, research and education.

More about the basic assumptions of t-test: normality and sample size

Most parametric tests start with the basic assumption on the distribution of populations. The conditions required to conduct the t-test include the measured values in ratio scale or interval scale, simple random extraction, normal distribution of data, appropriate sample size, and homogeneity of variance. The normality test is a kind of hypothesis test which has Type I and II errors, similar to the other hypothesis tests. It means that the sample size must influence the power of the normality test and its reliability. It is hard to find an established sample size for satisfying the power of the normality test. In the current article, the relationships between normality, power, and sample size were discussed. As the sample size decreased in the normality test, sufficient power was not guaranteed even with the same significance level. In the independent t-test, the change in power according to sample size and sample size ratio between groups was observed. When the sample size of one group was fixed and that of another group increased, power increased to some extent. However, it was not more efficient than increasing the sample sizes of both groups equally. To ensure the power in the normality test, sufficient sample size is required. The power is maximized when the sample size ratio between two groups is 1 : 1.

Presenting scientific work-news media theory in presentations, abstracts, and posters

The effective presentation of scientific work in abstracts, posters, and PowerPoint presentations resembles popular newspaper writing far more than it does traditional paper writing. In the former three, a good "hook" is essential and must be obvious and apparent up front. A slight element of truthful sensationalism is therefore often helpful. Research preparation and presentation and skills of all sorts are increasingly critical in the incrementally competitive world of academia. This paper will offer factual and practical guidelines with regards to preparing research in these formats since self-presentation is essential for career progression. It is not only what we do but how we are seen to be doing it, acceding to Batman's aphorism: "it's not who I am underneath but what I do that defines me."

Statistical evaluation of diagnostic tests: A primer for pediatric surgeons

BACKGROUND/PURPOSE

Diagnostic tests are of paramount importance for informing decision making in the surgical setting. Certain statistical methods are necessary to properly analyze data for diagnostic or prognostic tests involving biomarkers and risk factor data. Our goal is to provide a useful primer for the surgical researcher when performing a diagnostic research study in order to best analyze their data.

METHODS

We present the key concepts and statistics for diagnostic tests and receiver operating characteristic (ROC) curve analysis, and we illustrate each with hypothetical surgery research examples. We use hypothetical data regarding CT imaging and WBC count in their diagnostic ability in predicting acute appendicitis, an extremely common surgical condition, while reviewing the statistical concepts of sensitivity, specificity, positive and negative predictive value, positive and negative likelihood ratio, relative risk, odds ratio, and ROC curves. Then we will consider a hypothetical a risk factor analysis on 30-day readmission to illustrate how multiple predictors can be combined.

CONCLUSIONS

The statistical concepts presented are useful to the pediatric surgeon researcher in assessing the ability of diagnostic tests, which will translate into decision making and patient management implications in the clinical setting.

TYPE OF STUDY

Review Article LEVEL OF EVIDENCE: N/A.

Implementing online interventions in ICare: A biostatistical perspective

The implementation of research studies is a highly complex process. All decisions with respect to the study design impact the statistical analyses and interpretation of the results. Within the ICare research project (EU H2020 Grant agreement 634757) seven research trials are conducted to generate evidence on efficacy, effectiveness and the dissemination potential of online interventions targeting eating disorders, common mental health problems and resilience. Within the project a central biometrical unit was established to manage and coordinate data collection, processing and statistical data analysis. This allows for harmonized trial planning, conduct, data management processes and analysis strategies. The purpose of this article is to describe the common concepts underlying all seven ICare trials. This includes development of (adaptive sequential) study designs, handling of missing values, general data management and processing as well as data protection aspects.

Estimating a Dynamic Effect of Soda Intake on Pediatric Dental Caries Using Targeted Maximum Likelihood Estimation Method

An effect of soda intake on dental caries in young children (birth to 5 years) may vary over time. Estimating a dynamic effect may be challenging due to time-varying confounding and loss to follow-up. The purpose of this paper is to demonstrate utility of targeted maximum likelihood estimation (TMLE) method in addressing longitudinal data analysis challenges and estimating a dynamic effect of soda intake on pediatric caries. Data came from the Detroit Dental Health Project, a 4-year cohort study of low-income -African-American children and caregivers. The sample included 995 child-caregiver pairs who participated in 2002-03 (W1) and were followed up in 2004-05 (W2) and 2007 (W3). The outcome was counts of caries surfaces at W3, and the exposure was child's soda intake at W1 and W2. Time-varying covariates included caregiver's smoking status, oral health fatalism, and social support. Forty-three percent of children consistently consumed soda at W1 and W2, whereas 21% were nonconsumers throughout 2 surveys. The remaining 35% switched intake status between W1 and W2. Association between soda intake patterns and caries was tested using TMLE. Children with a consistent soda intake had 1.03 more caries lesions at W3 than those with consistently no soda intake (95% CI 0.09-1.97) on average. If soda was consumed only at W1 or W2, an estimated effect of soda on caries development at W3 was no longer statistically significant. In conclusion, consistent soda intake during the early childhood led to one additional caries tooth surface. The study highlights utility of TMLE in pediatric caries research as it can handle modeling challenges associated with longitudinal data.

Integrated Identification and Quantification Error Probabilities for Shotgun Proteomics

Protein quantification by label-free shotgun proteomics experiments is plagued by a multitude of error sources. Typical pipelines for identifying differential proteins use intermediate filters to control the error rate. However, they often ignore certain error sources and, moreover, regard filtered lists as completely correct in subsequent steps. These two indiscretions can easily lead to a loss of control of the false discovery rate (FDR). We propose a probabilistic graphical model, Triqler, that propagates error information through all steps, employing distributions in favor of point estimates, most notably for missing value imputation. The model outputs posterior probabilities for fold changes between treatment groups, highlighting uncertainty rather than hiding it. We analyzed 3 engineered data sets and achieved FDR control and high sensitivity, even for truly absent proteins. In a bladder cancer clinical data set we discovered 35 proteins at 5% FDR, whereas the original study discovered 1 and MaxQuant/Perseus 4 proteins at this threshold. Compellingly, these 35 proteins showed enrichment for functional annotation terms, whereas the top ranked proteins reported by MaxQuant/Perseus showed no enrichment. The model executes in minutes and is freely available at https://pypi.org/project/triqler/.

The Charlson Index Is Insufficient to Control for Comorbidities in a National Trauma Registry

BACKGROUND

The Charlson Comorbidity Index (CCI) is frequently used to control for confounding by comorbidities in observational studies, but its performance as such has not been studied. We evaluated the performance of CCI and an alternative summary method, logistic principal component analysis (LPCA), to adjust for comorbidities, using as an example the association between insurance and mortality.

MATERIALS AND METHODS

Using all admissions in the National Trauma Data Bank 2010-2015, we extracted mortality, payment method, and 36 International Classification of Disease, Ninth Revision-derived comorbidities. We estimated odds ratios (ORs) for the association between uninsured status and mortality before and after adjusting for CCI, LPCA, and separate covariates. We also calculated standardized mean differences (SMDs) of comorbidity variables before and after weighting the sample using inverse probability of treatment weights for CCI, LPCA, and separate covariates.

RESULTS

In 4,936,880 admissions, most (68.3%) had at least one comorbidity. Considerable imbalance was observed in the unweighted sample (mean SMD = 0.086, OR = 1.17), which was almost entirely eliminated by inverse probability of treatment weights on separate covariates (mean SMD = 0.012, OR = 1.36). The CCI performed similarly to the unweighted sample (mean SMD = 0.080, OR = 1.25), whereas two LPCA axes were better able to control for confounding (mean SMD = 0.04, OR = 1.31). Using covariate adjustment, the CCI accounted for 56.1% of observed confounding, whereas two LPCA axes accounted for 91.3%.

CONCLUSIONS

The use of the CCI to adjust for confounding may result in residual confounding, and alternative strategies should be considered. LPCA may be a viable alternative to adjusting for each comorbidity when samples are small or positivity assumptions are violated.