An introduction to causal inference

Three controversies in health data science

The routine operation of modern healthcare systems produces a wealth of data in electronic health records, administrative databases, clinical registries, and other clinical systems. It is widely acknowledged that there is great potential for utilising these routine data for health research to derive new knowledge about health, disease, and treatments. However, the reuse of routine healthcare data for research is not beyond debate. In this paper, we discuss three issues that have stirred considerable controversy among health data scientists. First, we discuss van der Lei's 1st Law of Medical Informatics, which states that data shall be used only for the purpose for which they were collected. Then, we discuss to which extent routine data sources and innovations in analytical methods alleviate the need to conduct randomised clinical trials. Finally, we address questions of governance, privacy, and trust when routine health data are made available for research. While we don't think that there is a definite "right answer" for any of these issues, we argue that data scientists should be aware of the arguments for different viewpoints, respect their validity, and contribute constructively to the debate. The three controversies discussed in this paper relate to core challenges for research with health data and define an essential research agenda for the health data science community.

Maternal Zika virus infection and newborn microcephaly-an analysis of the epidemiological evidence

PURPOSE

To evaluate whether existing data and evidence support a causal link between maternal Zika virus (ZIKV) infection and newborn microcephaly.

METHODS

I quantified and compared the prevalence of all and severe microcephaly in Brazil, during and before 2015-2016, to assess whether an outbreak has occurred, used time series analysis to evaluate if the presumed outbreak was linked to a previous outbreak of ZIKV infections, and quantitatively synthesized published data from observational studies testing this association.

RESULTS

The prevalences of microcephaly in 2015-2016 were similar or lower than background levels (prevalence ratio [PR] for all microcephaly: 0.19; 95% confidence intervals [CI]: 0.17, 0.20). Changes in the number of cases of ZIKV infections at times matching 11-18 weeks of pregnancy were not followed by changes in the number of microcephaly cases (PR for infection at 12 weeks: 1.02; 95% CI: 0.99, 1.05). In observational studies, the prevalence of microcephaly was not significantly increased in newborns of Zika-infected mothers (average PR: 1.30; 95% CI: 0.84, 2.02).

CONCLUSIONS

Existing evidence is insufficient to claim maternal ZIKV infection causes microcephaly. Although a public health response seems sensible, it should be consistent with existing knowledge and consider risks, potential benefits and harm, and competing priorities.

Subjective reality and brain topology: Inversion transformations on non-orientable atomic surfaces of membrane channels

Subject-object relations reflect the relation of phenomenology and physics and are at the centre of interest in brain research and neuro-psychology. The unresolved dichotomy behind this relation is one of the most challenging questions of our time. Setting out from causal modelling I suggest a particular topology for subject-object relations and argue that we can find a physical realization in living organism that provides a continuous transform between both domains. In a geometrical metaphor this transform has the topological properties of a one-sided surface or non-orientable flat. I argue that such a surface can be found within the electronic organization of atomic linings in the filter region of ion-conducting membrane proteins. Electron transfer along these atomic surfaces makes chiral induced spin changes to a promising signature of subject-object relations and has found experimental evidence in previous studies. I finally advocate the view that there is a basic dualism between subject and object which is physical on both sides and realized by an inversion relation along one-sided surfaces. The transition between these two aspects however is non-physical and hosts the phenomenology that characterizes subjectivity.

Mortality by Timing of Hip Fracture Surgery: Factors and Relationships at Play

In hip fracture care, it is disputed whether mortality worsens when surgery is delayed. This knowledge gap matters when hospital managers seek to justify resource allocation for prioritizing access to one procedure over another. Uncertainty over the surgical timing-death association leads to either surgical prioritization without benefit or the underuse of expedited surgery when it could save lives. The discrepancy in previous findings results in part from differences between patients who happened to undergo surgery at different times. Such differences may produce the statistical association between surgical timing and death in the absence of a causal relationship. Previous observational studies attempted to adjust for structure, process, and patient factors that contribute to death, but not for relationships between structure and process factors, or between patient and process factors. In this article, we (1) summarize what is known about the factors that influence, directly or indirectly, both the timing of surgery and the occurrence of death; (2) construct a dependency graph of relationships among these factors based explicitly on the existing literature; (3) consider factors with a potential to induce covariation of time to surgery and the occurrence of death, directly or through the network of relationships, thereby explaining a putative surgical timing-death association; and (4) show how age, sex, dependent living, fracture type, hospital type, surgery type, and calendar period can influence both time to surgery and occurrence of death through chains of dependencies. We conclude by discussing how these results can inform the allocation of surgical capacity to prevent the avoidable adverse consequences of delaying hip fracture surgery.

simcausal R Package: Conducting Transparent and Reproducible Simulation Studies of Causal Effect Estimation with Complex Longitudinal Data

The simcausal R package is a tool for specification and simulation of complex longitudinal data structures that are based on non-parametric structural equation models. The package aims to provide a flexible tool for simplifying the conduct of transparent and reproducible simulation studies, with a particular emphasis on the types of data and interventions frequently encountered in real-world causal inference problems, such as, observational data with time-dependent confounding, selection bias, and random monitoring processes. The package interface allows for concise expression of complex functional dependencies between a large number of nodes, where each node may represent a measurement at a specific time point. The package allows for specification and simulation of counterfactual data under various user-specified interventions (e.g., static, dynamic, deterministic, or stochastic). In particular, the interventions may represent exposures to treatment regimens, the occurrence or non-occurrence of right-censoring events, or of clinical monitoring events. Finally, the package enables the computation of a selected set of user-specified features of the distribution of the counterfactual data that represent common causal quantities of interest, such as, treatment-specific means, the average treatment effects and coefficients from working marginal structural models. The applicability of simcausal is demonstrated by replicating the results of two published simulation studies.

Applying Nonparametric Methods to Analyses of Short-Term Fine Particulate Matter Exposure and Hospital Admissions for Cardiovascular Diseases among Older Adults

Short-term exposure to fine particulate matter (PM_2.5) has been associated with increased risks of cardiovascular diseases (CVDs), but whether such associations are supportive of a causal relationship is unclear, and few studies have employed formal causal analysis methods to address this. We employed nonparametric methods to examine the associations between daily concentrations of PM_2.5 and hospital admissions (HAs) for CVD among adults aged 75 years and older in Texas, USA. We first quantified the associations in partial dependence plots generated using the random forest approach. We next used a Bayesian network learning algorithm to identify conditional dependencies between CVD HAs of older men and women and several predictor variables. We found that geographic location (county), time (e.g., month and year), and temperature satisfied necessary information conditions for being causes of CVD HAs among older men and women, but daily PM_2.5 concentrations did not. We also found that CVD HAs of disjoint subpopulations were strongly predictive of CVD HAs among older men and women, indicating the presence of unmeasured confounders. Our findings from nonparametric analyses do not support PM_2.5 as a direct cause of CVD HAs among older adults.

Socioeconomic and air pollution correlates of adult asthma, heart attack, and stroke risks in the United States, 2010-2013

Asthma in the United States has become an important public health issue, with many physicians, regulators, and scientists elsewhere expressing concern that criterion air pollutants have contributed to a rising tide of asthma cases and symptoms. This paper studies recent associations (from 2008 to 2012) between self-reported asthma experiences and potential predictors, including age, sex, income, education, smoking, and county-level average annual ambient concentrations of ozone (O3) and fine particulate matter (PM2.5) levels recorded by the U.S. Environmental Protection Agency, for adults 50 years old or older for whom survey data are available from the Centers for Disease Control and Prevention (CDC) Behavioral Risk Factor Surveillance System (BRFSS). We also examine associations between these variables and self-reported heart attack and stroke experience; all three health outcomes are positively associated with each other. Young divorced women with low incomes are at greatest risk of asthma, especially if they are ever-smokers. Income is an important confounder of other relations. For example, in logistic regression modeling, PM2.5 is positively associated (p<0.06) with both stroke risk and heart attack risk when these are regressed only against PM2.5, sex, age, and ever-smoking status, but not when they are regressed against these variables and income. In this data set, PM2.5 is significantly negatively associated with asthma risk in regression models, with a 10μg/m³ decrease in PM2.5 corresponding to about a 6% increase in the probability of asthma, possibly because of confounding by smoking, which is negatively associated with PM2.5 and positively associated with asthma risk. A variety of non-parametric methods are used to quantify these associations and to explore potential causal interpretations.

Causal inference in cumulative risk assessment: The roles of directed acyclic graphs

Cumulative risk assessments (CRAs) address exposures to multiple chemical and nonchemical stressors and often focus on characterization of health risks in vulnerable populations. Evaluating complex exposure-response relationships in CRAs requires the use of formal and rigorous methods for causal inference. Directed acyclic graphs (DAGs) are graphical causal models used to organize and communicate knowledge about the underlying causal structure that generates observable data. Using existing graphical theories for causal inference with DAGs, risk analysts can identify confounders and effect measure modifiers to determine if the available data are both internally valid to obtain unbiased risk estimates and are generalizable to populations of interest. Conditional independencies implied by the structure of a DAG can be used to test assumptions used in a CRA against empirical data in a selected study and can contribute to the evidence evaluations related to specific causal pathways. This can facilitate quantitative use of these data, as well as help identify key research gaps, prioritize data collection activities, and evaluate risk management alternatives. DAGs also enable risk analysts to be explicit about sources of uncertainty and to determine whether a causal effect can be estimated from available data. Using a conceptual model and DAG for a hypothetical community located near a concentrated animal feeding operation (CAFO), we illustrate the advantages of using DAGs for evaluating causality in CRAs. DAGs also can be used in conjunction with weight of evidence (WOE) methodology to improve causal analysis for CRA, which could lead to more effective interventions to reduce population health risks.

Identification of the joint effect of a dynamic treatment intervention and a stochastic monitoring intervention under the no direct effect assumption

The management of chronic conditions is characterized by frequent re-assessment of therapy decisions in response to the patient's changing condition over the course of the illness. Evidence most suitable to inform care thus often concerns the contrast of adaptive treatment strategies that repeatedly personalize treatment decisions over time using the latest accumulated data available from the patient's previous clinic visits such as laboratory exams (e.g., hemoglobin A1c measurements in diabetes care). The frequency at which such information is monitored implicitly defines the causal estimand that is typically evaluated in an observational or randomized study of such adaptive treatment strategies. Analytic control of monitoring with standard estimation approaches for time-varying interventions can therefore not only improve study generalizibility but also inform the optimal timing of clinical surveillance. Valid inference with these estimators requires the upholding of a positivity assumption that can hinder their applicability. To potentially weaken this requirement for monitoring control, we introduce identifiability results that will facilitate the derivation of alternate estimators of effects defined by general joint treatment and monitoring interventions in the context of time-to-event outcomes. These results are developed based on the nonparametric structural equation modeling framework using a no direct effect assumption originally introduced in a prior paper that inspired this work. The relevance and scope of the results presented here are illustrated with examples in diabetes comparative effectiveness research.

Analysis of Body Mass Index and Mortality in Patients With Colorectal Cancer Using Causal Diagrams

IMPORTANCE

Physicians and investigators have sought to determine the relationship between body mass index (BMI [calculated as weight in kilograms divided by height in meters squared]) and colorectal cancer (CRC) outcomes, but methodologic limitations including sampling selection bias, reverse causality, and collider bias have prevented the ability to draw definitive conclusions.

OBJECTIVE

To evaluate the association of BMI at the time of, and following, colorectal cancer (CRC) diagnosis with mortality in a complete population using causal diagrams.

DESIGN, SETTING, AND PARTICIPANTS

This retrospective observational study with prospectively collected data included a cohort of 3408 men and women, ages 18 to 80 years, from the Kaiser Permanente Northern California population, who were diagnosed with stage I to III CRC between 2006 and 2011 and who also had surgery.

EXPOSURES

Body mass index at diagnosis and 15 months following diagnosis.

MAIN OUTCOMES AND MEASURES

Hazard ratios (HRs) for all-cause mortality and CRC-specific mortality compared with normal-weight patients, adjusted for sociodemographics, disease severity, treatment, and prediagnosis BMI.

RESULTS

This study investigated a cohort of 3408 men and women ages 18 to 80 years diagnosed with stage I to III CRC between 2006 and 2011 who also had surgery. At-diagnosis BMI was associated with all-cause mortality in a nonlinear fashion, with patients who were underweight (BMI <18.5; HR, 2.65; 95% CI, 1.63-4.31) and patients who were class II or III obese (BMI ≥35; HR, 1.33; 95% CI, 0.89-1.98) exhibiting elevated mortality risks, compared with patients who were low-normal weight (BMI 18.5 to <23). In contrast, patients who were high-normal weight (BMI 23 to <25; HR, 0.77; 95% CI, 0.56-1.06), low-overweight (BMI 25 to <28; HR, 0.75; 95% CI, 0.55-1.04), and high-overweight (BMI 28 to <30; HR, 0.52; 95% CI, 0.35-0.77) had lower mortality risks, and patients who were class I obese (BMI 30 to <35) showed no difference in risk. Spline analysis confirmed a U-shaped relationship in participants with lowest mortality at a BMI of 28. Associations with CRC-specific mortality were similar. Associations of postdiagnosis BMI and mortality were also similar, but patients who were class I obese had significantly lower all-cause and cancer-specific mortality risks.

CONCLUSIONS AND RELEVANCE

In this study, body mass index at the time of diagnosis and following diagnosis of CRC was associated with mortality risk. Though evidence shows that exercise in patients with cancer should be encouraged, findings suggest that recommendations for weight loss in the immediate postdiagnosis period among patients with CRC who are overweight may be unwarranted.

Unexpected links reflect the noise in networks

BACKGROUND

Gene covariation networks are commonly used to study biological processes. The inference of gene covariation networks from observational data can be challenging, especially considering the large number of players involved and the small number of biological replicates available for analysis.

RESULTS

We propose a new statistical method for estimating the number of erroneous edges in reconstructed networks that strongly enhances commonly used inference approaches. This method is based on a special relationship between sign of correlation (positive/negative) and directionality (up/down) of gene regulation, and allows for the identification and removal of approximately half of all erroneous edges. Using the mathematical model of Bayesian networks and positive correlation inequalities we establish a mathematical foundation for our method. Analyzing existing biological datasets, we find a strong correlation between the results of our method and false discovery rate (FDR). Furthermore, simulation analysis demonstrates that our method provides a more accurate estimate of network error than FDR.

CONCLUSIONS

Thus, our study provides a new robust approach for improving reconstruction of covariation networks.

REVIEWERS

This article was reviewed by Eugene Koonin, Sergei Maslov, Daniel Yasumasa Takahashi.

Motor vehicle crashes during pregnancy and cerebral palsy during infancy: a longitudinal cohort analysis

OBJECTIVES

To assess the incidence of cerebral palsy among children born to mothers who had their pregnancy complicated by a motor vehicle crash.

DESIGN

Retrospective longitudinal cohort analysis of children born from 1 April 2002 to 31 March 2012 in Ontario, Canada.

PARTICIPANTS

Cases defined as pregnancies complicated by a motor vehicle crash and controls as remaining pregnancies with no crash.

MAIN OUTCOME

Subsequent diagnosis of cerebral palsy by age 3 years.

RESULTS

A total of 1 325 660 newborns were analysed, of whom 7933 were involved in a motor vehicle crash during pregnancy. A total of 2328 were subsequently diagnosed with cerebral palsy, equal to an absolute risk of 1.8 per 1000 newborns. For the entire cohort, motor vehicle crashes correlated with a 29% increased risk of subsequent cerebral palsy that was not statistically significant (95% CI -16 to +110, p=0.274). The increased risk was only significant for those with preterm birth who showed an 89% increased risk of subsequent cerebral palsy associated with a motor vehicle crash (95% CI +7 to +266, p=0.037). No significant increase was apparent for those with a term delivery (95% CI -62 to +79, p=0.510). A propensity score-matched analysis of preterm births (n=4384) yielded a 138% increased relative risk of cerebral palsy associated with a motor vehicle crash (95% CI +27 to +349, p=0.007), equal to an absolute increase of about 10.9 additional cases per 1000 newborns (18.2 vs 7.3, p=0.010).

CONCLUSIONS

Motor vehicle crashes during pregnancy may be associated with an increased risk of cerebral palsy among the subgroup of cases with preterm birth. The increase highlights a specific role for traffic safety advice in prenatal care.

Using observed sequence to orient causal networks

In learning causal networks, typically cross-sectional data are used and the sequence among the network nodes is learned through conditional independence. Sequence is inherently a longitudinal concept. We propose to learn sequence of events in longitudinal data and use it to orient arc directions in a network learned from cross-sectional data. The network is learned from cross-sectional data using various established algorithms, with one modification. Arc directions that do not agree with the longitudinal sequence were prohibited. We established longitudinal sequence through two methods: Probabilistic Contrast, and Goodman and Kruskal error reduction methods. In simulated data, the error reduction method was used to learn the sequence in the data. The procedure reduced the number of arc direction errors and larger improvements were observed with increasing number of events in the network. In real data, different algorithms were used to learn the network from cross-sectional data, while prohibiting arc directions not supported by longitudinal information. The agreement among learned networks increased significantly. It is possible to combine sequence information learned from longitudinal data with algorithms organized for learning network models from cross-sectional data. Such models may have additional causal interpretation as they more explicitly take into account observed sequence of events.

Within-Subject Mediation Analysis in AB/BA Crossover Designs

Crossover trials are widely used to assess the effect of a reversible exposure on an outcome of interest. To gain further insight into the underlying mechanisms of this effect, researchers may be interested in exploring whether or not it runs through a specific intermediate variable: the mediator. Mediation analysis in crossover designs has received scant attention so far and is mostly confined to the traditional Baron and Kenny approach. We aim to tackle mediation analysis within the counterfactual framework and elucidate the assumptions under which the direct and indirect effects can be identified in AB/BA crossover studies. Notably, we show that both effects are identifiable in certain statistical models, even in the presence of unmeasured time-independent (or upper-level) confounding of the mediator-outcome relation. Employing the mediation formula, we derive expressions for the direct and indirect effects in within-subject designs for continuous outcomes that lend themselves to linear modelling, under a large variety of settings. We discuss an estimation approach based on regressing differences in outcomes on differences in mediators and show how to allow for period effects as well as different types of moderation. The performance of this approach is compared to other existing methods through simulations and is illustrated with data from a neurobehavioural study. Lastly, we demonstrate how a sensitivity analysis can be performed that is able to assess the robustness of both the direct and indirect effect against violation of the "no unmeasured lower-level mediator-outcome confounding" assumption.

The Visual Causality Analyst: An Interactive Interface for Causal Reasoning

Uncovering the causal relations that exist among variables in multivariate datasets is one of the ultimate goals in data analytics. Causation is related to correlation but correlation does not imply causation. While a number of casual discovery algorithms have been devised that eliminate spurious correlations from a network, there are no guarantees that all of the inferred causations are indeed true. Hence, bringing a domain expert into the casual reasoning loop can be of great benefit in identifying erroneous casual relationships suggested by the discovery algorithm. To address this need we present the Visual Causal Analyst-a novel visual causal reasoning framework that allows users to apply their expertise, verify and edit causal links, and collaborate with the causal discovery algorithm to identify a valid causal network. Its interface consists of both an interactive 2D graph view and a numerical presentation of salient statistical parameters, such as regression coefficients, p-values, and others. Both help users in gaining a good understanding of the landscape of causal structures particularly when the number of variables is large. Our framework is also novel in that it can handle both numerical and categorical variables within one unified model and return plausible results. We demonstrate its use via a set of case studies using multiple practical datasets.

Beyond Mendelian randomization: how to interpret evidence of shared genetic predictors

OBJECTIVE

Mendelian randomization is a popular technique for assessing and estimating the causal effects of risk factors. If genetic variants which are instrumental variables for a risk factor are shown to be additionally associated with a disease outcome, then the risk factor is a cause of the disease. However, in many cases, the instrumental variable assumptions are not plausible, or are in doubt. In this paper, we provide a theoretical classification of scenarios in which a causal conclusion is justified or not justified, and discuss the interpretation of causal effect estimates.

RESULTS

A list of guidelines based on the 'Bradford Hill criteria' for judging the plausibility of a causal finding from an applied Mendelian randomization study is provided. We also give a framework for performing and interpreting investigations performed in the style of Mendelian randomization, but where the choice of genetic variants is statistically, rather than biologically motivated. Such analyses should not be assigned the same evidential weight as a Mendelian randomization investigation.

CONCLUSION

We discuss the role of such investigations (in the style of Mendelian randomization), and what they add to our understanding of potential causal mechanisms. If the genetic variants are selected solely according to statistical criteria, and the biological roles of genetic variants are not investigated, this may be little more than what can be learned from a well-designed classical observational study.

A Review of Graphical Approaches to Common Statistical Analyses: The Omnipresence of Latent Variables in Statistics

We provide a comprehensive review of simple and advanced statistical analyses using an intuitive visual approach explicitly modeling Latent Variables (LV). This method can better illuminate what is assumed in each analytical method and what is actually estimated, by translating the causal relationships embedded in the graphical models in equation form. We recommend the graphical display rooted in the century old path analysis, that details all parameters of each statistical model, and suggest labeling that clarifies what is given vs. what is estimated. We link in the process classical and modern analyses under the encompassing broader umbrella of Generalized Latent Variable Modeling, and demonstrate that LVs are omnipresent in all statistical approaches, yet until directly 'seeing' them in visual graphical displays, they are unnecessarily overlooked. The advantages of directly modeling LVs are shown with examples of analyses from the ActiveS intervention designed to increase physical activity.

[Poor quality of sleep associated with low adherence to highly active antiretroviral therapy in Peruvian patients with HIV/AIDS]

This cross-sectional study analyzed the association between poor quality of sleep and adherence to highly active antiretroviral therapy (HAART) in 389 Peruvian patients with HIV/AIDS. Poor quality of sleep was measured with the Pittsburgh Sleep Quality Index (PSQI) and adherence with the CEAT-VIH (Peruvian adaptation). A Poisson generalized linear model with robust standard errors was used to estimate prevalence ratios and 95%CI. A crude model showed that mild, moderate, and severe poor quality of sleep were associated with inadequate treatment adherence. In the adjusted model for variables associated in the bivariate analysis or variables theoretically associated with adherence, only moderate/severe poor quality of sleep remained associated (PR = 1.34, 95%CI: 1.17-1.54; and PR = 1.34, 95%CI: 1.16-1.57, respectively). The study concluded that moderate/severe poor quality of sleep was independently associated with adherence to HAART. Assessing quality of sleep may be helpful in the comprehensive evaluation of HIV patients.

Sequence of Functional Loss and Recovery in Nursing Homes

PURPOSE OF THE STUDY

This study provides benchmarks for likelihood, number of days until, and sequence of functional decline and recovery.

DESIGN AND METHODS

We analyzed activities of daily living (ADLs) of 296,051 residents in Veteran Affairs nursing homes between January 1, 2000 and October 9, 2012. ADLs were extracted from standard minimum data set assessments. Because of significant overlap between short- and long-stay residents, we did not distinguish between these populations. Twenty-five combinations of ADL deficits described the experience of 84.3% of all residents. A network model described transitions among these 25 combinations. The network was used to calculate the shortest, longest, and maximum likelihood paths using backward induction methodology. Longitudinal data were used to derive a Bayesian network that preserved the sequence of occurrence of 9 ADL deficits.

RESULTS

The majority of residents (57%) followed 4 pathways in loss of function. The most likely sequence, in order of occurrence, was bathing, grooming, walking, dressing, toileting, bowel continence, urinary continence, transferring, and feeding. The other three paths occurred with reversals in the order of dressing/toileting and bowel/urinary continence. ADL impairments persisted without any change for an average of 164 days (SD = 62). Residents recovered partially or completely from a single impairment in 57% of cases over an average of 119 days (SD = 41). Recovery rates declined as residents developed more than 4 impairments.

IMPLICATIONS

Recovery of deficits among those studied followed a relatively predictable path, and although more than half recovered from a single functional deficit, recovery exceeded 100 days suggesting time to recover often occurs over many months.

Amino acids, lipid metabolites, and ferritin as potential mediators linking red meat consumption to type 2 diabetes

BACKGROUND

Habitual red meat consumption was consistently related to a higher risk of type 2 diabetes in observational studies. Potentially underlying mechanisms are unclear.

OBJECTIVE

This study aimed to identify blood metabolites that possibly relate red meat consumption to the occurrence of type 2 diabetes.

DESIGN

Analyses were conducted in the prospective European Prospective Investigation into Cancer and Nutrition-Potsdam cohort (n = 27,548), applying a nested case-cohort design (n = 2681, including 688 incident diabetes cases). Habitual diet was assessed with validated semiquantitative food-frequency questionnaires. Total red meat consumption was defined as energy-standardized summed intake of unprocessed and processed red meats. Concentrations of 14 amino acids, 17 acylcarnitines, 81 glycerophospholipids, 14 sphingomyelins, and ferritin were determined in serum samples from baseline. These biomarkers were considered potential mediators of the relation between total red meat consumption and diabetes risk in Cox models. The proportion of diabetes risk explainable by biomarker adjustment was estimated in a bootstrapping procedure with 1000 replicates.

RESULTS

After adjustment for age, sex, lifestyle, diet, and body mass index, total red meat consumption was directly related to diabetes risk [HR for 2 SD (11 g/MJ): 1.26; 95% CI: 1.01, 1.57]. Six biomarkers (ferritin, glycine, diacyl phosphatidylcholines 36:4 and 38:4, lysophosphatidylcholine 17:0, and hydroxy-sphingomyelin 14:1) were associated with red meat consumption and diabetes risk. The red meat-associated diabetes risk was significantly (P < 0.001) attenuated after simultaneous adjustment for these biomarkers [biomarker-adjusted HR for 2 SD (11 g/MJ): 1.09; 95% CI: 0.86, 1.38]. The proportion of diabetes risk explainable by respective biomarkers was 69% (IQR: 49%, 106%).

CONCLUSION

In our study, high ferritin, low glycine, and altered hepatic-derived lipid concentrations in the circulation were associated with total red meat consumption and, independent of red meat, with diabetes risk. The red meat-associated diabetes risk was largely attenuated after adjustment for selected biomarkers, which is consistent with the presumed mediation hypothesis.

Reverse enGENEering of Regulatory Networks from Big Data: A Roadmap for Biologists

Omics technologies enable unbiased investigation of biological systems through massively parallel sequence acquisition or molecular measurements, bringing the life sciences into the era of Big Data. A central challenge posed by such omics datasets is how to transform these data into biological knowledge, for example, how to use these data to answer questions such as: Which functional pathways are involved in cell differentiation? Which genes should we target to stop cancer? Network analysis is a powerful and general approach to solve this problem consisting of two fundamental stages, network reconstruction, and network interrogation. Here we provide an overview of network analysis including a step-by-step guide on how to perform and use this approach to investigate a biological question. In this guide, we also include the software packages that we and others employ for each of the steps of a network analysis workflow.

Assumption Trade-Offs When Choosing Identification Strategies for Pre-Post Treatment Effect Estimation: An Illustration of a Community-Based Intervention in Madagascar

Failure (or success) in finding a statistically significant effect of a large-scale intervention may be due to choices made in the evaluation. To highlight the potential limitations and pitfalls of some common identification strategies used for estimating causal effects of community-level interventions, we apply a roadmap for causal inference to a pre-post evaluation of a national nutrition program in Madagascar. Selection into the program was non-random and strongly associated with the pre-treatment (lagged) outcome. Using structural causal models (SCM), directed acyclic graphs (DAGs) and simulated data, we illustrate that an estimand with the outcome defined as the post-treatment outcome controls for confounding by the lagged outcome but not by possible unmeasured confounders. Two separate differencing estimands (of the pre- and post-treatment outcome) have the potential to adjust for a certain type of unmeasured confounding, but introduce bias if the additional identification assumptions they rely on are not met. In order to illustrate the practical impact of choice between three common identification strategies and their corresponding estimands, we used observational data from the community nutrition program in Madagascar to estimate each of these three estimands. Specifically, we estimated the average treatment effect of the program on the community mean nutritional status of children 5 years and under and found that the estimate based on the post-treatment estimand was about a quarter of the magnitude of either of the differencing estimands (0.066 SD vs. 0.26-0.27 SD increase in mean weight-for-age z-score). Choice of estimand clearly has important implications for the interpretation of the success of the program to improve nutritional status of young children. A careful appraisal of the assumptions underlying the causal model is imperative before committing to a statistical model and progressing to estimation. However, knowledge about the data-generating process must be sufficient in order to choose the identification strategy that gets us closest to the truth.

Mediation analysis to estimate direct and indirect milk losses due to clinical mastitis in dairy cattle

Milk losses associated with mastitis can be attributed to either effects of pathogens per se (i.e., direct losses) or effects of the immune response triggered by intramammary infection (indirect losses). The distinction is important in terms of mastitis prevention and treatment. Regardless, the number of pathogens is often unknown (particularly in field studies), making it difficult to estimate direct losses, whereas indirect losses can be approximated by measuring the association between increased somatic cell count (SCC) and milk production. An alternative is to perform a mediation analysis in which changes in milk yield are allocated into their direct and indirect components. We applied this method on data for clinical mastitis, milk and SCC test-day recordings, results of bacteriological cultures (Escherichia coli, Staphylococcus aureus, Streptococcus uberis, coagulase-negative staphylococci, Streptococcus dysgalactiae, and streptococci other than Strep. dysgalactiae and Strep. uberis), and cow characteristics. Following a diagnosis of clinical mastitis, the cow was treated and changes (increase or decrease) in milk production before and after a diagnosis were interpreted counterfactually. On a daily basis, indirect changes, mediated by SCC increase, were significantly different from zero for all bacterial species, with a milk yield decrease (ranging among species from 4 to 33g and mediated by an increase of 1000 SCC/mL/day) before and a daily milk increase (ranging among species from 2 to 12g and mediated by a decrease of 1000 SCC/mL/day) after detection. Direct changes, not mediated by SCC, were only different from zero for coagulase-negative staphylococci before diagnosis (72g per day). We concluded that mixed structural equation models were useful to estimate direct and indirect effects of the presence of clinical mastitis on milk yield.

Causal Inference in the Age of Decision Medicine

Causal analyses and causal inference is a growing area of biostatics. In parallel, there is increasing focus on using genomic information to guide medical practice, i.e. personalized medicine or decision medicine. This perspective discusses causal inference in the context of personalized or decision medicine, including the assumptions and the concept that the task is different depending on whether the primary goal is the average response of treatment in the population or the ability to characterize the response for an individual or a subgroup. This perspective provides a tutorial of modern causal inference and then provides suggestions how application of specific kinds of causal inference would promote advances in translational sciences. The concept of the subpopulation causal effect is one path toward improved decision medicine. A dataset containing cardiovascular disease risk factor levels and genomic information is analyzed and different causal effects are estimated.

High-dimensional propensity score algorithm in comparative effectiveness research with time-varying interventions

The high-dimensional propensity score (hdPS) algorithm was proposed for automation of confounding adjustment in problems involving large healthcare databases. It has been evaluated in comparative effectiveness research (CER) with point treatments to handle baseline confounding through matching or covariance adjustment on the hdPS. In observational studies with time-varying interventions, such hdPS approaches are often inadequate to handle time-dependent confounding and selection bias. Inverse probability weighting (IPW) estimation to fit marginal structural models can adequately handle these biases under the fundamental assumption of no unmeasured confounders. Upholding of this assumption relies on the selection of an adequate set of covariates for bias adjustment. We describe the application and performance of the hdPS algorithm to improve covariate selection in CER with time-varying interventions based on IPW estimation and explore stabilization of the resulting estimates using Super Learning. The evaluation is based on both the analysis of electronic health records data in a real-world CER study of adults with type 2 diabetes and a simulation study. This report (i) establishes the feasibility of IPW estimation with the hdPS algorithm based on large electronic health records databases, (ii) demonstrates little impact on inferences when supplementing the set of expert-selected covariates using the hdPS algorithm in a setting with extensive background knowledge, (iii) supports the application of the hdPS algorithm in discovery settings with little background knowledge or limited data availability, and (iv) motivates the application of Super Learning to stabilize effect estimates based on the hdPS algorithm.

Statistical approaches for enhancing causal interpretation of the M to Y relation in mediation analysis

Statistical mediation methods provide valuable information about underlying mediating psychological processes, but the ability to infer that the mediator variable causes the outcome variable is more complex than widely known. Researchers have recently emphasized how violating assumptions about confounder bias severely limits causal inference of the mediator to dependent variable relation. Our article describes and addresses these limitations by drawing on new statistical developments in causal mediation analysis. We first review the assumptions underlying causal inference and discuss three ways to examine the effects of confounder bias when assumptions are violated. We then describe four approaches to address the influence of confounding variables and enhance causal inference, including comprehensive structural equation models, instrumental variable methods, principal stratification, and inverse probability weighting. Our goal is to further the adoption of statistical methods to enhance causal inference in mediation studies.

Assessment of network inference methods: how to cope with an underdetermined problem

The inference of biological networks is an active research area in the field of systems biology. The number of network inference algorithms has grown tremendously in the last decade, underlining the importance of a fair assessment and comparison among these methods. Current assessments of the performance of an inference method typically involve the application of the algorithm to benchmark datasets and the comparison of the network predictions against the gold standard or reference networks. While the network inference problem is often deemed underdetermined, implying that the inference problem does not have a (unique) solution, the consequences of such an attribute have not been rigorously taken into consideration. Here, we propose a new procedure for assessing the performance of gene regulatory network (GRN) inference methods. The procedure takes into account the underdetermined nature of the inference problem, in which gene regulatory interactions that are inferable or non-inferable are determined based on causal inference. The assessment relies on a new definition of the confusion matrix, which excludes errors associated with non-inferable gene regulations. For demonstration purposes, the proposed assessment procedure is applied to the DREAM 4 In Silico Network Challenge. The results show a marked change in the ranking of participating methods when taking network inferability into account.

[Causal analysis approaches in epidemiology]

Epidemiological research is mostly based on observational studies. Whether such studies can provide evidence of causation remains discussed. Several causal analysis methods have been developed in epidemiology. This paper aims at presenting an overview of these methods: graphical models, path analysis and its extensions, and models based on the counterfactual approach, with a special emphasis on marginal structural models. Graphical approaches have been developed to allow synthetic representations of supposed causal relationships in a given problem. They serve as qualitative support in the study of causal relationships. The sufficient-component cause model has been developed to deal with the issue of multicausality raised by the emergence of chronic multifactorial diseases. Directed acyclic graphs are mostly used as a visual tool to identify possible confounding sources in a study. Structural equations models, the main extension of path analysis, combine a system of equations and a path diagram, representing a set of possible causal relationships. They allow quantifying direct and indirect effects in a general model in which several relationships can be tested simultaneously. Dynamic path analysis further takes into account the role of time. The counterfactual approach defines causality by comparing the observed event and the counterfactual event (the event that would have been observed if, contrary to the fact, the subject had received a different exposure than the one he actually received). This theoretical approach has shown limits of traditional methods to address some causality questions. In particular, in longitudinal studies, when there is time-varying confounding, classical methods (regressions) may be biased. Marginal structural models have been developed to address this issue. In conclusion, "causal models", though they were developed partly independently, are based on equivalent logical foundations. A crucial step in the application of these models is the formulation of causal hypotheses, which will be a basis for all methodological choices. Beyond this step, statistical analysis tools recently developed offer new possibilities to delineate complex relationships, in particular in life course epidemiology.

Tenofovir or zidovudine in second-line antiretroviral therapy after stavudine failure in southern Africa

BACKGROUND

There is debate over using tenofovir or zidovudine alongside lamivudine in second-line antiretroviral therapy (ART) following stavudine failure. We analysed outcomes in cohorts from South Africa, Zambia and Zimbabwe

METHODS

Patients aged ≥16 years who switched from a first-line regimen including stavudine to a ritonavir-boosted lopinavir-based second-line regimen with lamivudine or emtricitabine and zidovudine or tenofovir in seven ART programmes in southern Africa were included. We estimated the causal effect of receiving tenofovir or zidovudine on mortality and virological failure using Cox proportional hazards marginal structural models. Its parameters were estimated using inverse probability of treatment weights. Baseline characteristics were age, sex, calendar year and country. CD4(+) T-cell count, creatinine and haemoglobin levels were included as time-dependent confounders.

RESULTS

A total of 1,256 patients on second-line ART, including 958 on tenofovir, were analysed. Patients on tenofovir were more likely to have switched to second-line ART in recent years, spent more time on first-line ART (33 versus 24 months) and had lower CD4(+) T-cell counts (172 versus 341 cells/μl) at initiation of second-line ART. The adjusted hazard ratio comparing tenofovir with zidovudine was 1.00 (95% CI 0.59, 1.68) for virological failure and 1.40 (0.57, 3.41) for death.

CONCLUSIONS

We did not find any difference in treatment outcomes between patients on tenofovir or zidovudine; however, the precision of our estimates was limited. There is an urgent need for randomized trials to inform second-line ART strategies in resource-limited settings.

Racial and ethnic variations in the effects of family history of colorectal cancer on screening compliance

BACKGROUND & AIMS

Individuals with a family history of colorectal cancer (CRC) have a higher risk of developing CRC than the general population, and studies have shown that they are more likely to undergo CRC screening. We assessed the overall and race- and ethnicity-specific effects of a family history of CRC on screening.

METHODS

We analyzed data from the 2009 California Health Interview Survey to estimate overall and race- and ethnicity-specific odds ratios (ORs) for the association between family history of CRC and CRC screening.

RESULTS

The unweighted and weighted sample sizes were 23,837 and 8,851,003, respectively. Individuals with a family history of CRC were more likely to participate in any form of screening (OR, 2.3; 95% confidence limit [CL], 1.7, 3.1) and in colonoscopy screening (OR, 2.7; 95% CL, 2.2, 3.4) than those without a family history, but this association varied among racial and ethnic groups. The magnitude of the association between family history and colonoscopy screening was highest among Asians (OR, 6.1; 95% CL, 3.1, 11.9), lowest among Hispanics (OR, 1.4; 95% CL, 0.67, 2.8), and comparable between non-Hispanic whites (OR, 3.1; 95% CL, 2.6, 3.8) and non-Hispanic blacks (OR 2.6; 95% CL, 1.2, 5.7) (P for interaction < .001).

CONCLUSIONS

The effects of family history of CRC on participation in screening vary among racial and ethnic groups, and have the lowest effects on Hispanics, compared with other groups. Consequently, interventions to promote CRC screening among Hispanics with a family history should be considered.

Making valid causal inferences from observational data

The ability to make strong causal inferences, based on data derived from outside of the laboratory, is largely restricted to data arising from well-designed randomized control trials. Nonetheless, a number of methods have been developed to improve our ability to make valid causal inferences from data arising from observational studies. In this paper, I review concepts of causation as a background to counterfactual causal ideas; the latter ideas are central to much of current causal theory. Confounding greatly constrains causal inferences in all observational studies. Confounding is a biased measure of effect that results when one or more variables, that are both antecedent to the exposure and associated with the outcome, are differentially distributed between the exposed and non-exposed groups. Historically, the most common approach to control confounding has been multivariable modeling; however, the limitations of this approach are discussed. My suggestions for improving causal inferences include asking better questions (relates to counterfactual ideas and "thought" trials); improving study design through the use of forward projection; and using propensity scores to identify potential confounders and enhance exchangeability, prior to seeing the outcome data. If time-dependent confounders are present (as they are in many longitudinal studies), more-advanced methods such as marginal structural models need to be implemented. Tutorials and examples are cited where possible.

Integrative genomic analysis of the human immune response to influenza vaccination

Identification of the host genetic factors that contribute to variation in vaccine responsiveness may uncover important mechanisms affecting vaccine efficacy. We carried out an integrative, longitudinal study combining genetic, transcriptional, and immunologic data in humans given seasonal influenza vaccine. We identified 20 genes exhibiting a transcriptional response to vaccination, significant genotype effects on gene expression, and correlation between the transcriptional and antibody responses. The results show that variation at the level of genes involved in membrane trafficking and antigen processing significantly influences the human response to influenza vaccination. More broadly, we demonstrate that an integrative study design is an efficient alternative to existing methods for the identification of genes involved in complex traits. DOI:http://dx.doi.org/10.7554/eLife.00299.001.

Reverse engineering cellular networks with information theoretic methods

Building mathematical models of cellular networks lies at the core of systems biology. It involves, among other tasks, the reconstruction of the structure of interactions between molecular components, which is known as network inference or reverse engineering. Information theory can help in the goal of extracting as much information as possible from the available data. A large number of methods founded on these concepts have been proposed in the literature, not only in biology journals, but in a wide range of areas. Their critical comparison is difficult due to the different focuses and the adoption of different terminologies. Here we attempt to review some of the existing information theoretic methodologies for network inference, and clarify their differences. While some of these methods have achieved notable success, many challenges remain, among which we can mention dealing with incomplete measurements, noisy data, counterintuitive behaviour emerging from nonlinear relations or feedback loops, and computational burden of dealing with large data sets.

HIV infection disrupts the sympatric host-pathogen relationship in human tuberculosis

The phylogeographic population structure of Mycobacterium tuberculosis suggests local adaptation to sympatric human populations. We hypothesized that HIV infection, which induces immunodeficiency, will alter the sympatric relationship between M. tuberculosis and its human host. To test this hypothesis, we performed a nine-year nation-wide molecular-epidemiological study of HIV–infected and HIV–negative patients with tuberculosis (TB) between 2000 and 2008 in Switzerland. We analyzed 518 TB patients of whom 112 (21.6%) were HIV–infected and 233 (45.0%) were born in Europe. We found that among European-born TB patients, recent transmission was more likely to occur in sympatric compared to allopatric host–pathogen combinations (adjusted odds ratio [OR] 7.5, 95% confidence interval [95% CI] 1.21–infinity, p = 0.03). HIV infection was significantly associated with TB caused by an allopatric (as opposed to sympatric) M. tuberculosis lineage (OR 7.0, 95% CI 2.5–19.1, p<0.0001). This association remained when adjusting for frequent travelling, contact with foreigners, age, sex, and country of birth (adjusted OR 5.6, 95% CI 1.5–20.8, p = 0.01). Moreover, it became stronger with greater immunosuppression as defined by CD4 T-cell depletion and was not the result of increased social mixing in HIV–infected patients. Our observation was replicated in a second independent panel of 440 M. tuberculosis strains collected during a population-based study in the Canton of Bern between 1991 and 2011. In summary, these findings support a model for TB in which the stable relationship between the human host and its locally adapted M. tuberculosis is disrupted by HIV infection.

Human tuberculosis (TB) caused by Mycobacterium tuberculosis kills 1.5 million people each year. M. tuberculosis has been affecting humans for millennia, suggesting that different strain lineages may be adapted to specific human populations. The combination of a particular strain lineage and its corresponding patient population can be classified as sympatric (e.g. Euro-American lineage in Europeans) or allopatric (e.g. East-Asian lineage in Europeans). We hypothesized that infection with the human immunodeficiency virus (HIV), which impairs the human immune system, will interfere with this host–pathogen relationship. We performed a nation-wide molecular-epidemiological study of HIV–infected and HIV–negative TB patients between 2000 and 2008 in Switzerland. We found that HIV infection was associated with the less adapted allopatric lineages among patients born in Europe, and this was not explained by social or other patient factors such as increased social mixing in HIV–infected individuals. Strikingly, the association between HIV infection and less adapted M. tuberculosis lineages was stronger in patients with more pronounced immunodeficiency. Our observation was replicated in a second independent panel of M. tuberculosis strains collected during a population-based study in the Canton of Bern. In summary, our study provides evidence that the sympatric host–pathogen relationship in TB is disrupted by HIV infection.

Networking development by Boolean logic

Eric Davidson at Caltech has spent several decades investigating the molecular basis of animal development using the sea urchin embryo as an experimental system^1,2 although his scholarship extends to all of embryology as embodied in several editions of his landmark book.³ In recent years his laboratory has become a leading force in constructing gene regulatory networks (GRNs) operating in sea urchin development.⁴ This axis of his work has its roots in this laboratory’s cDNA cloning of an actin mRNA from the sea urchin embryo (for the timeline, see ref. 1)—one of the first eukaryotic mRNAs to be cloned as it turned out. From that point of departure, the Davidson lab has drilled down into other genes and gene families and the factors that regulate their coordinated regulation, leading them into the GRN era (a field they helped to define) and the development of the computational tools needed to consolidate and advance the GRN field.

Estimating efficacy in the presence of non-ignorable non-trial interventions in the Helsinki Psychotherapy Study

In a randomised clinical trial with a longitudinal outcome, analyses of the efficacy of the study treatments may be complicated by both non-trial interventions, which have not been administered by the researcher, and sparsely measured outcome values. The delay between the change in outcome and the starting of the non-trial intervention may be much shorter than the time intervals between the actual measurements. We propose a model that accounts for the possible dynamic interdependence between the longitudinal outcome and time-to-event data. The model is based on discretising time into short intervals. This results in a missing data problem, which we tackle using Bayesian inference and data augmentation. The method is based on the assumption that decisions to initiate non-trial interventions are not confounded by unobservable factors. The Helsinki Psychotherapy Study data are used as an illustration. Different psychotherapies were compared, and possible episodes of psychotropic medication were viewed as non-trial interventions. Simulation studies suggest that our method provides reasonable estimates of the effects of both the study treatment and the non-trial intervention also showing some robustness against possible latent background factors. An application of marginal structural modelling, however, appeared to underestimate the differences between the treatments.

The causal mediation formula--a guide to the assessment of pathways and mechanisms

Recent advances in causal inference have given rise to a general and easy-to-use formula for assessing the extent to which the effect of one variable on another is mediated by a third. This Mediation Formula is applicable to nonlinear models with both discrete and continuous variables, and permits the evaluation of path-specific effects with minimal assumptions regarding the data-generating process. We demonstrate the use of the Mediation Formula in simple examples and illustrate why parametric methods of analysis yield distorted results, even when parameters are known precisely. We stress the importance of distinguishing between the necessary and sufficient interpretations of "mediated-effect" and show how to estimate the two components in nonlinear systems with continuous and categorical variables.