Claim causality with clarity


 Causal inference has attracted enormous attention in clinical neuroscience over the past years. The correct communication and interpretation of causal claims is crucial for the researchers. In this review, we briefly introduced the statistical basis for causal inference in terms of the potential outcome framework and distinguished the different interpretations in experimental and observational studies. Based on the recent efforts in formulating neuroscience research into a causal framework, we provide four recommendations to facilitate better communication of open and reproducible causal inference research.


The Rising Trend of Causal Related Research
Causal inference is one of the most popular topics in statistics, and its applications in both experimental and observational researc h hav e exponentiall y gr own.Figur e 1 shows the number of publications related to causal research as an indicator of its popularity in different research disciplines .T he data are obtained from a PubMed search using expressions such as "(causal OR causality) AND (discipline)" based on the texts of publication without further manual content validation, the full details of which are available on the GitHub repo: https:// github.com/Vincent-wq/causal _ liter ature _ tr end .As illustr ated in Fig. 1 , causal r elated r esearc h has the richest literature and the largest number of published papers.Clinical related causal research has the second largest number of papers published.Both neurology and psychiatry show similar escalating tr ends.Inter estingl y, the rising slope of neur ology exceeded that of psychiatry in 2014, which may indicate that the application of causal related analysis has become more widespread in neurology than in psychiatry.Ho w ever, it is har d to kno w the r easons for suc h c hanges without a detailed in-depth liter atur e r e vie w.Neur oima ging has r ecentl y enjo y ed a burst of applications in clinical pr actice, especiall y in neurology and psychiatry, yet it has the smallest number of published papers .T his ma y be related to the complexity and high-dimensional nature of neur oima ging data and modeling.In conclusion, the number of causal related publications is increasing.
The misuse and misinter pr etation of statistical methods have contributed to the r epr oducibility crisis ( Adler et al. , n.d. ;Baker, 2016 ;Open Science Collaboration, 2015 ;Wang et al., 2023 ).By analogy, the boosting of causal related research calls for better communication and interpretation of causal analysis .T he aim of this mini r e vie w is to r aise the awar eness of the clarity when r eporting and inter pr eting causal r elated r esearc h so that the misuse and misinter pr etation can be r educed.

Basics of Causal Inference
In this r e vie w, we limit our discussion of causal inference to the e v aluation of causal effects rather than the identification of causal mechanisms.Causal inference can be conducted by (i) formulating the research question in a causal framework; (ii) specifying assumptions based on which causal effects can be identified; and (iii) assessing the sensitivity to the violation of causal assumptions .T her e ar e two main causal infer ence fr ame works: the potential outcome (PO) fr ame work (Hernán & Robins, 2020 ) and the causal dia gr am fr ame w ork (Judea P earl, 2009 ).These tw o fr ame works ar e mathematicall y connected with differ ent established goals (Richardson & Robins, 2013 ).We will focus on the PO fr ame work in this r e vie w as most of the liter atur e r e vie wed falls under the umbrella of the PO framework.
First, we briefly r e vie w the k e y conce pts in the PO fr ame work as illustrated in Fig. 2 : (i) unit, the person or subject on whom the treatment will be operated; (ii) target population, a well-defined population of units whose causal effects are going to be estimated; (iii) sample, a random sample of N from the target population, the data collected from the sample being used for further analysis; (iv) tr eatment (interv ention/exposur e/manipulation), the effects of which the investigator would like to assess compared to no suc h tr eatment; and (v) outcome, the final observ ation after treatment (can be no treatment).The PO framework aims to answer the question "what would potentially happen to the same units or participants had they exposed to a different (counterfactual) condition (treatment)?"By definition, we can ne v er observ e the individual treatment effect (ITE) since we can onl y observ e the outcome from one treatment at a time (illustrated in Fig. 2 A).Most of the time, the av er a ge tr eatment effect (ATE) or av er a ge tr eatment effect in the treated (ATT) is the main causal effect we would like to estimate (as illustrated in Fig. 2 B).Stated formally, causal inference is to estimate the causal effect from the outcome of a tr eatment, interv ention, exposur e, or manipulation with observed confounders and/or covariates and unobserved confounders and/or co variates .Since the individual cannot sim ultaneousl y r eceiv e and not r eceiv e the tr eatment, we ar e unable to observ e the differ ence between the POs (icon with boundary) of r eceiving tr eatment and not r eceiving tr eatment for the same individual, i.e.ITE is unobservable; (B) A TE and A TT.A TE is the av er a ge tr eatment effect for the whole gr oup while ATT is the av er a ge tr eatment effect for the treated group, ATE = ATT for the ideal RCT (being in the control or treatment group is random and unrelated to the outcome), but they are not necessarily the same in the observational studies.We use the observed outcome to estimate ATE and ATT.
Traditional statistical inference draws conclusions based on associations, and the main differences between these traditional data analyses and causal inference lies in the causal assumptions , i.e .the identification conditions for causal effects.One basic assumption for causal inference is the stable unit treatment value assumption (SUTVA): "The potential outcomes for any unit do not vary with the treatments assigned to other units, and, for eac h unit, ther e ar e no differ ent forms or v ersions of eac h tr eatment le v el, whic h lead to differ ent potential outcomes" (Imbens & Rubin, 2015 ).SUTVA describes the basic properties of treatment unite and connects the intervention we observed with the causal interv ention of inter est, and it is a strong assumption about no interference and no multiple versions of a treatment, which contributes to a well-defined intervention.

Causality in Experimental and Observ a tional Researc h
The golden standard of estimating causal effects is the ideal randomized controlled trial (RCT) (Hernán & Robins, 2020 ), where RCT is a true random sample from the target population.In addition to the SUTVA, an ideal RCT with well-established random tr eatment assignment mec hanisms allows the infer ence of causal effects since it satisfies the following assumptions: (i) "Unconfoundedness" (or "Ignor ability," "Exc hangeability"), (ii) "Positivity" (or "ov erla p"), and (iii) "Consistency" (part of the SUTVA assumption) (Cole & Fr angakis, 2009 ).Specificall y, Unconfoundedness assumes the independence of treatment assignment and the outcomes, which implies that within the subpopulations defined by the values of observed co variates , the treatment assignment is r andom, i.e. tr eated and untr eated participants, censor ed and uncensored participants have equal distributions of POs.Consistency assumes that an individual's PO under the observed exposure history is precisely the observed outcome .P ositivity assumes that all the le v els of exposure for every combination of values of exposure and confounders occur among individuals in the population.Ho w e v er, these assumptions cannot always be met, and the ideal RCT can be compromised due to ethical, economical, protocol violations, and other limitations that endanger the estimation of causal effect.Ther efor e, clarifying causal assumptions and constructing a meaningful causal estimand to draw interpretable causal conclusions is highly challenging, especially for observational studies (Liu et al., 2021 ).
In observational studies, we can neither control nor be clear about the intervention assignment mechanisms, and it is common to violate some or all of the assumptions from before, which makes justification of causal assumptions essential.For example, assuming ther e ar e no unobserv ed confounders, failur e in randomized assignment of the treatment may cause imbalanced covariates between the treatment and control groups.As a result, statistical methods must be introduced to balance these two groups, and the typical procedures include regression, matching, pr opensity scor e-based methods (suc h as inv erse pr obability weighting) or their combination such as double robust (DR) estimators (Li et al., 2018 ).When r esearc hers ar e not confident that all confounders ar e full y observ ed and corr ectl y measur ed, instrumental variable techniques are introduced to circumvent these limitations.(Marinescu et al., 2018 ;Liu et al., 2021 ).
The k e y logic of causal infer ence in observ ational studies is to mimic a target experiment (trial) that produces similar results to an RCT in a hypothesized population.For example, quasiexperimental a ppr oac hes hav e been widel y used in economics and psychology.Liu, Marinescu and others have reviewed this family of methods including regression discontinuity design, differ ence in differ ence, and instrumental v ariables (IV) (Liu et al., 2021 ;Marinescu et al., 2018 ).An IV is a variable that is only associated with the exposure to the intervention but not with other factors associated with the outcome of interest.Using IV does not r equir e the assumption of unconfoundedness, but three other conditions should be met: namely, the relevance condition, the exclusion restriction, and the marginal exchangeability (Hernán & Robins, 2020 ).Regression discontinuity design is a special case of IV that uses the discontinuity feature of the running variable as IV.Another commonly used IV in life sciences is genetics, which is assumed to be r andoml y inherited from the parents, and the corr esponding a ppr oac h is called Mendelian r andomization (Bur gess & Thompson, 2021 ).All these models and a ppr oac hes r el y heavil y on strong assumptions and complex computations, which means the results can be very different on any meaningful violation of assumptions or any changes in the algorithms or computing envir onment.Sensitivity anal ysis is also necessary to assess such biases.

Causality in Clinical Neuroscience
The current causal inference framework from the statistics world has not been pr operl y tr anslated to face the c hallenges in clinical neur oscience r esearc h due to its intrinsic complexity including but not limited to the lack of RCT data sources due to ethical concerns or other factors such as cost, the justifications of causal assumptions for experiments other than an ideal RCT or observational studies, and the definition of an interv ention, whic h is mor e complicated than just taking or not taking a specific medicine, and it can be one of many types of br ain stim ulation, modulation, or e v en tar geted sur gery.In addition, Bar ac k et al. hav e called for more clarity about causality in neuroscience research since the w or d "causality" can r efer to as differ ent meanings in neur oscience (Bar ac k et al., 2022 ), some neur oscientists belie v ed that causes are the events that produce other events while others may think that causes are the factors that e v ents depend on.Such ambiguous definitions of "causes" impedes the communication and inter pr etation of causal analyses from different researchers.Taking clinical r esearc h as an example , Siddiqi et al. ha v e r e vie wed most of the available interventions in clinical neuroscience practice r egarding ma pping human br ain functions and hav e br ought about six criteria for a ppr aising causality ada pted fr om Br adford Hill criteria: counterfactual, specificity, experimental manipulation, dose-r esponse r elationship, coher ence, and r e v ersibility (Siddiqi et al., 2022 ).They also suggested that causal claims based on pur el y corr elation r esults should be a voided.T her e ar e v arious types of intervention used in clinical neuroscience, such as drugs, non-inv asiv e neur oima ging with stim uli, neur ofeedbac k, lesion, br ain stim ulation, etc .It is not easy to model all of these interventions with a unified causal fr ame work so that they are compar able, a binary v ariable (whether to use or not to use a specific type of intervention) is insufficient to ca ptur e the full information of these interventions (SUTVA assumption is very likely to be violated); a m ultiv ariate mec hanistic a ppr oac h might be helpful, such as dynamic causal modeling (Friston et al., 2003 ), which tries to ca ptur e the complex mechanism of how the specific intervention (experimental task design) changes the outcome with a dynamical biophysical forw ar d model.Reid et al. attempted to formulate functional connectivity estimates using a causal framework but ended up by using v a gue definitions and mixing differ ent le v els of concepts (Reid et al., 2019 ).For example, there is no clear definition of the "causal effect of interest," but a rather general term "target theoretical properties" was used.The definition of "confounding pr operties" mainl y includes artifacts during the imperfect measurement of functional connectivity, but there ar e so man y mor e confounding sources outside the measurement pr ocedur e, suc h as age, sex, and so on .Another emerging trend is the mining of large-scale observational imaging datasets and ima ging-deriv ed phenotypes with the Mendelian randomization a ppr oac h, i.e. using genome as an instrumental variable to e v aluate the potential causal relationships between ima ging-deriv ed phenotypes and neurological or psychiatric disorders (Guo et al., 2022 ;Taschler et al., 2022 ).In summary, solid statistical-based causal inference is still lacking in clinical neur oscience r esearc h, and we are still at the stage of formulating the questions properly with causal language, where the process can be benefited by interdisciplinary collaborations.

Claim Causality with Clarity
Clarity and sensitivity analyses are crucial for causal inference, especially in observational studies.To promote open and reproducible r esearc h (Jin et al., 2022 ) and to avoid further mis-claiming or misinter pr etation of causal anal ysis, we encour a ge the r esearc hers to r eport: (i) full details of the causal formulation of the r esearc h question and the reasoning behind the causal model (can be r epr esented by a dir ected acyclic gr a ph or DAG), including: (a) study type (whether this is a RCT or observational study), (b) welldefined causal effect of interest (e.g.A TE or A TT), including clear descriptions of tr eatment/interv ention/exposur e/manipulation, confounder selection and its r ationale, known unobserv ed confounders with corresponding assumptions about them, and (c) the observed outcome (continuous or binary, etc.); (ii) all the necessary assumptions condition on which the causation can be interpr etated, especiall y for observ ational studies , e .g. whether Unconfoundedness, positivity, and consistency ar e r easonable assumptions for this study; (iii) full details of the causal estimand, including (a) the statistical a ppr oac h and (b) the effect size of the causal effect, such as the estimation of A TE or A TT or causal odds ratios; (iv) the results of sensitivity analysis, for both the meaningful violations of the assumptions in (ii) and different model estimation algorithms.With all necessary information shar ed, r eaders and r e vie wers should be able to replicate and generalize such causal anal yses and hav e a better understanding of the strength of the causal claims.

Figure 1 :
Figure 1: Causal related research literature from PubMed search.

Figure 2 :
Figure 2: Visual illustrations of the basic concepts in causal inference.(A) ITE.Since the individual cannot sim ultaneousl y r eceiv e and not r eceiv e the tr eatment, we ar e unable to observ e the differ ence between the POs (icon with boundary) of r eceiving tr eatment and not r eceiving tr eatment for the same individual, i.e.ITE is unobservable; (B) A TE and A TT.A TE is the av er a ge tr eatment effect for the whole gr oup while ATT is the av er a ge tr eatment effect for the treated group, ATE = ATT for the ideal RCT (being in the control or treatment group is random and unrelated to the outcome), but they are not necessarily the same in the observational studies.We use the observed outcome to estimate ATE and ATT.