Causal Diagrams for Epidemiologic Research Sander Greenland, Judea Pearl,? and James M. Robins? Causal diagrams have a long history of informal use and, more recently, have undergone formal development for applications in expert systems and robotics. We provide an introduction 0 these developments and their use in epidemiologic research. Causal diagrams can provide a starting point for idenciing variables that must be measured and controlled to obtain ‘unconfounded effect estimates. They alo provide a method for Keywords: bias, causation, confounding, epidemiologic methods, itil evaluation of eadtionalepidemologieerteria for con- founding. In particular, they reveal certain heretofore unno- teed shortcomings of those criteria when used in considering smukiple potential confounders. We show how to modify the teadionalerieria fo comect those shorcomings. (Epidemiology 1999:1037-48) graphical methods, observational studies. ‘Summarisation of causal links via graphs or diagrams has long been used as an informal aid co causal analysis. Causal. graphs in the form of path diagrams are an integral component of path analysis! and structural equations modeling. In more recent times, che theory of directed acyclic graphs (DAGs) has been extended to application in expert-systems research. In these appli- cations, there is a pressing need for valid formal rules that allow an automated system or robot co deduce correctly the presence or absence of causal links given ‘correct background information and new data. The out- ‘growth of this research has been the development of a formal theory for evaluating causal effects using the language of causal diagrams."* Unlike path analysis and structural-equations modeling, this theory does not re- quite parametric assumptions such as linearity. ‘The theory of causal graphs is equivalent to the G- computation theory of Robins." Ir has a benefit, how- ‘ever, of providing a compact graphical as well as alge- braic formulation of assumptions and results, which may be easier for the general reader to comprehend. In ad- dition, i provides a novel perspective on «raditional epidemiologic criteria for confounder identification. ‘This perspective reveals how traditional criteria can be inadequate when multiple confounders must be consid- ered simultaneously. We describe the modifications re- aquired of tradicional criteria that enable their valid ex- tension to situations involving multiple confounders. From the ‘Deparment of Emil, UCLA School of Public Heath, Lt Angler CA,Coguive Spams Labor. Conpae Science Depart LOCE School of Enincering, Las Angles, CAs ad Departs of ode ‘Sly sd Besta Hanaeé School Pac Heh Buon MA. ‘Alles comspondence wo: Deparment of Epdemsogy, UCLA Schost of Pe Hest Ln Angel, CA 90095172 Sum Jl 16 197i version aceped March 25, 198 ‘S105 by Epa Renae le ‘We here provide a brief introduction to the theory of causal diagrams based on DAGs.** We pay special at- tention t0 its relation to nongraphical epidemiologic treatments of confounding. We show how diagrams can serve as a visual yer logically rigorous aid for sum- marizing assumptions about a problem and for identify ing variables that must be measured and controlled to ‘obtain unconfounded effect estimates given those at- sumptions. Thus, use of such graphs can aid in planning of data collection and analysis, in communication of results, and in avoiding subtle picfalls of confounder selection, Except where noted otherwise, the present paper will deal only with relations among variables in a given source population; that is, we will deal only with suc tural (systematic) relations among the underlying vari- ables of interest, 50 that issues of measurement error and random variation will not arise. We will also not present proofs of results, bur we will give references in which proofs can be found. A Rationale for Graphs ‘Any deduction about a causal relation must start from some set of assumptions, which we call the analysis model." For example, such a deduction may assume that uncontrlled confounding is negligible; chis assumption usually corresponds to a set of assumptions that various uncontrolled factors have negligible associations wich the study factor or study disease, given what has been, controlled. The subsets of these assumptions peztaining to causation, measurement, and selection correspond t0 the causal, measurement, and selection models, whereas the subset of assumptions pertaining to the probability distribution of the observations corresponds fo the star tistical model. These subsets of assumpcions may over lap. For example, the assumption that vaginal bleeding increases the chance of diagnosis of endometrial cancer and hence selection fora case-control study is common 3738 GREENLAND ET AL to all four subsets; ic isa probabilistic assumption about causation of diagnosis, measurement of disease, and se- lection for study."* Sometimes the assumptions of an analysis model are so obviously correct that they sound silly to explicate; an example is the assumption that the apparent deadliness of botulism in humans in noe attributable ro confound- ing factors. In such situations ic is rarely recognized that ‘model is present. But most epidemiologic research is plagued by uncercaincy about assumptions, in which case {cis imporeant to recognize and explicate fully the anal- ysis model and tis drawback of common sail mades is at they embody many parametric assumptions that are not known to be correct and may well be incorrect. Consies, for example, a convenconal corelsedout come logistic regression analysis of data on daily air pollution exposure an respiratory illaess in schoolchil dren, with sex entered asa covariate. The analysis model assumes (among other things) that the odds ratios relat- ing pollution co the illness do noe vary by sex and that the relation of the illness ods to pollution is exponen- tial, Neither ofthese assumptions is known to be correct, and texts of them often have little power to detect imporcane violations. Another drawback of common statistical models 1s that they cannot capture all types of assumptions. Con- tinuing the example, a conventional logistic-regression ‘model for illness cannot represent or impose the assump- tion that the sexes of the children are not affected by pollution exposure, even though this assumption is cor- rectand may be useful in an analysis of pollution effects. Causal diagrams are graphical models for causal rela- tions that can serve a role complementary to conven- tional models; they are also called influence diagrams, relevance diagrams, or causal networks. Such diagrams do not incorporate the strong parametric assumptions of conventional models; instead, they display assumptions about the web of causation that are not captured by conventional models. As with conventional models, some of the assumptions may be of unknown validicy: some may be untestable, but others may be tested (albeit with limited power) and may be varied to examine the sensitivity of inferences to reasonable variacions. Basics of Graph Construction A causal diagram can be constructed by abstracting che ‘causal assumptions embedded in a narrative description of che hypothesized relations among the scudy variables. To illustrate the idea, consider a proposed study of the relation of ancihistamine treatment to asthma incidence among first-grade children attending various public schools. Suppose our narrative asserts that pollution levels and sex are independene among first-grade public- school children: char sex influences alministraion of antihistamines only chrough ie celation co bronchial reactiviey but directly influences asthena esks: chat in- Justrial air pollution leads co asthma attacks only through its influence on antihistamine use and bron Epidemiology January 1999, Volume 10 Number 1 FIGURE 1. chial reactivicy; and that there is no important con- founder beyond air pollution, bronchial reactivity, and sex. For this example, lec A represent air pollution level, and let B,C, E, and D represent indicator for sex (B = 1 for boy, 0 for gil), bronchial reacrvicy, antihistamine treatment, and asthma, respectively. The assertions of the narrative are incorporated into Figure 1, which we will use to illustrate terminology, Any line or arrow connecting two variables in the araph is called an arc or an edge. Two variables ina graph are adjacent if they are dicectly connected by an arc; in Figure 1, A and C are adjacent, but A and D are not Single-headed arrows represent direct links from causes to effects; in Figure 1, the arrow linking A to C repre- sents a direct effect of A on C (chat is, an effect noc mediated by another variable in the diagram). In the example, the astertion that pollution afects asthma only through bronchial reactiviey and antihistamine use cor- responds to the absence of an arrow from A to Ds likewise, the assertion that sex affects antihistamine administration only through bronchial reactivity corre: sponds to the absence of an arrow from B co E. ‘The points on the graph representing the variables are called nodes or vertices. A path through the graph is any unbroken route traced out along or against arrows ot lines connecting adjacenc nodes; an example is the E- C-D path becween E and D, which passes only through C.A directed path from one node o another in the graph is one that can be traced through a sequence of single- headed arrows, always encering an arrow through the ail and leaving through the head; such a path is also called 2 causal path in causal graphs. In Figure 1, the path ACD is directed, bu E-C-D is not. A node within a pach is said co inercepe the path; in Figure 1, C intercepts the paths A-C-D and ECD. ‘A variable X is an ancestor or cause of another variable Y if there isa directed pach of arrows leading out of X neo Y; in such a case, Y is said to be a descendant of X ot affected by X. In Figure 1, A, B, and C are ancestors of E and D, and E and D are descendants of A, B, and C. A variable X is a parent of Y ina graph if there i a Single-headed arrow from X to Yin such a case, Yis said tw bea child of X or directly afectad by X. In Figure 1, A and C are parents of E, whereas C and E are children of A ‘Note chat the “parent-child” relation and the corte- sponding notion of "direct effece™ are not inherent bio-Epidemiology January 1999, Volume 10 Number 1 7 logic properties, bue merely express the limits of detail represented in the causal model. For example, to say that “antihistamine use E directly affects asthma occurrence D" means only that the intermediate steps through which antihistamine may alter asthma occurrence (for example, by decreasing capeure and elimination of anci- gens in the upper airways) are not elaborated in the rodel. In contrast, the more general concepts of ances- tor and descendant correspond ro the basic concepts of cause and effect. ‘A bidirectional (cwo-headed) arrow connecting two variables ina graph is often used to indicate that the C¥0 variables share one or more ancestors (that is, have a cause in common), but the ancestors and their interre- lations are not shown in the graph.® We will instead represent such unspecified common ancestors by the letter U, with dashed arrows. U may represent more than ‘one variable. For example, if we suspected actions taken at high-pollution-area schools reduced pollution expo- sure and also independently reduced asthma risk, we ‘would use Figure 2 instead of Figure 1; these conditions would be satisfied if children were kept indoors on high-pollution days, with consequent reduced exposure to particulate emissions and allergens. ‘A nondirectional are (an arc without arrowheads) is sometimes used to indicate that two variables are asto- ciated for reasons other than sharing an ancestor or affecting one another; Figure 3 gives an example, which we will discuss below. Here, we use dashed nondirec- tional arcs to represent relations whose source is not A —E——___—__>D FIGURE 3. CAUSAL DIAGRAMS FOR EPIDEMIOLOGIC RESEARCH 39 specified by the graph. In the example of Figure 1, che tieercion that pollution exposure and sex are marginally tunassociated follows from the absence of arcs of com- mon ancestors connecting A and B in Figure I "A graph is directed if all arcs between variables are arrows (single or double headed), that i, ifit contains no rondirectional arc. A graph is acyclic (or recursive) if no icected path in the graph forms a closed loop. Figures 1 tnd Z are examples of graphs that are directed and fcyclic. Inthe presene paper we will deal primarily with DAGs; we will also discuss other acyclic graphs arising from DAG, such as Figure 3. ‘We will also need the following terms. A path that ‘connects X co Y is backdoor path from X zo Y if has an arrowhead pointing to X. In Figure 1, all paths from E to D except the direct path are backdoor paths. A path collides at a variable X if the path enters and exits X through arrowheads, in which case X is called a colider fon the path’ A path is blacked if ic has one or more collides; otherwise itis unblocked. The backdoor path E-A.C-B-D in Figure | is blocked because it collides at CG; Cis the only collider on the path. In contrast, the backdoor path E-A-C-D is unblocked because neither A nor C are collides on this path. Either kind of path may include nondirectional ares; for example, the backdoor path E-A-B-D in Figure 3 is unblocked. Causation and Association in Graphs Graphical assumptions are qualitative and nonparamet- ric, in that they imply nothing about the specific func- tional form of the relations or distributions among the variables. In particular, the variables may be discrete or continuous, and dose-response relations may have any Shape. Production of an effect by a cause requires a ‘causal path from the cause to the effect, however; such paths are represented by directed paths in a graph. Thus, absence of a dicected path from X to Y represents the assumption that there is no effect of X on ¥. Every absence of a single-headed arrow represents @ basic null (no-direct-ffect) assumption encoded by the graph and implies absence of every causal effect that would be transmitted through that arow. For example, Figure 1 has no single-headed arrows from A to B, B to ‘A, Ato D, and Bt E (these single headed arrows are the only ones chat could be put back in without disrupe- ing the acyclic nature of the graph). Thus, the figure encodes the basic causal assumptions that A does not irecly affect B, B does not direcly affect A, A does not directly affect D, and B does noe directly affect E: the Fist assumption also implies thac there is no effect of A fon D that is transmitted through B, and so on. Tn some situations, one may wish to assume that co variables have no marginal (crude) association. Al- though this assumption would rarely be exaccly correct icis often a reasonable working hypothesis; for example, it is 2 common assumption in studies of genetic and environmental factors. This assumption follows from the absence of an unblocked path berween the varia- bles?* for example, Figure | implies that A and B have40 GREENLAND ET AL ‘no marginal asociation. Put another way, a marginal association becween two variables in a graph requires that there be an unblocked path berween them. Ina DAG, there are only two possible kinds of un- blocked paths becween variables: directed paths and backdoor paths through a shared ancestor. Thus, a mar- ginal astociation berween two variables in a DAG re- quires chat che DAG exhibit a causal pathway from one (0 the other or a common cause of both variables. The first condition holds when the association besween the variables is at least partly causal, and the second condi- tion holds when he association is at leat partly con- founded. Of course, both conditions may hold; chat is, the marginal association may be partly causal and parcly confounded, as with E and D in Figure l ‘The fact that every association berween two variables X and Y in a DAG is due to either a direct causal connection of a common cause may appeat to conflict with the usual epidemiologic incuitions. For example, associations may appear in a study population because of biases in selection into the population: Berkson's bias is a well-known example.” The conflict may be resolved by noting that the associations generated by selection bias are represented by nondirectional ars, as in Figure 3: maphs wich such ares ae by defirion no drected graphs. Directed graphs represent only causal relations in unselected populations; nonetheless, we will show holw such graphs can be used to deduce nondirected graphs for selected populations, such as Figure 3. The independence implied by the absence of un- blocked paths beeween two variables applies regardless of, the magnitude of any effector association shown in the sraph. On the other hand, the presence of an unblocked path berween two variables allows but does not auto- matically imply that the two variables are marginally associated. For example in Figure I there isadireee path and four backdoor paths berween E and D. Although cach path can induce an association, these associations might cancel one another out co yield no marginal E-D association, Standard epidemiologic examples occur when an effect is hidden by confounding, so thac one observes no crude association of exposure and disease, even though exposure affects disease risk. The presence or absence of a blocked path berween ‘wo variables is irrelevant to cheie marginal astociation. This irrelevance corresponds to the fact that the rar- ginal association between two causes of an effect (that i, two ancestors of a collider) is fixed by the time both causes have occurred, and the association cannot be changed by the fact that these causes may trigger a laeer event. For example, ifthe BRCAI gene and intrauterine diethylslbestrol (DES) exposure are marginally inde- pendence of one another, this independence is noc changed by the face that later ouccomes (such as breast cancer) may be affected by both variables. In other words, a blocked pach cannot contribute to a marginal association, because an event cannot alter conditions that were determined before the event occurred; in particular, brease cancer eannoe induce 2 marginal asso- Ciation beeween BRCA! and DES exposure. Epidemiology January 1999, Volume 10 Number 1 Confounding and Causal Graphs There are many different definitions of confounding and D FIGURE 4. logue of this process. Unforcunately, analytic adjustment for C has consequences quite different from those of physical control. Wich adjustmenc for C, we must shift four atention fom the erode A-B association to the A-B susociations within serata of C. In doing so, we descend into strata of C, in which A and B can be asociated (as in Figure 3), even though they have no cause (ancestor) in common and hence have no asscciation in the orig- inal directed graph. Unnecessary ADjusTMeNT AND HARMUL ADJUSTMENT Improper adjustment for a third variable can create confounding éven when exposure and disease share no common cause. Figure 5 gives an example; here, A and CC have no effect on D other than through E, and B and Crhave no effect on E. As in Figure 1, however, A and B would be associated within strata of C; hence, if one adjusted for C, one would have to adjust for A or B to ensure thar there is no confounding within the strata. ‘There is a broader lesson in Figure 5. Given that diagram, there is no confounding and hence no need to adjust for C, because there is no unblocked backdoor path from E'to D. We can nonetheless make A and B into confounders by adjusting for C. Thus, given Figure 5, unconfounded answers can be obcained by adjusting for nothing or by adjusting for anything other than C alone. More generally, adjustment for variables (such 35 C in Figure 5) that are not necessary to control may necessitate adjustment for even more variables. ‘Now consider Figure 6, in which we observe a variable F chats affected by E and D. There is no confounding of the effect of E on D in this diagram. Thus, given the E——_——_>D FIGURE 5.Epidemiology January 1999, Volume 10 Number 1 E—————_>D F FIGURE 6. diagram, the crude association of E and D should cepre- sent the effect of E on D without bias, and adjustment for Fis unnecessary. Furthermore, because Fis associated with boch E and D, the associations of E and D wich serata of F wold differ from the crude association and hence differ from the true effec; that is, they would be biased. This situation is similar to that of adjustment for Cin Figure 5, except that there is no other variable to conerol that would remove the bias. Thus, adjustment for F is not only unnecessary but irtemediably harmful (biasing). ‘A classic example of adjustment-induced bias oc- curred when, in studies of estrogen (E) and endometrial cancer (D), some researchers attempted to control for erection bias by stratifying on uterine bleeding (F). ‘which could be caused by cher estrogen or cancer, 358 Figure 6. The association between estrogen and cancer within levels of bleeding was drastically reduced by this Stratification; unforeunately, the reduction was more plausibly ateriburable to bias produced by stratification than to removal of derecrion bias-* ‘Another important example of adjustment-induced bias ean occur in arcemprs to adjust for the “healthy- worker survivor effect.”” This effect is the bias that ‘occurs when unmeasured health conditions (U) influ tence decisions to leave work (for example, quit, take disability leave, or retire) and also influence morality (D); when this happens, leaving work may become 25- sociated wich mortality, even if it has no effecr on ‘mortality. Suppose the ‘study exposure (E) is job-site fssignment, which influences worker decisions about when to leave work (L). Figure 7 illuscrates such a Scenario. In this figure, there is no confounding of the ED association, 30 that the crude association is unbi- ased. Because L is asociaced with E and D, however, control of L would most likely yield an E-D association different from the crude and therefore would be harmful, FIGURE 7. (CAUSAL DIAGRAMS FOR EPIDEMIOLOGIC RESEARCH 43 ‘Surmciest Sers oF Apjustuent VARIABLES Suppose now chat we have laid out out working assump- tions about causal relations in a DAG, and we wish to estimate the effect of E on D, where E and D are ‘ariables in the graph and E temporally precedes D. We iknow thac adjustment for a variable can produce bias if the variable isa descendant of E, We have also seen that fdjustmene for a variable along a backdoor path from E dnd Dean produce bias if that variable isa collider along the path, as C was in Figure 5. This finding is a subcle insight about analytic control added by causal graph theory. and we must account for ic in revising conven- tional criteria for adjustment. ‘Suppose we have a sec S of variables that we are considering for use in adjustment, We will refer to the Scrata defined by crost-clasification on all of the vari- ables in S as “srata of S". We define a set of variables S fs sufciens for confounding adjustment ofthe E-D relation if there is no confounding of the E-D relationship in any stratum ofS, Our usage of“suffcienc” will ehus refer only to contr of confounding. Concems about possible het- trogeneicy of effece measures (effect-measure modifica tion) may dicate finer straefication than that sufficient for confounding adjustment. For example, in a large randomized tral, one may have no concem about con- founding, in that the unadjusted (crude) estimates may bbe unconfounded summary effect estimates; nonetheless cone may wish to look for possible differences in effects gmong men and women, young and old patient, etc. ‘Gaastica. CATERA Given a DAG and a see S of variables in the graph that are not descendants (effects) of E or D, it can be shown that Sis sufficient for adjustment if upon adjustment for §, there is no unblocked backdoor path from E ro Dé This condition is equivalent to the following pair of graphical cries ‘Gl. Every unblocked backdoor path from E to D is incercepted by a variable in S, and ‘G2. Every unblocked path from E to D induced by adjustment for the variables in S is intercepted by @ variable in S. Céiterion G2 is the key addition to conventional wisdom. Ac firs glance, it might seem difficult to verify that this ericeron is satisfied; forcunately, iis equivalent to the following operational exicerion: Gi. Ifevery collider on a backdoor path from E to D is either in $ 0¢ has a descendanc in S, then S must also contain a noncollider along that path.* Ta Figures | and 5, che set containing only C does noe satisfy this criterion, whereas the secs (A, C) and (B, Cl do, "The above criteria lead co a graphical (and hence visual) algorichm for checking whether a set of variables isaufficiene for adjustmene given a particular DAG. This algorichm is known as the backdoor est for sufficiency." ‘This texe may be performed as follows. Civen a subset S'= (Sy, -- + Sa of variables that contains no descen- anc of E or D, perform the following series of manipu lations (steps) on the original graph:44 GREENLAND ET AL FIGURE 8. MA. Delete all arrows emanating from E (chat is, remove all exposure effects); ‘M2. Draw unditected arcs to connect every pair of variables that share a child that is either in S or has a descendanc in $ (that i pue in all arcs generated by control of S); and MG. In the new graph derived from steps MI and M2, see whether there is any unblocked path from E to D that does not pass through S. Sis sufficient ifthe answer to M3 is negative, that is, if every unblocked path from E to D in the graph resulting from MI and M2 passes through S. ‘ Applying MI and M? to Figure 1 with S = (4, C}, S = 1B, Ch, or S = (C} yields Figure 8. In this figure, every path from E to D passes through (a, Cl, s0 (A, Ch is suficienc; however, the unblocked path E-A-B-D does not pas through C, 50 {C} isnot sufficient. Applying MI and M2 to Figure 5 with S = (C} yields Figure 9, in which the unblocked path E-A-B-D does not pass through C, so (C] isnot sufficient. Applying Ml and M2 to Figure 5 with S = 0) (che empey set) yields Figure 10, in which there is no unblocked path from E ro D; thus, We need not adjust for anything given Figure 5 ‘When a graph contains many variables, it may be- come tedious to carry out M3. The back-door test may be simplified as follows. After we adjust for a sec S, any unblocked backdoor pach can only pass through vari- ables that are ancestors of E, D, or a variable in S. Therefore, before applying the test, we may simplify the graph by dropping all variables cutsde of E, D, and S that are not ancestors of Eor D or a variable in S. This, simplification may considerably reduce the number of backdoor paths from E to D. Apelying this simplification FIGURE 9. Epidemiology January 1999, Volume 10 Number 1 A — “ c FIGURE 10. B to Figure 5 with S = (} leads us to drop C from the graph. Dropping C yields Figure 11, in which there is no back: door path from E to D. Thus, the empey et is sufficient to adjust for confounding in Figure 5, which isto say that there is no confounding in the figure. This is s0, even though C would almost cerainly be marginally asoci- ated with E and associated with D condieional on E, and $9 it would appear to be a confounder of the E-D effect if A and B were ignored. Mxua, SuRCIENCY A sec of variables may be sufficient for adjustment, but it may be unnecessary to adjust forall of the variables in the sec. A set S of variables is minimally sufficient for adjustment if S is sufficient bur no proper subset of Sis, sufficient. In Figure 1, both (A, C} and (B, C) are rinimally sufficient secs, because some control is neces- sary and no single variable is sufficient for control. On the other hand, although (A, C} and (B, C} are both sufficient in Figure 5, neither is minimally sufficient, because no control is necessary (that is, che empcy set i also sufficient). To find a minimally sufficient set, we may sequently dele variables fom a silent et tuncl no more ean be dropped without the new sec fling the backdoor test.# Ie is important to recognize that a set S may be sufficient for confounding adjustmene, and yet adding variables to $ may yield asec S¥ chat is not sufficient. In ‘other words, control of a see of variables may yield an unbiased measure of effect, and yet control of more variables may create bias" Figure 5 is a simple example ‘of this phenomenon, in which no control (S = {}) is A E————____=>D FIGURE 11Epidemiology January 1999, Volume 10 Number 1 i‘ Xo | eee teeter FIGURE 12. sufficient, and yer control of C (S* = {C}) is not suffi- It is also important to recognize that there may be several different minimally sufficient sees, which may ‘not even overlap with one another, and which may be quite disparate in size and in difficulty of measurement. Figure 12 provides an example in which all of the causes, of D are associated with E only through a final common * path that passes through F. Boch (a, B, C) and (F} are minimally sufficient but have no variable in common. Grartical, Sepagarion ‘The graphical criteria for sufficiency (eriteria 1-3 above) involve a more general graphical concepe chat wll prove useful in evaluating confounding criteria and variable subsets. We say a set of variables S separates two other sets Rand T, or S blocks every path between R and T, if the following criceria are met: S1. Every unblocked path from R to T is intercepted by a variable in S, and 'S2. Every unblocked path from R to T generated by adjusmenc for the variables in $ i intercepted by 8 variable in S. (This concepe is usualy called “d-separation of Rand T by S"in the graphical iterature,™ where d stands for “directional”.) With this definition, we can say that Sis sufficient for control of confounding under given DAG if S contains no descendants of E or D, and S separates E from D in the graph obtained by deleting all arrows emanating from E. Thus, in Figure 8 (which is Figure 1 with the E-D arrow deleted), the set (A, C} separates the se ([E] from the set [D}, and so does che sec [B, C}, but the set [C] does not. In Figure 12, che set (FI separates (E) from {D} once the E-D arrow is deleted, and 50 does the seta, B, Ch, The concept of separation is useful beyond the back- door test, because it implies the seatisical concept of stratum-specific independence: if $ separates R and T, then every variable in R is unassociated with every variable in T within strata of S.!28 For example, in Figure 8, (A, C} separates (E} from (DI, and hence, under this graph, E and D are unassociated within A-C strata. CAUSAL DIAGRAMS FOR EPIDEMIOLOGIC RESEARCH 45 Although the converse is not necessarily correct, the exceptions involve perfect cancellations of associations and so may be of litcle practical relevance in epidemi- ology. Connection of Graphical and Statistical Criteria Suppose thac we have partitioned a sufficient subser of potential confounders in a problem into two subsets S and T, where no variable in S or Tis affected by E or D ‘Although there. is some variation in details, criteria based on statistical associations are often used to decide whether to treat S alone as sufficient for adjustment. ‘Commonly, S is treated as sufficient if every potential confounder C in T satisfies atleast one of the following UL Cand E are unassociated (independent) within the strata defined by S; or UL. Cis unassociated with D within the strata defined by E and S. For example, if S = (age, sex}, T = fbirch year), E = smoking, and D = myocardial infarction (MI), the cri- teria say thar birth year need not be controlled once age and sex are controlled if either (1) smoking and birth year are unassociated within age-sex strata ot (2) MI incidence and birch year are unassociated within smok- ingeage-sex strata. Some authors use a less restrictive version of criterion U2, in which C is unassociated with D among the unexposed. Such usage is justifiable ifthe effect measures of interes refer or are standardized to the exposed pop- ulation, as is often assumed.!®t42 Our version is needed, however, if one wishes C ro be unnecessary regardless of which standard (reference) distribution is chosen.!? Consider the statistical criteria applied to Figure 1 with S = {C} and T = {A, B). Aparc from examples with perfect cancellation, the above criteria would dececr the insufficiency of S. For example, within strata of C, E and A are associated because A affects E; and, within E-C scrara, A is associated wich D because A would be associated with B and B affects D. ‘Now consider Figure 5 with S = () (che empty et) and T =A, B, C). Apart from examples with perfect can- cellation, the above statistical criteria would fail co de- tect the sufficiency of S (that is, that no control is necessary), because C fails both criteria; E and C would be associated because both are affected by A; and C and Dwould be associated within strata of E because both are affected by B. Thus, the usual criteria for evaluating a set of confounders do not correspond perfectly with the graphical test; the criteria may noc idencfy all of the Sufticient sets and, in particular, may fail to identify minimally sufficient ses Te is possible, however, co generalize che usual stais- tical ericeria so that they idenefy al ofthe suficienc ses found by the backdoor test (see corollary 4.1 of Robins”). Suppose that S and T together are sufficiene for adjust- ment. Then T will be unnecessary given $ if T can be decomposed into two disjoine subsets T, and T; such that, within stata of S, both ofthe following criteria are met46 GREENLAND ET AL Ut*. 7, is unassociaced with E within stata defined by S; and U2*. T, is unassociated wich D within strata defined by E, S, and Tis ‘Note’ that variables in T, may be associaced with variables in Ty; the criteria do require, however, that T, and T, have no variable in common and that all vati- ables in T appear in either T, or Ts. To apply these criteria, we may begin by taking T, to be the largest subset ofS that is independent of E and so satisfies Ul*; we then check to see whether T; = TT), the remain- der of T after delecing T,, satisfies U2*. ‘Whereas the usual approach requites either criterion Ul or U2 to hold for each variable in T to ignore T, we requite both criteria UI* and U2* to hold. When con- sidering jut one variable beyond S, however, criteria UL and U2 are special cases of criteria UI* and U2*. To see this fact, first noce thar the empty set () is unassociated with everything (because it defines no strata), and hence ie vacuously satisfies both UI" and U2*. Hence, if C satisfies criterion Ul, T = (C} satisfies criteria UL? and U2* wich T, = (CY'and T, = fh if C instead satishes eviterion U2, T = (C} satisfies criteria UI* and U2* with T, = and 7, = (Ch. ‘As an example, suppose S = fage, sex}, T = {bicch year, height), E = smoking, and D = Ml. Then crceriat Ul* and U2* imply chat neither birch year nor height need be controlled if any one of the following four conditions hold: 1. Within age-sex strata, smoking is unassociated with the compound variable birch year-height (T, = {birch year, height), T, =); 2. Within smoking-age-sx strata, the compound vari- able birch year-height is unassociated with MI (T, = 0. rch year, height); Smoking and birth year are unassociated within age-sex strata, and height and MI are unassociated within smoking-age-sex-birth year strata (Ty = {bieth year), Ty = theight): or 4. Smoking and height are unassociated within age- seesran and bith yer and Mare unasociated within smoking-age-sex-heighe strata (T, = {height T; = {birth year). Each of these conditions is sufficient, even if birch Year is associated with height. Consider again Figure 5 wich $ = {) and T = (A. B, Ch. Using T, = {B) and T; = (A, Ch, we may show that T satisfies criceria UL" and U2* and hac S (no adjust- ment) is sufficient. First, B and E are unassociated, s0 Ut* is satisfied: second, within E-B sata, A-C is unas soctated with D, so U2* is satisfied. The preceding example illustrates a general theorem linking graphical and associational erievia. Suppose we have disjoine sets of variables S and T in a DAG thae coneain no descendane of E or D and cogether sais the backdoor test; then $ alone satisfies the backdoor cese only if T can be partitioned into disjoine subsets that satisfy criteria UU and U2*."=" Thus, graphical sulf- ciency implies satisfaction of our generalize statistical criteria. The advantage of che graphical approach is that Epidemiology January 1999, Volume 10 Number I FIGURE 13. ic offers a straightforward visual means of checking cri- teria, at least in moderace-szed graphs, and it translates inco an efficient algorithm for finding minimally suffi- ciene subsees of conctol variables This algorithm does ‘not depend on quantitative information and so can be applied in the design as well as the analysis stage of a seudy, 2s well as to situations in which some key poten- tial confounders have nor been measured. The graph also clearly displays the essencial assumptions for the algorithm, such asthe assumptions that S and T contain no variable affected by E or D. Ieis possible for S and T to satisfy UI* and U2* even though S fails the backdoor test In such a caze $ would ssll be sufficient for adjustment; thus, the backdoor test does not identify all sufficienc subsets in all sertings. Nonetheless, these case involve perfect cancellations of associations and so may be of little practical relevance f epidemiology. Ina subsequent paper, we hope to discuss the relation, of graphical criteria for nonconfounding to exchange- ability and collapsibiliy criteria“! Briefly, fS and T satisfy UL* and U2*, then the different exposure groups will be exchangeable within strata of S, and both the risk difference and risk ratio (bue not necessarily the odds ratio) will be collapsible over T? Thus, if S pastes the backdoor test there will be no confounding within strata of S, and, given stratification on S, T will contain no confounder according to the definitions in our previous writings 626 Discussion A closed loop in a causal diagram represents reciprocal causation (feedback). Graphs with such loops are called calc or nencecarve, We have limited ou discusion ro acyclic graphs, because cyclic graphs are interpretable a3 shorthand representations of acyclic graphs with multi- ple time-specific measurements of the variables. For ex- ample the statement "X can cause Y and ¥ can cause X° could be represented in a cyclic graph as simple 23 Figure 1B. Nonetheless, Because caver must precede efits the statement can only be meaningfully ineerpreted as saying that “he occurrence of X ean cause a subsequent occurrence of Y, and vice vera.” An acyclic represen- ‘ation ofthis statement requires time-specific versions of each variable, such as in Figure 14. Cyclic graphs can have a deceptively simple appearance relative to the complex time-dependent processes they represent. With Proper translation into acyclic form, however, such graphs can be analysed sing che methods given here: such translation will also reveal exactly whae time-series of measurements are needed for unconfounded effect estimation. For discussions of graphical mechods for time-dependent variables and ditect effects, see Robins? and Pearl and Robins.”Epidemiology January 1999, Volume 10 Number 1 be Yy xy Y FIGURE 14. Causal diagrams provide results equivalent to those obsainable using the more established counterfactual causal models of philosophy" and statistics”; see Rob- ins and Galles and Peatl® for descriptions of the con- nection, Although we have found counterfactual models useful in our previous work,™ graphical models seem to have a broad intuitive appeal, and we are enthusiastic about their use in teaching principles for identification and control of confounding. Nonetheless, we must em- phasize that no approach solves the central epistemo- logic problem of inferring causation from nonexpeti- mental (observational) data. As realized by Hume centuries ago and reinforced by many authors since? all causal inference is based on assumptions that cannot be derived from observations alone. Even if we obcain randomized-trial data, we must assume that no bias occurred in allacation and compli- ance, or that any such bias can be handled by adjustment procedures; such assumptions are not always correct. In ronesperimental studies, our inferences are furcher lim- ited by assumptions (often wishful) that we have ade- quately measured all_of the importanc selection and confounding factors. Graphs can only help convey those assumptions and indicate what measurements are suffi cient for adjustment, given those assumpdons. Because the graphical techniques used in this paper presume a qualitative understanding ofthe causal seruc- ture behind the data, chey are less ambitious than those used in some of the graphical literature, which attempe to infer such structures from data.2'28 The methods used for such inferences involve assumptions that may not be warranted in typical epidemiologic studies. Full desrip- tions and critiques of these methods are somewhat in- volved and appear elsewhere." Because the meth- ‘ods are highly controversial, their use for drawing inferences (such as in the TETRAD software of Scheines ea?) should be cautioned by awareness ofthe controversy ‘We have limited our presentation to graphical anal- sts of confounding. Of equal or greater importance (© validity are problems of measurement error, which in most cases (and contrary to popular lore) cannot be validly handled by simple approaches, such as stratifca- tion on indicators of “information qualicy."® Graphs can, however, be used (0 convey and analyze assumnp- tions about measuremene processes. Important applica- tions arise when a measurement can be affected by variables other than the one supposedly being measured. Asa simple example, consider a case-control study of sleep position E and sudden infant death syndrome D. CAUSAL DIAGRAMS FOR EPIDEMIOLOGIC RESEARCH 47 Suppose the measurement of E is parental response F to ‘a question about sleep position; response bias would then correspond to the presence of an arrow from the study courcome D to the exposure measurement F, as in Figure 6. Ina future article, we hope to describe the application of graphical models to validity questions thac involve simultaneous consideration of diferent sources of bias In closing, we wish to emphasize thac the methods we have discussed are purely qualitative and so donot address the quantitative problem of determining how ‘much contro is necessary to reduce bias o an acceptable level. Such determinations require much more informa- tion than demanded for purely graphical analysis, such as information about strength of associations among con- founders and exposure, as well as strength of confounder cffecs on disease. This information is rarely known with any accuracy; asa consequence, estimates of residual bias are rately if ever accurate. One way to address this problem is to conduct a sensitivity analysis of bias, ?-" although such analyses may themselves depend on and hence be sensitive to various simplifying assumptions. ‘These issues exemplify che complexities one can expect to encounter in a quantitative analysis of bias Acknowledgments We ae indebrd wo Clarice Wenher,Alsader Walter, Pee Mat sa {he eer forthe excep el commen. References 1, Woah. Congion and custom, J Ag Rey 192120557585. 2 Danan OD. Smuenal Eqns Made New Yous Acalenic Pes bie 4 Foal Bsbblsc Rennig in nein Sens Sn Fc CA Morn Kauann (88. 4. Speaker, Darl AP, Lain S, Cowell RG. Brean ach in pe an (wh Soca Sea Se 198)3219-28). [er Da 19 ge tol Rabins M. Comment Bomar 1958295-698. 9: Robins JM. Coun afeence fom conpes longline Beane ‘Mead Lins Vale Matcng nis Aplenton easly. New Tork ‘Springer Veg, 199769-107. 10. Maren OS, Ck EF Conon: Eence and Detection. Am) Ep del 1891 114395-00) 1, Greelnd 5, bis JM Mensiabiy,exchangbliy, an epdemislg ‘conbuading it Epcot WER ISAIAD. 12, Robie ]M Morera The fiunditon fone in ee (py Moth Mang 94714 309-916 13, Wher CR Tina deer dines econfunting Am Epes bounce 14, Robs IM, Cerin 5. The nie of mall ast nal nleence fom unerprmensl dna) Eee 9863922 15, Grenlnd 8: News RR. Am alps of econ by a prone ‘Srestons nthe sa of etme snd endowed ance} Chon Dit aaea—os 16, Kary Ml Gene epdemlg Ch. 3, Ine Rchman Kevan tk, Sem Epuemilog.Phaipha: Lppncer aver (998815317 11, Beane Lemans of the aprcanon of label ep thes Bane Bal 60 18, Whegemare AS Cll in miamersinelcncngny sh.) Snr Soe B 197840328-040 19. Kim KPa )-A computation tale Gr combined cna sd Sate48 GREENLAND ET AL in iene ne Po th Int Co Asi Ineiecs 20, Begoneh WW, Wenbey CR, Tarlo JA. Nonhercie lic wel (au cont io or tng cep in poplar ce (Scola, Sx Med ISIE 21, Bog CE Thang 2. Seal ons of molecular glenn es fmplying coe teen Cancer Epi Bomar & Prey ATS ts 12, Peal}, Commene: pope male, couaiy a inevenson St Se odes 13, Lunn Dew AP Laren BN, Line HO: tndepenence pei (danced son fee Neer 199020491595 124 Tam] at A Pe] ding inal eprint TELA Gonete Soence Bepromene Techies Rep 8-234 Las Ange les UCLA Row 1987 25. Vena T Peat Cal nerd: seman and exresivenes. ‘Shocker # vec TS, Kal A i Unerainyin Arf ne free 4 New Yok Evie, 19069 76 14 fein Ged Malem pense, 2a ot Pb paca Raven 198. 11, PER Ree Mt. Pb evalu of equencl plans fom cul Itosel woh Raden able Is Bermand. Hands Ss Unceraing 0 Secs tigctS erence Hom Kors, S. 28 Leva. Couto} Phin 1973:70556-567. Repel with pons Lite B Ploegh Papen vl 2 New Yorks Ox, 1986 19, Rakin DB Comnens Ney (192) sd cura ieferenc i expen el cern sai St Si 19905472180 5a, Gale, Peat] Anions chancercson of eunetenals. Found Set foesist-te 31 Peat]. Veg. A theory of need cation, ns Alle JA Bs Be Exnleral is Prope of Keowee Repracnion and Renee, 2 a M % 3 m “ 2 Epidemiology Jaruary 1999, Volume 10 Number 1 Prceting of th od Icon Cones Sun Fanci, CA: Mac tea Eine DBL ‘Eheinar Spine P, Gimour C. Mek C, Rehankon T. The TETRAD ‘powes creme ta neo ual nel pcan. Mal Behav Ret Format ample, Flan O. The ga ley (wth Suvi) Be) Pl St {Dee HTATii2) (Denanon (OFTHE Robes JM: Waserman On tbe igen of fering causation fom {ooraton wit bck ourd raed fs Cooper C. Clyne Ce ‘Eitotoe and Cormpraton Camlge: MA: MITAAAL Pres, 1999 ( en Post TETRAD and SEM Male Behow Res 9983311908, Gimar © A revew orn wos on he uatins feu inference. fe Bckim Vic Tomer SP (le) Calc Cast. South Be 1 {Gnvete of Nae Dame Pes, 1997201-28- eeinan 5, Pom scion cauacon va resin (wih cuion) [er bckn VR Tomer SP Gis), Casey in Ces, South Berd N Unive of Nowe Dae Pe 1997113182. (reel §, Rane Jo Conuning ard mice, Am) Ede ma 1985132995 58, ‘Greil Se meth rest ana ad exer sent. (GLP is Rotman H, Ceenand 5 eds. Mover Epes, Phlb- Sipad Upgioewe Raven, B38500-38 Fob JM Cameron ar nnsomglance i egualence al Sak Med (pp17269-302- obi Romi A SchainD. Seiya fo leon Bia {mney asd cal ence mes I Haloran ME, Sx Male in Epemoloy- New York Spang Verng 1989 i rem Grelond $ The sey of sera anaes. 1997 Pree of (Ge Boner Secson ofthe Ameren Snead Asoeacon, Aen, VAs Amenan Sonal Ancestor, 5819-2