International Journal of Epidemiology, 2017, 1–9

doi: 10.1093/ije/dyw328
Essay Review

Essay Review

Causal inference—so much more than statistics

Neil Pearce1,2* and Debbie A. Lawlor3,4
1
Department of Medical Statistics and Centre for Global NCDs, London School of Hygiene and Tropical
Medicine, London, UK, 2Centre for Public Health Research, Massey University, Wellington, New
Zealand, 3MRC Integrative Epidemiology Unit at the University of Bristol, Bristol, UK and 4School of
Social and Community Medicine, University of Bristol, Bristol, UK
*Corresponding author. Department of Medical Statistics and Centre for Global NCDs, London School of Hygiene and
Tropical Medicine, London, UK. E-mail: [Link]@lshtm.
Accepted 19 October 2016

Introduction that the same is true in epidemiology, and that whereas there
are debates about the relative prominence of these tools (as
It is perhaps not too great an exaggeration to say that
illustrated in recent papers and correspondence in the IJE1,9–
Judea Pearl’s work has had a profound effect on the theory 15
), it is essential that biostatisticians and epidemiologists alike
and practice of epidemiology. Pearl’s most striking contri-
bution has been his marriage of the counterfactual and are familiar and comfortable with these tools.
probabilistic approaches to causation.1 The resulting tool- Given the complex nature of some of the concepts and
kit, particularly the use of counterfactual concepts and dir- methods covered, particularly for those who are not familiar
ected acyclic graphs (DAGs) has been extended by some with them, the book is remarkably accessible and clearly
epidemiologists to remarkable effect,2,3 so that some prob- written. Chapter 1 introduces the fundamental concepts of
lems which were previously almost intractable can now be causality, including the causal model. Chapter 2 explains
solved relatively easily. What we previously tried to under- how causal models are reflected in data, and how one might
stand using words, probabilities and numerical examples search for models that explain a given data set; graphical
can now be explored using causal diagrams, so that mind- methods–in particular causal directed acyclic graphs
bending problems such as Berkson’s Bias can be explained (DAGs)–are introduced. Chapter 3 is concerned with how to
and understood relatively easily.4,5 make predictions using causal models. Chapter 4 then intro-
However, like War and Peace or Finnegan’s Wake, al- duces the concept of counterfactuals, and discusses how we
though most epidemiologists have by now heard of Pearl’s can compute them and what sorts of questions we can answer
work, we suspect that relatively few have read it, at least not using them. The companion website [[Link]/go/
in the form of the original texts.6,7 It is therefore of consider- Pearl/Causality] is a valuable resource and provides answers
able interest that Pearl, together with Madelyn Glymour and to the many study questions throughout the book that help
Nicholas Jewell, has now produced a primer Causal Inference with learning and understanding (it is not straightforward to
in Statistics.8 Their motivation, set out in the preface, is that register with Wiley for this and you are initially taken to a
‘statisticians are invariably motivated by causal questions’ but site that appears to advertise the book only, but if you can ne-
that the ‘peculiar nature of these questions is that they cannot gotiate the site, it will help you get the most out of the book).
be answered, or even articulated, in the traditional language
of statistics’. They note that the development of new tools for
causal inference over the decade has not excited statistical Key concepts
educators and that they are ‘essentially absent from statistics There are a number of key concepts and tools which are
textbooks, especially at the introductory level’. We would add clarified in the book, but we will focus here on three: (i)

C The Author 2017. Published by Oxford University Press on behalf of the International Epidemiological Association
V 1
This is an Open Access article distributed under the terms of the Creative Commons Attribution License ([Link] which permits
unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
2 International Journal of Epidemiology, 2017, Vol. 0, No. 0

the relationship between causality and statistics; (ii) con- generalizability is mainly conceptualized in terms of sam-
cepts of causality; and (iii) causal DAGs. pling from larger (infinite) populations, and also much of
randomized controlled trial (RCT) theory in which the
focus is on effect estimation rather than aetiological
The relationship between causality and statistics–
understanding.
Simpson’s Paradox and the importance of context
The book starts with a simple example of Simpson’s
Paradox showing how the results of a drug study in pa- Concepts of causality
tients with an (unspecified) illness may look quite different Given current debates,1,9–15 it is also of considerable inter-
depending on whether the findings are stratified by gender; est as to how causality is conceptualized by Pearl et al.:8
if not, the drug appears to be decrease survival, whereas it For our purposes, the definition of causation is simple,
actually increases survival within men and within women. if a little metaphorical. A variable X is a cause of a vari-
This ‘confounding by gender’ can be readily addressed able Y if Y in any way relies on X for its value. . . X is a
using stratification or any other form of adjustment, such cause of Y if Y listens to X and decides its value in re-
as multiple regression. However, the same data are then re- sponse to what it hears.
presented with the name of one of the variables changed.
This is compatible with definitions that have been used in
The potential stratification variable is now high/low post-
epidemiology for many years16 (see for example
treatment blood pressure (BP), and it is known that the
Lilienfeld,17 who stated that ‘a factor may be defined as a
drug can lower blood pressure. The results are the same in
cause of a disease, if the incidence of the disease is dimin-
the two examples (i.e. whether the strata are gender or
ished when exposure to this factor is likewise diminished’)
post-treatment blood pressure, the drug decreases survival
as well as in some recent papers in the IJE.1,12
in aggregate data but improves it when stratified by gender
Note that there is no requirement here for any sort of
or post treatment BP, with exactly the same magnitude and
intervention, or in fact any specification of how the value
direction of results in both cases); only one variable name
of X may change (or be changed). All that is required is
has been changed. But, in the former example, the correct
that if the value of X were different, then the value of Y
result lies in the sex-stratified (segregated) results, whereas
in the latter example it lies in the non-stratified by post- might also be different as a result.
treatment blood pressure data (i.e. the aggregated results). It is particularly noteworthy that this inclusive pluralist
Moreover, there is no statistical method which can help us concept of causation inherently involves causes which have
to identify which of the two scenarios apply to a particular been questioned in recent debates. In particular, causation
data set or analysis approach (aggregate or stratified). This is not restricted to specific actions (e.g. exercising 1 h/day),
can only be decided by information from outside the data and ‘states’ such as ethnicity, gender, sex, obesity, hyper-
set (e.g. that gender is a potential confounder and that the tension and high cholesterol levels can also be causes. As
drug in part may increase survival by reducing BP). with other causes of disease, some ‘states’ may be direct
Importantly, Pearl et al. use this example to illustrate causes (e.g. the risk of breast cancer depends on the value
the more general point that ‘causation is not merely an as- of the variable ‘sex’), whereas others may only affect the
pect of statistics; it is an addition to statistics, an enrich- risk of disease in certain contexts (e.g. in the context of
ment that allows statistics to uncover workings of the sexism or racism). Furthermore, all of these different types
world that traditional methods cannot’. Thus, we need to of causes can be represented in DAGs, and we can attempt
understand how and why causes influence their effects. to estimate their causal effects (with varying degrees of suc-
This is not only essential in deciding how to analyse the cess) while controlling for confounding and other sources
data in a particular study in a particular population. It is of bias. Of course, one may wish to identify subgroups of
also only by understanding how and why causes have their causes with particular characteristics (e.g. states, actions),
effects that we can also understand why causes may not which are more or less prone to various types of bias.
have the same effects in other contexts. Thus, generaliz- However, these represent differences between various
ability is a scientific process, not simply a matter of statis- types of causes; not between causes and ‘non-causes’.
tics (interestingly the book is titled Causal Inference in
Statistics, thus implying that causal inference can involve
statistics and vice versa, but they are not the same thing). Directed acyclic graphs (DAGs)
This emphasis on the context in which causes occur (‘the DAGs are increasingly used in epidemiology, but in our ex-
causal story behind the data set’ as Pearl et al. refer to it), perience they are not universally taught to epidemiologists.
contrasts with much frequentist theory in which Even among early and mid-career epidemiologists, there
International Journal of Epidemiology, 2017, Vol. 0, No. 0 3

appears to be a bimodal distribution of those who feel that 5. Pearl et al., like others, use the DAG concept of
all epidemiological research questions should be addressed ‘back door path(s)’ to define confounding. A back
using DAG(s) and those who seem to avoid them at all door path is a series of arrows that link the expos-
costs. We agree with others9 that DAGs are useful tools, ure with the outcome; back door paths have an
but are neither necessary nor sufficient for causal inference. arrow going into the exposure at one end, and an
Nevertheless, they can be an extremely valuable way of arrow going into the outcome at the other end of
illustrating the context (story) in which a causal question is the path. Some back door paths are shown in
being asked; in particular, they can illustrate the assump- Figure 1. To remove confounding, we want to
tions being made in causal analyses, and help us question block all back door paths.
their validity. For those less familiar with their use we pro- The meaning of arrows and drawing DAGs:
vide a brief description of their key features in Box 1.
Arrows are drawn between any two variables accord-
ing to the following criteria:
Using DAGs to decide what to adjust for and what not to
adjust for–confounding and collider bias
1. An arrow from one variable to a second indicates
Epidemiologists are very familiar with the concept of con-
that you assume that it is plausible that the first
founding; many lay people also understand this concept, as
variable causes the second.
‘to confound’ has a straightforward (non-technical) mean-
2. Where there is no arrow between one variable and
ing (‘to fool’) which describes the problem of assuming a second, this indicates that you assume that there
causality in the presence of uncontrolled ‘confounders’. is no causal relationship between the first and se-
When DAGs are drawn appropriately they can clarify our cond variable.
assumptions about confounders, and can point to situ-
ations where observed and unobserved confounders can be The absence of an arrow between two variables is

controlled for. For example, when a confounding path very important:

Indeed, if we think about confounding, the absence of

an arrow is as important as the presence of one. For
Box 1. Our summary of the basics of directed acyclic example, if we have an arrow from a variable to the
graphs when used in causal inference outcome of interest, but no arrow (or path made up of
Features of DAGs: a series of arrows) from that variable to the exposure,
then we are assuming that the variable is not a con-
1. Arrows (also known as ‘edges’ or ‘arcs’) connect founder. If in reality the variable is related to the ex-
‘nodes’ which represent variables. posure, then any observed association between expos-
2. Arrows between nodes are directed. That is, only ure and outcome might be biased as an estimate of
single-headed arrows can be included in a DAG. causal effect due to residual confounding.
3. Relationships are acyclic. That is, there are no ser-
ies of arrows connecting nodes (i.e. no ‘paths’)
that lead back to a node (variable) already in the
path. The assumption is that a variable (in a given
(back door path) includes unobserved variables that do not
population at a given time) cannot cause itself.
influence exposure through any other path, the path may
4. Ideally, every variable that influences two or more
be blocked by controlling for observed confounders
other variables is shown in the DAG. In particular,
(Figure 1), assuming that these are accurately measured
the focus should be on those variables that influence
the exposure and outcome. Though Pearl et al. in and appropriately adjusted for.
this book show situations where causal inference This primer also illustrates how conditioning (adjusting)
may be made without observing and adjusting for on some variables–’colliders’–may introduce bias. Unlike
all potential confounders (e.g. where a confounding the word ‘confounding’, ‘collider bias’ is not so intuitive
path can be blocked by conditioning on just one and has no corresponding ‘lay’ meaning (it makes sense
variable in the path) and even where none of the only with the use of DAGs). A collider is a node (represent-
key confounders is observed [by using definite ing a variable) that has two arrows coming into it on a
(known) causal mediators], these unobserved con- path. Where a collider occurs, that back door path is
founders need to be depicted in the graph (they are blocked (Figure 1); there is thus no need to adjust for the
an essential part of the story/context). collider as that path is already blocked. Importantly, ad-
justing on a collider opens up such a back door path, and
4 International Journal of Epidemiology, 2017, Vol. 0, No. 0

Addicve
personality Socio-economic posion (SEP)

Smoking Age at
pregnancy

Obesity Pre-eclampsia (PE)

Gestaonal age

Figure 1. Illustrative example–directed acyclic graph for the hypothesis that obesity is causally related to pre-eclampsia
Deciding what we should and should not adjust for on the basis of this DAG:

Scenario 1

Assume that current knowledge does not imply a plausible effect of addictive personality on smoking or obesity, or that there is a direct relationship of
SEP to PE risk, so these relationships (all shown with dashed arrows) are not included in the DAG in Scenario 1. We can make appropriate decisions
about what needs to be adjusted for and what should not be adjusted for to obtain a valid estimate of the causal effect of obesity on PE if we assume that
our DAG is correct [i.e. there are no other variables (nodes) or arrows that should be included] and that all variables are measured accurately (with little
or no misclassification). We want to adjust for confounding–i.e. we want to block all back door paths. In this scenario there are four unblocked.

Unblocked backdoor paths

• PE-Age at pregnancy-SEP-Smoking-Obesity;
• PE-Age at pregnancy-SEP-Obesity;
• PE-Age at pregnancy-Smoking-Obesity;
• PE-Smoking-Obesity.
Because age at pregnancy is in the first three paths, we can block all three of those by adjusting for age at pregnancy only: assuming our DAG is cor-
rect and pregnancy age is accurately measured and so adjusting on it can fully block those paths. The last path does not include age; to block that we
must control for smoking.

There is also one blocked path

• PE-Age at pregnancy -Smoking-SEP-Obesity; this is blocked because age and SEP collide on smoking.
However, we have said above that we have to adjust for smoking. When we do that, this path is unblocked and a spurious association between
Pregnancy age and SEP is generated. In this scenario we are going to adjust for pregnancy age, which will block this path even when we adjust for
smoking. To conclude, if we assume the DAG is correct and pregnancy age, obesity and PE are accurately measured (and there are no other sources
of bias), then adjusting for pregnancy age and smoking will provide a valid causal estimate.

Scenario 2

New research/knowledge provides evidence that: (a) Addictive personality is relevant to our causal understanding of obesity on PE and must be
added to the DAG as shown with dashed arrows (related to smoking and obesity) and (b) SEP is directly related to PE, also added to the DAG with a
dashed arrow. This introduces one new unblocked path (in addition to the ones above):

• PE-SEP-Smoking-Addictive personality-Obesity.

We do not have a measure of Addictive personality or SEP, but we can block this path by adjusting for smoking (assuming our DAG is correct and no
misclassification or other bias). We also still need to adjust for age and smoking to block the paths described above but now, when we adjust for
smoking, we unblock the following blocked back door paths:

• PE-SEP-Addictive personality-Obesity;

Because we generate a spurious association between SEP and Addictive personality, if we do not have a measure of either of these in our dataset,
the question is:

• Should we adjust for smoking to deal with confounding or should we not adjust for it because to do so would introduce collider bias?

The DAG cannot answer that–the answer lies in background knowledge and/or simulation studies that provide evidence for whether bias would be
greatest with or without adjustment for smoking.

Should we adjust for gestational age in either scenario?

Very often in perinatal epidemiology gestational age is conditioned on–frequently this is done by excluding women who do not have a term delivery
(i.e. where the baby is born before 37 weeks of completed gestation) either in the study design or analyses. In any analyses where exposure and out-
come influence gestational age (as in this example, and commonly for many questions in this field), we should not do this. To do so potentially intro-
duces a spurious association between Obesity and PE. In this specific case, that spurious association would be inverse and so this ‘collider’ bias
could produce an effect estimate that is weaker than any true positive effect (biased towards the null). Note that whereas SEP would rarely be a plaus-
ible cause of ‘Age’ in this example, it is plausible to assume that SEP influences the age at which women start their family and hence become preg-
nant, with young women more likely to be from lower SEP and older from higher SEP.19
International Journal of Epidemiology, 2017, Vol. 0, No. 0 5

thereby produces a spurious association between the two is influenced by both of them since obese women are likely
variables (e.g. exposure and disease) that it ‘connects’. to have shorter duration pregnancies and those with PE are
Pearl et al. explain collider bias by using a theme that runs more likely to have their pregnancy induced or ended early
throughout the book, in which they define conditioning (or by caesarean section). The importance of recognizing this
adjusting) as ‘filtering’ by the value(s) of the conditioning is that many studies in perinatal epidemiology do restrict
variable. In a very clear and simple way they point out that to term pregnancies only (either through excluding women
if Z is a collider for X and Y (i.e. the variable Z is influ- who deliver preterm from being in the study or from being
enced by X and Y; written in the book as Z ¼ X þ Y), and in analyses), without considering whether this might intro-
X and Y are independent of each other, and no other vari- duce bias.
ables influence X, then conditioning on Z is the same as fil-
tering on participants with the same value of Z. To take Front door paths and the possibility of not having to meas-
Pearl et al.’s simple additive example, if we know (only) ure confounders. In section 3.4, Pearl et al. suggest that an
that X ¼ 3 for any participants that tells us nothing about unconfounded causal effect can be estimated using obser-
the value of Y for those participants. But if we also condi- vational data, even when there are back door paths that
tion (filter) on Z (as well as knowing that X ¼ 3) within cannot be blocked (because of unmeasured confounders).
each stratum of Z, we now know the value of Y (if Z ¼ 10, This is done using a front door path. A front door path is
Y must ¼ 7; if Z ¼ 5, Y must ¼ 2; if Z ¼ 1, Y must ¼ -2. . . where there is one (or more) mediator(s) between the ex-
and so on); by conditioning on (adjusting for) Z we have posure and outcome and where there are no confounders
generated a spurious association between X and Y. of the exposure-mediator or mediator-outcome (Figure 2).
This fits with Simpson’s Paradox as illustrated in The concept is that if there are unmeasured confounders
Chapter 1 of the book. Gender in the first example in between X (exposure) and Y (outcome) but no confound-
Chapter 1 is a confounder and should be adjusted for, ers between X and a mediator (M) or between M and Y,
whereas post-treatment BP (the second example in Chapter then the (unadjusted) associations of X and M and M and
1) is a collider (influenced both by the drug and by recov- Y can provide the causal effect of X on Y. It feels like
ery from the (unspecified) illness that the participants were alchemy!
suffering from) and should not be adjusted for. The example that Pearl et al. use to demonstrate this
In reality, few researchers would adjust for post- refers to an old argument that smoking does not cause lung
treatment BP in a study exploring the effect of a drug on an cancer but rather that there are genes which influence both
unspecified illness. Therefore, to illustrate collider bias fur- smoking and (independently) lung cancer risk, and thus
ther we use a more plausible example in Figure 1. This confound the association of smoking with lung cancer.
shows a DAG that might be drawn and used to inform They present a thought experiment in which ‘tar deposits
what we should (and should not) adjust for to explore the in the lung’ are a mediator between smoking and lung can-
causal effect of obesity on pre-eclampsia (PE) risk. The cer, and show using a DAG (Figure 2), that an uncon-
DAG shows our assumptions that: socioeconomic position founded causal effect can be estimated despite having no
(SEP) is at least plausibly causally related to obesity, smok- measure of the genetic confounder. If the DAG presented
ing and age (at pregnancy), but not (directly) to pre- by Pearl et al. is correct, we agree that using this front door
eclampsia, in scenario 1 that smoking is related to obesity approach could provide a valid causal effect estimate.
and PE; that age is related to smoking, obesity and PE; and However, the example is fictional, and we struggle to im-
that both obesity and PE are related to gestational age at agine any situation in which there are not confounders be-
birth of the infant. These assumptions are based, to some tween an exposure and a mediator, or mediator and
extent, on research findings,18–20 but the DAG is also sim- outcome or misclassification of the mediator that is corre-
plified for illustrative purposes and does not show all lated with misclassification of the exposure.21–23 For us
plausible influences on all variables represented in the this front door approach is theoretically interesting but not
DAG (see later discussion on limitation of DAGs). This likely to be widely applicable.
DAG suggests that we can adjust solely for age at preg- Mendelian randomization (MR), using genetic variants
nancy and smoking to prevent confounding (including by in genes that encode the nicotinic acetylcholine receptor as
SEP; see Figure 1). Thus, if we did not have a measure of instrumental variables (IV), suggests a causal effect of
SEP in our dataset, assuming that all other variables are ac- greater intensity of smoking on lung cancer (Figure
curately measured and the DAG is correct, we can obtain 2b).24,25 However, instrumental variable analyses (which
an estimate of the causal effect of obesity on PE risk. By Pearl et al. mention only in passing) have very different
contrast, we should not adjust for gestational age at birth DAGs from that shown in Figure 2a, and a different set of
as this is a collider on the path between PE and obesity (it assumptions (Figure 2b) from the more conventional
6 International Journal of Epidemiology, 2017, Vol. 0, No. 0

A U: Genotype B U: Genotype

X: Smoking M: Tar deposit Y: Lung cancer Z: Genotype X: Smoking Y: Lung cancer

This DAG is adapted from Pearl et al. (page 66)1 and used by them to This DAG shows an example of a genec instrumental
illustrate how a frontdoor path can be used to test unconfounded variable analysis (Mendelian randomizaon).27-31
causal effects. Z: genec instrumental variable; U: unmeasured
U: unmeasured confounders; X: exposure; M: mediator; Y: outcome. confounders; X: exposure; Y: outcome.
Causal queson: does smoking cause lung cancer? Causal queson: does smoking cause lung cancer?
Pearle et al. suggest the causal effect of exposure (X) on outcome (Y) An unconfounded causal effect esmate can be
can be esmated using the (unadjusted) associaons of X- mediator esmated if the assumpons shown in the DAG are
(M) and of M-Y. This is true if M (tar deposits in the lung) can be correct:
measured and the assumpons shown in the DAG are correct: • Z is robustly related to X.
• There are no confounders between X and M. • Confounders of X-Y are unrelated to Z.
• There are no confounders between M and Y. • There is no path from Z to Y other than through X.
• There is no path from X to Y other than through M.

Figure 2. Pearl et al.’s front door and Mendelian randomization methods for testing unconfounded causal effects. A. Example of a front-door path to
test unconfounded causal effects. B. Example of a Mendelian randomization approach to test unconfounded causal effects.

approaches used in most of this book. These assumptions characteristic A at time one (At1) influences characteristic B
bring their own potential sources of bias. However, genetic at a later time (Bt1þx) which then goes on to influence char-
variants are often valid IVs, and recent developments that acteristic A at a subsequent time [At1þy (where y is > x)], and
provide valuable sensitivity analyses of the potential viola- so on, these relationships can be represented in a DAG with
tion of the IV assumptions when using MR, mean that MR no violation of its directed and acyclic properties. The DAG
provides the potential for better causal inference in obser- depicting these relationships treats characteristics at differ-
vational studies.26–30 ent time points as distinct nodes. However, causal processes
cannot always be defined as directed and acyclical. This ‘lin-
Limitations of DAGs ear’ approach to causality contrasts with complexity re-
We often find DAGs are useful for being explicit about as- search involving non-linearity and feedback loops which
sumptions of the causal context and helping researchers to cannot be readily summarized in a DAG.31
better determine what should and should not be adjusted DAGs are also non-parametric, i.e. they make no as-
for. However, their limitations should also be considered. sumptions about the nature or form of the causal relation-
Clearly, they can only ever be as good as the context ships they depict, or even the direction (causative or
(background information) that is used to draw them. For ex- preventive) of potential effects. Statistical interaction or ef-
ample, if they are drawn solely on the basis of available data fect modification can also be difficult to depict, although
rather than showing all key variables whether observed or some methods have been proposed for doing this.3,32
unobserved, then causal effect estimates may be (residually) Perhaps the largest limitation of DAGs is that they can
confounded. Perhaps more importantly, their use to guide be used to indicate possible sources of bias but cannot eas-
analyses also depends strongly on the accuracy of the avail- ily indicate how likely or how strong the biases may be. In
able data. This is true of all epidemiology, but may be par- one recent example relating to Berkson’s Bias,4, 5 DAGs
ticularly true when DAGs are used to imply that ‘causal’ were extremely powerful in helping to identify the nature
analyses are straightforward and can determine complex of the bias, but not its strength. Berkson’s Bias produces
causal paths, such as mediation with multivariable extremely biased results when a study involves prevalent
approaches applied to observational data.21–23 cases, a situation which cannot be easily represented by
By their very nature DAGs assume that relationships are DAGs. If a study involves incident cases, the DAG remains
directed and acyclical. This will be true for many common the same, but (in this particular case) the bias becomes triv-
biological and epidemiological processes, but there are also ial.4 In our experience, creative colleagues can use DAGs
many exceptions in which truly cyclical or bidirectional rela- to identify possible ‘collider bias’ in virtually any analysis,
tionships exist. It may be possible to resolve this with tem- but this tells us little about whether the bias is likely to be
poral knowledge. For example, if it is plausible that large enough to be of practical importance.
International Journal of Epidemiology, 2017, Vol. 0, No. 0 7

Related to this, in some situations the distinction be- we could go more ‘distal’, to add potential causes of the
tween what to adjust for and what not to adjust for is not proximal common causes of exposure and outcome [i.e. dis-
simple even with a well-drawn DAG (Figure 1). For ex- tal ancestors of the main exposure (obesity) and outcome
ample, let us assume that following well-conducted re- (PE)]. Where or when to stop is not clear. Software such as
search, it is clear that addictive personality is related to DAGitty and the suite of DAG functions in R (dagR) can
both smoking and obesity and therefore should be added deal with the most complex of DAGs and provide investiga-
to the DAG in Figure 1. Furthermore, new evidence sug- tors with a minimum set of variables that should allow them
gests it is plausible that SEP influences preeclampsia risk to deal with potential confounding without resulting in col-
through mechanisms that do not involve either maternal lider bias. However, some studies using these packages fail
age at pregnancy or her smoking. This also needs adding to to appropriately take account of theoretical context, but ra-
the DAG. However, we do not have data on either addic- ther control for a large number of variables without clear
tive personality or SEP; now our conclusions about what reasoning and assume that this produces valid causal esti-
we should and should not adjust for are more complex. mates from purely observational data.34
Above, before consideration of this new knowledge, we
noted that we need only adjust for age at pregnancy and
smoking. However, with the addition of this new knowl-
Integrating diverse types of knowledge to answer
edge, smoking is now a collider on the back door path PE-
causal questions
SEP-addictive personality-obesity and if we adjust for it we The use of methods such as triangulation, in which the aim
open that back door path (by generating a spurious associ- is to integrate evidence from several approaches, that are
ation between addictive personality and SEP). (see chosen because they are sufficiently different to be likely to
Scenario 2; Figure 1). The question of whether the correct have different and unrelated key sources of bias that would
(or best) causal estimate is with or without adjustment for be unlikely to produce the same result (due to these
smoking cannot be answered from the DAG; though we biases),35 may also be particularly important and even cru-
would suggest that adjusting for it, given its proximal rela- cial, along with evidence from time trends and ecological
tionships to obesity and pre-eclampsia, is likely to be most studies. Going back to Pearl et al.’s front door example dis-
important.33 In situations like this, the relatively new con- cussed above, evidence that smoking was a causal factor
cept of collider bias can lead to a tendency to not adjust for for lung cancer (rather than being confounded by genes or
a variable if there is a possibility of collider bias (‘collider other factors) came several decades ago from such an inte-
anxiety’4), even if the collider bias is likely to be very weak grative approach (including time trends in lung cancer inci-
whereas the uncontrolled confounding may be relatively dence and mortality),1,36 rather than a theoretically correct
strong. Greenland described this situation in a seminal but unrealistic DAG.
paper in 2003.33 Although it will depend on the relative Thus, in epidemiology, the assessment of whether some-
strengths of all associations between confounders and col- thing is a cause is usually addressed through a process of
lider with exposure and outcome, in most situations more integrating diverse types of knowledge, even if this is rarely
proximal confounding will be more important to control acknowledged.37 Even when a particular study appears to
for. Greenland usefully provides suggestions for how one be decisive, there are always assumptions, theories and
might undertake sensitivity analyses to test this, though contextual background information–from previous add-
they require appropriate contextual information to add itional studies–that are necessary for a definitive judgement
value.33 to be made.38 Thus, every process of causal identification
These limitations highlight a general issue that the DAGs and explanation involves evidence of a variety of types and
used throughout this book, as in the many methodological from a variety of sources, and no single study is definitive.
papers that advocate their use, are extremely simple (in This is partly due to the Duhem/Quine’s thesis’ that a the-
order to illustrate specific methodological issues) and rarely ory always relies on (but does not explicitly use) auxiliary
reflect the reality of the numerous auxiliary hypotheses hypotheses, and if some consequences of the theory turn
related to the main causal question (see below for more dis- out to be false, one of the auxiliary hypotheses rather than
cussion). The DAG we show in Figure 1 is more complex the theory may be incorrect.39 The fact that leaves may be
than many in the primer, but it is a simple representation of observed to fly upwards in the wind does not necessarily
the relationships that those of us working clinically and/or refute the law of gravity but may instead refute auxiliary
academically in this area know are relevant. A, by no means hypotheses (e.g. that there are no other forces operating
exhaustive, list of variables that ought also to be added to that are stronger than gravity). Similarly, every epidemio-
the DAG includes parity, change of partner, multiple preg- logical study involves the auxiliary hypothesis that no un-
nancy, placental function and fetal growth. For each of these controlled bias is occurring, and it may be this auxiliary
8 International Journal of Epidemiology, 2017, Vol. 0, No. 0

hypothesis that is falsified rather than the main hypothesis This book thus represents a major resource for epidemi-
of interest. As Pearl et al. point out in Chapter 3, even in a ologists to learn the use of methods (e.g. structural causal
randomized controlled trial, a valid test of a theory (inter- models and DAGs) which have had major effects on the
vention) can only be obtained if a number of auxiliary con- theory and practice of epidemiology in recent years. Our
ditions are met (full and/or unbiased participation, lack of own experience in teaching is that these methods are ex-
misclassification, lack of contamination of the comparison tremely useful and would benefit from being introduced at
group, etc.), and even a ‘perfect’ trial (which almost never an early stage of introductory epidemiology courses, pro-
exists) is intended (by design) to produce false-positive re- vided that they are used ‘in context’ (i.e. studying the dis-
sults 5% of the time (noting that most RCTs are designed tribution and determinants of health in populations) rather
to have sufficient power to detect a clinical/public health than as a set of generic methods. They are not particularly
meaningful difference at the conventional 5% level of sig- difficult except to those who have been trained using dif-
nificance). Thus, interpretation of even the best possible ferent concepts and methods. If they are used (carefully
trials always involves auxiliary information. These issues and appropriately) from the beginning, then new students
are considerably more acute in observational studies, but can grasp these concepts relatively easily–just as a teenager
they are not unique to epidemiology. This is how most sci- can usually use a modern cellphone easily whereas older
ence works.39 generations may struggle. However, the limitations of
Although any individual study can usually be represented these methods should also be considered in this teaching,
in terms of counterfactual contrasts, which can in turn be and they should always be used as part of the epidemiolo-
represented in DAGs, it is difficult if not impossible to repre- gical toolkit to address real-world problems (problem-
sent the overall process of epidemiological discovery and based epidemiology41–43) rather than being used ‘out of
causal inference using these methods. Even if the available context’ as a set of generic methods.
evidence is assessed at one particular point in time, the task
of combining a wide variety of evidence from a wide variety Acknowledgements
of sources continues to be a matter of judgement,10 albeit Dr Laura Howe and Prof. Kate Tilling (both of the University of
one that can be aided by particular considerations such as Bristol) made valuable comments on earlier drafts of the section on dir-
ected acyclic graphs. Prof. Jan Vandenbroucke (University of Leiden)
those of Hill.37 None of this activity–the real ‘causal infer-
provided useful comments on an earlier draft of the paper. Brice LB
ence’–can be captured adequately in methods which focus
Kuimi (MSc), doctoral student at McGill University Canada, noticed
on causal inference in a single study with a single DAG. an error in our DAG shown in Figure 1 and our discussion of it; we are
Some of the commentaries in this issue suggest that DAGs grateful for his eagle-eyed smartness that we were able to correct that
do take account of all such relevant knowledge,11 but error. The views expressed in this paper are those of the authors and
Krieger and Davey Smith challenge this.34 not necessarily of any funding body or people acknowledged.

Concluding remarks Funding

The Centre for Global NCDs is supported by the Wellcome Trust
Pearl et al. note in their preface that over the past decade Institutional Strategic Support Fund (097834/Z/11/B). The MRC
the methods covered in this primer have resulted in a Integrative Epidemiology Unit is supported by the University of
‘transformative shift of focus in statistics research, accom- Bristol and UK Medical Research Council (MC_UU_1201/5). D.A.L.
panied by unprecedented excitement about the new prob- is a National Institute of Health Research Senior Investigator (NF-SI-
lems and challenges. . .’. This has been accompanied by a 0166-10196). The research leading to these results has received fund-
ing from the European Research Council under the European Union’s
number of excellent textbooks that develop Pearl’s work
Seventh Framework Programme (FP7/2007-2013) / ERC grant agree-
further (e.g. references 2 and 40). One of us (see references ment no 668954 and ERC grant agreement no 669545.
1, 10 and 23) has been highly critical of the naive use of
Conflict of interest: None declared.
these methods and of the accompanying claims that they
form a complete and sufficient theory of causal inference,
References
rather than merely a useful set of tools which are appropri-
ate in some situations but not others.9 However, we recog- 1. Vandenbroucke J, Broadbent A, Pearce N. Causality and causal
inference in epidemiology–the need for a pluralistic approach.
nize the value and power of these methods when used
Int J Epidemiol 2016; in press;doi:10.1093/ije/dyv341.
appropriately and cautiously, together with other
2. Hernan MA, Robins JM. Causal inference. [Link]
approaches such as triangulation.35 The problem is how to [Link]/miguel-hernan/causal-inference-book/ (8 November 2016,
use these new methods critically and appropriately, rather date last accessed).
than being captured by them in a manner which redefines 3. VanderWeele T. Explanation in Causal Inference. New York,
and restricts what epidemiology is.1 NY: Oxford University Press, 2015.
International Journal of Epidemiology, 2017, Vol. 0, No. 0 9

4. Pearce N, Richiardi L. Commentary:Three worlds collide: objective measures of tobacco exposure. J Natl Cancer Inst
Berkson’s Bias, selection bias and collider bias. Int J Epidemiol 2012;104:740–48.
2014;43:521–24. 25. Taylor AE, Davies NM, Ware JJ, VanderWeele T, Davey Smith
5. Snoep JD, Morabia A, Hernandez-Diaz S, Hernan MA, G, Munafo MR. Mendelian randomization in health research:
Vandenbroucke JP. Commentary: A structural approach to Using appropriate genetic variants and avoiding biased esti-
Berkson’s fallacy and a guide to a history of opinions about it. mates. Econ Hum Biol 2014;13:99–106.
Int J Epidemiol 2014;43:515–21. 26. Davey Smith G, Ebrahim S. ‘Mendelian randomization’: can
6. Pearl J. Causality: Models, Reasoning and Inference. 1st edn. genetic epidemiology contribute to understanding environmental
Cambridge, UK: Cambridge University Press, 2000. determinants of disease? Int J Epidemiol 2003;32:1–22.
7. Pearl J. Causality: Models, Reasoning and Inference. 2nd edn. 27. Davey Smith G, Ebrahim S. Mendelian randomization: pro-
Cambridge, UK: Cambridge University Press, 2009. spects, potentials, and limitations. Int J Epidemiol
8. Pearl J, Glymour M, Jewell N. Causal Inference in Statistics: A 2004;33:30–42.
Primer. Chichester, UK: Wiley, 2016. 28. Davey Smith G, Lawlor DA, Harbord R, Timpson N, Day I,
9. Blakely T, Lynch J, Bentley R. DAGs and the restricted potential Ebrahim S. Clustered environments and randomized genes: A
outcomes approach are tools, not theories of causation. fundamental distinction between conventional and genetic epi-
Int J Epidemiol 2016; in press. demiology. Plos Med 2007;4:1985–92.
10. Broadbent A, Vandenbroucke JP, Pearce N. Formalism or plural- 29. Lawlor DA. Two-sample Mendelian randomization: opportuni-
ism? A reply to commentaries on Causality and Causal Inference ties and challenges. Int J Epidemiol 2016;45:908–15.
in Epidemiology. Int J Epidemiol 2016; in press. 30. Lawlor DA, Harbord RM, Sterne JAC, Timpson N, Davey Smith
11. Daniel R, De Stavola B, Vandsteelandt S. The formal approach G. Mendelian randomization: Using genes as instruments for
to quantitative causal inference in epidemiology: misguided or making causal inferences in epidemiology. Stat Med
misrepresented? Int J Epidemiol 2016; in press. 2008;27:1133–63.
12. Krieger N, Davey Smith G. The tale wagged by the DAG: 31. Pearce N, Merletti F. Complexity, simplicity and epidemiology.
broadening the scope of causal inference and explanation for epi- Int J Epidemiol 2006;35:515–19.
demiology. Int J Epidemiol 2016; in press. 32. VanderWeele T. On the distinction between interaction and ef-
13. Robins JM, Weissman MB. Counterfactual causation and street- fect modification. Epidemiology 2009;20:863–71.
lamps: what is to be done? Int J Epidemiol 2016; in press. 33. Greenland S. Quantifying biases in causal models: Classical con-
14. VanderWeele TJ. Explanation in causal inference: developments founding vs collider-stratification bias. Epidemiology
in mediation and interaction. Int J Epidemiol 2016; in press. 2003;14:300–06.
15. VanderWeele TJ. On causes, causal inference and potential out- 34. Krieger N, Davey Smith G. FACEing reality: productive tensions
comes. Int J Epidemiol 2016; :in press. between our epidemiological questions, methods, and mission.
16. Broadbent A. Philosophy of Epidemiology. Basingstoke, UK: Int J Epidemiol 2016; submitted for publication.
Palgrave Macmillan, 2013. 35. Lawlor DA, Tilling K, Davey Smith G. Triangulation in aetiolo-
17. Lilienfeld AM. Epidemiological methods and inferences in studies gical epidemiology. Int J Epidemiol 2016; in press.
of noninfectious diseases. Public Health Reps 1957;72:51–60. 36. Cornfield J, Haenszel W, Hammond EC, Lilienfeld AM, Shimkin
18. Freathy RM, Kazeem GR, Morris RW et al. Genetic variation at MB, Wynder EL. Smoking and lung cancer: recent evidence and
CHRNA5-CHRNA3-CHRNB4 interacts with smoking status to a discussion of some questions. J Natl Cancer Inst
influence body mass index. Int J Epidemiol 2011;40:1617–28. 1959;22:173–203.
19. Lawlor DA, Mortensen L, Andersen AMN. Mechanisms under- 37. Hill AB. The environment and disease:association or causation?
lying the associations of maternal age with adverse perinatal out- Proc R Soc Med 1965;58:295–300.
comes:a sibling study of 264 695 Danish women and their 38. Cartwright N. Hunting causes and using them: approaches in
firstborn offspring. Int J Epidemiol 2011;40:1205–14. philosophy and economics. Cambridge, UK: Cambridge
20. Macdonald-Wallis C, Tilling K, Fraser A, Nelson SM, Lawlor University Press, 2007.
DA. Established pre-eclampsia risk factors are related to patterns 39. Pearce N, Crawford-Brown D. Critical discussion in epidemi-
of blood pressure change in normal term pregnancy: findings ology: problems with the Popperian approach. J Clin Epidemiol
from the Avon Longitudinal Study of Parents and Children. 1989;42:177–84.
J Hypertens 2011;29:1703–11. 40. VanderWeele TJ, Hernan M. Causal effects and natural laws: to-
21. Blakely T. Commentary: Estimating direct and indirect effects– wards a conceptualization of causal counterfactuals for nonma-
fallible in theory, but in the real world? Int J EpidemiolI nipulable exposures, with applications to the effects of race and
2002;31:166–67. sex. In: Berzuini C, Dawid P, Bernardinell C (eds). Causality:
22. Blakely T, McKenzie S, Carter K. Misclassification of the medi- Statistical Perspectives and Applications. Hoboken, NJ: Wiley,
ator matters when estimating indirect effects. J Epidemiol 2012.
Community Health 2013;67:458–66. 41. Galea S. An argument for a consequentialist epidemiology. Am J
23. Pearce N, Vandenbroucke JP. Causation, mediation and explan- Epidemiol 2013;178:1185–91.
ation: Essay review of Explanation in Causal Inference. Int 42. Pearce N. Traditional epidemiology, modern epidemiology, and
J Epidemiol 2016; in press. public health [comment]. Am J Public Health 1996;86:678–83.
24. Munafo MR, Timofeeva MN, Morris RW et al. Association be- 43. Pearce N. Epidemiology as a population science. Int J Epidemiol
tween genetic variants on chromosome 15q25 locus and 1999;28:S1015–18.