CONCEPTS & SYNTHESIS

EMPHASIZING NEW IDEAS TO STIMULATE RESEARCH IN ECOLOGY

Ecology, 101(4), 2020, e02962
Published 2019. This article is a U.S. Government work and is in the public domain in the USA.

Scientist’s guide to developing explanatory statistical models using

causal analysis principles
1,3 2
JAMES B. GRACE AND KATHRYN M. IRVINE
1
Wetland and Aquatic Research Center, U.S. Geological Survey, 700 Cajundome Boulevard, Lafayette, Louisiana 70506 USA
2
Northern Rocky Mountain Science Center, U.S. Geological Survey, 2327 University Way Suite 2, Bozeman, Montana 59715 USA

Citation: Grace, J. B., and K. M. Irvine. 2020. Scientist’s guide to developing explanatory statistical models
using causal analysis principles. Ecology 101(4):e02962. 10.1002/ecy.2962

Abstract. Recent discussions of model selection and multimodel inference highlight a gen-
eral challenge for researchers: how to convey the explanatory content of a hypothesized model
or set of competing models clearly. The advice from statisticians for scientists employing multi-
model inference is to develop a well-thought-out set of candidate models for comparison,
though precise instructions for how to do that are typically not given. A coherent body of
knowledge, which falls under the general term causal analysis, now exists for examining the
explanatory scientific content of candidate models. Much of the literature on causal analysis
has been recently developed, and we suspect may not be familiar to many ecologists. This body
of knowledge comprises a set of graphical tools and axiomatic principles to support scientists
in their endeavors to create “well-formed hypotheses,” as statisticians are asking them to do.
Causal analysis is complementary to methods such as structural equation modeling, which
provides the means for evaluation of proposed hypotheses against data. In this paper, we sum-
marize and illustrate a set of principles that can guide scientists in their quest to develop
explanatory hypotheses for evaluation. The principles presented in this paper have the capacity
to close the communication gap between statisticians, who urge scientists to develop well-
thought-out coherent models, and scientists, who would like some practical advice for exactly
how to do that.
Key words: causal analysis; causal diagrams; explanatory models; multimodel averaging; multimodel
comparison; path analysis; regression; science methodology; structural equation modeling.

using regression models. Burnham and Anderson (2002)

INTRODUCTION
highlighted this pervasive issue in their classic text on
Building regression models using a set of candidate multimodel inference (pp. 440–441), where they criticize
explanatory variables is a common practice in ecology. ecologists for not doing enough “critical thinking” when
In recent years, Burnham and Anderson (2002) have crafting candidate models for evaluation. In the absence
championed an approach for choosing amongst alterna- of a carefully crafted set of competing models, model
tive models using Akaike information (AIC) and multi- selection cannot lead to scientifically interpretable find-
model inference. As a result, the use of information ings. The proposed remedy, given repeatedly in the Burn-
criteria and multimodel comparisons has become com- ham and Anderson writings, is simply for the scientist to
mon practice in many disciplines, particularly in the “think hard” about the candidate models. Scientists
environmental sciences (Symonds and Moussalli 2011). might wonder, however, “How am I supposed to ‘think
A careful reading of the literature on model comparison hard’ about my candidate model set? Where is the chap-
and selection reveals persistent mention by statisticians ter in Burnham and Anderson or other statistics books
of fundamental difficulties with ascribing explanatory that present the fundamental rules for thinking hard
interpretations to the coefficient estimates obtained about candidate models in a set so that scientific inter-
pretability is supported?” The unambiguous message
being conveyed by statisticians is that it is the responsi-
Manuscript received 29 August 2019; revised 29 October bility of the scientist to bring to the analysis a set of can-
2019; accepted 26 November 2019. Corresponding Editor:
Brian D. Inouye. didate models with clearly distinguishable scientific
3
E-mail: [email protected] content. The reason for this is that the necessary expert

Article e02962; page 1

 Article e02962; page 2 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

knowledge to permit interpretation lies with the scien- some unmeasured common cause? Perhaps X1 has a cau-
tist, not the statistician. This raises the question of where sal influence on X2? Perhaps the other way around? As
scientists can find a set of general rules for developing shown in Fig. 1B, there are many possible reasons for
well-reasoned explanatory hypotheses. three predictors to be correlated, each implying a differ-
We propose that the answer lies in the logical system ent scientific explanation for the nature of the relation-
referred to as causal analysis (Pearl 2000, 2009, Hern
an ships between predictors and response. Because
and Robins 2020). Recent advances in the field of causal scientific knowledge about possible explanations for the
analysis provide the core principles for how to evaluate intercorrelations cannot be encoded in the multiple
the causal content of competing hypotheses. Some regression equation because of its simplicity, the hypoth-
aspects of modern causal analysis will be familiar to esis represented by a multiple linear regression model
ecologists through the related literature on structural (Fig. 1A) is incomplete regarding potential mechanisms
equation modeling (SEM; Shipley 2000a, 2016, Grace underlying observed correlations, and thereby, interpre-
2006), whereas other elements are relatively new and tationally ambiguous.
ecological examples are scarce. In this paper, we first Compounding the problem of interpreting regression
illustrate the challenges that accompany attempts at coefficients is the fact that the correlations among pre-
drawing scientific inferences from conventional regres- dictors have profound effects on the estimated regression
CONCEPTS & SYNTHESIS

sion models. We then describe and illustrate a set of coefficients. To show how this is manifested in regression
principles to guide scientists when developing explana- studies, let us turn to a specific example. In the fall of
tory hypotheses for evaluation. 2003, Keeley et al. (2005) established 1,000-m2 plots in
each of 90 sites to investigate the ecological effects of a
series of wildfires that burned through southern Califor-
THE PROBLEM WITH DRAWING SCIENTIFIC INTERPRETATIONS
nia. The investigators measured a set of variables that
FROM MULTIPLE REGRESSION MODELS
might explain variations in postfire vegetation recovery
The challenge to drawing causal conclusions by throughout the region. Measurements included: land-
employing conventional regression methods is tied, in scape position variables (distance from the coast and ele-
part, to the limited capacity of such models to represent vation), prefire stand age (estimated from ring counts of
complete hypotheses. This problem can be made more stems), fire severity (based on skeletal remains of
tangible through graphical representation (Fig. 1). As shrubs), and vegetation cover the spring following the
revealed in Fig. 1A, the predictors in a multiple regres- fires. The intercorrelations among variables in the wild-
sion model can be (and typically are) intercorrelated. fire study (Table 1) were only moderately strong; yet,
However, no causal explanations for the correlations intercorrelation still presents a substantial problem for
among predictors are specified as part of the regression interpreting the contributions of individual factors, as
model. Why are X1 and X2 correlated? Is it because of we will show.

FIG. 1. (A) Graphical depiction of a regression model. Observed variables are in boxes and e represents errors of prediction.
Double-headed arrows represent intercorrelations and regression relationships by unidirectional arrows. (B) Six potential causal
hypotheses for the three correlated predictors (X1–X3). U variables represent unmeasured causes.
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 3

TABLE 1. Bivariate correlations among variables used in the TABLE 2. All-subsets regression models for vegetation recovery
wildfire example. following wildfire.

vegcover firesev age elev coastdist Explanatory variables included in the model
vegcover 1 Model Fire Preburn age Distance
firesev
0.437 1 no. severity of stand Elevation from coast
age
0.350 0.454 1 1
elev 0.218
0.117 0.093 1 2
0.084
coastdist 0.243
0.278
0.278 0.606 1 3
0.088
4 0.027
Using conventional regression, all-subsets model com- 5 0.875
parison might be employed (Fisher et al. 2018). In the 6
0.067
0.048
7
0.080 0.021
case of four explanatory variables, there are 16 possible
8
0.077 0.475
regression models that could be specified (Table 2). Can-
9
0.094 0.031
didate models were estimated and ranked (Table 3;
10
0.077 0.568
Appendix S1, Data S1), and ultimately a single model
11 0.014 0.634

CONCEPTS & SYNTHESIS

selected for scientific interpretation (model 12 in this 12
0.059
0.058 0.025
case). The reader can observe in Table 2 that the unstan- 13
0.063
0.043 0.373
dardized partial regression coefficients obtained from 14
0.078 0.018 0.155
different models exhibit substantial variations in magni- 15
0.097 0.033
0.093
tude depending on the other predictors included. In the 16
0.060
0.062 0.029
0.201
case of distance from the coast, for example, values actu-
ally shift from positive (0.875 in model 5) to negative Notes: Numbers in cells are raw coefficients relating predic-
tors to vegetation cover in the various models. Coefficients for a
(
0.201 in model 16). The mathematical reasons for predictor can vary among models, depending on the other vari-
these shifts are well understood (Cohen et al. 2003). ables included (e.g., distance from coast).
Each regression coefficient is adjusted (also known as
partialed) based on all of the other predictors in the methods for evaluating those hypotheses against data.
model. The fundamental mathematical rule is that the Before diving into those key parts of our presentation,
bivariate associations between Xs and Y are taken to be we must first provide some essential background for
fixed values regardless of the other predictors in a understanding causal effects.
model. When predictors are strongly correlated, adding
or subtracting a predictor from a regression model can
OPERATIONAL DEFINITION OF A CAUSE–EFFECT RELATION-
lead to large changes in the regression coefficients for
SHIP—COUNTERFACTUALS AND POTENTIAL OUTCOMES
the included predictors, including reversals of sign. In
regression models, such as Fig. 1A, understanding the One approach to thinking about cause–effect rests on
scientific implications of such shifts in coefficients is the concept of counterfactuals. The counterfactual ques-
extremely difficult and there is a long history of failed tion is, “What would have happened if. . . ?” Behind this
efforts (reviewed by Elwert and Winship 2014, Pearl and question is an implied manipulation of the sort conceived
MacKenzie 2018). For the interested reader, we provide of in an idealized experiment, that is, “What would have
an illustration of exactly why coefficients change when been the fate of an individual or sample if they had been
variables are added to a model in Appendix S2, Data S2. in the control group instead of the treatment group?” This
Although the problem with interpreting coefficients conceptualization of the problem allows causal effects to
from regression models has long been recognized, it was be defined using either experimental or observational
recently brought to the fore by a series of papers dis- data, which is a major part of its appeal.
cussing problems with the practice of model averaging Consider an ecological experiment investigating the
(Cade 2015, Fieberg and Johnson 2015, Banner and effects of pesticides on frogs. Our concept of a causal
Higgs 2017). Among those papers, Cade (2015) provided effect assumes each individual frog has two potential out-
the bluntest assessment when he said that averaging comes, the individual’s potential response after exposure
coefficients over models with different sets of predictors to pesticide and their potential response to nonexposure
produces “muddled” inference because the scientific (Neyman 1923, Rubin 1974). Only one of these will be
meaning of coefficients shifts whenever a predictor is observed. Let us denote the treatment variable as Tmt
added or removed. In this paper, we seek to clarify, from (e.g., exposure to pesticide or not). The question the
the perspective of the scientist, why regression models experiment is attempting to answer is “What would have
are difficult to use for explanatory modeling and what happened to individual i (or individual sample unit i),
can be done to remedy that situation. To accomplish which did not receive the treatment (i.e., was assigned to
this, we provide investigators with a work flow for the control group, Tmt = 0), if that individual had been
explanatory modeling, a set of principles for developing assigned to the pesticide exposure group (Tmt = 1).”
causal hypotheses, and references to statistical modeling The causal effect of the treatment is estimated by the
 Article e02962; page 4 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

difference between the two outcomes (treated or not

EXPLANATORY MODELING IN ACTION
treated). In this two-treatment case, each individual sub-
ject has two potential outcomes: In this paper, we make a fundamental distinction
between the hypothesis and the inferential model. The
• Yi1 : potential outcome if individual i receives Treat- structure of the hypothesis and its logical requirements are
ment = 1 (exposed to pesticide). independent of any particular set of data. Inferential mod-
• Yi0 : potential outcome if same individual receives els, however, are inevitably dependent on the available data
Treatment = 0 (not exposed). and the statistical machinery used for estimation and eval-
uation. Based on this distinction, we place causal analysis,
In the case of explanatory models based on observa- a set of procedures for examining hypotheses independent
tional data, we can exploit Pearl’s (2000) introduction of of data, into the broader context of explanatory modeling,
a new mathematical operator, the do-operator, to express which encompasses all the information that influences the
the idea of a hypothetical manipulation. With this oper- interpretation of the results (Fig. 2). Causal analysis
ator, we can unambiguously represent the probability for describes the scientific assumptions upon which explana-
an individual’s hypothetical response when a treatment tory interpretations depend, whereas the latter must con-
variable Tmt is set to the specific value t (which is 0 or 1 sider the imperfections of the data, the limits of our
CONCEPTS & SYNTHESIS

in this case) as P[Y|do(Tmt = t)]. Notice the ability to knowledge, and the need to accommodate computational
manipulate the variable Tmt physically is not required relationships in models. Because it depends so critically on
(Bollen and Pearl 2013), though the ability to imagine knowledge outside of an individual study, explanatory
the value of Tmt being set to some value by humans or modeling is best thought of as a process that builds confi-
nature is implied. Note that the common statistical nota- dence in our understanding through a series of investiga-
tion “P(Y|X = x)” means the “probability of observing tions. In this paper, our presentation focuses on causal
Y, given that X is observed to equal x,” whereas Pearl’s analysis to serve our objective of showing how to think
notation “P(Y|do(X = x))” means “the probability of hard about the scientific content of candidate models.
observing Y, given that X is forced to equal x.” In gen- More about the details associated with explanatory model-
eral P(Y|X = x) is not equal to P(Y|do(X = x)). Pearl’s ing, moving from theory to models to results to interpreta-
do-calculus tells us when the two are equal and when tions, is outlined in Grace et al. (2012) and Shipley (2016:
they are not equal (Pearl et al. 2016: Chapter 3). Chapter 8). Many of the general points we present are
This conceptualization of causal relations reveals a widely applicable. However, the specific techniques herein
problem, but also a possible solution to that problem. will most commonly be useful in observational studies
The problem is that we can only observe one of the and field experiments with important covariate effects.
potential outcomes for any given individual frog during An overview of explanatory modeling is presented in
an experiment. Therefore, individual-level causal effects Fig. 2. The first step, assembling background information
are counterfactual and not directly observable, even in about the system under study, is critical to success and
randomized experiments. For this reason, the focus is represents the mechanistic foundation required for scien-
typically on the average causal effect, ACE, where tific interpretations. Step 2 is the process of causal analy-
sis, which describes the requirements for drawing
ACE ¼ E ½Y1 
E ½Y0  (1) scientific (cause–effect) interpretations, our primary focus
in this paper. Step 3 refers to the subsequent step where
averaged across all i. When the consistency assump- models are confronted with data and their testable impli-
tion (Pearl 2009:99) holds, cations are evaluated. This step is described by the litera-
ture on structural equation modeling (Appendix S3, Part
E ½Y1  ¼ E ½Y jT ¼ 1 and E ½Y0  ¼ E½Y jT ¼ 0 (2) S1). Step 4 relates to drawing interpretations with proper
respect for assumptions. Finally, confident interpretations
In words, Eq. 2 says that the potential outcomes for ultimately depend on the consistency of findings and our
the total population, E[Y1] or E[Y0] are being inferred knowledge of mechanisms, which we refer to as sequential
from the subpopulations assigned to the treatments, E learning (Step 5). Although this step is widely recognized
[Y|T = 1] and E[Y|T = 0]. For observational studies, this as important, it is often omitted from formal discussions
assumption permits causal effects to be estimated from of the requirements for drawing causal inferences.
data as long as confounding effects are controlled for
(i.e., the treatment subpopulations are comparable). This
FOUR KEY PRINCIPLES OF CAUSAL ANALYSIS—THE PRAC-
is a critical and also powerful assumption that determi-
TICE OF “THINKING HARD”
nes our ability to infer causal information from observa-
tional (but also experimental) data. Note that this
Principle 1: Causal networks provide a powerful and
conceptualization can be extended to the situa-
convenient interpretive structure for causal analysis
tion where X is continuous, in which case the causal
effect is captured in a parameter or set of parameters The first of the four principles for causal analysis we
(for complex functions). present in this paper (Table 4) deals with the merits of
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 5

CONCEPTS & SYNTHESIS

FIG. 2. Flow chart showing major steps in the process of explanatory modeling.

TABLE 3. Ranking of candidate models based on AICc.

Model K AICc DAICc AICcWt Cum.Wt LL Predictors

m12 5 31.41 0.00 0.39 0.39
10.35 F+A+E
m16 6 33.51 2.10 0.14 0.52
10.25 F+A+E+C
m6 4 33.83 2.42 0.12 0.64
12.68 F+A
m7 4 33.96 2.54 0.11 0.75
12.74 F+E
m2 3 34.94 3.52 0.07 0.81
14.33 F
m13 5 34.95 3.54 0.07 0.88
12.12 F+A+C
m8 4 35.32 3.90 0.06 0.94
13.42 F+C
m14 5 36.07 4.66 0.04 0.97
12.68 F+E+C
m9 4 37.70 6.28 0.02 0.99
14.61 A+E
m15 5 39.90 8.49 0.01 1.00
14.59 A+E+C
m10 4 42.02 10.61 0.00 1.00
16.78 A+C
m3 2 42.23 10.82 0.00 1.00
17.98 A
m5 3 48.52 17.11 0.00 1.00
21.12 C
m4 3 49.66 18.25 0.00 1.00
21.69 E
m11 4 49.97 18.56 0.00 1.00
20.75 E+C
m1 2 51.88 20.47 0.00 1.00
23.87 Intercept only
Notes: Also presented are K = number of parameters in the model, including intercept and error variance (thus, K = number of
predictors + 2); AICc = sample-size adjusted Akaike information criterion; DAICc = numeric difference between the AICc for a
model and the minimum value for the model set; AICcWt = probability that an individual model is the best in the set of candidates;
Cum.Wt = cumulative value of AICcWt values; LL = log likelihood for a model. Predictor abbreviations: F, fire severity; A, age of
stand; E, elevation, C, distance from coast.

representing hypothesized relationships among a set of Understanding the structure of causal diagrams.—A cau-
variables using probabilistic causal networks. There are sal diagram is a pictorial graph of cause–effect relation-
two essential components of causal networks. One is the ships that is used to represent scientists’ ideas about how
causal diagram, which allows us to convey the logic of a system works (Robins 1986, Pearl and Verma 1991,
our hypotheses, and the second is an appropriate equa- Greenland et al. 1999). The diagram is meant to repre-
tional framework for estimating network relationships. sent a translation of the scientist’s ideas and hypotheses
 Article e02962; page 6 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

TABLE 4. Four key principles for causal analysis. variables in M except for X and Y and those variables
along the directed path (causal chain) between X and Y
Principle Description being discussed (W in this case). Importantly, variables
1 Causal networks provide a powerful and convenient and links omitted from a causal diagram also represent
interpretive structure for causal analysis. a set of explicit assumptions with empirical implications.
2 Many elements of network structure can be tested There are several definitional features of causal dia-
with appropriate data, such as both omitted and
included links. grams. First, directed arrows imply cause–effect relation-
3 Confounding due to model misspecifications can ships. This means that if we physically change the values
bias parameter estimates. Analysis of causal of the variable at the base of an arrow (e.g., W in Fig. 3),
diagrams can provide strategies for addressing this while holding constant all other predictors, it can alter
issue.
the values of the variable at the tip of the arrow (e.g., Y
4 The inclusion of mediators can encode testable
hypotheses about mechanisms and strengthen in Fig. 3). There are two corollaries that go with this def-
inferences. inition of arrows. One is that there is understood to be a
finite passage of time between the cause and the
response. Thus, the arrows represent movement of infor-
mation from some time in the past to present. Another
CONCEPTS & SYNTHESIS

corollary is that we can force changes to the quantity at

the tip of the arrow (e.g., Y) without influencing the
quantity at the base of the arrow (W) (also referred to as
the assumption of asymmetry). This means that chang-
ing present conditions should not change past condi-
tions. Thus, arrows represent implied experimental
manipulations and routinely we use thought experiments
to analyze the logical relations in causal diagrams. When
we can be more rigorous, we use physical experiments to
test our assumptions (Gough and Grace 1999). Note
that in regression models (Fig. 1A), the arrows simply
represent computations—if we see a particular value of
FIG. 3. Causal diagram consisting of four observable vari- X1, we expect to see a particular value of Y. This concep-
ables, X, Y, Z, and W connected by four directed arrows repre- tual distinction between the predictive interpretation of
senting cause–effect relationships. U variables represent
unspecified, unobserved causes of variation. By convention, regression and causal relationships is sometimes
models and causal diagrams are distinguishable by the fact that described as the difference between “seeing” (an associa-
in causal diagrams, variables are not outlined as they are in tion) and “doing” (hypothesizing responses to manipula-
models (Fig. 1). tions); Lindley (2002).
A second attribute of directed relations in causal dia-
about the data-generating mechanisms into a set of grams is that they are transitive, meaning that if X causes
observational expectations. W and W causes Y, then X is a cause of Y. There is suffi-
Consider a simple example involving four random cient confusion about this point (particularly when peo-
variables, X, Y, Z, and W (Fig. 3). Also shown in Fig. 3 ple start thinking in terms of “true” causes) that we offer
are four unspecified causes of variation (U variables), a memory aid for the transitive property, “relative to
which create the random component of the variables. some response, a cause of a cause is a cause.” This
Causal diagrams are based in a particular graph form, phrase reminds us that distal causes are just as axiomatic
referred to as a directed acyclic graph—directed because as proximal ones. This idea will help us clarify causal
all arrows are unidirectional, and acyclic because it con- logic when we return to the wildfire example later in the
tains no causal loops. Variables directly connected by a paper.
link are adjacent and those not directly connected are A third attribute of causal diagrams is known as the
nonadjacent. Pathways can be traced through a series of Markov condition (Hitchcock 2019). This means that we
links either following the flow of causation (directed only need to know proximate causes to compute the
pathways) or ignoring directionality (undirected path- expected values for a response variable. Relative to
ways). Directed pathways passing through two or more Fig. 3, if the diagram is a true reflection of the data-gen-
links represent indirect effects in the model, whereas erating mechanism, we can compute the expected values
those conveyed by a single link constitute direct effects for Y from the values of W and Z. Any influences from
in the model. To say that W is a direct cause of Y in cau- X on Y can be seen to be conveyed through indirect
sal model M means “a change in W will induce a change effects; therefore, we can say that X is conditionally
in Y even if we hold constant all other variables in M ignorable given W and Z.
except for W and Y.” To say that X is an indirect cause Finally, causal diagrams are nonparametric in the
of Y in causal model M through W means a change in X sense that no specific statistical assumptions about the
will induce a change in Y if we hold constant all other type of responses, linkage functions, or error
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 7

distributions are implied. As a result, the interpretation One might argue that the most notable separation
of a causal diagram is independent of statistical details. between statistical modeling and scientific modeling is
reflected in the difference between the equational form
Understanding how associations among variables are gen- for regression vs. causal networks.
erated.—To understand the logic of causal diagrams, we This dichotomy can be traced back to competing
need to understand the various ways that associations world views established by Ronald Fisher and Sewall
between variables can be created. If the causal diagram Wright in the 1920s. Fisher, whose dominant influence
in Fig. 3 is true, a fundamental expectation is that all over the development of statistics remains to this day,
adjacent variables will show conditional associations. was the father of the regression equation (Aldrich 2005).
Considering the nonadjacent pairs of variables, there are There are many types of regression models, but they all
two primary ways associations can be created indirectly: have the same general weaknesses with regard to
(1) through causal chains and (2) through common explanatory interpretation. Regression models are based
causes. (There are actually other ways that associations on the fundamental equational form
among variables can accidentally be created, which we
will describe later in the paper.) Causal chains (e.g., X ? Y ¼ f ðXÞ (3)
W ? Y) produce correlations between nonadjacent vari-

CONCEPTS & SYNTHESIS

ables (X and Y in this case) through indirect, compound where Y is a vector of response variables (or single
pathways. The expected value for a bivariate association response variable), X is a vector of predictors, and f is
between X and Y simply as a result of the three-variable one of many functional forms, ranging from General
chain is the product of the strengths of effects along the Linear Models to Generalized Linear Mixed Models.
chain. For more complex relations, as in Fig. 3, where The limiting features of this equation were discussed
there are two causal chains connecting X to Y in the dia- above in relation to Figure 1.
gram (X ? W ? Y and X ? Z ? Y), we expect the total Causal networks are based on a fundamental equa-
bivariate association between X and Y to be the sum of tional form developed by Wright (1921) that can be seen
the indirect effects. Nonadjacent variables connected by as an extension of the regression equation
a common cause, for example, W X ? Z, will also be
expected to exhibit a bivariate association. Note that for Y ¼ f ðX; YÞ: (4)
this type of pathway there is an implied historical effect
of X on W and Z. The expected value for the bivariate here, Y is a vector of the response variables embedded
association between W and Z is also the product of the within a network of relationships, X is a vector of exoge-
strengths of effects along the connecting pathway. nous variables that define the boundaries for our model,
It is worth pointing out that there are special situa- and f is a set of functional forms, again with no inherent
tions where variables are inherently nonindependent limit as to type. Importantly, it is the network of equa-
quantities. For example, if W is the proportion of a plant tions that permits and requires a priori scientific knowl-
community biomass made up of legumes and Z is the edge.
proportion made up of grasses, then both variables are
derived quantities with a common denominator, total
Principle 2: Many elements of network structure can be
community biomass. The consequence is that W and Z
tested with appropriate data, including both omitted and
will be inherently nonindependent and be associated
assumed links
even in a null model (where all links are nonsignificant).
This is a fundamental problem that can contribute com- Although adopting a network structure for our
putational association to any pair of variables, including hypotheses is an essential step, it is important to point
adjacent ones. Adjustments for this fundamental issue in out that many aspects of hypothesized networks can be
data analysis are possible (Pawlowsky-Glahn and Buc- tested, given appropriate data. This testing can lead to
cianti 2011), but beyond the scope of the present discus- the discovery of previously unsuspected mechanisms and
sion. For another example, it is possible for two sets of the specification of new models that require additional
variables to be correlated because they follow a common testing and verification. Although we will illustrate this
developmental program. Again, this special case is process below, it is important to make the point that the
beyond the scope of our presentation. enterprise described in this paper is not part of so-called
assumption-based modeling. Many network-style mod-
Network equations: mathematical representation of causal els that ecologists might encounter are deductive expres-
networks.—Generally, scientists represent information sions of accumulated assumptions (see, for example, the
about systems using networks (ecosystem models, food journal Ecological Modelling). The same description can
webs, metabolic models) and those who apply science for be applied to most theoretical models. Explanatory sta-
engineering and decision making rely on information tistical models, however, provide methods for ruling out
networks. Statistical models, in contrast, have tradition- a large number of possibilities using the principles
ally been reductionist, adopting the simplest representa- described next. Omitted links from hypothesized models
tion possible in order to isolate a single effect of interest. represent strong assumptions that create testable
 Article e02962; page 8 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

implications for data. Included links, which only claim sampling scheme, and (5) through various roles the vari-
nonignorability, represent less strong, but nonetheless able might play in censoring or truncating the data or
testable, premises. The directionality of arrows (i.e., of the sample. The rules that follow are meant to apply to
causal relationships), however, is one type of assumption all these different situations.
that requires theoretical justification and is not testable
from observational data alone (though can be tested d-separation rule 1.—Given a causal chain, X ? W ? Y,
through manipulative experiments). complete or full mediation means once we regress Y on
W, X provides no capacity to explain additional varia-
tion in Y (i.e., the effect of X on Y is explained through
Principle 3: Confounding because of model
W). Therefore, Y is independent (⊥) of X, given that we
misspecifications can bias parameter estimates. Analysis
have conditioned Y on W (|W).
of causal diagrams can provide strategies for addressing
this issue
Y ?X jW (5)
Much of the methodological work on causal inference
has dealt with ways of addressing potential impacts from In d-separation parlance, X and Y are said to be d-sep-
model misspecification (models that do not match the arated when Y is conditioned on W.
CONCEPTS & SYNTHESIS

true data-generating process). Until fairly recently, very

little work focused on a formal analysis of the problem. d-separation rule 2.—Given a common-cause pathway,
Rules for recognizing potential sources of bias and devis- such as W X ? Z, conditioning W and Z, the descen-
ing remedies using causal diagrams now exist. These dent variables, on the common cause variable X renders
rules have been summarized in axiomatic form using a W and Z d-separated (conditionally independent):
principle called d-separation (Pearl 1988).
The formal definition of d-separation gives the W ?ZjX (6)
requirements for two nonadjacent variables embedded in
a graph (e.g., W and Z in Fig. 3) to be conditionally
independent. Conditional independence is manifested d-separation rule 3.—Considering an undirected path
through the absence of residual association, which including a collider, such as W ? Y Z, conditioning
means that if we were to include a directed link or error the ancestral variables (W and Z) on the collider (Y) can
correlation between W and Z, analyses with an appropri- create an association between otherwise independent
ate data set would find the link to be unsupported. To causes of the collider. Thus,
break d-separation down, the first two (of the three)
rules of d-separation tell us what we need to do to iden- W ZjfX ; Y g (7)
tify (permit estimation of) causal effects properly. The
third d-separation rule relates to things to not do to where the symbol ∦ means nonindependent. To make
avoid inadvertently creating false signals. this somewhat nonintuitive principle clear, a computa-
The elements of the d-separation criterion include the tional demonstration is given at the end of Appendix S4
two diagram structures that generate associations for the interested reader.
between nonadjacent variables mentioned earlier: (1) Note that we would not normally condition on a col-
causal chains and (2) common causes. It also addresses a lider if we were using a causal diagram to build our model
third diagram structure, referred to as collider relation- because the direction of causality should be clear. How-
ships. A collider variable is one receiving two or more ever, there are multiple ways one might accidentally con-
incoming arrows. In Fig. 3, W ? Y Z is our only col- dition antecedent variables based on a collider in other
lider pathway, with Y being the collider variable. It is contexts. One situation is in regression studies where ana-
characteristic that tracing pathways through colliders lysts use preconditioning to remove confounding effects
involves a reversal in causal direction. The implications from some relationship of interest (for example, a treat-
of collider relationships can be understood by recogniz- ment-response study where there is adjustment for
ing that the collider Y is a descendent of its parent vari- unequal treatment effectiveness). Another situation is
ables W and Z. where data collection is restricted by an outcome, such as
Conditioning is a key concept for understanding causal looking for associations among causes of death using a
analysis generally and d-separation specifically. This sample of dead animals (in this case, being dead, and
word describes a number of statistical operations, so its therefore part of the sample, is a descendent of the causes
use can be a source of confusion. The variety of ways to of death). There are legions of examples where inadver-
“condition on a variable” involve (1) including the vari- tently conditioning on a collider has introduced bias into
able in a model, (2) preconditioning (residualize) the analyses (Elwert and Winship 2014, Pearl and Mackenzie
variables of prime interest in order to remove the influ- 2018). Awareness of this issue is one reason the use of
ences of variables that are to be omitted from the model, causal diagrams is spreading into fields that have histori-
(3) stratifying the data by levels of the variable, (4) sam- cally relied on regression to address treatment-response
pling from a population using the variable in the questions, such as epidemiology and econometrics.
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 9

For more complicated diagrams than examined here, framework for estimation, evaluation, and summarizing
the d-separation criterion goes a step further and findings.
describes how to find the minimum sets of conditioning There are many modern treatments of SEM (Kline
variables to achieve d-separation. The implementation 2016), some directed at natural scientists (Grace et al.
of d-separation was first described in Shipley (2000b). 2015, Shipley 2016). It is beyond our purpose in this
Several software packages are now available for working paper to provide an in-depth introduction. Rather, we
through all the possibilities prior to fitting a model with offer a high-level view for scientists not already familiar
data (Textor et al. 2011, Marchetti et al. 2015). Routines with the methodology, as well as executable code for the
also exist in some SEM software (Lefcheck et al. 2018). examples in this paper in the Appendices. A survey of
Both Shipley (2016) and Kline (2016) provide thorough what ecologists find helpful about SEM can be found in
discussions of conditioning sets based on d-separation. Laughlin and Grace (2019).
Perhaps the most important point to make about
SEM is that it is a scientific framework for explanatory
Principle 4: The inclusion of mediation relationships, in
modeling rather than a specific statistical technique.
the form of causal chains, encodes testable mechanistic
There are no statistical assumptions that are inherent to
explanations
SEM; those depend on the implementations supported

CONCEPTS & SYNTHESIS

One of the most common scientific applications of a by particular software programs. Thus, SEM is perfectly
network structure is to permit formal evaluation of compatible with the nonparametric description of
mediation hypotheses. By permitting the specification of Pearl’s (2009) structural causal model, the primary
a mediation relationship in a causal chain, it is possible source for modern causal analysis. Related to that, the
to represent mechanistic explanations within models. methodology continuously evolves to incorporate proce-
For example, the diagram in Fig. 3 encodes the hypothe- dures supportive of explanatory modeling. This paper is
sis that X affects Y through two mediators, W and Z. As an illustration of that point through our demonstration
we shall see, once we transition to a fully specified statis- of how modern causal analysis integrates with SEM
tical model and data are used for estimation (Step 3 in (Fig. 2). Specialized regression procedures that are
Fig. 2), tests of conditional independence allow us to sometimes presented as alternatives to SEM, such as
discover that the specified causal chain is either clearly instrumental variables, simultaneous equations, two-
supported (all links are deemed biologically meaning- stage least squares, and hierarchical Bayesian models are
ful), unsupported (no evidence for some links), or all implementable within the SEM framework. This
incomplete (there is evidence that additional links are potential can be understood by reference back to the
needed). When unsupported, we may conclude that there form of the equations (Eq. 4), which allow the full suite
is no empirical evidence for the proposed mediation of regression-type statistical models as elements in an
mechanism. When incomplete, we conclude that there is SEM model. Another distinct attribute of SEM is that it
some additional mechanism whereby X influences Y that permits the evaluation of integrative system-level
is not through W or Z. This latter finding represents the hypotheses, as opposed to reductionist ones. It is cer-
discovery of a new mechanism that was unsuspected and tainly true that more complex models pose more risks
only revealed via tests of mediation hypotheses. Many for violations of assumptions. Nonetheless, the evalua-
other surprises can also be revealed through the use of tion of causal-network-style hypotheses can yield more
causal networks, a few of which we illustrate below in scientific insights and more complete explanations when
the second example presented. applied successfully.

CONFRONTING HYPOTHESES WITH DATA—STRUCTURAL Example 1: A return to the wildfire recovery example
EQUATION MODELING
Our explanatory modeling work-flow process requires
Our primary objective in this paper is to show scien- explicit consideration of expert knowledge in order to
tists how to “think hard” about their hypotheses to pro- construct plausible hypotheses to convey in a causal dia-
mote explanatory analyses (Fig. 2, Steps 1 and 2). Our gram. Following the data collection and initial examina-
four principles of causal analysis should assist in repre- tions of relationships conducted by Keeley et al. (2005),
senting causal hypotheses and establishing their empiri- Grace and Keeley (2006) compiled a carefully consid-
cal claims. Now, we need to say something about ered list of biological assumptions relevant to the mea-
methods for confronting those hypotheses with data so sured variables (Table 5). This table is an example of
we can complete the illustration through the use of Step 1 in Fig. 2 and we feel a useful companion to the
examples (Fig. 2, Steps 3 and 4). No longer are we deal- causal diagram. In this demonstration, we first consid-
ing with diagrams, but instead, fully specified models. ered a na€ıve causal diagram (Fig. 4A) that considered all
We use the general term explanatory modeling in this possible links (denoted by the corresponding numbers in
paper to describe the intention. For the context in which Table 5). We used our na€ıve causal diagram (Fig. 4A)
we are working, the statistical methodology of structural and assembled expert knowledge (Table 5) to arrive at
equation modeling (SEM) provides a well-established an informed diagram (Fig. 4B). In this example, we start
 Article e02962; page 10 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

TABLE 5. Expert opinion relevant to an initial hypothesis of how vegetation cover following wildfire could be explained by the
measured predictors.

Number Interpretation
1 In southern California, where the wildfires took place, elevation generally increases as one moves eastward from the
coastal lowlands to the interior highlands (see Keeley et al. 2005, Fig. 1). Thus, increasing the distance from the coast
tends to cause an increase in elevation (expected effect: positive).
2 Increasing elevation should produce more mesic conditions, which in turn, would reduce the frequency of wildfires and
thereby increase the average ages of forest stands (because age is determined by time since last stand-replacing fire).
The net relationship expected is thus increased age with increasing elevation (expected effect: positive).
3 As stands age, they accumulate more biomass (alive and dead) that can serve as fuel during a wildfire. Therefore,
increasing stand age should lead to more severe fires when burning finally takes place (expected effect: positive).
4 Independent from other effects, moving away from the coast should result in less wildfire suppression and larger fires
because of reduced urbanization and population densities (expected effect: negative).
5 Fire severity depends not only on fuel, but also humidity. Thus, increasing elevation should contribute to less severe fires,
all other things equal (expected effect: negative).
6 Once the effects of stand age and elevation are taken into account, it is not obvious, based on a priori information, that
additional processes related to distance from the coast should have a major influence on wildfire severity (expected
CONCEPTS & SYNTHESIS

effect: undetectable).
7 Once effects of elevation, stand age, and fire severity are controlled, it is not obvious, based on a priori information, that
vegetation recovery will be influenced by distance from the coast (expected effect: undetectable)
8 In the study region, a marked, orographic gradient leads to a strong positive effect of elevation on precipitation, which
would lead to faster plant regrowth and a more rapid recovery of the vegetation (expected effect: positive).
9 A negative effect of stand age on vegetation recovery might be observed because of the fact that increasing overstory
dominance in older stands will lead to a reduced understory. If the understory (including the seed bank) is a strong
contributor to vegetation recovery, an independent effect of stand age on vegetation recovery might be expected
(expected effect: negative).
10 There is a very strong a priori expectation that increased fire severity will result in reduced vegetation because of the
damage to the regenerating tissues of plants and possibly effects on soil that reduce water penetration (expected effect:
negative).

with a single hypothesis for consideration, which is our explanations and thereby avoid the problem of infi-
derived from our summary of expert knowledge. We do nite regress in explanation.
this knowing that we will have the opportunity to com- The causal diagram in Fig. 4B was translated into a
pare the support for this hypothesis against all the alter- structural equation model using the software package
native hypotheses based on the same causal ordering of piecewiseSEM (Lefcheck et al. 2018). As shown in
predictors. Appendix S3, Data S3, estimation produces a set of d-
The backbone of the causal diagram is a directed separation tests, one for each omitted link in the model.
chain relating vegetation recovery (V) to fire severity (F) These tests evaluate whether there is evidence to include
to stand age (A) to elevation (E) to distance from the a missing link, and thus, whether it should be added to
coast (C). This sequence of variables represents an the model. Global fit of the initial model suggested it
ordering from proximal to distal potential drivers of veg- was a plausible explanation for the data. Further, tests
etation response following fire. In an earlier review of suggested equivocal support for some of the links, how-
this paper, we were asked to explain how it can be ever, so simpler models were also considered. During
argued that the various predictors in this example actu- the evaluation process, individual d-separation tests
ally constitute causes. The rule related to this point is suggested the possibility of local violations of condi-
very basic and can be called the intervention-response tional independence. We ultimately considered four
rule: if we can manipulate something and induce a alternative models, omitting and adding various links.
response in another system property, then the thing The model selected from that process is shown in
manipulated qualifies as a cause of the thing that Fig. 4C. The scientific conclusion we draw from the
responds. Confusion comes about when someone starts results is that vegetation recovery varies widely in the
wanting to distinguish what they think of as “true” landscape, primarily as a function of fire severity, which
causes. Let us consider the most distal cause in the in turn is influenced by greater fuel accumulation in
chain, distance from the coast. Intuitively, it might seem older stands of woody plants. Distance from the coast
that this is a placeholder for some “actual” cause. A sim- (dismissed as important by all-subsets regression) is
ple statement shows that location qualifies as a causal shown to influence recovery, but only indirectly through
variable. That statement is, “If you think location in effects on both stand age and fire severity, ultimately
space is not an actual cause, then stand closer to the fire being quite important in the system. The explanation
and see if your opinion changes.” In reality, there are achieved is based on careful and explicit considerations,
many potential mediators that make up a causal chain. something we could not easily achieve using an all-sub-
We usually work with mediators that are meaningful for sets regression approach.
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 11

FIG. 4. (A) Naive causal diagram with all possible links. (B) Hypothesis based on available expert opinion (Table 5). (C) Model
informed by data and d-separation tests (Appendix S3). Dashed lines represent negative effects, solid lines positive effects.

is a binary (0/1) variable. As with regression, we are gener-

Example 2: In situ experimental study of a marine food
ally limited in ANOVA, ANCOVA, and MANOVA by
web

CONCEPTS & SYNTHESIS

the equational form of the model, Y = f(X), where the X
Models developed using experimental data are also vector includes the treatment variable and any variables
often evaluated using model comparisons, and such data included as covariates. In this case, Y can be a single
allow us to illustrate the utility of the principles in this response variable or any subset of the three responses.
paper further. For this example, we draw upon a study of Let us consider some possible rationale for the options
food webs in marine seagrass beds (Whalen et al. 2013). shown in Fig. 5. Diagram A is the classic ANOVA.
In this system, the dominant herbivores are small inverte- Because of randomization, we might reason that covari-
brates (specifically, microcrustaceans), which graze on the ates like macroalgae (Mac) and seagrass density (Gra),
epiphytic algae that grow on seagrasses while also being should be independent from the treatment assignment;
important prey for fish. A global decline in seagrass beds thus there is no reason one must include the covariates
has raised concern that reduced populations of larger fish in the diagram, as the effect of treatment should be the
(due to commercial harvest) is resulting in an increase in same regardless. Comparison to results from the
small fish and subsequent reduction in microcrustaceans ANCOVA (Diagram E) is expected to confirm this
(a trophic cascade). The microcrustaceans have been assumption, though one should always check, as ran-
hypothesized to serve in a protective role for seagrasses by dom assignment does not guarantee comparable control
grazing on the epiphytic algae that grow on seagrass and test groups. Diagram B could be justified because
leaves and that impair plant photosynthesis and survival. microcrustaceans (Gammarids and Caprellids) are a dis-
Convincing field experimental demonstration of this top- tinctly different type of response from epiphytes. We
down process has been lacking. In this study, the authors might, for example, see one or both microcrustacean
examined the responses of microcrustaceans and their groups respond, but no response from epiphytes. For
predominant food supply, epiphytic algae, to treatment this reason, we might want to perform separate analyses.
with slow-release insecticide in seagrass beds. Experimen- Diagram C represents a classic MANOVA with all three
tal plots were established by sinking plastic poles into the responses treated as a single response. For this particular
sediment and suspending insecticide-impregnated blocks, study, it might seem more logical to interpret results
or insecticide-free blocks, within the seagrass canopy from Diagrams A and B instead of C because the
20 cm above the sediment surface. Forty experimental response types are clearly different.
plots were located randomly within the study area and Diagram D represents an interesting possibility.
were randomly assigned to receive insecticide or not. Because epiphytes constitute the food supply for micro-
Sampling of microcrustaceans and epiphytes were con- crustaceans, then an investigator might use epiphytes as
ducted at the end of the experiment. Microcrustaceans a covariate to adjust for differential food supplies among
were divided by taxa and abundance was estimated. Epi- plots. They could either perform an ANCOVA individu-
phyte abundance was also measured. ally for the Gammarids and Caprellids, or look at their
Given the three response variables (epiphyte biomass, joint response as depicted in Diagram D. There are mer-
Gammarid crustaceans, and Caprellid crustaceans) and its and demerits to either approach. Diagrams E, F, and
two covariates (macroalgae and seagrass density) mea- G represent different versions of ANCOVA with epi-
sured, there are a variety of different analyses that an phytes as the focal response. Diagram E is a standard
investigator might perform. First to come to mind is the ANCOVA, including the two covariates abundance of
suite of options that include ANOVA, MANOVA, macroalgae (Mac) and density of the seagrass in a plot
ANCOVA, and MANCOVA. We present causal diagrams (Gra). In Diagram F, the logic shifts from that used in
for these classic models in Fig. 5. The primary differences Diagram E. Because the treatment method (slow-release
from our earlier example to keep in mind are (1) treat- insecticide) cannot kill 100% of the microcrustaceans in
ment was randomly assigned and (2) the treatment “Trt” the open plots, experimental control is incomplete in this
 Article e02962; page 12 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4
CONCEPTS & SYNTHESIS

FIG. 5. Causal diagrams representing (A) ANOVA 1, (B) ANOVA 2, (C) MANOVA 1, (D) MANCOVA 1, (E) ANCOVA 1, (F)
ANCOVA 2, (G) ANCOVA 3. Trt = treatment, Epi = epiphyte abundance, Gam = Gammarid abundance, Cap = Caprellid abun-
dance, Mac = macroalgae abundance, and Gra = seagrass density. Results for these models can be found in Appendix S4.

study. Controlling the variable responses of microcrus- to explain the response of epiphytes to treatment com-
tacean by including these variables as covariates might pletely. As a result, there is no link directly from treat-
seem to be a useful way to remove their variations from ment to epiphytes in the final model. Finding support
the assessment of treatment effects on epiphytes. Dia- for full mediation is a highly desirable outcome in an
gram G represents a commonly employed approach, experiment, because we are not scientifically interested
which is to control for as many possible influences as in the effects of insecticide and prefer a model where
possible. This cursory discussion of possibilities is meant artificial treatments are conditionally ignorable. Sec-
to point out that causal hypotheses will have to be ond, empirical evaluation of the implied conditional
brought to bear in order to decide which of these models independences from the initial hypothesis (Fig. 6A)
will produce defensible results. revealed three nonindependences that could be
It is instructive to compare the hypotheses that can be resolved by adding links (Fig. 6B, added links shown
examined using classical statistical models (Fig. 5) to the with asterisks). This produced evidence to suggest pre-
possibilities that emerge from adopting a causal network viously unanticipated biological discoveries related to
perspective. For this comparison, we refer the reader to the system under study.
a causal diagram representing a dual mediation hypothe- The positive effects discovered by the analysis (links 7,
sis in Fig. 6A. Here, impacts of treatment on Gam- 8, and 9) can be interpreted as indications that macroal-
marids and Caprellids, the two dominant groups of gae (to a large degree) and eelgrass density (to a lesser
microcrustaceans, are hypothesized to explain the effect degree) provide protective refuges for microcrustaceans.
of treatment on epiphytes. This model also includes By promoting microcrustaceans, macroalgae have indi-
macroalgae and seagrass density as covariates poten- rect negative effects on epiphytes (through the pathways
tially influencing epiphytes. This diagram was treated as 7 ? 3 and 8 ? 4). Within the total food web, this could
the initial hypothesis, which was subsequently evaluated translate into macroalgae facilitating seagrasses by har-
using the available data. The structural equation model boring microcrustaceans, the grazers of epiphytes (which
presented in Fig. 6B summarizes results for the final are the enemies of seagrasses). We believe scientists will
model once d-separation tests were performed and links find this system-level set of results substantially more
were added to resolve d-separation violations informative than results from traditional ANCOVA.
(Appendix S4, Data S4).
There are numerous, interesting scientific findings
INTERPRETATIONS AND CONSIDERATIONS
revealed by the SEM (Fig. 6B; see also Whalen et al.
2013). First, we are able to evaluate the full-mediation Explanatory modeling requires adequate expert
hypothesis formally. The two mediation pathways knowledge to defend scientific interpretations. Appropri-
(those combining links 1 ? 3 and 2 ? 4) were found ate data are also required and frequently limit the
April 2020 EXPLANATORY MODELING AND CAUSAL ANALYSIS Article e02962; page 13

FIG. 6. (A) Causal diagram representing initial hypothesis and (B) final structural model supported by the data (see
Appendix S4). Solid lines in (B) represent effects of positive sign and dashed lines represent effects of negative sign. Linkages in B
not seen in A (7, 8, and 9; indicated with asterisks) represent important effects uncovered because of failed d-separation (conditional

CONCEPTS & SYNTHESIS

independence) tests.

conclusions that can be drawn. There are many ways our Subsequent SEM studies (Keeley et al. 2008) have
models and interpretations can deviate from the truth, enhanced our confidence in the general inferences drawn
to either major or minor degrees. Here we list three from the original study. That said, we would not claim
increasingly challenging types of assumptions upon that all our parameter values are unbiased causal esti-
which interpretations depend. (1) Foundational assump- mates without further evidence to support such inferences.
tions—arrows in models indeed represent directional
cause–effect relationships. Generally, this must be sup- CONCLUSIONS AND FUTURE DIRECTIONS
ported by a priori expert knowledge of mechanisms. (2)
Parameter estimates approximate true values to the Causal understanding is ultimately about understand-
degree that overall conclusions are not confounded. ing mechanisms. The majority of studies that scientists
Again, expert knowledge of the system under investiga- conduct are directed to that end. Classical statistical
tion needs to be sufficient to defend general conclusions. models do not easily accommodate the explicit incorpo-
Further, it is important that investigators be aware of ration of mechanisms into hypotheses, limiting their
the ways that confounding can occur and the need to capacity for explanatory application. Adopting a causal
guard against the omission of critically important com- network framework and the principles of causal analysis
mon causes. (3) Inferring that one or more parameter for hypothesis development and evaluation greatly
estimates are unbiased (true) estimates of causal effects increases possibilities for the development of explana-
constitutes a stricter assumption. Both omitted con- tory models and clear expression of their logic. We hope
founders and measurement error, along with a host of this presentation provides a guide toward that future by
more technical issues, can bias estimates to various helping scientists see their own responsibilities and
degrees (though remedies to even these challenges exist; opportunities in quantitative analysis.
Bollen 2019). For these reasons, appropriate caution is ACKNOWLEDGEMENTS
required when drawing conclusions.
We thank Brian Cade, Katherine Banner, Megan Higgs, Dar-
Suitably cautious language for expressing findings will
ren Johnson, Lori Randall, Billy Schweiger, Magdalena Steiner,
vary depending on the above-mentioned factors. As one Lara Volery, and Rachel Korn for reviews of earlier drafts of
illustration, we present the language Grace and Keeley the manuscript. Brian Inouye, Bill Shipley, and an anonymous
(2006) used when reporting their findings from the wild- reviewer generously provided very helpful suggestions. JG was
fire recovery study. supported by the USGS Land Change Science and Ecosystems
Programs. Any use of trade, firm, or product names is for
We infer from the SEM results that postfire rich- descriptive purposes only and does not imply endorsement by
the U.S. Government.
ness in this system is strongly influenced by local
conditions and that these conditions are, in turn,
predictably related to landscape-level conditions. LITERATURE CITED
For example, we observed that older stands of Aldrich, J. 2005. Fisher and regression. Statistical Science
shrubs were characterized by more severe fires, 20:401–417.
which were associated with a low recovery of plant Banner, K. M., and M. D. Higgs. 2017. Considerations for
cover and low richness. These results may have assessing model averaging of regression coefficients. Ecologi-
implications for the use of prescribed fire in this cal Applications 27:78–93.
Bollen, K. A. 2019. Model implied instrumental variables
system if these findings extrapolate to prescribed
(MIIVs): An alternative orientation to structural equa-
burns as we would expect. tion modeling. Multivariate Behavioral Research 54:31–46.
 Article e02962; page 14 JAMES B. GRACE AND KATHRYN M. IRVINE Ecology, Vol. 101, No. 4

Bollen, K. A., and J. Pearl. 2013. Eight myths about causality Lefcheck, J., J. Byrnes, and J. B. Grace. 2018. Package ‘piece-
and structural equation models. Pages 301–328 in S. L. Mor- wiseSEM’. https://cran.microsoft.com/web/packages/piece
gan, editor. Handbook of causal analysis for social research. wiseSEM/piecewiseSEM.pdf
Springer, Dordrecht, The Netherlands. Lindley, D. 2002. Seeing and doing: The concept of causation.
Burnham, K. P., and D. R. Anderson 2002. Model selection International Statistical Review 70:191–214.
and multimodel inference. Second edition. Springer, New Marchetti, G. M., M. Drton, and K. Sadeghi. 2015. ggm: Func-
York, New York, USA. tions for graphical Markov models. R package version, 2.
Cade, B. S. 2015. Model averaging and muddled multimodel https://cran.r-project.org/web/packages/ggm/ggm.pdf
inferences. Ecology 96:2370–2382. Neyman, J. 1923. On the applicability of probability theory to
Cohen, J., P. Cohen, S. G. West, and L. S. Aiken 2003. Applied agricultural experiments. Essay on principles. Section 9. Sta-
multiple regression/correlation analysis for the behavioral tistical Science 5:465–480.
sciences, Third edition. Routledge, New York, New York, USA. Pawlowsky-Glahn, V., and A. Buccianti. 2011. Compositional
Elwert, F., and C. Winship. 2014. Endogenous selection bias: data analysis. John Wiley & Sons, New York, New York,
The problem of conditioning on a collider variable. Annual USA.
Review of Sociology 40:31–53. Pearl, J. 1988. Probabilistic reasoning in intelligent systems.
Fieberg, J., and D. H. Johnson. 2015. MMI: Multimodel infer- Morgan Kaufmann, San Francisco, California, USA.
ence or models with management implications? Journal of Pearl, J. 2000. Causality. Cambridge University Press, Cam-
Wildlife Management 79:708–718. bridge, UK.
Fisher, R., S. K. Wilson, T. M. Sin, A. C. Lee, and T. J. Lan- Pearl, J. 2009. Causality. Second edition. Cambridge University
CONCEPTS & SYNTHESIS

glois. 2018. A simple function for full-subsets multiple regres- Press, Cambridge, UK.
sion in ecology with R. Ecology and Evolution 8:6104–6113. Pearl, J., and D. MacKenzie. 2018. The book of why: The new
Gough, L., and J. B. Grace. 1999. Effects of environmental science of cause and effect. Basic Books, New York, New
change on plant species density: comparing predictions with York, USA.
experiments. Ecology 80:882–890. Pearl, J., and T. Verma. 1991. A theory of inferred causation.
Grace, J. B. 2006. Structural equation modeling and natural sys- Pages 441–452inJ. F. Allen, R. Fikes, and E. Sandewall editors.
tems. Cambridge University Press, Cambridge, UK. KR ’91: Proceedings of the Second International Conference
Grace, J. B., and J. E. Keeley. 2006. A structural equation model on Principles of Knowledge Representation and Reasoning.
analysis of postfire plant diversity in California shrublands. Morgan Kaufman, San Francisco, California, USA.
Ecological Applications 16:503–514. Pearl, J., M. Glymour, and N. P. Jewell 2016. Causal inference
Grace, J. B., D. R. Schoolmaster, Jr., G. R. Guntenspergen, A. in statistics: a primer. John Wiley & Sons, West Sussex, UK.
M. Little, B. R. Mitchell, K. M. Miller, and E. W. Schweiger. Robins, J. M. 1986. A new approach to causal inference in mor-
2012. Guidelines for a graph-theoretic implementation of tality studies with sustained exposure periods—Application
structural equation modeling. Ecosphere 3:art73. to control of the healthy worker survivor effect. Mathematical
Grace, J. B., S. M. Scheiner, and D. R. Schoolmaster Jr.2015. Modelling 7:1393–1512.
Structural equation modeling: building and evaluating causal Rubin, D. B. 1974. Estimating causal effects of treatments in
models. Pages 169–200 in G. A. Fox, S. Negrete-Yankelevich, randomized and nonrandomized studies. Journal of Educa-
and V. J. Sosa, editors. Ecological statistics: From principles tional Psychology 66:688.
to applications. Oxford University Press, Oxford, UK. Shipley, B. 2000a. Cause and correlation in biology: A user’s
Greenland, S., J. Pearl, and J. M. Robins. 1999. Causal diagrams guide to path analysis, structural equations, and causal infer-
for epidemiologic research. Epidemiology 10:37–48. ence. Cambridge University Press, Cambridge, UK.
Hern
an, M. A., and J. M. Robins 2020. Causal inference: What Shipley, B. 2000b. A new inferential test for path models based
if. Chapman & Hall/CRC, Boca Raton, Florida, USA. on directed acyclic graphs. Structural Equation Modeling
Hitchcock, C.2019. Causal models. In E. N. Zalta, editor. The Stan- 7:206–218.
ford encyclopedia of philosophy (summer 2019 edition). https:// Shipley, B. 2016. Cause and correlation in biology: A user’s
plato.stanford.edu/archives/sum2019/entries/causal-models guide to path analysis, structural equations and causal infer-
Keeley, J. E., C. J. Fotheringham, and M. Baer-Keeley. 2005. ence with R. Second edition. Cambridge University Press,
Factors affecting plant diversity during post-fire recovery and Cambridge, UK.
succession of Mediterranean-climate shrublands in Califor- Symonds, M. R., and A. Moussalli. 2011. A brief guide to
nia, USA. Diversity and Distributions 11:525–537. model selection, multimodel inference and model averaging
Keeley, J. E., T. Brennan, and A. H. Pfaff. 2008. Fire severity in behavioural ecology using Akaike’s information criterion.
and ecosystem responses following crown fires in California Behavioral Ecology and Sociobiology 65:13–21.
shrublands. Ecological Applications 18:1530–1546. Textor, J., J. Hardt, and S. Kn€ uppel. 2011. DAGitty: a graphical
Kline, R. B. 2016. Principles and practice of structural equa- tool for analyzing causal diagrams. Epidemiology 22:745.
tion modeling. Fourth edition. Guilford Press, New York, Whalen, M. A., J. E. Duffy, and J. B. Grace. 2013. Temporal
New York, USA. shifts in top-down versus bottom-up control of epiphytic
Laughlin, D. C., and J. B. Grace. 2019. Discoveries and novel algae in a seagrass ecosystem. Ecology 94:510–520.
insights in ecology using structural equation modeling. Ideas Wright, S. 1921. Correlation and causation. Journal of Agricul-
in Ecology and Evolution 12:28–34. tural Research 10:557–585.

SUPPORTING INFORMATION
Additional supporting information may be found in the online version of this article at http://onlinelibrary.wiley.com/doi/
10.1002/ecy.2962/suppinfo