Principal Components Analysis in
Demonstrating Causation
by Tammy Metzger
When the Supreme Court held that the Federal Rules of Evidence superseded the Frye1 common law “general acceptance” test for the admissibility of scientific testimony, they intended to liberalize Rules 702 and 703 regarding expert testimony. Daubert v. Merrell Dow Pharmaceuticals, Inc.2 The Majority specifically stated that they did not intend to open the doors to pseudoscience, but that trial court judges should screen scientific evidence by its methodology, thereby acting as a gatekeeper.3 However, just the opposite has occurred. In the years following Daubert, more scientific evidence has been excluded from courtrooms since the Supreme Court tried to open the doors to cutting edge, unpublished and otherwise not generally accepted research.4 Establishing causation in a toxic tort case is
difficult because the cause of the injury is not so directly
obvious.5
Instead of a precise, known moment of contact and immediate injury,
toxic exposures occur over an extended and unknown period of time
and result in an increased probability of disease. The injury
may take months or years to surface, there may be multiple causal
agents as well as interacting effects, therefore, scientific
testimony is required to establish factual causation. Because
of these difficulties, expert witnesses are essential to a toxic
tort case, and in many cases, they provide the only evidence of
causation. Critics have asserted that in deciding the
admissibility of scientific evidence, judges have collapsed legal
standards of proof into the more rigorous scientific standard,
thereby unfairly barring plaintiff’s claims.6
While some courts have specifically noted the difference between the
two standards,7
many have chosen to exclude expert testimony that does not show
causation to a 95% certainty.8
In Allen v. Pennsylvania Eng’g Corp.,9
The 5th Circuit excluded plaintiff’s causation evidence
showing a correlation between ethylene oxide and brain cancer,
saying numerous studies found no correlation and the agencies’
threshold of proof is lower than that required in tort law.
One possible answer is to apply a statistical method which is useful with epidemiological and other “unsupervised” data sets, which are not under strict laboratory controls. It also can explain things in terms of “more likely than not” with beloved 95% certainty. Principal components analysis (PCA) is a decorrelation technique often used for exploratory data analysis and is especially useful when there are multiple factors. This analysis can sometimes be useful when there are hidden dependencies between different object measures. PCA is commonly utilized in fields such as astronomy, neural networks, psychology, forestry, geochemistry and systems engineering to decrease the number of axes, thereby removing variables not of interest, in the data field in order to facilitate interpretation.10 For example, in ecology field studies, species and environmental data are analyzed via PCA to reveal various aspects of community structure, such as ecological gradients and relationships between species and their environment. This paper explains how principal components
analysis can be used to introduce causation evidence when the
standard scientific tests are not 95% certain of the causation
mechanism but do show with 95% confidence that a given factor is
most likely the cause of a given result. For example, if
toxicologists are 75% certain that chemical XYZ causes fish
mortality at a given exposure, this evidence may be excluded as
invalid knowledge to prove that XYZ actually killed any fish.
But a statistical test that demonstrates with a 95% certainty that
XYZ more likely than not (say 60% probability) caused the fish to
die should be admissible according to the Daubert standard.
To prove causation, evidence must be both
admissible and it must support the burden of proof. Following
the Supreme Court’s Daubert decision, federal trial judges have
restricted11
the admissibility of scientific evidence according to their new “gatekeeping”
role, which requires that they assess the reasoning and methodology
underlying the expert’s opinion and determine whether it is
scientifically valid and applicable to a particular set of facts.
Courts have held conflicting views of what constitutes causation,12
whether probabilistic evidence alone is sufficient to support a
finding of cause in fact, without an explanatory mechanism,13
and whether certain types of epidemiologic, toxicologic and
statistical studies are ever admissible. This section
describes legal and scientific methods of causation, the current
difficulty in linking these two approaches and a method to
scientifically demonstrate causation at the civil burden of proof
standard.
For most of this century American courts relied
on the Frye test to determine the admissibility of expert
testimony, that it “be sufficiently established to have gained
general acceptance in the particular field in which it belongs.”14
Frye had excluded testimony of a “crude” precursor to the
polygraph technique. This test served to exclude unreliable
testimony but it also barred new discoveries.
1.
Daubert v. Merrell Dow
Pharmaceuticals, Inc. In 1993, the Supreme Court changed this federal
standard and ruled that the 1975 Federal Rules of Evidence superceded the Frye test.15
Rule 702 provided for expert testimony to be admissible if
“scientific, technical, or other specialized knowledge will assist
the trier of fact” and the witness is “qualified as an expert by
knowledge, skill, experience, training, or education.”16
The Supreme Court characterizes the Federal Rules of Evidence as
having a “liberal thrust” and “relaxing the traditional barriers to
‘opinion’ testimony.”17 In comparison, they describe the Frye test as “rigid” and
“austere.”18
Furthermore, they espouse the American adversarial system of
“vigorous cross-examination, presentation of contrary evidence, and
careful instruction on the burden of proof are the traditional and
appropriate means of attacking shaky but admissible evidence.”19
In Daubert, the guardians of two children
sued over birth defects allegedly caused by Bendectin, an antinausea
drug manufactured by Merrell Dow Pharmaceuticals. The 9th
Circuit district court followed the Frye test in excluding expert
testimony that Bendectin had caused the birth defects because it was
not “generally accepted” within the epidemiological field. The
defendant’s expert witness testified that numerous studies indicated
no birth defects with Bendectin. The petitioner’s experts
provided in vitro (test tube), in vivo (animal studies),
pharmacological studies that indicated similar chemical structure to
other drugs known to cause birth defects and a reanalysis of
epidemiological research. The 9th Circuit Court of
Appeals affirmed the district court’s summary judgment for the
defendants,20
stating that expert opinion based on scientific evidence must be
based on techniques that are “generally accepted” as reliable in the
relevant scientific community.21 The Supreme Court interpreted Rule 702, testimony
of experts, to supersede the Frye general acceptance test and
replace it with a requirement of reliability and relevancy. In
dicta, they suggested a guideline for courts to determine the
scientific methodology’s reliability, which include whether it has
been generally accepted within the particular scientific community,
peer review and publication, the level of error and whether it has
been tested.22
Furthermore, relevancy includes “fit,” for example the phases of the
moon can be introduced to suggest the brightness on a given night
but not a person’s state of mind.23
The Supreme Court also stressed that it is the methodology, not the
conclusion that is to be judged.24
While expanding the types of scientific evidence that should be admitted, the Supreme Court clarified that this does not mean that “the Rules themselves place no limits on the admissibility of purportedly scientific evidence. Nor is the trial judge disabled from screening such evidence. To the contrary, under the Rules the trial judge must ensure that any and all scientific testimony or evidence admitted is not only relevant, rut reliable.”25 Following Daubert, courts had differing
opinions as to whether the new test applied to technical,
non-scientific, testimony and to what standard appellate courts were
to apply in review. The Supreme Court cleared up these matters
in two more important rulings, General Electric v. Joiner26
and Kumho Tire Co., Ltd. V. Carmichael.27
2.
General Electric Co. v. Joiner Joiner involved a smoker who was exposed
to polychlorinated biphenyls (PCBs) and subsequently contracted lung
cancer. Robert Joiner’s case relied heavily on expert
testimony that PCBs alone can cause lung cancer. This
testimony was based on animal studies, which the District Court
ruled were an insufficient basis for the expert opinions because the
differing exposure levels and pathways between the mice and workers.28
Furthermore, the experts would not definitively say that the PCBs
had caused the cancer in the exposed workers. The
epidemiologic studies were also excluded since one was inconclusive
and the other made no mention of PCBs. The Eleventh Circuit Court of Appeals reversed,
holding “because the Federal Rules of Evidence governing expert
testimony display a preference for admissibility, we apply a
particularly stringent standard of review to the trial judge’s
exclusion of expert testimony.”29
The Court of Appeals further opined that a district court should
limit its role to determining the “legal reliability of proffered
expert testimony, leaving the jury to decide the correctness of
competing expert opinions.”30
On Certiorari, the Supreme Court reversed,
holding that the standard for review for evidentiary matters is
abuse of discretion.31
Furthermore, in response to plaintiffs argument that Daubert
specified that the methodology, not the conclusion, is to be the
basis of the admissibility of the expert’s testimony, the justices
held that “conclusions and methodology are not entirely distinct
from one another. . . A court may conclude that there is simply too
great a gap between the data and the opinion proffered.”32
The Court examined the evidence and determined that the trial court
had not abused its discretion in excluding the expert testimony.33
3.
Kuhmo Tire Co. v. Carmichael The third case in the Daubert trilogy, Kuhmo,34 dealt with nonscientific expert testimony. This case is about the influence of an allegedly defective tire in a fatal minivan accident. The plaintiff’s expert attributed the crash to a defective tire based on his experience as an engineer with extensive experience examining failed tires.35 The District Court excluded the expert testimony by holding that technical testimony was subject to the Daubert test. The visual-inspection test failed the Daubert test of peer review or publication, known or potential rate of error, and its general acceptance within the relevant scientific community.36 As a matter of law, upon a de novo standard of
review, the 11th Circuit held that Daubert only
applied in the scientific context37
and remanded the case back to the district court for consideration
as to whether the testimony was sufficiently reliable and relevant
to assist the jury.38
The court was concerned that the expert’s methodology was unsound
because he had made his conclusions before ever examining the tire.39
The Supreme Court unanimously rejected the 11th
Circuit’s opinion and held that the trial court’s gatekeeping
obligation extends to all expert testimony, for example, including
that of a perfume tester.40
Furthermore, the court unanimously held that the appellate court
erred in applying a de novo standard of review and held that the
proper standard is an abuse of discretion.41
The Supreme Court, as in Joiner, looked at
the evidence and ruled that the trial court did not abuse its
discretion since the expert did not follow his own methodology in
coming to his conclusion that the tire caused the accident.42
The court was bothered by several weaknesses in Carlson’s testimony,
including the fact that he had no idea “whether the tire had
traveled more than 10, or 20, or 30, or 40, or 50 thousand miles.”
43 Yet he was certain that the tire caused the
accident despite evidence to the contrary. Interestingly, the Court seems to have backed away from establishing guidelines, instead, clarifying that the Daubert test depends on the circumstances of the case, is flexible and “may” bear on the judge’s gatekeeping determination.44 The Court clarified the judge’s gatekeeper role as ensuring that “an expert, whether basing testimony upon professional studies or personal experience, employs in the courtroom the same level of intellectual rigor that characterizes the practice of an expert in the relevant field.45
B. Scientific
Evidence of Causation in Fact Science is both a method for discovery and the
resulting body of knowledge. Although day-to-day science is
conducted in a more intuitive and less structured process,46
it is often described as following the Scientific Method, which
consists of systematically observing phenomena, recording facts,
formulating physical laws from the generalization of the phenomena,
and developing a theory that is used to predict new phenomena.
This Method is useful for creating new paradigms,47
when necessary; however, typical scientific studies fill in small
gaps within the current paradigm through observation,
classification, experimentation and hypothesis testing.48
Statistical methods use hypothesis testing to
determine if a relationship between variables is simply the result
of chance. If a relationship is established, the next step is
to determine the strength of that relationship, which is discussed
in the epidemiological measures of risk section of this paper.
Plaintiff’s experts usually testify about studies that meet both of
these elements as distinct steps. This is not the only way to
apply scientific methodologies, but it is the way it is currently
done in toxic torts.
1. Statistical
Techniques and Significance Testing in Assessing Scientific
Integrity While traditional, physical sciences have used
calculus and other particular mathematics, modern science
increasingly relies on statistics of data to explain phenomena.
Statistics describe data, often summarizing numerous observations,
and are commonly used in psychology, astronomy, toxicology,
climatology, economics and physics, just to name a few.
Through regression models, analyzing correlations between variables
and other multivariate techniques, statistics can be used to infer
causation by describing the likelihood that observed relationships
are random. In order to say whether a distribution of
variables is random or not, one must first select the degree of
accuracy required of this conclusion. For example, if a craps
shooter throws four straight 7s, what is the appropriate reaction of
the pit boss? How sure is he that the shooter is cheating?
Is he willing to immediately throw her out of the casino? Should he
stop the game to examine the dice? Perhaps he prefers to watch
the next few throws before he’s certain enough to risk making the
superstitious players unhappy by interrupting their lucky streak. People intuitively make statistical decisions
every day that are weighted by the risks of each circumstance.
Statisticians use p-values to choose these desired certainties in
relationships, typically at the 95% level of confidence that the
observation is not random, which is equal to a p-value of 0.05.
A lower p-value indicates a higher chance that something is going on
besides random error, so the scientists would reject the null
hypothesis49
and find that there is significant evidence that there is a
relationship. Simple linear regression models are used to infer
causation from association. A line is drawn through a set of
data so that the sum of the differences from each data point to the
line squared is minimized. The line is mathematically
described as a slope, which relates the variables. For
example, a regression model that roughly predicts a person’s salary
can be estimated based on their experience. Salary = $15,000 + $2,000(years of experience) In this model, a person with no experience is
predicted to make $15,000 while a person with 10 years of experience
is predicted to make $35,000. This simple linear model would
have relatively low statistical significance since it will have a
high error associated with each prediction. If the model were
improved by adding multiple regressions for the salary differences
of men and women, it would much more accurately predict salaries and
would have a higher level of confidence. This is because the
salary of men with the same number of years of experience is more
than that of women. This difference causes more variability in
the model, thus reducing its predictive value and producing results
that appear more random. Regression model predictions are typically
described by the confidence interval of the resulting estimate.
Higher confidence in an estimate is obtained when the data more
closely fit the model, there are more data to use in the model,50
there are few outliers,51
and there are no other influencing factors. Sometimes
confounding52
variables that are missing from the analysis are the actual cause of
observed relationships, so a significant relationship is not
necessarily causal.
2.
Epidemiological Measures of Risk
Epidemiologists study the incidence, distribution and etiology of
disease in populations and the influence of the environment and
lifestyle on disease patterns; for example, how tobacco has affected
the health of a particular population.
Epidemiologists focus on general causation, things that can
cause disease, rather than specific causation, such as the cause of
disease in a specific individual.
These studies are uncontrolled, meaning not under laboratory
conditions, since that would obviously be unethical treatment of
humans.53
The most common measures used to estimate the association
between exposure and risk are relative risk, odds ratio and
attributable proportion of risk (APR).
Since relative risk is most frequently used in litigation, only this
measure will be described.
Relative risk is the ratio of the incidence of disease in exposed individuals compared
to the incidence in unexposed individuals.54
Most courts have held that a relative risk of 2.0 or greater
is sufficient to prove causation by a preponderance of the evidence,
since this amounts to a doubling of risk.55
Example 10% of people with a given exposure contract a
specific disease compared with 5% of normal population, therefore RR
= .10/.5=2.
C.
Collapsed Legal and Scientific Standards in Practice
There has been much debate within the legal community whether
plaintiffs are unfairly held to a higher burden of proof when their
main evidence of causation is scientific.56
Some courts have held that scientific evidence should be
admitted at a relaxed standard so that it does not raise the
plaintiff’s burden of proof beyond a preponderance of the evidence
requirement.57
This is problematic because scientists are reluctant to make
assertions based on lower levels of certainty than what is required
in their professional journals.
Although much scientific work with results below the 95% certainty
now required in most courtrooms is considered valid and is published
in scientific journals, most lower certainty techniques are used
primarily for exploratory research, where these results are further
investigated via more robust methods, such as regression models.
After Daubert, some
courts have continued to distinguish between scientific certainty
and legal sufficiency,58 however,
most courts have insisted on evidence that meets the relevant
scientific methodology of the given field59 following
Daubert.
While it is true that science is not
amenable to reduced significance standards,60 that has
not deterred suggestions for modified approaches, such as reduced
certainty for scientific evidence.61
This idea, however, does not fully use the capabilities of
science to contribute to legal proceedings because p-values of 0.5
mean virtually nothing to scientists.
There are limitless ways to work with science and its strict
methodologies without transferring an undue burden of proof onto a
plaintiff. The trick is
to keep in mind the purpose of the investigation.
Instead of aiming to unveil universal truths, science should
be applied to determine probabilities in legal, “more likely than
not”, terms.
D.
Using Principal Components Analysis (PCA) to Demonstrate
Causation
Principal components analysis, an exploratory data analysis
technique, is useful for civil litigation because it interprets
complex environmental data more clearly than multiple regressions
and it specifies the association between each variable and the
outcome. The goal in PCA
is not always to perfectly model a phenomena, but to quickly
understand relationships between the independent variables and the
outcome at whatever completeness the researcher wishes to discern,
including 51%.
Most commonly used in discerning environmental relationships,62 PCA is
also used in various scientific fields such as astronomy, neural
networks, psychology, forestry, geochemistry, climatology and
systems engineering.63
1.
Introduction to Principal Components Analysis
Principal components analysis64 is a
multivariate statistical technique that works by considering only a
smaller number of these variables, which the researcher can
designate. Each
variable’s influence on the entire dataset’s variability is
calculated, along with the cumulative impact of a given group of
principal components.
Mr. Kendall65
originally proposed PCA to analyze data that does not work well in
regression models. The
technique obtains the principal components66 of a set
of explanatory variables, calculating their regression upon a
dependent variable, and projecting the resulting parameters back
into the terms of the original variates.
The principal components are orthogonalized, which eliminates
collinearities67 in the
dataset.
If you imagine the dataset as a cloud, graphed in as
many dimensions are there are descriptive variables, the first
principle component is the line that would extend furthest through
this cloud, thereby catching the most variability.
If you imagine that this line represents the line from each
corner of your mouth, as is done in lipreading programs, the form of
the lips can be described most efficiently from the first principal
component, the basic line across the mouth.
This is how PCA is used to compress data.
This is also useful to remove factors that are confusing the
analysis. The main
component should not be removed, but lets say that the person
speaking is nervously tapping and this motion is causing their face
and lips to tremble, thereby interfering in the speech recognition
program. PCA can remove
this “noise,” which is how it is used in climate studies, where
numerous subtle effects distort the more easily understandable major
climate influences, like latitude, elevation and El Niño events.
While scientists usually group these to explain as much of the outcome as
possible, they could also be grouped or singled out to explain 51%
of an outcome, such as the variables that explain cancer.
Although there are many factors which contribute to lung
cancer, a PCA would be able to separate out the factor or factors
that more likely than not caused a specific cancer.
Furthermore, other contributing variables that may be missing
from an individual’s case, such as the possibility that they never
smoked, could be included in the calculation to add confidence to a
prediction of the present cancer causing variables, such as work
exposure, via Bayes Theorem.
Intuitively, the chance that the work exposure caused the
cancer increases when other causes are eliminated.68 Standard
statistical programs describe significance parameters, such as the
confidence interval, covariance and standard deviation of these
principal components.
Factor analysis is a commonly used statistical technique that
measures an unobservable element by combining the measured ones by
using linear combinations of variables to explain sets of
observations of many variables, just like principle components. For
example, factor analysis is useful for intelligence tests, where
observed variables are various test score results.
Psychologists do not care about the various tests, but the
intelligence they were designed to describe.
The difference between the two is that in PCA, the observed
variables themselves are the focus of interest, whereas in factor
analysis, it is the unobservable element that is of interest.
The principle components approach simplifies the
interpretation of confounding variables, whereas factor analysis
disregards those observed variables and looks at the underlying
factor.
2. Hypothetical Application of
Principal Components Analysis to Lung Cancer
Since this paper merely aims to lay out the general concept of how
to use PCA for legal evidence, the numerous steps required to
produce a resulting table like the example below have been omitted.
Perhaps the weakest aspect of this approach is the actual
assignment of the principal components to the nearest independent
variable. Although it
involves some subjectivity, scientific studies utilizing PCA are
generally accepted in numerous fields and authoritative
publications.69
Suppose a PCA is run on 20,000 people from Texas who have lung
cancer and the analysis includes data on whether they smoke, live
with someone who smokes (both measured in packs per day), smog
levels outside their home and known work exposure to asbestos and
other high risk materials (both combined indices).
Example 1.
A particular plaintiff from Texas claims that he contracted
lung cancer because he and his wife both smoked cigarettes for half
a century, so he sues several tobacco companies under a market share
liability claim.
According to the PCA, the fact that he and his wife both smoke
explains 55% of the incidence of lung cancer.
This is calculated by adding components 1 and 3 while
skipping component 2 since the tobacco industry is not entirely
responsible for this variable.
Example 2.
A plaintiff who smokes but does not live in a smoggy area.
Here, only 40% of the variance is explained by smoking.
Some jurisdictions will allow this evidence to be introduced
to demonstrate a substantial factor of causation, but it would be
prudent to more rigorously combine the additional information to
establish a preponderance of the evidence that “but for” the
smoking, the plaintiff would not have contracted lung cancer.
Bayes’ Theorem describes how to link distinct probabilities of
specific conditions into the total combined probability given those
conditions.
Bayes’ Theorem
P(C/E) =
P(C ) * P(E/C )
where, P(C )
the current probability plaintiff did not contract cancer
from cigarettes E
new piece of probabilistic evidence P(C/E)
new
probability of something, P(C ), given E
(P(E/C)
the probability of E given C (probability that smog causes
cancer = 0.2) P(not C)
the probability that C is not true (1-C = 0.4) P(E/not C).
the probability of E if C were not true (since we have
evidence that the P(C/E) =
0.6 * 0.2 = 0.75
The combined information gives a new estimate that smoking is
75% likely the cause of the cancer for a person who smokes and lives
outside a smoggy area.
This could be further improved by accounting for the missing smoking
housemate and possibly the effects of work exposure.
3. Possible Challenges to using PCA to show Causation
One problem with this analysis is that the 9th Circuit
Court of Appeals added to the Supreme Court’s
Daubert opinion that
research done for litigation is a factor in excluding evidence.
In Response, this is unfair since much research is done
specifically for litigation, as in environmental assessments to
determine who is a potentially responsible party for the clean-up.
Under Daubert and
Kuhmo, most of these
environmental studies should be admitted as long as the methodology
is sound and follows generally accepted principles that were not
developed solely for litigation.
Many jurisdictions require particular evidence to show causation for the
individual plaintiff, therefore, statistics alone may be admissible
yet insufficient. Here,
a mechanism is needed, such as the etiology of the injury.
However, courts need to know that not every mechanism is
understood. The expert
should discuss possible mechanisms, based on scientific knowledge,
and leave that factual question for the jury.
Also, additional statistical tests, such as time-series
analysis, relating to the timing might sufficiently support
association evidence in proving causation.
Opposing counsel may challenge the methodology of exploratory
PCA evidence as unscientific and only suggestive of how the
scientist should proceed to conduct a meaningful investigation.
Although PCA is used in many fields and the results are
published in reputable journals, it has not yet made an appearance
in toxic tort litigation.
Hopefully this will soon change.
Principal components analysis works well in the context of civil
litigation, where a preponderance of the evidence standard of proof
is sought, since PCA can separate out the individual influence of
each independent variable.
Furthermore, it is especially well-suited for environmental
datasets, where there are many confounding influences.
Since this technique is generally accepted in the scientific
community and used in numerous scientific fields, it should be
admissible under Rule 702 when used appropriately.
Specific PCA studies, if reasonably conducted, should follow both
the law and spirit of Daubert to admit reliable, relevant and sound scientific evidence.
This approach is both scientifically and legally more
credible than other suggestions to avoid the collapsing standards
problem, where plaintiffs’ are sometimes held to an unfair, higher
standard of proof when their primary proof of causation is
scientific evidence.
This combined approach of PCA and Bayes’ Theorem is useful for
environmental and toxic torts, where there is not conclusive
laboratory experimental research to support the real-world evidence
of causation, such as epidemiological associations.
PCA is commonly used to discover ecological and climate
influences and it should be applied in toxic tort litigation to
bring polluters to justice and thereby reduce disease and resource
degradation.
The Supreme Court intended to liberalize the Federal Rules of
Evidence so that cutting edge technology would be available in the
courtrooms. PCA is
precisely the type of technique that is needed in the courtrooms to
facilitate the fair resolution of conflicts.
PCA combines a scientifically rigorous method with the
flexibility to find causation at the civil burden of proof, which is
an ideal balance to ensure justice and fairness in toxic tort
litigation. JuriSense, LLC Seal Beach, CA (800) 891-6592 info@jurisense.com Home | Research | Expert Testimony | Jury Selection | Graphics | CLE | Tammy Metzger | Contact | Papers | Blog |