MENDELIAN RANDOMIZATION FOR CAUSAL INFERENCE OF
BLOOD-RELATED RISK FACTORS FOR INFECTIOUS DISEASES
Aditya Mittal
James Logan High School
Union City, CA 94587
adityamittal307@gmail.com
July 1, 2020
ABSTRACT
This study determines significant causal relationships between blood-related risk factors and infec-
tious diseases using two-sample Mendelian Randomization in order to research possible causes of
infectious diseases and develop treatments to combat these infectious diseases. Currently, in the
biomedical field, there is little to no way to differentiate between correlation and causation when
determining the relationship between exposures, the factors that could affect the chance of the dis-
ease occurring, and outcomes, the diseases. To provide a statistical causal inference, two-sample
Mendelian Randomization is used against different risk factors for infectious diseases. For more
significant analysis, a random-forest classifier named MR-MoE was used to determine the optimal
Mendelian Randomization model for each exposure-outcome association. Significant causal rela-
tionships between blood-related risk factors and infectious diseases found from this genetic study
are the following: haemoglobin count and bleeding gums, lymphocyte counts and tonsillectomy,
haemoglobin count and tonsillectomy, and neutrophil count and wheezing in the chest. Due to the
large cost and ethical concerns associated with human trials, Mendelian Randomization is a useful
method for research that will undoubtedly influence genetic research for years to come.
1 Introduction
When testing for the causal relationship between an exposure and outcome using human trials, humans provide hundreds
of outside factors, otherwise known as confounding factors, that could skew the trial data and provide false results. To
reduce statistical bias, it is necessary to test genetic data instead of conducting human trials. To analyze the genetic
variants statistically, Mendelian Randomization is used, which uses genetic variants to determine the causal effect of an
exposure on a disease.
Mendelian Randomization provides an effective way to determine relationships between two genetic factors without
conducting a traditional trial [
1
]. This process extracts genetic variants or instrumental variables from the provided
exposures and outcomes, and then harmonizes the data into a single dataset upon which the randomization could be
run upon. The result of this randomization produces two key statistics that are instrumental in determining the causal
relationship: the p-value and the standard error. The p-value provides the probability that the null hypothesis – the
hypothesis that the risk factor has no affect on the outcome – is true, which normally requires it to be extremely low
for the result to be accepted. Furthermore, the standard error is the standard deviation of a statistic that is caused by
experimental uncertainty. These two statistics are ultimately the final results of the randomization experiment.
However, there are many limitations to the Mendelian Randomization model, such as the assumptions that must be
made before testing the casual relationship between an exposure and outcome. There are three main assumptions:
1. All instrumental variables (genetic variants) must be significantly associated with the treatment.
2. All instrumental variables must not be correlated with confounding factors.
3. Instrumental variables cannot be correlated with another exposure with a causal relationship to the outcome.
A PREPRINT - JULY 1, 2020
Figure 1: Diagram highlighting the conditions and relationships between variables [4].
This diagram highlights two key conditions that cannot occur: a relationship between the genetic variants and confounding factors
and a relationship between the genetic variants and the outcome. The latter is commonly referred to as horizontal pleiotropy, and
reduction of this is essential for a successful Mendelian Randomization model.
The third assumption is commonly the most difficult assumption to validate in genetic studies, as there are hundreds of
confounding factors that could affect the model. Although there are many limitations to the randomization model, when
properly run, it will provide an accurate result in a cost-effective and ethical manner. Human trials could cost billions of
dollars to perform properly with a large enough sample size, and some trials provide moral implications that are simply
not there with Mendelian Randomization.
The importance of the outcomes that are being tested are varied, but all stem out of infectious diseases. Combined, these
diseases account for thousands of deaths every year, which cannot be treated by traditional methods. Because there
are very limited treatments that can be applied to these bacteria, this research aims to expand the array of treatments
available by testing a set of exposures across different forms of infectious diseases. These exposures include the
following: lymphocyte counts, neutrophil count, and mean cell hemoglobin count. Exposures were chosen based on
whether they were blood cells; the aforementioned examples are just some of the numerous blood cells tested in this
experiment. The infectious diseases were chosen due to their abundance in the general population, and all pertain to
dental or lung problems.
The randomization models also provide an interesting point of discussion in the study. This is because when these
models are tested with the same exposures and outcomes, they produce different p-values and statistic error calculations
from one another. To provide the best results, a random-forest algorithm is ran to determine the optimal model that will
be used for later analysis.
2 Methods
2.1 Overview
As this paper heavily relies on knowledge of Mendelian Randomization, it is necessary to delve into greater depth
about what exactly it does. There are two kinds of Mendelian Randomization, but for this experiment, two-sample
Mendelian Randomization will be used, due to its statistical power and its greater applicability in this experiment [
2
].
Mendelian Randomization uses genetic variants as an instrumental variable, a third variable used to determine the
causal relationship between two variables with unexpected behavior, to provide a relationship between a risk factor and
a disease [
3
]. Over the years, Mendelian Randomization has become more optimized as more genetic-wide association
studies (GWAS) have been run and compiled. As explained above, Mendelian Randomization has many conditions that
must be met before the randomization can be run, and Figure 1 shows this in greater detail.
Each of these genetic variants is a segment of the genetic code where the nucleotides are altered for different groups
of people. This is commonly referred to as a single nucleotide polymorphism or SNP. If an SNP is valid, then the
causal relationship between an exposure and outcome (denoted by
β
) is determined by the SNP-outcome association
2
A PREPRINT - JULY 1, 2020
(denoted by
Γ
) and the SNP-exposure association (denoted by
γ
) [
3
]. This provides an estimated value that can be used
to experimentally determine β. The equation is as follows:
ˆ
β
ratio
=
ˆ
Γ
ˆγ
(1)
Using this equation for
β
, it is possible to now experimentally determine it using the two-stage least squares (TSLS)
method on the SNP-exposure-outcome dataset. There are three steps to achieve this:
1. First, a linear regression is performed on the exposure through the SNP via the model [3].
X|G
j
= γ
0
+ γ
j
G
j
+
XY
(2)
Taking the slope of this line will provide
γ
j
, which is a necessary component of Equation 1. The graph contains
the relationship between the exposure and instrumental variables, which is why the slope provides this key
term. However, this is the γ
j
for a single SNP G
j
, not the model as a whole.
2. Next, a linear regression is formed on the fitted values from Equation 2 (
ˆ
X
j
) and the outcome [3].
Y |
ˆ
X
j
= Γ
0
+ Γ
j
ˆ
X
j
+
XY
(3)
Taking the slope of this line would provide
Γ
j
, which is another important component of Equation 1. The
graph of the linear regression contains the relationship between the outcome and the SNP, which is why the
slope provides the key term. However, this is the Γ
j
for a single SNP G
j
, not the model as a whole.
3. Next, use Equation 1 to find the (
ˆ
β
ratio
)
j
for each SNP j.
After experimentally determining
β
j
for each SNP
G
j
, the last step is to find
β
for the model using the
β
j
for each SNP.
This changes drastically from model to model, as shown in the next section; however, one model commonly used is
the inverse variance weighted model, also known as IVW. IVW generates the causal effect (denoted by
ˆ
β
IV W
) in the
following way [5]:
ˆ
β
IV W
=
P
j
ˆγ
2
j
σ
¯
Y
2
j
ˆ
β
j
P
j
ˆγ
2
j
σ
¯
Y
2
j
(4)
where
ˆ
β
j
is the estimated ratio between the exposure and outcome for the the
j
th SNP and
σ
¯
Y
2
j
is the standard deviation
of the exposure for the jth SNP.
Now that the methodology behind Mendelian Randomization is clarified, it is necessary to clarify where the GWAS
data will come from that will provide these SNPs to perform the randomization upon. A platform containing a database
of GWAS known as MRBase was used to provide this crucial data. TwoSampleMR’s R package was used extensively
to run the two-sample Mendelian Randomization and allows for an efficient evaluation of the relationships produced
in these association studies [
6
]. This dataset was used to find potentially interesting risk factors towards infectious
diseases, and perform two-sample Mendelian Randomization if there was, indeed, a causal relationship between the risk
factor and outcome. Its 11 billion SNP-outcome associations from over 1600 GWAS provide an ample amount of data
to produce a statistically-backed relationship [6].
However, due to the overwhelming amount of Mendelian Randomization models out there that produce different
statistics, the data provided by two-sample Mendelian Randomization can be confusing and difficult to determine the
optimal model that fits the dataset. Each model fits the dataset in a different way due to the varying assumptions each
model takes. To fix this, a classifier known as MR-MoE chooses the most appropriate model for a specific causal
relationship analysis [
7
]. This classifier runs a random-forest algorithm consisting of several weak learners that assigns a
score to each of the Mendelian Randomization models based on how well it performed in testing the causal relationship.
It is important to note that MR-MOE will assign the best model depending on the particular dataset (exposure-outcome
pair) being tested, not the best model for every exposure-outcome pair.
Because the random-forest algorithm is used extensively, learning about the algorithm will help immensely in under-
standing what MR-MoE does. A random-forest algorithm is a supervised machine learning algorithm that uses multiple
small decision trees together to form a plausible hypothesis. Each decision tree is built on an overlapping subset of
the data, and is fed in training input to train the model. The most frequent answer from these trees forms the general
hypothesis provided by the random-forest algorithm. Some disadvantages of this algorithm are the extensive number
3
A PREPRINT - JULY 1, 2020
of hyperparameters (parameters assigned by humans before training the model), such as the pruning rate, forest size,
subset sample, and number of iterations [
8
]. As the model grows more complex, however, random-forest algorithms
work well in making sure the data is less prone to overfitting, where the model performs well with training data but
terribly with testing data.
2.2 Mendelian Randomization Models
The data provided by the Mendelian Randomization has a wide array of models that each produce different p-values
and standard errors. To choose the best model, a random-forest classifier is run that helps determine the best model
by assigning a score. The original two-sample Mendelian Randomization dataset contains eleven different models
that the Mendelian Randomization is run upon: simple mean, FE IVW, RE IVW, FE Egger, RE Egger, simple median,
weighted median, penalised mean, simple mode, weighted mode, and penalised mode. Many of the models use the
ratio between the
ˆγ
(effect size between the exposure and the SNPs) and
ˆ
Γ
(the effect size between the outcome and
the SNPs). Additionally, a term frequently used is heterogeneity, which is the phenomenon where the production of
the same phenotype is caused by different genetic mechanisms. More information about the types of models is shown
below:
1. Simple Mean:
A jackknife approach is taken with the data where there is a random selection with replacement
of the set of instrumental variables. This will create a sampling distribution, which over hundreds of rounds of
jackknife estimates, can help obtain a final estimate based on the mean of the distribution [7].
2. FE IVW:
The FE IVW model weights SNPS by the inverse variance of the SNP’s effect on the outcome. It
then runs a weighted fixed-effect meta-analysis on the resulting weights for the SNPs [9].
3. RE IVW:
Unlike the FE IVW model, the IVW random effects meta analysis reduces the horizontal pleiotropy,
so that there is increased heterogeneity around the line created by the linear regression model [7].
4. FE Egger:
The FE Egger model runs a weighted linear regression of the ratio between
ˆγ
j
and
ˆ
Γ
j
for each of
the instrumental variables
j
. The slope of this line provides the necessary causal effect [
10
]. Additionally, it
reduces horizontal pleiotropy by allowing a non-zero intercept, which allows for the total effect to influence an
outcome in a certain direction [7].
5. RE Egger:
The random effects Egger model allows heterogeneity around the slope while accounting for
directional horizontal pleiotropy. It makes an assumption that the horizontal pleiotropy effects are not related
to the SNP-exposure effects [7].
6. Simple median:
The simple median model takes the median of the ratio between
ˆγ
j
and
ˆ
Γ
j
for each of
the instrumental variables
j
[
10
]. This gives the advantage that only half of the instruments need to remain
unbiased for the model to run accurately [7].
7. Weighted median:
The weighted median model takes the weighted median of the ratio between
ˆγ
j
and
ˆ
Γ
j
for each of the instrumental variables
j
[
10
]. The weighted median takes into account the contribution of each
instrumental variable by the inverse of the variance [7].
8. Penalized median:
The penalized median model takes the median of the ratio between
ˆγ
j
and
ˆ
Γ
j
for each of
the instrumental variables
j
. However, this ratio is changed to penalize some of the weights based on their
significance [
11
]. The penalization works by penalizing any instrument that contributes to the heterogeneity
statistic [7].
9. Simple mode:
The simple mean model clusters SNPS into groups based on similarity to causal effect estimates.
Then, it estimates the causal relationship by taking the cluster with the largest number of SNPs [10].
10. Weighted mode:
The weighted mode model weights the SNPs by the inverse variance of the SNP’s effect on
the outcome. Then, it estimates the causal relationship by taking the cluster with the largest number of SNPs
[10].
11. Penalized mode:
The penalized mode performs relatively the same as the simple mode model; however, the
ratio is changed to penalize some of the weights based on their significance [
10
]. The penalization works by
penalizing any instrument that contributes to the heterogeneity statistic [7].
2.3 Terminology
When attempting to analyze the data, it is important to understand key terms that occur inside the data. Here are some
terms and their definitions:
4
A PREPRINT - JULY 1, 2020
Figure 2: Flowchart of the analysis project.
This flowchart represents the steps taken to run the experiment, and includes files, datasets, and descriptions of each. First, exposures
and outcomes were extracted from MRBase and harmonized into a singular database containing rows of associated SNPs. Next,
two-sample Mendelian Randomization was run on the dataset, and the best model was chosen using the MR-MoE classifier for each
exposure-outcome association. Then, the best model for each of the exposure-outcome association was compiled into a singular
dataset. Additionally, scatterplots of each association were created to visualize the exposure-outcome relationship.
5
A PREPRINT - JULY 1, 2020
1. id.exposure: The ID of the exposure inside MRBase.
2. id.outcome: The ID of the outcome inside MRBase.
3. method: The Mendelian Randomization model used to generate the data.
4. nsnp: The number of single-nucleotide polymorphisms that are pertinent to the model.
5. b:
ˆ
β
ratio
or the causal relationship between an exposure and outcome.
6. se: The standard error or the experimental uncertainty of the model.
7. ci_low:
Using the standard error, the lower bound of the confidence interval is determined by multiplying a
constant by the standard error, and then subtracting it from
ˆ
β
ratio
.
8. ci_upp:
Using the standard error, the higher bound of the confidence interval is determined by multiplying a
constant by the standard error, and then adding it to
ˆ
β
ratio
.
9. pval:
The p-value of the data or the probability that there is no relationship between the exposure and outcome.
10. MOE:
The score produced by the MR-MoE model by running a random-forest algorithm on the Mendelian
Randomization model.
11. steiger-filtered: Steiger filtering restricts SNPs that have a stronger relationship to the outcome than they do
to the exposure [12]. This boolean value shows whether the model implements Steiger filtering.
12. outlier-filtered:
Outlier filtering systematically removes SNPs that have drastically different data to that of the
other SNPs, so that it will not affect the model [
13
]. This boolean value shows whether the model implements
outlier filtering.
2.4 Analysis Pipeline
During this experiment, multiple steps were taken to achieve the results achieved. First, viable outcomes were extracted
from the MRBase repository. To determine the viability of the outcomes, a variety of different factors were taken
into consideration, such as the number of cases, the number of controls, and the number of SNPs associated with the
outcome. Next, viable exposures were chosen based on these outcomes from the MRBase repository. To determine
the viability of the exposures, many factors were considered, including the sample size, its relevancy to the outcomes
being tested, and the number of SNPs associated with the exposure. Using these viable exposures and outcomes, more
data was extracted from MRBase, including the IDs, traits, sample sizes, and number of cases. This data was split into
datasets, based on whether it was exposure or outcome data. For each exposure-outcome association, the exposure and
outcome data were harmonized together to form a singular dataset comprising of rows of SNPs or genetic variants.
Next, for each exposure-outcome association, the two-sample Mendelian Randomization was run on the harmonized
dataset, producing a dataset consisting of different Mendelian Randomization methods, their p-value, and their standard
error. Then, the random forest classifier was run on the dataset to provide scores for each of the models. A compilation
of datasets was made for each exposure-outcome association that contains the scores for each model within these
associations. Additionally, in this step, scatterplots were graphed where each data point is a SNP
j
mapped based on
its
γ
j
and
Γ
j
. A collection of these graphs was made for each exposure-outcome association. Lastly, with the former
compilation of datasets, the model with the highest MOE score for each exposure-outcome association was chosen
and placed into a new dataset. This dataset contains the optimal Mendelian Randomization model, its MOE score, its
pvalue, and its standard error each association tested. This dataset was then sorted by p-value, and the names of the
exposures and outcomes were appended for easier readability of the dataset. A detailed diagram of the experiment is
seen in Figure 2.
A link to the comprehensive list of findings and the code used to run the experiment can be found here:
https:
//github.com/adityamittal13/MR_Infectious_Diseases.
3 Results
Below is a comprehensive list of the most significant findings from the experiment, and a discussion into the details of
each relationship. In the experiment, "multiple testing" was accounted for, which is the problem where more erroneous
inferences are likely to occur when more inferences are made. For each infectious disease, six plausible risk factors
are being tested, which means that mutliple testing will be used by a factor of six. Although there are six infectious
diseases being tested, each infectious disease tested is treated as its own experiment, reducing the multiple testing from
a factor of thirty-six to a factor of six. To account for multiple testing, the p-values have been adjusted accordingly.
6
A PREPRINT - JULY 1, 2020
Figure 3: Scatterplot of the relationship between haemoglobin concentration and bleeding gums.
Each dot represents a single SNP mapped based on its effect on both haemoglobin concentration and bleeding gums. From this
scatterplot, a line of best fit is made for each Mendelian Randomization method. Since this line has a positive slope, it indicates a
positive effect.
3.1 Mouth/teeth Dental Problems: Bleeding Gums and Haemoglobin Count
The most significant risk factor associated with bleeding gums from this experiment, is haemoglobin concentration, with
a multiple-testing adjusted p-value of
1.4 10
3
and a standard error of
4.6 10
3
.
β
is
1.7 10
2
, which indicates a
positive relationship between haemoglobin concentration and bleeding gums, which therefore shows an increase in
haemoglobin concentration entails an increase in the chance of bleeding gums. This result can be further shown in the
SNP-exposure association vs. SNP-outcome association graph in Figure 3.
Using the MR-MoE, it was determined that the weighted mode model would provide the most accurate p-value. Looking
at Figure 3, it is observed that the weighted mode is the red line of best fit. Using this knowledge, it is possible to
experimentally determine
β
from the graph, as the SNP-exposure effect is
γ
, and the SNP-outcome effect is
Γ
. Using
Equation 1,
β
is the slope of the line, which is positive, further backing the claim that there is a causal relationship. Due
to the large number of SNPs (256) associated with the causal relationship, the causal relationship is backed by the large
amount of statistical data associated with it.
To further understand the causal relationship, it is necessary to delve into the biological aspect of the risk factor and
outcome. Haemoglobin is an iron-containing protein that is crucial in the delivery of oxygen from the lungs to the
tissues in the body. There are many other functions of haemoglobin, including the carrying carbon dioxide to the lungs
as carbaminohaemoglobin, nitric oxide metabolism, and the buffering of hydrogen ions in the conversion from carbon
dioxide to bicarbonate [14].
Bleeding gums is one of the most common types of gum disease and is largely caused by gingivitis and periodontitis.
Gingivitis is caused by bacterial plaque that accumulates near the gingival sulcus, which is an area between the teeth
and surrounding gingival tissue. Due to poor hygiene and other etiologic factors, plaque accumulates near this area,
7
A PREPRINT - JULY 1, 2020
Figure 4: Graph of the patient’s haemoglobin concentration in gm% before and after the operation [17].
The graph shows the patient’s haemoglobin concentration in gm% during the five months before the operation at a consistently lower
level, while the patient’s haemoglobin concentration during the six months after the operation is at a consistently higher level.
progressing the infection and causing an increase in microbial colonization. Some species specifically involved in the
progression of gingivitis are Streptoccous, Fusobacterium, Aclinomyces, VeiUonella, and Treponema and possibly
Bacteroides, Capnocytophaga, and Eikenella [
15
]. Periodontitis, on the other hand, is an inflammatory disease inside
the periodontum, which are specialized tissues to support the teeth. This disease is the progression of the destruction of
these tissues, and is caused by a combination of over 300 species of bacteria [16].
Although the relationship between haemoglobin count and bleeding gums has been explored, it has only been observed
as bleeding gum’s effect on haemoglobin levels. Much research points towards the hypothesis that bleeding gums cause
a lower haemoglobin count. For example, a case study from the Journal of Clinical and Diagnostic Research shows
an 18-year old male patient that suffers from general gingivitis and localized periodontitis, which are general causes
of bleeding gums as indicated above [
17
]. When studying haemoglobin concentration against the patient’s periodic
episodes of gum bleeding, it was found that the patient normally bleeded at lower haemoglobin levels. Additionally, as
shown in Figure 4, the patient had lower haemoglobin levels before the operations, and higher levels after the operation,
which indicates that bleeding gums was indeed the cause of low haemoglobin levels. The reason this study cannot be
used to study the causal relationship of haemoglobin count on bleeding gums is because there were many confounding
factors that could have altered the relationship. The patient was suffering from severe Glanzmann’s Thrombasthenia,
a rare blood clotting disorder that impairs the platelets, causing abnormal bleeding. Since this disease affected the
platelets, blood would not be able to clot, which means that the patient’s bleeding gums would have been exacerbated.
The cause of the bleeding gums may have been from the disorder, low haemoglobin levels or other confounding
factors, which is why it is imposssible to establish a causal relationship without more data. This is a prime example of
correlation vs. causation, because although low haemoglobin levels correlate to bleeding gums, there is no proof that
they are the cause of it through this case study.
3.2 Operation: Tonsillectomy and Lymophcyte Count
When determining risk factors that provided a causal relationship to tonsillectomy, there were two that provided
significant p-values: lymphocyte count and haemoglobin count. The former will be talked about in this section, while
the latter will be talked about in the next section. The multiple-testing adjusted p-value for this exposure-outcome
association was
9.6 10
3
, while the standard error was
8.3 10
3
.
β
is
2.7 10
2
, which indicates a positive
relationship between lymphocyte counts and tonsillectomy. To clarify, this means that a higher lymphocyte count
will increase the chance of tonsillectomy. A more detailed result can be shown in the SNP-exposure association vs.
SNP-outcome association graph in Figure 5.
From MR-MoE, it was determined that the penalized mode model provided the most accurate p-value and model to use
for the causal relationship. Because this scatterplot graphs
γ
against
Γ
, it is possible to experimentally determine
β
by using Equation 1, which takes the slope of the line to find
β
. The slope is positive, which further backs a positive
relationship between the risk factor and the disease.
As done above, an in-depth biological analysis of both the exposure and outcome is necessary to understand what
exactly the causal relationship entails. Lymphocytes are white blood cells that play a vital part in executing the immune
8
A PREPRINT - JULY 1, 2020
Figure 5: Scatterplot of the relationship between lymphocyte count and tonsillectomy.
Each dot represents a single SNP mapped based on its effect on both lymphocyte count and tonsillectomy. From this scatterplot, a
line of best fit is made for each Mendelian Randomization method. Since this line has a positive slope, it indicates a positive effect.
system’s functions. There are three main types of lymphocytes: B cells, T cells, and NK cells. B cells provide a ready
defense against pathogens; they are used for a wide variety of functions, including antibody production, which are
the proteins produced by the immune system to repel antigens [
18
]. T cells, on the other hand, help suppress extreme
immunological responses by the host, act to destroy cells in the body taken over by viruses, and work to maintain
immunological unresponsiveness to self-antigens [
19
]. NK cells or natural killer cells respond quickly to antigens, but
specialize in the killing of cancer and virus-infected cells.
Tonsillectomy is the removal of tonsils at the back of the throat. Part of the lymphatic system, the tonsils are soft tissue
located at the pharynx, and has tissue similar to lymph nodes. Tonsillectomy is commonly used for treating problems
related to rare diseases in the tonsils. It is also a common procedure to remove tonsillitis if the patient doesn’t respond to
other treatments or to prevent recurring episodes of it [
20
]. Additionally, it is generally only administered for tonsillitis
if the bacterial infectious disease causing tonsillitis will not improve with antibiotic treatment. For enlarged tonsils, a
tonsillectomy will commonly be performed if there is difficulty breathing or difficulty swallowing. For other diseases, a
tonsillectomy will be performed if there is a cancerous tissue or recurring bleeding from the tonsils [20].
This relationship has been researched based on tonsillectomy’s impact on lymphocyte counts. For example, a study
conducting human trials found that shortly after undergoing tonsillectomy, many patients had a decrease in cellular
immunity; however, the cellular immunity of the patients came back to normal after a few hours [
21
]. This shows that
tonsillectomy causes a short-term decrease in lymphocyte counts but causes no change long-term. The long-term result
shows the body’s reaction to the surgery as it tries to obtain homeostasis within the body. However, the short-term result
shows a direct relationship between tonsillectomy and reduced lymphocyte count.
Another study finds a relation between an increase in lymphocyte counts and recurrent acute tonsillitis, which is a
common reason for tonsillectomy, as referenced above. The study discovered that there was a significant increase in
9
A PREPRINT - JULY 1, 2020
Figure 6: Scatterplot of the relationship between haemoglobin count and tonsillectomy.
Each dot represents a single SNP mapped based on its effect on both haemoglobin count and tonsillectomy. From this scatterplot, a
line of best fit is made for each Mendelian Randomization method. Since this line has a negative slope, it indicates a negative effect.
T cells with infectious processes in tonsils in comparison to obstructive pathology [
22
]. This shows that lymphocyte
count (T cells) does, in fact, have a relationship with tonsillectomy, as it directly affects tonsillitis. This study directly
complements the finding produced by the Mendelian Randomization experiment.
3.3 Operation: Tonsillectomy and Haemoglobin Count
The latter finding about tonsillectomy relates haemoglobin count to tonsillectomy. The multiple-testing adjusted p-value
for this finding is
3.6 10
3
, with a standard error of
2.2 10
3
.
β
is
7.5 10
3
, which indicates a negative
relationship between haemoglobin count and tonsillectomy; therefore, from this information, a lower haemoglobin
count leads to a higher chance of tonsillectomy. This is further proved by Figure 6, which can be used to experimentally
determine
β
due to the fact that the graph plots
γ
vs.
Γ
. Using Equation 1,
β
is the slope, and since the slope is negative,
there must be a negative relationship between the two. For greater detail, reference Figure 6.
Using the MR-MoE classifier, the weighted median was determined as the optimal model for this specific exposure-
outcome association. In Figure 6, the weighted median is the dark green line of best fit. To understand what the causal
relationship entails, it is necessary to understand what tonsillectomy and haemoglobin count are. Information about
both of these (Section 3.1 and Section 3.2 respectively) will be found above.
This relationship has been analyzed from tonsillectomy’s effects on haemoglobin levels. Research seems to indicate
that tonsillectomy has little to no effect on haemoglobin if the operation is successful, from a trial involving children
undergoing tonsillectomy [
23
]. The number of cases of recurrent tonsillectomy in this trial was 44, as compared to
40 control cases, which shows there is a good sample size to make a hypothesis upon. Another study proved that
complications such as a post-tonsillectomy haemorrhage will cause a reduction of haemoglobin levels, but the levels
10
A PREPRINT - JULY 1, 2020
Figure 7: Scatterplot of the relationship between neutrophil count and wheezing.
Each dot represents a single SNP mapped based on its effect on both neutrophil count and wheezing. From this scatterplot, a line of
best fit is made for each Mendelian Randomization method. Since this line has a positive slope, it indicates a positive effect.
will quickly stabilize within hours of the haemorrhage [
24
]. Although this does not support the finding made by the
experiment, these are interesting studies that could provide insight into why the experiment’s finding holds true.
Past research suggests that there is a plausible connection between lower haemoglobin levels and tonsillectomy. For
example, a study of children about to undergo adenoid-tonsillectomy found that they were largely iron-deficient [
25
].
Additionally, they had low hematocrit levels, which determine the proportion of blood that contains red blood cells.
Because haemoglobin is an essential component to many red blood cells, a reduction in hematocrit levels means that
there must have been a reduction in haemoglobin. This shows a correlation between low haemoglobin and tonsillectomy
as a large proportion of the blood samples taken from these children expressed this finding. This result supports the
relationship between haemoglobin levels and tonsillectomy made in the experiment.
3.4 Wheezing and Neutrophil Count
Another relationship explored in this study was the relationship between neutrophil count and wheezing. The multiple-
testing adjusted p-value for this relationship was
4.9 10
2
, while the standard error was
3.2 10
3
.
β
is
8.7 10
3
,
which indicates a positive relationship between neutrophil count and wheezing. This means that an increase in neutrophil
count will lead to a higher chance of wheezing. More information about this relationship can be referenced from Figure
7.
From MR-MoE, the simple mean Mendelian Randomization model seems to provide the optimal result. Although
Figure 7 shows varying slopes for the lines of best fit, the statistical data provides a positive causal relationship. The
reason why some of these slopes are negative rather than positive, could be because there is a breaking of the Mendelian
Randomization assumptions when running these models. To further prove a positive causal relationship, the confidence
intervals are from
2.3 10
3
to
1.5 10
2
, which indicates that experimental uncertainty does not play a role in the
11
A PREPRINT - JULY 1, 2020
sign of the slopes of the graph. Additionally, because the data was not steiger-filtered or outlier-filtered, there could
be SNPs that negatively alter the line of best fit. Therefore, the overall direction of causality must be positive due to
the data provided above. Because the data consists of 135 SNPs, the causal relationship produced by the Mendelian
Randomization model is backed by a large amount of statistical evidence.
To understand the causal relationship, it is once again necessary to understand what the exposure and outcomes are.
Neutrophils are white blood cells in the immune system that check the lymphatic system and bloodstream systematically
for antigens. They are the first to arrive at the infection, and quickly start microbial functions once they reach. These
include producing antimicrobial products and proinflammatory cytokines that will help contain the infection [
26
]. This
will allow the immune system time to generate a response to the antigen. Neutrophil count is commonly measured
using the absolute neutrophil count, which provides a measure of the number of neutrophil granulocytes that are present
in the blood [27]. Having a high percentage of neutrophils is called neutrophilia, which will be discussed later.
Wheezing in the chest is a lung sound produced by a narrowed or compressed airway. It is not a clear defined ailment
and is subjective to the patient. For instance, some refer to it as difficult breathing, while some refer to it as a whistling
type of breathing. It is normally caused from localized airway narrowing and the obstruction from the larnyx to the
small bronchi. This can be caused by a number of things, including foreign infections, external compression, or
bronchoconstriction. Wheezes are caused by oscillations of nearly closed airway walls that produce high velocity air.
This air causes pressure that barely opens the airway lumen. The airway then closes, which is what causes the wheezing
sound (the cycle of opening and closing). Wheezing is commonly heard during expiration which is where airways
normally narrow [28].
Having a high percentage of neutrophils can cause damage to fragile body systems, even though the purpose of
neutrophils is to protect the body from damage. An example of an excess of neutrophils causing damage is seen in
its role in causing acute lung disease [
29
]. Due to the fragile tissue in the lung, an excess of neutrophils can cause
irreparable damage if they come in too aggressively. A dysregulation of neutrophil activation is likely one of the reasons
that an excess of neutrophils can occur. Although not the primary reason for addressing this study, acute lung disease
does in fact cause wheezing symptoms, which shows an indirect relationship between high neutrophil count and a
higher chance of wheezing [30].
A high percentage of neutrophils, or neutrophilia, has been associated with other lung conditions, including bronchiolitis
obliterans syndrome [
31
]. Bronchiolitis obliterans syndrome is a condition that affects the bronchioles, the lung’s
smallest airways. This causes inflammation in these airways, which leads to extensive scarring that eventually blocks
the airway [
32
]. Diseases such as these are often the causes of wheezing, and due to the multiple studies found that
already relate neutrophilia to different lung diseases, it is possible to form a causal pathway using this research between
nuetrophilia and a higher chance of wheezing.
4 Discussion
Four distinct exposure-outcome associations were analyzed and statistically proven to have a causal relationship:
haemoglobin count and bleeding gums, tonsillectomy and lymphocyte count, tonsillectomy and haemoglobin count,
and wheezing and neutrophil count. Using these causal relationships, it is now possible to form a plausible connection
between certain blood cells and infectious diseases. From these connections, treatments can be developed after further
analyzing the results.
Although these results are backed statistically, trials must be done to support the practicality of these conclusions in the
real world. When testing the causal relationship between haemoglobin and bleeding gums, a mice study could be done.
Mice will have haemoglobin injected into them, causing a temporary spike in haemoglobin levels within their system.
Over the course of a few days, their gums will be examined to check for any signs of bleeding. Additionally, a sample
will be taken of the gums at the end of the study and checked for bacteria that eventually leads to bleeding gums. As
some of the causes of bleeding gums take months or years to develop, taking these samples will provide knowledge into
whether bleeding gums will develop in the near future.
Additionally, a human study could be done. Patients with bleeding gums could be examined, and their haemoglobin
levels will be measured. Patients without bleeding gums will also have measured haemoglobin levels. Although this is
relatively simpler compared to the mice study, this data could have hundreds of confounding factors that could affect
the data. To prevent genetic variation that could skew the data, testing will be localized to one region and one ethnicity.
However, even with this data, it is difficult to determine which direction the causal relationship between haemoglobin
and bleeding gums is occurring. As a result, the mice study would be the preferred method of genetic study, but the
human trials could also be used.
12
A PREPRINT - JULY 1, 2020
From a biological perspective, the relationship between haemoglobin levels and bleeding gums seems to contradict
that of the experiment’s findings. From one biological perspective, a negative relationship between haemoglobin levels
and bleeding gums seems to make sense. Due to haemoglobin’s importance in the transfer of oxygen, a reduced
concentration will cause a reduction in the amount of oxygen available in the mouth. Many of the aforementioned
bacteria that contribute to periodontis and gingivitis thrive in an oxygen-free environment, which shows the relationship
between lack of oxygen and dental disease. In the case study, the patient was observed having darker gums than average,
which occurs with lack of oxygen. Therefore, it seems plausible to link haemoglobin concentration to bleeding gums
biologically; however, this isn’t supported by the two-sample Mendelian Randomization study, which provides an
interesting point. Although it seems as though there is much evidence against the finding in the study, most of the
evidence is just biological hypotheses formed from human trials, while Mendelian Randomization is a statistical and
objective study. Therefore, Mendelian Randomization can be used to refute opposing biological hypotheses.
Additionally, from another biological perspective, a negative relationship between haemoglobin levels and bleeding
gums seems to make sense. In one study, mice with low haemoglobin levels had anemia due to a mutation in the Sec15l1
gene, which shows a causal relationship between low haemoglobin levels and anemia [
33
]. People who are anemic
have a likelihood of bleeding gums, due to a deficiency in iron. This iron deficiency causes a reduction in platelet
formation and more bleeding, which is further supported by the case study previously aforementioned. Therefore, this
provides a causal pathway between low haemoglobin levels and bleeding gums, which isn’t supported by the Mendelian
Randomization.
There isn’t much literature to prove the positive causal relationship between haemoglobin levels and bleeding gums,
which makes the result interesting, especially with so many studies seeming to point towards a negative relationship.
Further research must be done to prove this positive relationship. Although much research points towards a negative
relationship, as previously mentioned, genetic variants provide an objective way of analyzing the genome, while human
trials, such as the ones mentioned above, can only be used to produce plausible biological hypotheses. As a result, the
data provided in Figure 3 should be taken with more credibility than the other trials above.
Further research can also be done on haemoglobin’s effect on other mouth conditions, such as tooth decay or gingivitis.
Haemoglobin is the key transporter of oxygen throughout the body, so a reduction of haemoglobin will cause a reduction
of oxygen in the mouth. Since many bacteria thrive in an oxygen-free environment, a reduction in haemoglobin could
cause a wide variety of mouth diseases caused by these bacteria. Since a causal relationship between haemoglobin
and bleeding gums has already been analyzed and proven, it is probable that other mouth diseases will be linked to
haemoglobin levels.
When testing the causal relationship between lymphocyte counts and tonsillectomy, a human trial can be conducted
with a greater degree of accuracy than the aforementioned human trial. Instead of selecting patients that have already
went through a tonsillectomy, patients about to undergo a tonsillectomy will be selected for this experiment. Before the
tonsillectomy, immediately after the tonsillectomy, and a week after the tonsillectomy, their lymphocyte counts will
be measured. The control’s lymphocyte counts will only be measured once as they are not undergoing an operation
that could affect their lymphocyte count. If there is a higher lymphocyte count for those undergoing tonsillectomy,
the causal relationship is proven. Additionally, examining the lymphocyte count after the tonsillectomy will provide
knowledge on how quickly the body regains homeostasis after the operation.
From a biological perspective, this causal relationship seems to make sense. Normally, with a high lymphocyte count,
it means that the patient’s body is working to repel a foreign disease. If the disease originates from the tonsils, there
is plausible cause for a tonsillectomy to be performed. Therefore, using this causal pathway, it is possible to make
a biological relationship between high lymphocyte counts and tonsillectomy, which directly correlates back to the
findings made.
The same human trial could be done with haemoglobin counts to prove the causal relationship between haemoglobin
counts and tonsillectomy. If there is a lower haemoglobin count for those undergoing tonsillectomy as compared to
the controls, the causal relationship is proven. Additionally, the haemoglobin levels should stabilize after the study, as
proven by several aforementioned studies.
From a biological perspective, this causal relationship seems to make sense. A reduction in haemoglobin normally
causes a reduction in red blood cells, which can cause anemia and other diseases within the patient’s body. Due to
haemoglobin’s importance in delivering oxygen through the body, a reduction of haemoglobin will cause a reduction
of oxygen in the body. As a result, there will less oxygen in the mouth. As previously discussed in Section 3.1, this
can lead to periodontitis and bleeding gums due to the buildup of bacteria in the mouth caused by this lack of oxygen;
however, there can also be other infectious diseases that could affect the mouth, such as tonsillitis. Some of these
problems would eventually necessitate a tonsillectomy, as the tonsil diseases will keep recurring due to the reduction of
13
A PREPRINT - JULY 1, 2020
oxygen in the mouth. This causal pathway seems to directly support the idea that a reduction in haemoglobin will cause
tonsillectomy.
Further research can be done on tonsillectomy’s effect on germ disease. Tonsils are an essential part of the body’s
immune system, because they stop germs from entering the body through the nose or mouth. Because of this, the
removal of tonsils could cause a wide array of other germ diseases, because the body doesn’t have that barrier to protect
against germs anymore. A two-sample Mendelian Randomization could be ran by making tonsillectomy the risk factor
and germ diseases as the outcome.
When testing the causal relationship between neutrophil counts and wheezing in the chest, a human trial can also
be conducted. People with wheezing will have their neutrophil counts measured, and this will be compared to the
controls. If the people with wheezing have a higher neutrophil count than the controls, the causal relationship is proven.
However, another step can be taken to further prove the relationship. Since there are already neutrophil inhibitors in
the market, using these inhibitors should reduce wheezing in the chest. If the number of times a patient undergoes
wheezing reduces, the causal relationship is further proven.
Research has already been done towards potential neutrophil inhibitors that could reduce the number of cases of acute
lung injury. Most inhibitors in the market focus towards inhibiting neutrophil elastase, a destructive enzyme that
disturbs the function of the lung permeability barrier and releases pro-inflammatory cytokines [
34
]. These actions lead
to common symptoms of acute lung disease, suggesting a strong relationship between neutrophil elastase and acute
lung disease. Neutrophil elastase inhibitors, such as sivelestat, attenuate the degree of lung damage caused by these
enzymes and reduce neutrophil accumulation inside of the lungs [
35
]. Trials done on rats shows that sivelestat reduces
lipopolysaccharide-induced acute lung injury and inhibits inflammation within the lungs [
36
]. This is a plausible
treatment that could greatly help reduce the number of wheezing cases caused by high neutrophil accumulation.
As the cost of genetic sequencing reduces with the breakthrough of new technologies, genetic studies are becoming
more common in biological research, due to the objectivity of the results found with it. As a result, Mendelian
Randomization will provide an increasingly useful role in the future, as more researchers shift towards this method of
finding relationships between exposures and outcomes. Its reliable and cost-effective method of determining causal
relationships between a risk factor and disease make it an important method of research to be used in the years to come.
References
[1]
Debbie A Lawlor. Commentary: Two-sample mendelian randomization: opportunities and challenges. Interna-
tional Journal of Epidemiology, 45(3):908–915, June 2016.
[2]
Alexander Teumer. Common methods for performing mendelian randomization. Frontiers in Cardiovascular
Medicine, 5, May 2018.
[3]
Jack Bowden and Michael V. Holmes. Meta-analysis and mendelian randomization: A review. Research Synthesis
Methods, 10(4):486–496, April 2019.
[4]
Peggy Sekula, Fabiola Del Greco M, Cristian Pattaro, and Anna Köttgen. Mendelian randomization as an approach
to assess causality using observational data. Journal of the American Society of Nephrology, 27(11):3253–3265,
August 2016.
[5]
Jack Bowden, George Davey Smith, Philip C. Haycock, and Stephen Burgess. Consistent estimation in mendelian
randomization with some invalid instruments using a weighted median estimator. Genetic Epidemiology, 40(4):304–
314, April 2016.
[6]
Gibran Hemani, Jie Zheng, Benjamin Elsworth, Kaitlin H Wade, Valeriia Haberland, Denis Baird, Charles Laurin,
Stephen Burgess, Jack Bowden, Ryan Langdon, Vanessa Y Tan, James Yarmolinsky, Hashem A Shihab, Nicholas J
Timpson, David M Evans, Caroline Relton, Richard M Martin, George Davey Smith, Tom R Gaunt, and Philip C
Haycock. The MR-base platform supports systematic causal inference across the human phenome. eLife, 7, May
2018.
[7]
Gibran Hemani, Jack Bowden, Philip Haycock, Jie Zheng, Oliver Davis, Peter Flach, Tom Gaunt, and
George Davey Smith. Automating mendelian randomization through machine learning to construct a puta-
tive causal map of the human phenome. August 2017.
[8]
Vladimir Svetnik, Andy Liaw, Christopher Tong, J. Christopher Culberson, Robert P. Sheridan, and Bradley P.
Feuston. Random forest: a classification and regression tool for compound classification and QSAR modeling.
Journal of Chemical Information and Computer Sciences, 43(6):1947–1958, November 2003.
14
A PREPRINT - JULY 1, 2020
[9]
Cosetta Minelli, M. Fabiola Del Greco, Diana A. van der Plaat, Jack Bowden, Nuala A. Sheehan, and John
Thompson. The use of two-sample methods for mendelian randomization analyses on single large datasets. May
2020.
[10]
Liang-Dar Hwang, Deborah A Lawlor, Rachel M Freathy, David M Evans, and Nicole M Warrington. Using a
two-sample mendelian randomization design to investigate a possible causal effect of maternal lipid concentrations
on offspring birth weight. International Journal of Epidemiology, 48(5):1457–1467, July 2019.
[11]
Jessica M. B. Rees, Angela M. Wood, Frank Dudbridge, and Stephen Burgess. Robust methods in mendelian
randomization via penalization of heterogeneous causal estimates. PLOS ONE, 14(9):e0222362, September 2019.
[12]
Jie Zheng, Monika Frysz, John P. Kemp, David M. Evans, George Davey Smith, and Jonathan H. Tobias. Use of
mendelian randomization to examine causal inference in osteoporosis. Frontiers in Endocrinology, 10, November
2019.
[13]
Yoonsu Cho, Philip C. Haycock, Eleanor Sanderson, Tom R. Gaunt, Jie Zheng, Andrew P. Morris, George Davey
Smith, and Gibran Hemani. Exploiting horizontal pleiotropy to search for causal pathways within a mendelian
randomization framework. Nature Communications, 11(1), February 2020.
[14]
Caroline Thomas and Andrew B Lumb. Physiology of haemoglobin. Continuing Education in Anaesthesia
Critical Care & Pain, 12(5):251–256, October 2012.
[15] Roy C. Page. Gingivitis. Journal of Clinical Periodontology, 13(5):345–355, May 1986.
[16]
Max A. Listgarten. Pathogenesis of periodontitis. Journal of Clinical Periodontology, 13(5):418–425, May 1986.
[17]
Abhishek Ghosh. Total extraction as a treatment for anaemia in a patient of glanzmann’s thrombasthenia with
chronic gingival bleed: Case report. JOURNAL OF CLINICAL AND DIAGNOSTIC RESEARCH, 2016.
[18]
Gabriel J. Tobón, Jorge H. Izquierdo, and Carlos A. Cañas. B lymphocytes: Development, tolerance, and their
role in autoimmunity—focus on systemic lupus erythematosus. Autoimmune Diseases, 2013:1–17, 2013.
[19] Shimon Sakaguchi, Tomoyuki Yamaguchi, Takashi Nomura, and Masahiro Ono. Regulatory t cells and immune
tolerance. Cell, 133(5):775–787, May 2008.
[20] Mayo Clinic Staff. Tonsillectomy: About, 2018.
[21]
˙
Irfan Kaygusuz, Ahmet Gödekmerdan, Turgut Karlida
ˇ
g, Erol Kele¸s, ¸Sinasi Yalçin,
˙
Ibrahim Aral, and Mücahit
Yildiz. Early stage impacts of tonsillectomy on immune functions of children. International Journal of Pediatric
Otorhinolaryngology, 67(12):1311–1315, December 2003.
[22]
M.A López-González, B Sánchez, F Mata, and F Delgado. Tonsillar lymphocyte subsets in recurrent acute
tonsillitis and tonsillar hypertrophy. International Journal of Pediatric Otorhinolaryngology, 43(1):33–39,
February 1998.
[23]
E. Mira, M. Benazzo, L. Asti, A. Marchi, P. Spriano, and R. Losi. Iron status in children undergoing tonsillectomy
and its short-term modification following surgery. Acta Oto-Laryngologica, 105(sup454):261–264, January 1988.
[24]
P. G. Herreen and M. C. R. Wiese. Concealed post-tonsillectomy haemorrhage associated with the use of the
antiemetic tropisetron. Anaesthesia and Intensive Care, 29(4):421–422, August 2001.
[25]
A. Busuttil, A. I. G. Kerr, and R. W. Logan. Iron deficiency in children undergoing adenoid-tonsillectomy. The
Journal of Laryngology & Otology, 93(1):49–58, January 1979.
[26]
Fumitaka Hayashi, Terry K. Means, and Andrew D. Luster. Toll-like receptors stimulate human neutrophil
function. Blood, 102(7):2660–2669, October 2003.
[27]
Christoph P. Hornik, Daniel K. Benjamin, Kristian C. Becker, Daniel K. Benjamin, Jennifer Li, Reese H. Clark,
Michael Cohen-Wolkowiez, and P. Brian Smith. Use of the complete blood cell count in early-onset neonatal
sepsis. The Pediatric Infectious Disease Journal, 31(8):799–802, August 2012.
[28]
Wilbur Dallas Hall, John Willis Hurst, and Henry Kenneth Walker. Clinical Methods. Butterworths, Boston, 3rd
edition, 1990.
[29]
Warren L. Lee and Gregory P. Downey. Neutrophil activation and acute lung injury. Current Opinion in Critical
Care, 7:1–7, February 2001.
[30]
Craig R. Weinert, Cynthia R. Gross, James R. Kangas, Caron L. Burry, and William A. Marinelli. Health-
related quality of life after acute lung injury. American Journal of Respiratory and Critical Care Medicine,
156(4):1120–1128, October 1997.
[31]
L. Zheng. Airway neutrophilia in stable and bronchiolitis obliterans syndrome patients following lung transplanta-
tion. Thorax, 55(1):53–59, January 2000.
15
A PREPRINT - JULY 1, 2020
[32]
D Heng, LD Sharples, K McNeil, S Stewart, T Wreghitt, and J Wallwork. Bronchiolitis obliterans syndrome:
incidence, natural history, prognosis, and risk factors. The Journal of heart and lung transplantation : the official
publication of the International Society for Heart Transplantation, 17(12):1255—1263, December 1998.
[33]
Jackie E Lim, Ou Jin, Carolyn Bennett, Kelly Morgan, Fudi Wang, Cameron C Trenor, Mark D Fleming, and
Nancy C Andrews. A mutation in sec15l1 causes anemia in hemoglobin deficit (hbd) mice. Nature Genetics,
37(11):1270–1273, October 2005.
[34]
Kazuhito Kawabata, Tetsuya Hagio, and Shozo Matsuoka. The role of neutrophil elastase in acute lung injury.
European Journal of Pharmacology, 451(1):1–10, September 2002.
[35]
Akihiro Sakashita, Yoshihiro Nishimura, Teruaki Nishiuma, Kaori Takenaka, Kazuyuki Kobayashi, Yoshikazu
Kotani, and Mitsuhiro Yokoyama. Neutrophil elastase inhibitor (sivelestat) attenuates subsequent ventilator-
induced lung injury in mice. European Journal of Pharmacology, 571(1):62–71, September 2007.
[36]
Satoshi Hagiwara, Hideo Iwasaka, Kazumi Togo, and Takayuki Noguchi. A neutrophil elastase inhibitor, sivelestat,
reduces lung injury following endotoxin-induced shock in rats by inhibiting HMGB1. Inflammation, 31(4):227–
234, June 2008.
16