We used Mendelian randomization (MR) to assess whether causal effects
exist between depression and a number of other traits and disorders. MR uses
genetic variants as a proxy for environmental exposures, assuming that: i) the
genetic variants are associated with the exposure; ii) the genetic variants are
independent of confounders in the exposure-outcome association; iii) the genetic
variants are associated with the outcome only via their effect on the exposure,
i.e. there is no horizontal pleiotropy whereby the variants affect both exposure
and outcome independently. Individual genetic variants may be weak instruments
for assessing causality, particularly if they have only small effect sizes.
Using multiple genetic variants can increase the strength of the instrument, but
also increases the risk of violating the MR assumptions.
We performed bidirectional, two-sample MR between our meta-analysis
results for depression and all available traits which had a significant genetic
correlation with depression (identified in the previous section). Traits
directly related to or including depression (major depressive disorder,
depressive symptoms and PGC cross disorder) were excluded due to potential bias.
The genetic instruments for depression consisted of the independent, genome-wide
significant variants, their effect sizes and standard errors, as estimated in
our genome-wide meta-analysis. Summary statistics from genome-wide association
studies for the other traits were sourced from either publicly available
datasets or from the MR-Base database 11 (link).
Overlapping datasets for the exposure and the outcome can lead to bias and
inflation of causal estimates 68 (link). To
mitigate this, when the source of the other trait included UK Biobank, 23andMe
or any of the studies that contributed to PGC_139k, these studies were removed
from our meta-analysis of depression; for example, the neuroticism trait from
van den Berg, et al. 69 (link) was assessed
against a meta-analysis of depression using UK Biobank and 23andMe_307k only.
Where UK Biobank, 23andMe and PGC_139k were all included in the genome-wide
association study of the other trait then an alternative study was sought for
that other trait.
All analyses were performed using the MR-Base
“TwoSampleMR” v0.4.9 package11 (link) in R. To avoid bias in the MR estimates due to linkage
disequilibrium (r2), clumping was applied using the
“clump_data” function with an r2 < 0.001.
Genetic variants were required to be available in both the exposure and outcome
traits and were harmonised using the default parameters within the TwoSampleMR
package. Following this harmonisation, we only examined causal relationships
where there were at least 30 instrumental genetic variables.
Directional horizontal pleiotropy, where a genetic instrument has an
effect on an outcome independent of its influence on the exposure, can be a
problem in MR analysis, particularly when multiple genetic variants of unknown
function are used. We therefore firstly tested for directional horizontal
pleiotropy using the MR Egger intercept test, as previously described by
Hagenaars, Gale, Deary and Harris 70 . If
the MR Egger intercept test had a significant P-value
(P < 0.05) then it was excluded from the analysis.
However, no tests were excluded due to directional horizontal pleiotropy.
The second analysis conducted was a variant heterogeneity test for
global horizontal pleiotropy. Variant heterogeneity is an important metric, but
high heterogeneity doesn’t necessarily mean bias or unreliable results;
for example, every instrumental variable could have horizontal pleiotropic
effects but if they have a mean effect of 0 then there will be no bias, just
larger standard errors due to more noise. For analyses that had evidence of high
variant heterogeneity (P < 0.05), additional sensitivity
MR tests were conducted. The sensitivity tests that were used were the MR Egger
test and the weighted median test to examine whether the effect estimate was
consistent.
The principal MR test of a causal effect was conducted using
inverse-variance weighted (IVW) regression. This method is based on a regression
of the exposure and the outcome which assumes the intercept is constrained to
zero, and produces a causal estimate of the exposure-outcome association. Where
there was no evidence of global horizontal pleiotropy (P≥ 0.05) from the second analysis (see previous paragraph), a FDR
adjusted67
P-value < 0.01 from the IVW test was required for
evidence of a causal effect. Where there was evidence of global horizontal
pleiotropy (P < 0.05) from the second analysis, additional evidence was
also sought from the sensitivity tests (MR Egger test and the weighted median
test). To ensure that a causal effect was not driven by a single variant a
‘leave one variant out’ IVW regression analysis was conducted with
the least significant observed P-value used to assess whether
significance was maintained. We tested the causal effect of depression on 24
other traits, and the causal effect of 9 other traits on depression.
exist between depression and a number of other traits and disorders. MR uses
genetic variants as a proxy for environmental exposures, assuming that: i) the
genetic variants are associated with the exposure; ii) the genetic variants are
independent of confounders in the exposure-outcome association; iii) the genetic
variants are associated with the outcome only via their effect on the exposure,
i.e. there is no horizontal pleiotropy whereby the variants affect both exposure
and outcome independently. Individual genetic variants may be weak instruments
for assessing causality, particularly if they have only small effect sizes.
Using multiple genetic variants can increase the strength of the instrument, but
also increases the risk of violating the MR assumptions.
We performed bidirectional, two-sample MR between our meta-analysis
results for depression and all available traits which had a significant genetic
correlation with depression (identified in the previous section). Traits
directly related to or including depression (major depressive disorder,
depressive symptoms and PGC cross disorder) were excluded due to potential bias.
The genetic instruments for depression consisted of the independent, genome-wide
significant variants, their effect sizes and standard errors, as estimated in
our genome-wide meta-analysis. Summary statistics from genome-wide association
studies for the other traits were sourced from either publicly available
datasets or from the MR-Base database 11 (link).
Overlapping datasets for the exposure and the outcome can lead to bias and
inflation of causal estimates 68 (link). To
mitigate this, when the source of the other trait included UK Biobank, 23andMe
or any of the studies that contributed to PGC_139k, these studies were removed
from our meta-analysis of depression; for example, the neuroticism trait from
van den Berg, et al. 69 (link) was assessed
against a meta-analysis of depression using UK Biobank and 23andMe_307k only.
Where UK Biobank, 23andMe and PGC_139k were all included in the genome-wide
association study of the other trait then an alternative study was sought for
that other trait.
All analyses were performed using the MR-Base
“TwoSampleMR” v0.4.9 package11 (link) in R. To avoid bias in the MR estimates due to linkage
disequilibrium (r2), clumping was applied using the
“clump_data” function with an r2 < 0.001.
Genetic variants were required to be available in both the exposure and outcome
traits and were harmonised using the default parameters within the TwoSampleMR
package. Following this harmonisation, we only examined causal relationships
where there were at least 30 instrumental genetic variables.
Directional horizontal pleiotropy, where a genetic instrument has an
effect on an outcome independent of its influence on the exposure, can be a
problem in MR analysis, particularly when multiple genetic variants of unknown
function are used. We therefore firstly tested for directional horizontal
pleiotropy using the MR Egger intercept test, as previously described by
Hagenaars, Gale, Deary and Harris 70 . If
the MR Egger intercept test had a significant P-value
(P < 0.05) then it was excluded from the analysis.
However, no tests were excluded due to directional horizontal pleiotropy.
The second analysis conducted was a variant heterogeneity test for
global horizontal pleiotropy. Variant heterogeneity is an important metric, but
high heterogeneity doesn’t necessarily mean bias or unreliable results;
for example, every instrumental variable could have horizontal pleiotropic
effects but if they have a mean effect of 0 then there will be no bias, just
larger standard errors due to more noise. For analyses that had evidence of high
variant heterogeneity (P < 0.05), additional sensitivity
MR tests were conducted. The sensitivity tests that were used were the MR Egger
test and the weighted median test to examine whether the effect estimate was
consistent.
The principal MR test of a causal effect was conducted using
inverse-variance weighted (IVW) regression. This method is based on a regression
of the exposure and the outcome which assumes the intercept is constrained to
zero, and produces a causal estimate of the exposure-outcome association. Where
there was no evidence of global horizontal pleiotropy (P≥ 0.05) from the second analysis (see previous paragraph), a FDR
adjusted67
P-value < 0.01 from the IVW test was required for
evidence of a causal effect. Where there was evidence of global horizontal
pleiotropy (P < 0.05) from the second analysis, additional evidence was
also sought from the sensitivity tests (MR Egger test and the weighted median
test). To ensure that a causal effect was not driven by a single variant a
‘leave one variant out’ IVW regression analysis was conducted with
the least significant observed P-value used to assess whether
significance was maintained. We tested the causal effect of depression on 24
other traits, and the causal effect of 9 other traits on depression.
