Package 'SUMnlmr'

Hunt dataset containing the allele score-associations with bmi and all cause mortality, by 100 quantiles of residual BMI. As used in the paper "Body mass index and all cause mortality in HUNT and UK Biobank studies: linear and non-linear mendelian randomisation analyses" https://doi.org/10.1136/bmj.l1042

Description

Hunt dataset containing the allele score-associations with bmi and all cause mortality, by 100 quantiles of residual BMI. As used in the paper "Body mass index and all cause mortality in HUNT and UK Biobank studies: linear and non-linear mendelian randomisation analyses" https://doi.org/10.1136/bmj.l1042

Usage

bmi_acm

Format

A data frame with 100 rows and 5 variables:

beta_bmi

linear association between BMI and allele score, adjusted for age, sex and age-squared

se_bmi

standard error for the beta_bmi term from the linear regression

beta_acm

Cox proportional hazards regression association between all cause mortality and allele score, adjusted for age and sex

se_acm

standard error for the beta_acm term from the Cox regression

mean_bmi

Average BMI in stratum

...

---

Generation of individual level data

Description

generates individual level data with a single genetic variant

Usage

create_ind_data(
  N,
  gpar = 0.3,
  par1 = 1,
  par2 = 0,
  beta0 = 0,
  beta1 = 3,
  beta2 = 7,
  confound = 0.8
)

Arguments

N	number of individuals to create
gpar	genetic parameter; used to create g: single genetic snp, from a binomial distribution with n=2 and p = gpar.
par1	power parameter for fractional poly generation. See details
par2	power parameter for fractional poly generation. . See details.
beta0	covariate parameter. See details
beta1	covariate parameter. See details
beta2	covariate parameter. See details
confound	confounding parameter,c. See details.

Value

data A data-frame containing the values of g, the genetic variate; X, the exposure; and a variety of Y, the outcome values, including y.bin, a binary outcome generated from a logistic model of X.

Note

This function generates a database with genetic relationships suitable for evaluating non-linear MR relationships. A unknown covariate,u, is generated as a N(0,1) variable. Error terms are generated: Ex ~exp(1) and for Ey ~ N(0,1) $X= 2+ 0.25*g +u + E_x$ Outcomes are as follows

• Linear: $Y=b_0+ b_1 X + cU +E_y$

• Quadratic $Y = b_0 + b_1 X + b_2 X^2 + cU +E_y$

• Squareroot $Y = b_0 + b_1 \sqrt{X} + cU +E_y$

• Log $Y = b_0 + b_1 \log(X) + cU +E_y$

• Threshold $Y = b_0+ b_1 X + cU +E_y$ if$X>b_2$ and $Y = b_0 + cU +E_y$ otherwise

• fracpoly $Y = b_0 + b_1 X^{p_1} + b_2 X^{p_2} + cU + E_y$ with the usual adaptions for p=0 or p_1=p_2

Author(s)

Amy Mason

---

Creation of summarized Mendelian randomization local estimates

Description

create_nlmr_summary Takes individual level data and creates summarized dataset, ready to save and share for summarized non-linear analysis

Usage

create_nlmr_summary(
  y,
  x,
  g,
  covar = NULL,
  gxe_covar = NULL,
  gxe_interaction = NULL,
  strata_method = "ranked",
  x_corrected = NULL,
  x_strata = NULL,
  family = "gaussian",
  controlsonly = FALSE,
  q = 10,
  prestrat = 1,
  strata_bound = c(0.2, 0.1, 0.8, 0.9),
  extra_statistics = FALSE,
  report_GR = FALSE,
  report_het = FALSE,
  return_assignment = FALSE,
  seed = 1234
)

Arguments

y	vector of outcome values.
x	vector of exposure values.
g	the instrumental variable.
covar	An optional matrix of covariates used in deriving the stratification and the genetic associations with the exposure and the outcome within the strata. When using "interaction" method, these covariates are only used in the genetic association calculation with the outcome and exposure in each strata, not in the creation of the strata.
gxe_covar	An optional matrix of covariates used in deriving the stratification when using the "interaction" method. See details.
gxe_interaction	An optional matrix of covariates used as effect modifiers in deriving the stratification when using the "interaction" method. See details.
strata_method	What method to use for determining strata. There are three options "residual", "ranked", "interaction". See details for a longer explanation of these methods. By default this is set to "ranked".
x_corrected	An optional numeric vector used if method is "ranked". This will use these alternative values in the ranked method in place of the exposure in the stratification calculation. This can also be used as a way of creating strata with your own interaction model - supply the interaction corrected exposure values here.
x_strata	An optional numeric vector used to replace the stratification calculation with the supplied strata assignment. Only use this if you have precalculated the strata and just want the genetic associations within those strata.
family	A description of the error distribution and link function to be used in the model. It must be one of "gaussian", "binomial" or "coxph". The "gaussian" option should be used for continuous outcome data and will fit linear regression models. The "binomial" option should be used to fit logistic regression models to binary outcome data. The "coxph" option should be used to fit a Cox proportional hazards model for a time-to-event outcome - in this case the outcome y must be a Surv object.
controlsonly	Only applied if family is "binomial" or "coxph". If TRUE, the genetic associations with the exposure are calculated in the controls only.
q	The number of quantiles the exposure distribution is to be split into. Within each quantile a causal effect will be estimated, known as a localised average causal effect (LACE). The default is deciles (i.e. 10 quantiles).
prestrat	Only applied if method is "ranked". The proportional size of pre-strata in the doubly-ranked method. If prestrat = 1 (default), then pre-strata will contain the number of individuals equal to the number of strata, and 1 individual from each pre-stratum is selected into each stratum. If prestrat = 10, then pre-strata contain 10 times the number of individuals as the number of strata, and 10 individuals from each pre-stratum are selected into each stratum. Larger pre-strata can improve the differentiation between pre-strata, although if pre-strata are too large such that the instrument values vary strongly within pre-strata, then the benefit of the doubly-ranked method is lost.
strata_bound	This controls what range to use for the LACE estimates in the graphical display. By default, this is taken conservatively as the 10th and 90th percentile of internal strata and the 20th and 80th for the bottom of the lowest stratum and top of the highest stratum. It is supplied as a vector of percentiles (lower bottom stratum, lower other, higher top stratum, higher other). This only impacts the "max" and "min" values for the summary table. This can be overridden in piecewise_summ_mr() by using the xbreaks argument to hardset different breakpoints or replacing default with c(0,0,1,1) to return to true max and minimums.
extra_statistics	This will add a second output reporting extra statistics for each stratum. These include the true max and min of the strata (regardless of strata_bound setting) and the F statistic and p-value for the regressions.
report_GR	This will add the Gelman-Rubin statistics for each stratum to the output. Note this only works if the strata method is "ranked".
report_het	This will add p-values for assessing the heterogeneity of the instrument - exposure relationship. The first column is the p-value of the Cochran Q heterogeneity test (Q); the second column is the p-value from the trend test (trend).
return_assignment	Reports the strata assignment of each participant, to assist in examining stratum distributions.
seed	The random seed to use when generating the quantiles (for reproducibility). If set to NA, the random seed will not be set.

Details

Stratification models

Given vectors of an exposure x, an outcome y and an instrument g, as well as a matrix of covariates E, this package implements several possible models for stratifying over the exposure values.

The simplest of these the residual method ("residual"), as described in Staley and Burgess,2017 https://doi.org/https://doi.org/10.1002/gepi.22041. where we regress x on the genetic instrument and the covariate matrix E

$x = \beta_0 + \beta_g g + \beta_E E$

where ⁠\beta_E⁠ is a suitable vector of coefficients, and calculate the residual value

$x_{resid} = x- (\beta_0 + \beta_g G + \beta_E E)$

Once the strata are formed, the genetic associations with x and y are calculated, including the covariates matrix E e.g.

$x = \beta_1 + \beta_x g + \beta_{E1} E$

$y = \beta_2 + \beta_y g + \beta_{E2} E$

with $\beta_x$ and $\beta_y$ returned for each stratum. This method performs poorly unless the genetic effects on the exposure are constant. This assumption is often implausible, and thus this approach is not recommended. See Burgess, 2023 https://doi.org/https://doi.org/10.1159/000531659.

The second option is "ranked", using Haodong Tian's double ranked method to calculate strata. See Tian et al, 2022 <doi: https://doi.org/10.1101/2022.06.28.497930> for more details on this method. However, there are concerns about this method's application, particularly in UK Biobank, with thanks to Hamiliton et al, 2023 <doi: https://doi.org/10.1007/s10654-024-01113-9> for their examples of this. In particular, this method performs poorly when there is a GxE interaction, as explained in Zhao et al, 2026 <doi: https://doi.org/10.64898/2026.01.22.26344640>

The final method "interaction" is an extension of the ranked method developed specifically to mitigate GxE-induced bias in that method, as detailed in Zhou et al, 2026 <doi: https://doi.org/10.64898/2026.01.22.26344640>. In this method, an interaction model is first fitted using two additional matrices, passed using the gxe_covar (F, covariates) and gxe_interaction (H, effect modifiers).

$x = \beta_0 + \beta_g g + \beta_{F} F + \beta_{H} H + \beta_{g \times H} (g \times H)$

Then an interaction corrected value of X, $X - \beta_(g \times H)(g \times H)$ is used to form strata using the ranked method. Within the strata, the associations with exposure and outcome are then calculated using the usual covar covariate matrix E e.g.

$x = \beta_1 + \beta_x g + \beta_{E1} E$

$y = \beta_2 + \beta_y g + \beta_{E2} E$

with $\beta_x$ and $\beta_y$ returned for each stratum.

The two matrices supplied in gxe_covar and gxe_interaction are used to allow flexibility in choice of models for correcting potential GxE interactions.

Value

A list containing summary data and optional diagnostics.

• summary: data frame with one row per stratum containing bx, by, bxse, byse, xmean, xmin, and xmax.

• Dropped: always returned; integer giving the number of individuals removed due to missing values in any of y, x, g, covar, gxe_covar, gxe_interaction, x_corrected, or x_strata. A message is also printed when this value is greater than zero.

• strata_statistics: returned when extra_statistics = TRUE; per-stratum min/max of x and y, mean x, and exposure-model F statistic.

• GR_max: maximum Gelman–Rubin statistic for ranked strata.

• GR_results: full Gelman–Rubin output when report_GR = TRUE and strata_method = "ranked" or "interaction".

• Heterogeneity_results: heterogeneity/trend test p-values when report_het = TRUE and strata_method = "ranked" or "interaction".

• Strata_assignment: strata assignment for each participant when return_assignment = TRUE

Author(s)

Amy Mason

---

generation of summary level data

Description

create_summary_data generates semi-summarized level data

Usage

create_summary_data(
  Ytype,
  q = 10,
  keep = FALSE,
  family = "gaussian",
  controlsonly = "TRUE",
  ...
)

Arguments

Ytype	The relationship between X and Y; can be "linear", "quad", "sqrt", "log", "threshold" or "fracpoly"
q	The number of quantiles.
keep	Whether to retain the individual level data as well as the summary data
family	a description of the error distribution and link function to be used in the model. For piecewise_mr this can be a character string naming either the gaussian (i.e. "gaussian" for continuous data) or binomial (i.e. "binomial" for binary data) family function.
controlsonly	whether to estimate the gx association in all people, or in controls only. This is set to TRUE as default. It has no effect if family is set to "gaussian"
...	parameters passed to mr_create_data for control of X-Y relationship

Value

summary A data-frame containing the semi-summarised beta_X and beta_Y values, mean of X_0, mean of X, max of X and min of X for each quantile.

Author(s)

Amy Mason

---

Instrumental variable analysis using fractional polynomials based on summary data

Description

frac_poly_summ_mr performs instumental variable analysis by fitting fractional polynomial models to localised average causal effects using meta-regression.

Please note that if you provide synthetic data that are too regular (eg the by associations are all zero or the xmean values are exactly 1, 2, 3, ...), the function may error as several fractional polynomials provide the same fit.

Usage

frac_poly_summ_mr(
  by = NULL,
  bx = NULL,
  byse = NULL,
  bxse = NULL,
  xmean = NULL,
  summ = NULL,
  method = "FE",
  d = "both",
  powers = c(0, -2, -1.5, -1, -0.5, 1, 2),
  pd = 0.05,
  average.exposure.associations = FALSE,
  ci = "model_se",
  nboot = 100,
  fig = FALSE,
  family = "binomial",
  offset = 0,
  pref_x = "x",
  pref_y = "y",
  ref = NA,
  ci_type = "overall",
  breaks = NULL,
  ylim_lower = NA,
  ylim_upper = NA,
  xlim_lower = NA,
  xlim_upper = NA,
  seed = 335
)

Arguments

by	vector of gene-outcome associations.
bx	vector of gene-exposure associations.
byse	vector of standard errors of gene-outcome associations.
bxse	vector of standard errors of gene-exposure associations.
xmean	average value of the exposure in each stratum (or whatever summary of the exposure level in the stratum is desired).
summ	a dataframe contains bx, by, bxse, byse, xmean. This is the summary output of the create_nlmr_summary function; either as the full output or just passing the summary part.
method	meta-regression method parsed to the rma package. The default is fixed-effects ('FE').
d	fractional polynomial degree. The default is degree 1. The other options are: 1, 2, or 'both'.
powers	fractional polynomial powers to test.
pd	p-value cut-off for choosing the best-fitting fractional polynomial of degree 2 over the best-fitting fractional polynomial degree 1. This option is only used if d='both'. The default is 0.05.
average.exposure.associations	TRUE means that the bx estimates are averaged across strata, FALSE means that they are not. Default option is FALSE.
ci	the type of 95\% confidence interval. There are three options: (i) using the model standard errors ('model_se'), (ii) using bootstrap standard errors ('bootstrap_se'), (iii) using bootstrap percentile confidence intervals ('bootstrap_per'). The default is the model standard errors.
nboot	the number of bootstrap replications (if required). The default is 100 replications.
fig	a logical statement as to whether the user wants the results displayed in a figure. The default is false.
family	a character string named either 'gaussian' (for continuous data) or binomial (for binary data) or cox (for survival data) family function. This only affects the plotting function - whether the y-axis is log-transformed or not, and the graph's default label.
offset	offset on the x-axis (default is zero).
pref_x	the prefix/label for the x-axis. The default is 'x'.
pref_y	the prefix/label for the y-axis. The default is 'y'.
ref	the reference point for the figure. This is the value of the function that represents the expected difference in the outcome compared with this reference value when the exposure is set to different values. If ref = NA (the default option), then it is set to the mean of x.
ci_type	the type of confidence interval to be displayed on the graph. The default is 'overall' where confidence intervals are presented as bands across the range of x. The alternative option is 'quantile' where the confidence intervals are presented as error bars at the mean in each quantile of x.
breaks	breaks on the y-axis of the figure.
ylim_lower	lower limit for the y-axis of the figure.
ylim_upper	upper limit for the y-axis of the figure.
xlim_lower	lower limit for the x-axis of the figure.
xlim_upper	upper limit for the x-axis of the figure.
seed	The random seed to use when generating the bootstrap samples (for reproducibility). If set to NA, the random seed will not be set.

Value

model the model specifications. The first column is the number of quantiles (q); the second column is the position used to relate x to the LACE in each quantiles (xpos); the third column is the type of confidence interval constructed (ci); the fourth column is the number of bootstrap replications performed (nboot).

powers the powers of the chosen polynomial.

coefficients the regression estimates. The first column is the regression coefficients (beta); the second column is the standard errors of regression coefficients (se); the third column is the lower confidence interval (lci); the fourth column is the upper confidence interval (uci); the fifth column is the p-value (pval).

lace the localised average causal effect estimate in each quantile. The first column is the regression coefficients (beta); the second column is the standard errors of regression coefficients (se); the third column is the lower confidence interval (lci); the fourth column is the upper confidence interval (uci); the fifth column is the p-value (pval).

xcoef the association between the instrument and the exposure in each quantile. The first column is the regression coefficients (beta); the second column is the standard errors of regression coefficients (se).

p_tests the p-value of the non-linearity tests. The first column is the p-value of the test between the fractional polynomial degrees (fp_d1_d2); the second column is the p-value from the fractional polynomial non-linearity test (fp); the third column is the p-value from the quadratic test (quad); the fourth column is the p-value from the Cochran Q test (Q). The first column is the p-value of the Cochran Q heterogeneity test (Q); the second column is the p-value from the trend test (trend).

figure ggplot command to produce a figure.

Author(s)

Stephen Burgess [email protected], leaning heavily on James R Staley [email protected]

---

Artificial genetic & phenotype data, for purposes of package tests

Description

Simulated individual-level data generated by create_ind_data, used in package examples and tests.

Usage

generated_data

Format

A data frame with 10000 rows and 12 variables:

g

gene count of 0, 1 or 2. Distributed binomial with n=2 and p=0.3

u

"unmeasured" confounder, Uniform(0,1)

errorX

Error term for X, ~Exp(1)

errorY

Error term for Y, ~N(0,1)

X

Exposure. X = 2 + 0.25*g + u + errorX

linear.Y

Linear outcome. Y = X + 0.8u + errorY

quadratic.Y

Quadratic outcome. Y = 2X^2 + X + 0.8u + errorY

sqrt.Y

Square root outcome. $Y = \sqrt{X} + 0.8u + errorY$

log.Y

Log outcome. $Y = \log(X) + 0.8u + errorY$

threshold.Y

$X + 0.8u + errorY$ if $X > 2$, $0.8u + errorY$ otherwise

y.bin

Binary outcome. Bernoulli with probability $logistic(X + 0.8u)$, centred

fracpoly.Y

Fractional polynomial outcome. $Y = X + 2/X + 0.8u + errorY$

Details

Generated with set.seed(1234) and create_ind_data(N = 10000, beta2 = 2, beta1 = 1). Dataset was regenerated on 2026-05-06 to add the y.bin column.

---

Gelman–Rubin (GR) Statistics

Description

getGRvalues calculates the the Gelman–Rubin (GR) statistics for each strata (i.e. LACE-strata) as described in Haodong's paper on doubly-ranked methods in Text S1

Usage

getGRvalues(X = X, Zstratum = Zstratum, NC = 2, tl = 1.02)

Arguments

X	vector of coarsened exposures
Zstratum	vector of pre-strata index column
NC	number of chains - a lower NC is encouraged
tl	threshold level, should be >1

Value

max_GR is the maximum G-R statistic accross all strata

statified_strata is a list of the strata where the GR-statistic meets the threshold limit

GR_values is a table of the a and b values for each strata

Author(s)

Haodong Tian for underlying code, Amy Mason for conversion to R function [email protected]

Examples

#Toy data example------------------------------------------------------------------
 N<-10000  #N is the total ssample size
Z<-rnorm( N,0,0.5   )  # instrument
U1<-rnorm(N,0,1)
U2<-rnorm(N,0,1)   #U2 is confounder
h_X<-function(Z){ 0.5*Z  }
X<-h_X(Z)+U2+U1
fitX<-lm( X~Z   )
RR<-as.numeric(summary(fitX)$r.squared[1])
No<-1000   #No is the number of pre-strata (i.e. IV-strata)
X_<-X  # true exact X
X<-round(X)  #coarsened by round
dat<-cbind(   Z, X, X_)  # X: coarsened exposure; X_: exact exposure
dat_order<-dat[ order(dat[,1]  ),  ]  #ordered by Z
dat_order<-cbind(  dat_order   ,     sort(rep( 1:No, N/No  )))
dat_order<-as.data.frame( dat_order )
names(dat_order)<-c(  'Z_order', 'X','X_' ,'ZStratum')
getGRvalues( X=dat_order$X, Zstratum = dat_order$ZStratum, NC=2, tl = 1.02  )
#--------------------------------------------------------------------------------------

---

IV-free exposure

Description

iv_free computes the IV-free exposure.

Usage

iv_free(
  y,
  x,
  g,
  covar = NULL,
  gxe_covar = NULL,
  gxe_interaction = NULL,
  q = 10,
  controlsonly = T,
  family = "gaussian"
)

Arguments

y	vector of outcome values.
x	vector of exposure values.
g	the instrumental variable.
covar	a matrix of covariates.
gxe_covar	a matrix of covariates for use in the interaction correction
gxe_interaction	a matrix of covariates to use as interaction terms in the interaction correction
q	the number of quantiles the exposure distribution is to be split into. The default is deciles (i.e. 10 quantiles).
controlsonly	whether to estimate the gx association in all people, or in controls only. This is set to TRUE by default, but only takes effect when family is "binomial" or "coxph".
family	the outcome family; one of "gaussian" (default), "binomial", or "coxph". Controls-only filtering is only applied for "binomial" or "coxph" outcomes.

Value

xcoef	the association between the exposure and the instrument
x0	the IV-free exposure if covar is supplied OR the interaction free exposure if an interaction term is supplied.
x0q	the quantiles of x0.

Author(s)

Amy Mason [email protected] based on similar function in nlmr by James R Staley

---

UK Biobank dataset containing the allele score-associations with ldl-cholesterol and CAD diagnosis, by 10 quantiles of residual ldl.

Description

UK Biobank dataset containing the allele score-associations with ldl-cholesterol and CAD diagnosis, by 10 quantiles of residual ldl.

Usage

LDL_CAD

Format

A data frame with 10 rows and 8 variables:

bx

linear association between ldl-cholesterol and allele score, unadjusted

by

logistic regression coefficent between CAD diagnosis and allele score, unadjusted

bxse

standard error for the bx term from the linear regression

byse

standard error for the by term from the logistic regression

xmean

Average ldl in stratum

xmin

Minimum ldl in stratum

xmax

Maximum ldl in stratum

...

---

UK Biobank dataset containing the allele score-associations with ldl-cholesterol and CAD diagnosis adjusted for age, sex and first 10 principle components, by 10 quantiles of residual ldl.

Description

UK Biobank dataset containing the allele score-associations with ldl-cholesterol and CAD diagnosis adjusted for age, sex and first 10 principle components, by 10 quantiles of residual ldl.

Usage

LDL_CAD_covar

Format

A data frame with 10 rows and 8 variables:

bx

linear association between ldl-cholesterol and allele score, adjusted

by

logistic regression coefficent between CAD diagnosis and allele score, adjusted

bxse

standard error for the bx term from the linear regression

byse

standard error for the by term from the logistic regression

xmean

Average ldl in stratum

xmin

Minimum ldl in stratum

xmax

Maximum ldl in stratum

...

---

Lower Coefficient value

Description

function to obtain the lower coefficient values (i.e. b in Text S1) for each pre-strata

Usage

lowfun(vec)

Arguments

vec	input vector

---

Piecewise linear figure

Description

piecewise_figure plots the piecewise linear function.

Usage

piecewise_summ_figure(
  xcoef,
  coef,
  xmean,
  lci,
  uci,
  xbreaks,
  family = "gaussian",
  ref = mean(xmean),
  pref_x = "x",
  pref_x_ref = "x",
  pref_y = "y",
  breaks = NULL,
  ci_fig = "point"
)

Arguments

xcoef	the association between the exposure and the instrument.
coef	the coefficients of the localized causal effects.
xmean	mean of each x stratum (for plotting point estimates)
lci	upper confidence interval range for each causal effects
uci	upper confidence interval range for each causal effects.
xbreaks	breakpoints in x (for plotting line estimates)
family	a description of the error distribution and link function to be used in the model. This is a character string naming either the gaussian (i.e. "gaussian" for continuous data) or binomial (i.e. "binomial" for binary data) family function.
ref	the reference point for the figure. This is the value of the that represents the expected difference in the outcome compared with this reference value when the exposure is set to different values. The default is the mean of x.
pref_x	the prefix/label for the x-axis. The default is "x".
pref_x_ref	the prefix for the reference value displayed on the y-axis. The default is "x".
pref_y	the prefix/label for the y-axis. The default is "y".
breaks	breaks on the y-axis of the figure.
ci_fig	point confidence intervals, or as ribbon ("point" or "ribbon")

Value

the plot of the piecewise linear function.

Author(s)

Amy Mason [email protected], leaning on work by James Statley and Matt Arnold

---

Instrumental variable analysis using piecewise linear method based on summary data

Description

piecewise_summ_mr performs instumental variable analysis by fitting piecewise linear functions to localised average causal effects

Usage

piecewise_summ_mr(
  by = NULL,
  bx = NULL,
  byse = NULL,
  bxse = NULL,
  xmean = NULL,
  xmin = NULL,
  xmax = NULL,
  summ = NULL,
  xbreaks = NULL,
  family = "gaussian",
  average.exposure.associations = FALSE,
  ci = "model_se",
  nboot = 1000,
  fig = T,
  ref = mean(xmean),
  pref_x = "x",
  pref_x_ref = "x",
  pref_y = "y",
  breaks = NULL,
  ci_fig = "point",
  seed = 875
)

Arguments

by	vector of gene-outcome associations.
bx	vector of gene-exposure associations.
byse	vector of standard errors of gene-outcome associations.
bxse	vector of standard errors of gene-exposure associations.
xmean	average value of the original exposure in each iv-free strata (or whatever summary of the exposure level in the stratum is desired).
xmin	min value of the original exposure in each stratum (see note)
xmax	max value of the original exposure in each stratum (see note)
summ	a dataframe contains bx, by, bxse, byse, xmean. This is the summary output of the create_nlmr_summary function; either as the full output or just passing the summary part.
xbreaks	break points for the stratum x values (see note)
family	a character string named either 'gaussian' (for continuous data) or binomial (for binary data) or cox (for survival data) family function. This only affects the plotting function - whether the y-axis is log-transformed or not, and the graph's default label. This should match whichever option was used in creating the summary data.
average.exposure.associations	TRUE means that the bx estimates are averaged across strata, FALSE means that they are not. Default option is FALSE.
ci	the type of 95\% confidence interval. There are four options: (i) using the model standard errors ('model_se'), (ii) using bootstrap standard errors ('bootstrap_se'), (iii) using bootstrap percentile confidence intervals ('bootstrap_per') The default is the model standard errors.
nboot	the number of bootstrap replications (if required). The default is 1000 replications.
fig	a logical statement as to whether the user wants the results displayed in a figure. The default is false.
ref	the reference point for the figure. This is the value of the function that represents the expected difference in the outcome compared with this reference value when the exposure is set to different values. The default is the mean of x.
pref_x	the prefix/label for the x-axis. The default is "x".
pref_x_ref	the prefix for the reference value displayed on the y-axis. The default is "x".
pref_y	the prefix/label for the y-axis. The default is "y".
breaks	breaks on the y-axis of the figure.
ci_fig	setting confidence interval type. "point" places error bars at the mean of each stratum; "ribbon" draws upper and lower piecewise lines.
seed	The random seed to use when generating the bootstrap samples (for reproducibility). If set to NA, the random seed will not be set.

Value

model the model specifications. The first column is the number of quantiles (q); the second column is the position used to relate x to the LACE in each quantiles (xpos); the third column is the type of confidence interval constructed (ci); the fourth column is the number of bootstrap replications performed (nboot).

powers the powers of the chosen polynomial.

coefficients the regression estimates. The first column is the regression coefficients (beta); the second column is the standard errors of regression coefficients (se); the third column is the lower confidence interval (lci); the fourth column is the upper confidence interval (uci); the fifth column is the p-value (pval).

Note

The non-linearity tests uses 'method="DL"' to calculate the p-value for the hetrogeneity trend. The fractional polynomial equivalent function allows you to set the method, meaning you may get different results.

There is no option for covariates; they would need to be applied at an earlier stage in the individual data, using the mr_summarise function.

The min and max of x stratum values are used to choose the appropiete range for fitting of each causal estimate. In the code for summarising data, this is set at the 10% point, and the 90% of each stratum; 20% and 80% in the external ends of the end strata. The first lower value and all upper value are used to set the break points for the estimates in the graph. Alternatively you can hardset this using xbreaks.

Author(s)

Amy Mason, leaning heavily on work by James Statley and Matt Arnold

---

Create plot of Fractional Polynomial Fit

Description

summary method for class 'frac_poly_mr'.

Usage

## S3 method for class 'frac_poly_mr'
print(x, ...)

Arguments

x	an object of class 'frac_poly_mr'.
...	additional arguments affecting the summary produced.

Author(s)

James R Staley [email protected]

---

Print Summary Fractional Polynomial Fits

Description

print.summary method for class 'frac_poly_mr'.

Usage

## S3 method for class 'summary.frac_poly_mr'
print(x, ...)

Arguments

x	an object of class 'frac_poly_mr'.
...	additional arguments affecting the summary produced.

Author(s)

James R Staley [email protected]

---

Print summary of piecewise linear fits

Description

print summary method for class "piecewise_mr".

Usage

## S3 method for class 'summary.piecewise_summ_mr'
print(x, ...)

Arguments

x	an object of class "piecewise_mr".
...	Arguments to be passed to or from other methods,

Author(s)

Amy Mason [email protected]

---

Repeat rows

Description

This function creates a matrix of a repeated vector where each row is the same.

Usage

reprow(x, n)

Arguments

x	vector to be repeated
n	number of repeats

Author(s)

James R Staley [email protected]

---

Summarizing Fractional Polynomial Fits

Description

summary method for class 'frac_poly_mr'.

Usage

## S3 method for class 'frac_poly_mr'
summary(object, ...)

Arguments

object	an object of class 'frac_poly_mr'.
...	additional arguments affecting the summary produced.

Author(s)

James R Staley [email protected]

---

Summary of piecewise linear fits

Description

summary method for class "piecewise_summ_mr".

Usage

## S3 method for class 'piecewise_summ_mr'
summary(object, ...)

Arguments

object	an object of class "piecewise_summ_mr".
...	Arguments to be passed to or from other methods,

Author(s)

Amy Mason [email protected]

---

Upper Coefficient value

Description

function to obtain the upper coefficient values (i.e. a in Text S1) for each pre-strata

Usage

upfun(vec)

Arguments

vec	input vector

---