---
title: 'simdata: NORTA based simulation designs'
author: "Michael Kammer"
date: "`r Sys.Date()`"
output:
rmarkdown::html_vignette:
fig_width: 8
fig_height: 6
toc: yes
toc_depth: 2
number_sections: yes
html_document:
toc: yes
toc_depth: '2'
df_print: paged
vignette: |
\usepackage[utf8]{inputenc}
%\VignetteIndexEntry{simdata: NORTA based simulation designs}
%\VignetteEngine{knitr::rmarkdown}
bibliography: Demo_references.bib
---
```{r message=FALSE, warning=FALSE, include=FALSE}
library(simdata)
library(fitdistrplus)
library(dplyr)
library(ggplot2)
library(patchwork)
library(ggcorrplot)
knitr::opts_chunk$set(
error = TRUE,
eval = requireNamespace("NHANES", quietly = TRUE)
)
```
# Introduction
This document describes the workflow to define NORmal-To-Anything (NORTA)
based simulation designs using the `simdata` package. The method is very
useful to re-create existing datasets through a parametric
approximation for usage in simulation studies. It is also quite easy to use,
and allows the definition of presets for sharing simulation setups.
General details of the methodology and further references are given in e.g.
@NORTA1 and @NORTA2.
In this vignette we will prefix all relevant function calls by `::` to show
the package which implements the function - this is not necessary but only
done for demonstration purposes.
## Outline of NORTA {#outline}
The goal of the NORTA procedure is to produce identically independently
distributed (iid) samples from random variables with a given
correlation structure (Pearson correlation matrix) and given marginal
distributions, thereby e.g. approximating existing datasets.
Following @NORTA2, we want to sample iid replicates of the random vector
$X = (X_1, X_2, \ldots, X_k)$. Denote by $F_i(s) = P(X_i \leq s)$ the
distribution functions (i.e. the marginal distributions) of the components of
$X$, and by $\Sigma_X$ the $k \times k$ correlation matrix of $X$. Then NORTA
proceeds as follows:
- Generate multivariate standard normal random vectors (i.e mean 0, variance 1)
$Z = (Z_1, Z_2, \ldots, Z_k)$ with a correlation matrix $\Sigma_Z$.
- Compute the random vector $X$ via $X_i := F_i^{-1}(\Phi(Z_i))$, where $\Phi$
denotes the distribution function of the standard normal distribution, and
$F_i^{-1}(t) := \inf\{x: F_i(x) \geq t\}$ is the quantile function of $X_i$.
The resulting vector $X$ then has the desired marginal distribution. To obtain
the target correlation structure $\Sigma_X$, the correlation matrix $\Sigma_Z$
for the first step has to be chosen appropriately. This can be achieved
via solving univariable optimisation problems for each pair of variables
$X_i$ and $X_j$ in $X$ and is part of the `simdata` package.
## Caveats of NORTA
The NORTA procedure has some known limitations, which may lead to discrepancies
between the target correlation structure and the correlation structure obtained
from the sampling process. These are, however, partly alleviated when using
existing datasets as templates, or by special techniques within `simdata`.
- Not all combinations of given marginal distributions and target correlation
are feasible by the nature of the variables. This is not an issue when an
existing dataset is used as template, since that demonstrates that
the combination exists.
- The optimisation procedure to obtain $\Sigma_Z$ may lead to a matrix which
is not positive definite (since the optimisation is only done for pairs of
variables), and therefore not a proper correlation matrix. To
alleviate this, the `simdata` package ensures positive definiteness by using
the closest positive definite matrix instead. This may lead to discrepancies
between the target correlation and the achieved correlation structure.
- NORTA cannot reproduce non-linear relationships between variables. This
may lead to issues for continuous variables and categorical variables with more
than two categories when the goal is to faithfully re-create a real dataset that
features non-linear relations.
- The optimisation procedure to obtain $\Sigma_Z$ may take a while to compute
when the number of variables increases. This is alleviated through the fact
that this computation has to be done only a single time, during definition
of the simulation design. All further simulation iterations only use the
optimisation result and are therefore not subject to this issue.
- When applied to an existing dataset, NORTA relies on the estimation of the
target correlation matrix and marginal distributions. More complex
data (e.g. special marginal distributions, complex correlation structure)
therefore requires more observations for an accurate representation.
## Comparison to other methods
NORTA is well suited to re-create existing datasets through an explicit
parametric approximation. Similar methods exist, that achieve this through
other means. A particularly interesting alternative is the generation
of synthetic datasets using an approach closely related to multiple imputation,
and is implemented in e.g. the `synthpop` R package (@Synthpop). Its' primary
aim is to achieve confidentiality by re-creating copies to be shared for
existing, sensitive datasets.
In comparison, `synthpop` potentially offers more flexible data generation than
NORTA, thereby leading to a better approximation of an original dataset.
However, `synthpop` is also more opaque than the explicit, user defined
specification of correlation and marginal distributions of NORTA. This also
entails that `synthpop` can be generally used more like a black-box approach,
which requires little user input, but is also less transparent than the
manual curation of the simulation setup in NORTA. Furthermore, NORTA allows
easy changes to the design to obtain a wide variety of study designs from a
single template dataset, whereas `synthpop` is more targeted at re-creating
the original dataset.
Both methods therefore have their distinct usecases and complement each other.
# Workflow in `simdata`
Given the [outline of the method](#outline), all the user has to specify to
define a NORTA design on $k$ variables are
- A $k \times k$ target correlation matrix
- The $k$ marginal distributions for each variable, given as quantile functions
These can be estimated from existing datasets of interest. `simdata` offers a
helper function to automate this process, but the user can also specify
the required input manually. We demonstrate both use cases in the
[example below](#example).
## Quantile functions for some common distributions
The required marginal distributions are given as quantile functions. R provides
implementations of many standard distributions which can be directly used, see
the help on `distributions`. The quantile functions use the prefix "q", as in
e.g. `qnorm` or `qbinom`. Further implementations can be found in the
packages `extraDistr`, `actuar` and many others (see
https://CRAN.R-project.org/view=Distributions).
# Example {#example}
In this example we will setup a NORTA based simulation design for a dataset
extracted from the National Health And Nutrition Examination Survey (NHANES),
accessible in R via several packages (we use the `NHANES` package in this
demo).
## Load dataset
First we will load the dataset and extract several variables of interest,
namely gender ('Gender'), age ('Age'), race ('Race'), weight ('Weight'),
bmi ('BMI'), systolic ('BPsys') and diastolic blood pressure ('BPdia'). These
variabes demonstrate several different kinds of distributions. For a detailed
description of the data, please see the documentation at
https://www.cdc.gov/nchs/nhanes.htm and for the `NHANES` R package on CRAN.
Here we are not concerned with the exact codings of the variables, so we will
remove labels to factor variables and work with numeric codes. Further, we will
only use data from a single survey round to keep the dataset small.
```{r eval=TRUE, echo=TRUE, message=FALSE, warning=FALSE}
if (requireNamespace("NHANES", quietly = TRUE)) {
df = NHANES::NHANES %>%
filter(SurveyYr == "2011_12") %>%
select(Gender, Age, Race = Race1, Weight,
BMI, BPsys = BPSys1, BPdia = BPDia1) %>%
filter(complete.cases(.)) %>%
filter(Age > 18) %>%
mutate(Gender = if_else(Gender == "male", 1, 2),
Race = as.numeric(Race))
print(head(df))
} else {
message("Package 'NHANES' not available.")
}
```
## Estimate target correlation
Using this dataset, we first define the target correlation `cor_target` and
plot it.
```{r echo=TRUE}
cor_target = cor(df)
ggcorrplot::ggcorrplot(cor_target, lab = TRUE)
```
## Define marginal distributions
Further, we define a list of marginal distributions `dist` representing the
individual variables. Each entry of the list must be a function in one
argument, defining the quantile function of the variable. The order of the
entries must correspond to the order in the target correlation `cor_target`.
### Automatically
`simdata` offers the helper function `simdata::quantile_functions_from_data()`
to automate estimation of quantile functions from the available data. It does so
non-parametrically and implements two approaches, one more suited for
categorical data, and the other more suited to continuous data. In practice
the parameter `n_small` can be used to determine a number of unique values
required to use the latter approach, rather than the former. See the
documentation for more details.
```{r echo=TRUE, results='hide', warning=FALSE}
dist_auto = quantile_functions_from_data(df, n_small = 15)
```
### Manually
We use the `fitdistrplus::fitdist` function to find
appropriate distribution candidates and fit their parameters. Decisions
regarding the fit of a distribution can be made using e.g. the
Akaike information criterion (AIC) or Bayesian information criterion (BIC)
displayed by the summary of the fit object returned by the function (the lower
their values, the better the fit).
In case a parametric distribution doesn't fit very well, we instead make use
of a density estimate and use this to define the marginal quantile function.
- Gender: a binomial distribution with $P(2) \approx 0.5$.
- Age: the distribution is not very "nice".
- We approximate it using a kernel density estimate using the
`stats::density` function.
- Note that the boundaries of the distribution can be more or less smoothed
with the `cut` parameter.
- To obtain a quantile function, first we integrate the density, normalize
it, and then use `stats::approxfun` to derive a univariable quantile
function.
- Race: a categorical distribution with 5 categories specified by probabilities
- Frequencies of categories were identified by a simple call to `table()`
on the full dataset.
- Can also be implemented using the categorical distribution from the
package `LaplacesDemon` implemented via `qcat`
- Weight: gamma distribution parameters estimated using `fitdistrplus::fitdist`
- BMI and systolic blood pressure: log-normal distribution parameters
estimated using `fitdistrplus::fitdist`
- Diastolic blood pressure: normal distribution parameters estimated using
`fitdistrplus::fitdist` after removing zero values from the data
The code to implement these marginal distributions is shown below.
```{r echo=TRUE, results='hide', warning=FALSE}
dist = list()
# gender
dist[["Gender"]] = function(x) qbinom(x, size = 1, prob = 0.5)
# age
dens = stats::density(df$Age, cut = 1) # cut defines how to deal with boundaries
# integrate
int_dens = cbind(Age = dens$x, cdf = cumsum(dens$y))
# normalize to obtain cumulative distribution function
int_dens[, "cdf"] = int_dens[, "cdf"] / max(int_dens[, "cdf"])
# derive quantile function
# outside the defined domain retun minimum and maximum age, respectively
dist[["Age"]] = stats::approxfun(int_dens[, "cdf"], int_dens[, "Age"],
yleft = min(int_dens[, "Age"]),
yright = max(int_dens[, "Age"]))
# race
dist[["Race"]] = function(x)
cut(x, breaks = c(0, 0.112, 0.177, 0.253, 0.919, 1),
labels = 1:5)
# weight
fit = fitdistrplus::fitdist(as.numeric(df$Weight), "gamma")
summary(fit)
dist[["Weight"]] = function(x) qgamma(x, shape = 16.5031110, rate = 0.2015375)
# bmi
fit = fitdistrplus::fitdist(as.numeric(df$BMI), "lnorm")
summary(fit)
dist[["BMI"]] = function(x) qlnorm(x, meanlog = 3.3283118, sdlog = 0.2153347)
# systolic blood pressure
fit = fitdistrplus::fitdist(as.numeric(df$BPsys), "lnorm")
summary(fit)
dist[["BPsys"]] = function(x) qlnorm(x, meanlog = 4.796213, sdlog = 0.135271)
# diastolic blood pressure
fit = fitdistrplus::fitdist(as.numeric(df %>%
filter(BPdia > 0) %>%
pull(BPdia)), "norm")
summary(fit)
dist[["BPdia"]] = function(x) qnorm(x, mean = 71.75758, sd = 11.36352)
```
### What to use?
Both, the automatic and the manual way to specify marginals may be useful.
The automatic way works non-parametrically which may be useful when a real
dataset should be re-created, while the manual way allows to
specify marginals parametrically which may be useful when the data is defined
from purely theoretical specifications.
## Simulate data
Now we can use `simdata::simdesign_norta` to obtain designs using both the
manual and automated marginal specifications.
After that, we simulate datasets of the same size as the
original data set using `simdata::simulate_data`, and compare the resulting
summary statistics and correlation structures.
```{r}
# use automated specification
dsgn_auto = simdata::simdesign_norta(cor_target_final = cor_target,
dist = dist_auto,
transform_initial = data.frame,
names_final = names(dist),
seed_initial = 1)
simdf_auto = simdata::simulate_data(dsgn_auto, nrow(df), seed = 2)
# use manual specification
dsgn = simdata::simdesign_norta(cor_target_final = cor_target,
dist = dist,
transform_initial = data.frame,
names_final = names(dist),
seed_initial = 1)
simdf = simdata::simulate_data(dsgn, nrow(df), seed = 2)
```
## Results
Summary statistics of the original and simulated datasets.
```{r message=FALSE, warning=FALSE}
summary(df)
summary(simdf_auto)
summary(simdf)
```
Correlation structures of the original and simulated datasets.
```{r echo=FALSE, message=FALSE, warning=FALSE}
ggcorrplot::ggcorrplot(cor(df), title = "Original", lab = TRUE)
ggcorrplot::ggcorrplot(cor(simdf_auto), title = "Simulated (automated)", lab = TRUE)
ggcorrplot::ggcorrplot(cor(simdf), title = "Simulated (manual)", lab = TRUE)
```
We may also inspect the continuous variables regarding their univariate and
bivariate distributions. The original data is shown in black, the simulated
data is shown in red. (Note that we only use the first 1000 observations to
speed up the plotting.)
```{r echo=FALSE, message=FALSE, warning=FALSE}
vars = c("Age", "Weight", "BMI", "BPsys", "BPdia")
limits = list(
Age = c(10, 90),
Weight = c(0, 225),
BMI = c(10, 80),
BPsys = c(65, 240),
BPdia = c(0, 140)
)
plist = list()
for (i in seq_along(vars)) for (j in seq_along(vars)) {
vari = vars[i]
varj = vars[j]
if (i == j) {
p = ggplot(df, aes_string(x = vari)) +
geom_density() +
geom_density(data = simdf, color = "red") +
coord_cartesian(xlim = limits[[vari]])
} else if (i < j) {
p = ggplot(df[1:1000, ], aes_string(x = vari, y = varj)) +
geom_point(alpha = 0.04) +
coord_cartesian(xlim = limits[[vari]], ylim = limits[[varj]])
} else {
p = ggplot(simdf_auto[1:1000, ], aes_string(x = varj, y = vari)) +
geom_point(color = "red", alpha = 0.04) +
coord_cartesian(xlim = limits[[varj]], ylim = limits[[vari]])
}
plist = append(plist, list(p + theme_bw(base_size = 7)))
}
p = patchwork::wrap_plots(plist) +
patchwork::plot_annotation(title = "Simulated (automated)")
print(p)
plist = list()
for (i in seq_along(vars)) for (j in seq_along(vars)) {
vari = vars[i]
varj = vars[j]
if (i == j) {
p = ggplot(df, aes_string(x = vari)) +
geom_density() +
geom_density(data = simdf, color = "red") +
coord_cartesian(xlim = limits[[vari]])
} else if (i < j) {
p = ggplot(df[1:1000, ], aes_string(x = vari, y = varj)) +
geom_point(alpha = 0.04) +
coord_cartesian(xlim = limits[[vari]], ylim = limits[[varj]])
} else {
p = ggplot(simdf[1:1000, ], aes_string(x = varj, y = vari)) +
geom_point(color = "red", alpha = 0.04) +
coord_cartesian(xlim = limits[[varj]], ylim = limits[[vari]])
}
plist = append(plist, list(p + theme_bw(base_size = 7)))
}
p = patchwork::wrap_plots(plist) +
patchwork::plot_annotation(title = "Simulated (manual)")
print(p)
```
From this we can observe, that the agreement between the original data and
the simulated data is generally quite good. Both, automated and manual
specification work equally well for this dataset. Note, however, that e.g. the
slightly non-linear relationship between age and diastolic blood pressure
cannot be fully captured by the approach, as expected. Furthermore, the
original data shows some outliers, which are also not reproducible due to the
parametric nature of the NORTA procedure.
# R session information {#rsession}
```{r echo=FALSE}
sessionInfo()
```
# References