Code tips and hacks

2025-10-21

library(pacman)

R

Plotting

shadowtext Plot with outline text in different colors

See here: https://cran.r-project.org/web/packages/shadowtext/vignettes/shadowtext.html OR

ggfx: https://ggfx.data-imaginist.com/

p_load(shadowtext)

or use

Make a blank plot object

p_load(ggplot2, patchwork)

blank_plot <- wrap_elements(
    ggplot() + geom_blank() +
    theme(
        panel.border = element_blank(),
        plot.background = element_blank()
    )
)

blank_plot <- wrap_elements(ggplot() + theme_void())

Descriptive

Make quick and beautiful tables with gtsummary

p_load(gtsummary)

Summarize a dataset quickly with skimr

p_load(skimr)

Modeling

Easy summary of statistical test

https://easystats.github.io/report/

p_load(report)

Modeling with constraints:

In case anyone needs to do this, I figured out how to fit a linear model with constraints on the coefficients. The nls function for non-linear least squares fitting supports setting bounds on the coefficients. For example, let’s say we have a simple linear model y as a function of x:

library(parameters)

x <- rnorm(1000)
y <- x + rnorm(1000)
fit <- lm(y ~ x)
parameters(fit)

First, we rewrite this model using nls. The formula is a bit different because we need to name the coefficients explicitly, and supply starting values for the iterative optimization. It’s called non-linear least squares because it allows non-linear functions, but you’re still allowed to specify a linear relationship. (In general, it’s important to choose an appropriate starting point, but in the case of a linear model I’m pretty sure the optimization is convex, so any starting value will eventually converge to the global optimum.)

nlfit <- nls(
  y ~ a * x + b,
  start = c(a = 0, b = 0)
)
parameters(nlfit)

In this case, we get the same result as lm, because there’s no constraints. Next, let’s say that for some reason, we want to slope (a) to be exactly 0.9, and then we want to estimate the intercept (b) conditional on the slope being 0.9. We specify the start point as well as the lower and upper bounds to be 0.9 for a, and for b we specify the bounds as +/-Inf since we want it to be unbounded. We also have to set algorithm = “port” because this is the algorithm that supports coefficients bounds.

nlfit <- nls(
  y ~ a * x + b,
  start = c(a = 0.9, b = 0),
  lower = c(a = 0.9, b = -Inf),
  upper = c(a = 0.9, b = Inf),
  algorithm = "port"
)
parameters(nlfit)

One important note: I don’t know if you can use the p-values for any coefficients with bounds, because I’m pretty sure the significance is still computed while ignoring the bounds. So the p-value for a in this model is meaningless. However, I believe the p-value for b is still valid, but it is conditional on the bounds of a.Lastly, if you want confidence intervals on the fitted line (e.g. for plotting in the style of geom_smooth), you can get them with investr::predFit:

library(investr)
nlpr <- predFit(nlfit, newdata = data.frame(x = sort(x)), interval = "confidence")
head(nlpr)

(Normally you would get confidence intervals with predict, but the nls method for this function does not support confidence intervals.) (edited)

Slightly related to this, I also learned recently it is possible to set monotonic constraints in XGBoost - useful for clinical models where you expect risk to increase monotonically with risk factor https://xgboost.readthedocs.io/en/stable/tutorials/monotonic.html

Bash

retcon a screen session

Cool minerva trick of the day: Have you ever accidentally started a long-running command outside of a screen session? It turns out there’s a command you can use to transfer an already running command to a new shell. Just do module load reptyr and then use reptyr PID inside the screen session to transfer your running command into that screen window. PID is the process ID of the running command. You can get it using ps if you know how to use that, or else just do module load htop; htop and then scroll through the list of commands with the arrow keys until you find the command you want to transfer.

Code Tips and Tricks - Running Log