p-curve Analysis in Python
You can install pypcurve with pip:
pip install pypcurve
First and foremost, read the user guide to the p-curve. It is crucial that users understand what p-curve can and cannot do, that they know which statistical results to select, and that they properly prepare the disclosure table.
pypcurve only requires a list of statistical results, stored in a list (or an array). Similar to the p-curve app, pypcurve accepts the following formats of statistical tests:
In addition, pypcurve will accept raw p-values:
This is not recommended though: p-values are often weirdly rounded, so enter the statistical result instead if it is reported in the paper.
For this example, I will assume that your tests have been properly formatted, and stored in a column called “Tests” of a .csv file.
from pypcurve import PCurve import pandas as pd df = pd.read_csv("mydata.csv") pc = PCurve(df.Tests)
If all your tests are properly formatted, there will be no error, and pcurve will be initialized properly.
Next, you can print the summary of the p-curve, as you would see it using the web-app:
This will output the p-curve plot, as well as the table summarizing the binomial and Stouffer tests of the
p-curve analysis. You can get the plot alone, or the table alone, using the methods
You can use pycurve to estimate the power of the design that generated the statistical tests:
pc.estimate_power()will return the power estimate, and the (lower, upper) bounds of 90% confidence interval.
pc.plot_power_estimate()will plot the power estimate (as the webapp does).
You can directly access the results of the p-curve analysis using three methods:
pc.get_stouffer_tests()will recover the Z and p-values of the Stouffer tests
pc.get_binomial_tests()will recover the p-values of the binomial tests
pc.get_results_entered()will recover the statistical results entered in the p-curve, and the pp-values and z scores associated with the different alternatives to which they are compared.
You can also directly check if the p-curve passes the cutoff for evidential value, and the cutoff for
inadequate evidence (as defined in
using the properties
The app is still in beta, so please take care when interpreting the results. I have tested pypcurve against the p-curve app using multiple examples: There are occasional minor deviations between the two, because of the way R (vs. Python) compute the non-central F distribution.
First beta release.