Parametric vs. Non-Parametric Tests: When to Use Which

1. The Fundamental Choice: Assumptions vs. Flexibility

At the heart of many statistical tests are assumptions about the data's underlying distribution. This is the main dividing line between our two types of tests.

• Parametric Tests: These tests assume that your data was drawn from a specific probability distribution, which can be described by a set of parameters. The most common assumption is that the data follows a normal distribution, characterized by its mean ($\mu$) and standard deviation ($\sigma$). The t-test you learned about is a prime example.
• Non-Parametric Tests: These tests are "distribution-free." They don't make assumptions about the underlying distribution of the data. This makes them more flexible and robust.

So, why not always use non-parametric tests? The main trade-off is statistical power. If the assumptions of a parametric test are met, it is more powerful—meaning it's more likely to detect a statistically significant effect if one truly exists.

To get a clear overview of the factors that guide this choice, let's watch a short video.

As the video explained, the decision hinges on a few key checks:

1. Normality: Is the data approximately normally distributed?
2. Sample Size: A larger sample size (often n > 30) can sometimes make parametric tests robust even with skewed data, thanks to the Central Limit Theorem.
3. Outliers: Parametric tests are sensitive to extreme values, whereas non-parametric tests are more robust.
4. Data Type: Non-parametric tests are often the only choice for ordinal data (like rankings or satisfaction ratings).

2. The "Magic" of Non-Parametric Tests: Using Ranks

You might be wondering how a test can work without making assumptions about the data's distribution. The most common technique is to transform the data by ranking it. Instead of using the raw values, the test uses their position in a sorted list. This simple transformation makes the test robust to outliers and the underlying distribution.

Let's watch another short video that illustrates this concept beautifully.

By converting values to ranks, the exact magnitude of outliers becomes less important. For example, in a dataset [1, 2, 3, 100], the value 100 would have a huge influence on the mean. But its rank is simply 4. This ranking mechanism is the key to the robustness of non-parametric tests.

3. A Practical Decision-Making Framework

Now let's put this all together into a step-by-step process. When you're faced with a dataset and need to choose a test, you can follow this logic.

Step 1: Understand Your Goal and Data

• Question: Are you comparing independent groups, paired groups, or looking for a relationship (correlation)?
• Data Type: Is your data continuous (interval/ratio), ordinal, or nominal?

Step 2: Check the Assumptions for a Parametric Test

The article "Parametric vs. Non-Parametric Tests and When to Use Them" provides a great summary of these assumptions and the decision criteria.

The most critical assumption to check is normality. You can assess this both visually and with a statistical test.

• Visual Check: A histogram or a Q-Q (Quantile-Quantile) plot. A Q-Q plot plots the quantiles of your data against the quantiles of a theoretical normal distribution. If the data is normal, the points will fall along a straight 45-degree line.
• Statistical Check: The Shapiro-Wilk test is a common choice. Its null hypothesis is that the data is normally distributed. Therefore, if you get a small p-value (e.g., < 0.05), you reject the null hypothesis and conclude your data is likely not normal.

Here is a quick Python example showing how to generate a Q-Q plot, which is a very useful skill for your data science toolkit.

import numpy as np
import statsmodels.api as sm
import matplotlib.pyplot as plt




# Generate some data that is NOT normal (e.g., skewed)
np.random.seed(42)
non_normal_data = np.random.exponential(size=100)




# Generate some data that IS normal
normal_data = np.random.normal(loc=0, scale=1, size=100)




# Create Q-Q plots
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 5))

sm.qqplot(non_normal_data, line='45', ax=ax1)
ax1.set_title('Q-Q Plot for Non-Normal Data (Exponential)')

sm.qqplot(normal_data, line='45', ax=ax2)
ax2.set_title('Q-Q Plot for Normal Data')

plt.show()

Notice how the points for the normal data stick closely to the red line, while the points for the non-normal data deviate significantly.

Step 3: Choose Your Test

Based on the checks above, you can now make an informed choice. The flowchart below is a great visual summary of the entire decision process.

4. Common Parametric/Non-Parametric Pairs

To make this more concrete, here is a table of common statistical tasks and their corresponding parametric and non-parametric tests. You've already encountered the t-test and will soon learn about correlation.

If you need to implement these non-parametric tests in Python, the scipy.stats library has you covered. The article "How to Calculate Nonparametric Statistical Hypothesis Tests in Python" by Jason Brownlee is an excellent hands-on guide for this, which you can save for future reference. For example, running a Mann-Whitney U test is just one line of code:

from scipy.stats import mannwhitneyu



# data1, data2 are your two independent samples
stat, p_value = mannwhitneyu(data1, data2)

Conclusion

Great work today! You now have a solid framework for one of the most common decisions in statistical analysis: choosing the right tool for the job. This prevents you from applying tests incorrectly and drawing invalid conclusions.

Key Takeaways:

• Parametric tests are powerful but rely on strict assumptions about your data's distribution (e.g., normality).
• Non-parametric tests are more flexible and robust because they don't assume a specific distribution, often by using ranks instead of raw data.
• The choice involves a trade-off: power vs. flexibility. Always prefer a parametric test if your data meets its assumptions.
• You can check for normality using visual tools like Q-Q plots and statistical tests like the Shapiro-Wilk test.

Preview of the next lesson:

In this lesson, we briefly mentioned the Pearson and Spearman tests for measuring relationships. In our next session, we will dive deep into this topic. We'll learn how to calculate and interpret correlation coefficients, and more importantly, discuss the critical distinction between correlation and causation—a concept that is absolutely fundamental to thinking like a data scientist.