Calculating Global Spatial Autocorrelation with Moran's I

The Formula Behind the Statistic

Before we dive into the Python code, it's valuable to understand what the Moran's I calculation actually represents. While the formula can look intimidating, its logic is quite intuitive. It essentially compares the similarity of neighboring values to the overall variation in the data.

A fantastic video by ritvikmath breaks this down into simple, logical steps using a grid-based example. This will give you a strong conceptual model of what the pysal library will be doing under the hood.

After watching, the key takeaway should be this: Moran's I is high and positive when neighboring locations have similar values (high-high or low-low), and this similarity is greater than what you'd expect from random chance.

Calculating Moran's I in Python

Now, let's translate this concept into code. For your project, you'll have a GeoDataFrame where each row is a district (a polygon) and one of the columns is the violation_rate you've calculated. The process involves two main steps: defining the spatial weights and then running the test.

We'll use the pysal ecosystem of libraries, specifically libpysal for weights and esda (Exploratory Spatial Data Analysis) for the Moran's I statistic.

Step 1: Create the Spatial Weights Matrix

First, we must formally define "neighbors". Since you are working with district polygons, the most intuitive approach is contiguity. Two districts are neighbors if they share a border. The pysal library offers two common types:

• Rook Contiguity: Polygons are neighbors if they share a line segment.
• Queen Contiguity: Polygons are neighbors if they share a line segment or a single point (a corner).

Queen is the more common and inclusive choice, so we'll use that. The following reading provides a clear, practical example of creating a weights matrix and then using it to compute Moran's I.

As you can see from the reading, the code to perform this complex statistical test is surprisingly concise:

import libpysal.weights
import esda

# Assuming 'gdf' is your GeoDataFrame with a 'violation_rate' column
# and a 'geometry' column

# 1. Create the spatial weights matrix (using Queen contiguity)
w = libpysal.weights.Queen.from_dataframe(gdf)

# Optional but good practice: row-standardize the weights
w.transform = 'R'

# 2. Calculate Moran's I
y = gdf['violation_rate']
moran = esda.Moran(y, w)

# 3. Print the results
print(f"Moran's I: {moran.I}")
print(f"p-value: {moran.p_sim}")

This simple block of code is the core of your global spatial autocorrelation analysis.

Interpreting the Results

The moran object contains two crucial pieces of information: the statistic I and the p-value p_sim.

1. Moran's I (moran.I): This value typically ranges from -1 to +1.

   • Positive I (approaching +1): Indicates positive spatial autocorrelation, or clustering. Districts with high violation rates tend to be near other districts with high rates, and districts with low rates are near other low-rate districts. This is evidence of hotspots and coldspots.
   • Negative I (approaching -1): Indicates negative spatial autocorrelation, or dispersion. High-rate districts are systematically located next to low-rate districts (a checkerboard pattern). This is less common in social data.
   • I near 0: Indicates no spatial pattern, or spatial randomness. The location of a district gives no information about the violation rates of its neighbors.

2. The p-value (moran.p_sim): This tells you if your result is statistically significant. The test's null hypothesis is that the values are randomly distributed across your map. The p-value is the probability of observing a Moran's I value as extreme as yours (or more extreme) if the null hypothesis were true.

   • A low p-value (e.g., < 0.05) gives you evidence to reject the null hypothesis. You can conclude that the spatial pattern you observe (clustering or dispersion) is unlikely to be the result of random chance.
   • A high p-value (e.g., > 0.05) means you cannot reject the null hypothesis. You don't have enough evidence to say the pattern is not random.

The following image provides a great visual summary of how to interpret the results based on the Z-score (which is related to the Moran's I value) and the p-value.

This diagram illustrates how the Moran's I statistic relates to a standard normal distribution. A high positive score (right tail) indicates significant clustering, a high negative score (left tail) indicates significant dispersion, and a score in the middle suggests a random pattern.

The workshop from UCLA provides a very intuitive explanation of how this p-value is generated through permutation, which powerfully illustrates the concept of statistical significance.

This simulation is why the p-value is often named p_sim—it's a simulated p-value. For your project, a significant positive Moran's I (e.g., I = 0.45, p = 0.001) would be a major finding, confirming that the clusters of high and low violation rates on your map are statistically meaningful.

This example shows a Moran's I of 0.47 and a Z-score of 45.92, which corresponds to a p-value of less than 0.01. This allows us to reject the null hypothesis (H₀) of a random spatial pattern and conclude that the data is significantly clustered.

Conclusion

You have now bridged the gap from visual pattern recognition to rigorous statistical testing. By computing Global Moran's I, you can make a definitive statement about the overall spatial structure of meal quality violations in your data.

Here are the key takeaways from this lesson:

• Moran's I is a global statistic that quantifies the overall clustering or dispersion in your dataset with a single number.
• The calculation requires a spatial weights matrix (W), typically based on Queen contiguity for polygon data, to define which locations are neighbors.
• The esda library in Python makes the computation straightforward: esda.Moran(values, weights).
• Interpretation requires checking both the I value (for direction and strength of correlation) and the p_sim value (to confirm statistical significance and rule out randomness).

With this tool, you can now answer the question: "Is there a significant geographic pattern to meal quality violations?" However, this is a global statistic. It tells you if clustering exists on the map as a whole, but it doesn't tell you where the specific clusters are.

In our next lesson, we will address that exact question by moving from global to local analysis. We will compute Local Moran's I (LISA) to identify the specific districts that form statistically significant hotspots and coldspots.