Calculate Cook’s Distance in SAS ?

The Mathematical Foundation: Defining Cook’s Distance

As a foundational metric, Cook’s distance serves as a composite measure that encapsulates both the discrepancy (outlier status) and the position (leverage) of an observation. This combination is essential because an observation must be both unusual (a large residual) and far from the center of the predictor variables (high leverage) to exert significant influence on the regression model coefficients. It is calculated by considering the sum of squared differences between the predicted responses when all observations are included versus when the specific observation ($i$) is excluded.

The rigorous definition of the metric is encapsulated in the following formula:

D_i = (r_i² / p*MSE) * (h_ii / (1-h_ii)²)

Understanding the components of this formula is critical for grasping why certain points are deemed highly influential:

• r_i is the i^th residual. This term reflects how poorly the model predicts the observed value for point $i$. A larger residual contributes to a larger Cook’s Distance.
• p is the number of coefficients (or predictors plus intercept) in the regression model. This acts as a scaling factor, normalizing the influence based on model complexity.
• MSE is the Mean Squared Error. This is an estimate of the variance of the error terms.
• h_ii is the i^th leverage value. This measures how far the observation’s predictor values are from the mean of the predictor values. High leverage indicates the observation has the potential for high influence.

Interpreting Cook’s Distance: Thresholds and Influence

Essentially, the Cook’s distance calculation quantifies the total change in all fitted values of the model that occurs when the $i^{text{th}}$ observation is hypothetically removed from the dataset. It is a comprehensive measure that provides a single value representing the overall impact of that specific data point. If the observation is not influential, its removal will cause negligible changes to the overall regression estimates, resulting in a Cook’s Distance close to zero.

The magnitude of the value directly corresponds to the degree of influence: the larger the calculated value for Cook’s distance, the more influential the given observation is on the established parameter estimates. Identifying these influential points is critical because they can distort estimates of the slope and intercept, leading to biased conclusions about the relationship between variables.

While there is no universally fixed cutoff, a commonly cited rule of thumb suggests that any observation with a Cook’s distance value greater than $4/n$ (where $n$ represents the total number of observations in the sample) should be flagged as highly influential and warrants further investigation. Some literature also suggests using a cutoff value of 1.0, though the $4/n$ criterion is often preferred for smaller to moderately sized datasets as it scales relative to the sample size. The following example demonstrates the necessary steps to calculate and interpret Cook’s distance for each observation in a linear regression model using SAS.

Prerequisites for Calculation in SAS

Before diving into the actual calculation of Cook’s Distance, it is essential to ensure that the data is correctly loaded and structured within the SAS environment. We utilize the DATA step to create a simple dataset suitable for a linear regression model. This dataset contains two variables: ‘x’ (the independent predictor) and ‘y’ (the dependent response). For this illustrative case, we use a relatively small sample size ($n=12$), which makes the impact of individual observations particularly noticeable.

It is important to remember that while Cook’s Distance is most frequently associated with ordinary least squares (OLS) linear regression, the underlying principles of measuring influence apply broadly across various generalized linear models, though the computational implementation might vary. The data preparation step shown below is crucial for setting up the environment required by PROC REG.

Step-by-Step Example: Preparing the Dataset

For our demonstration, we define and input the sample data directly into a SAS dataset named my_data. The following code demonstrates the creation and initial inspection of this dataset, confirming the structure of our 12 observations:

/*create dataset*/
data my_data;
    input x y;
    datalines;
8 41
12 42
12 39
13 37
14 35
16 39
17 45
22 46
24 39
26 49
29 55
30 57
;
run;

/*view dataset*/
proc print data=my_data;

Upon execution, the PROC PRINT command yields a basic tabular visualization of the data points, ensuring data integrity before proceeding with the regression modeling phase.

This dataset now forms the basis upon which we will fit a simple linear model, analyzing how the ‘x’ variable relates to the ‘y’ variable, and subsequently diagnosing the influence of each row.

Utilizing PROC REG for Cook’s Distance Calculation

The efficiency of SAS shines when calculating influence diagnostics. The primary tool for this is the PROC REG procedure. This procedure not only handles the fitting of the model but also allows for the computation of various diagnostic statistics, including Cook’s Distance. To obtain these statistics, we must utilize the OUTPUT statement within the procedure block.

When employing the OUTPUT statement, we specify a new dataset (here, cooksData) where the results will be stored. Crucially, we use the COOKD= option to assign the calculated Cook’s Distance value for each observation to a new variable (named cookd in this example). This process is far more efficient than calculating the components (residuals, leverage, MSE) manually. The code below illustrates the necessary syntax to fit a simple linear model and simultaneously append the influence statistics to a new dataset:

/*fit simple linear regression model and calculate Cook's distance for each obs*/
proc reg data=my_data;
    model y=x;
    output out=cooksData cookd=cookd;
run;

/*print Cook's distance values for each observation*/
proc print data=cooksData;

It is worth noting that PROC REG offers other important diagnostic options within the OUTPUT statement, such as RESIDUAL= (for standard residuals), RSTUDENT= (for R-Student residuals), and H= (for leverage values), all of which contribute to a comprehensive influential observation analysis.

Analyzing the Tabular Output

The execution of the PROC REG and subsequent PROC PRINT statements generates a new dataset, cooksData, which contains all original variables (x and y) alongside the newly calculated influence measure, cookd. This final table is paramount for a detailed, observation-by-observation assessment of influence.

The printed output integrates the original data points with their respective influence scores:

By examining the cookd column, we can immediately highlight points that exhibit higher influence. For instance:

• Cook’s distance for the first observation is calculated to be 0.36813.
• Cook’s distance for the second observation is significantly lower at 0.06075.
• Cook’s distance for the third observation is negligible, measuring 0.00052.

Comparing these values to the accepted threshold of $4/n$, where $n=12$ in this example, establishing a cutoff of $4/12 approx 0.333$. We immediately observe that the first observation exceeds this critical value. Further scanning reveals other points that may also breach this threshold, signaling potential issues with the stability of the model parameters.

Visual Diagnostics: Interpreting Cook’s Distance Plots

Beyond tabular output, the PROC REG procedure automatically generates a series of diagnostic plots if the ODS graphics system is active in SAS. These visualizations are often more intuitive for quickly identifying influential points across the entire dataset. One of the most useful visualizations is the plot of Cook’s Distance versus the observation number, which we see here:

In this graphic representation, the horizontal axis displays the sequence of observation numbers (indices 1 through 12), while the vertical axis represents the calculated Cook’s Distance for each corresponding point. Crucially, the plot includes a reference line, or cutoff line, which is placed precisely at the influential threshold of $4/n$.

As established, with $n=12$, this cutoff line is positioned at approximately $0.33$. Visual inspection clearly shows that three distinct observations rise above this critical line. This graphical evidence strongly corroborates the tabular findings, identifying these points as potentially highly influential to the estimation of the regression coefficients. These observations possess a combination of high leverage and large residuals, driving their disproportionate influence.

Handling Influential Observations and Conclusion

The detection of highly influential observations through Cook’s Distance is a diagnostic step, not the final conclusion. When points exceed the threshold (like the three points identified in our example), data analysts must carefully investigate their nature. Potential causes for high influence include data entry errors, measurement errors, or genuinely unique cases that do not follow the general linear trend established by the rest of the data. Ignoring these influential points can lead to serious overfitting or misleading predictions.

Strategies for mitigating the impact of these influential points typically involve several options. The first step is always to verify the data for correctness. If the data is correct, analysts might consider transformation of variables, using robust regression techniques that are less sensitive to outliers, or, in rare, justified cases, excluding the observation and noting the exclusion in the final report. However, exclusion should always be justified on theoretical or domain-specific grounds, not merely statistical convenience.

In summary, Cook’s Distance provides a powerful, single metric for assessing data influence within regression models. By leveraging the automated capabilities of PROC REG in SAS, analysts can quickly identify and visualize points that disproportionately affect parameter estimates, ensuring the final statistical model is both stable and representative of the underlying data structure.

To deepen your expertise in advanced diagnostics and statistical programming, consider exploring the following resources and tutorials which explain how to perform other common tasks in SAS: