community
cancel
Showing results for 
Search instead for 
Did you mean: 

Alteryx Designer Knowledge Base

Definitive answers from Designer experts.
Community v19.6

Looks aren't everything... But the latest Community refresh looks darn good!

Learn More

Analyzing Partial Effects in Alteryx

Alteryx
Alteryx
Created on
Tools for Analyzing Partial Effects in Alteryx

 

If you are building a predictive model, inevitably you will want to analyze the effect that your independent variables have on your dependent variable. This article is meant to shed some light on the Alteryx-specific options for this type of analysis. The options for analyzing these effects will vary depending on what type of model you are using, so I will run through the different options for each predictive tool. The attached v11.7 workflow has examples for each model.

 

1. Coefficients

The first and probably easiest way to analyze the partial effect can be done by simply viewing the coefficient. This method of directly interpreting the coefficient as the partial effect that an independent variable has on a dependent variable is typically only appropriate for the Linear Regression model.

 

Using the examples in the attached workflow, we see that this is relatively easy:

 

A. Linear Regression

We can find the coefficients in the R Output or the I Output of the Linear Regression tool.

R output:1 - Linear Output.png

 

 

From these results, the coefficients can be interpreted as follows:

  1. BatAge: As the average age of a batter increases by 1 year, runs per game is expected to decrease by 0.01431 (C.P.).
  2. H: As the number of hits in a season increases by 1, runs per game is expected to increase by 0.00548 (C.P.).

 

2. Plots

 

Analyzing the Effect Plots that are included in some of the predictive tools is relatively easy and intuitive. This option is provided for the following tools in Alteryx: Logistic Regression tool, Naïve Bayes Classifier tool, Neural Network tool, Boosted Model tool, and the Spline Model tool.

 

A. Logistic Regression (Conditional Density Plot)

 

The I output of the Logistic tool provides an interactive Conditional-Density Plot. This tool allows you to view the conditional probability of your dependent variable as it relates to non-categorical independent variable. It is worth noting that the Conditional-Density Plot reveals the probability of a response of No (or 0).

 

You can find this option by clicking on the Conditional-Density Plots option of the I input.

 

I Output:

 

4 - Logistic Conditional-Density Option.png

 

Since this is an interactive output, you can hover your mouse over the graph to reveal the probability that someone does not donate relative to the number of degrees they have.

 

5 - Logistic Conditional Density Plot 1.png

 6 - Logistic Conditional Density Plot 2.png

 

From these results, we see that the estimated probability that someone does not donate is approximately 0.5451542 (or 54.51542 %) when someone has 1-1.5 degrees (C.P.). We can also see that the probability that someone does not donate is approximately 0.3674242 (or 36.74242 %) when someone has 1.5-2.5 degrees, (C.P.).

 

B. Naïve Bayes Classifier (Effect Plots)

 

The R output of the Naïve Bayes Classifier tool provides an Effect Plot for each predictor variable used in the model. It is worth noting that the Effects Plots reveal the probability of a Yes (or 1).

 

These graphs are not interactive, so the probabilities are not exact. This option is automatically included in the R output.

 

R Output:

 

 

7 - Naive Bayes Effects Plot.png

 

 

This graph reveals that the estimated probability that someone does respond is approximately 0.58 (58%) when someone has 1 degree (C.P.) and that the probability that someone does respond is approximately 0.35 (35%) when someone has 1.5-2.5 degrees (C.P.).

 

NOTE: If you are estimating a model that includes many independent variables, you may have to click on the Records arrow in order to view the Effects Plots.

 

8 - Include Effects Plot.png

 

 

C. Neural Network (Effect Plot)

 

The R output of the Neural Network tool provides an Effect Plot for each predictor variable used in the model. These graphs are not interactive, so the probabilities are not exact. This option must be selected in the configuration of the tool.

 

Include Effects Plots:

  

NN Include Effects Plots.png

 

 R Output:

 

9 - NN Effects Plot.png 

This graph reveals that the estimated probability that someone does respond is approximately 0.41 (41%) when someone has 1 degree (C.P.) and that the probability that someone does respond is approximately 0.47 (47%) when someone has 1.5-2.5 degrees (C.P.).

 

 

D. Boosted Model (Effect Plot)

 

The R output of the Boosted Model tool provides an Effect Plot for each predictor variable used in the model. These graphs are not interactive, so the probabilities are not exact. This option must be selected in the configuration of the tool.

 

Include Effects Plots:

 

10 - Include Effects Plot.png 

R Output:

11 - 11 Boosted Model Effects.png

This graph reveals that the estimated probability that someone does respond is approximately 0.46 (46%) when someone has 1 degree (C.P.) and that the probability that someone does respond is approximately 0.62 (62%) when someone has 1.5-2.5 degrees (C.P.).

 

E. Spline Model (Effects Plot)

 

The R output of the Spline Model tool provides an Effect Plot for each predictor variable used in the model. These graphs are not interactive, so the probabilities are not exact. This option must be selected in the configuration of the tool.

 

Include Effects Plots:

 

12 - Include Effects Plot.png

 

R Output:

 

13 - Spline Model Effects.png

 

This graph reveals that the estimated probability that someone does respond is approximately 0.45 (45%) when someone has 1 degree (C.P.) and that the probability that someone does respond is approximately 0.62 (62%) when someone has 1.5-2.5 degrees (C.P.).

 

3. Scoring

 

When all else fails, you can always score your data! This process is relatively simple:

 

1. Score the original values of your data

2. Score your data after the desired change in the independent variable has been made

3. You can view the change in the score for each individual

4. You can view the average change in score across all individuals (Average Partial Effect APE)

 

Since the Decision Tree tool, Forest Model tool, Count Regression tool, Support Vector Machine tool, and Gamma Regression do not have effects plots and the coefficients cannot be directly interpreted, this option has been demonstrated with these tools in the attached workflow.

 

A. Decision Tree

 

14 - Decision Tree Individuals (Revised).png

 

When we compare the individual scores, we see that records 1 through 7 are unchanged after we increase degrees by 1, but record 8 decreases from 26.663636 to 1647619.

 

14 - Decision Tree Score (Revised).png

 

 

 

The output of our Summarize tool reveals that the Average Partial Effect (APE) is -3.443006, or predicted MPG decreases by 3.443006 average when a Cylinder is added.

 

B. Forest Model

 

15 - Forest Model Indiviuals (Revised).png

 

When we compare the individual scores, we see that record 1 increases from 0.216 (21.6%) to 0.24 (24.0%) when the number of degrees is increased by 1.

 

15 - Forest Model Score (Revised).png

 

 

 

The output of our Summarize tool reveals that the Average Partial Effect (APE) is 0.010182, or the probability of a response increases 1.0182 percentage points on average when a Degree is added.

 

C. Support Vector Machine

 

16 - SVM Individuals (Revised).png

 

When we compare the individual scores, we see that record 1 increases from 0.533969 (53.3969%) to 0.587181 (58.7181%) when the number of degrees is increased by 1.

 

16 - SVM Score (Revised).png

 

 

 

The output of our Summarize tool reveals that the Average Partial Effect (APE) is 0.05625, or the probability of a response increases 5.625 percentage points on average when a Degree is added.

 

D. Count Regression

17 - Count Data Individual (REvised).png

When we compare the individual scores, we see that the predicted number of claims for record 1 increases from 1.658907 to 1.671996 when the average cost of a claim is increased by 1.

 

17 - Count Score (Revised).png

 

 

 

The output of our summary tool reveals that the Average Partial Effect (APE) is 0.573603, or the predicted number of claims increases by 0.573603 on average when a average cost is increased by 1.

 

E. Gamma Regression

18 - Gamma Regression Individual.png

 

 

When we compare the individual scores, we see that the predicted average cost increases from 203.770014 to 203.807904 when the number of claims is increased by 1.

 

18 - Gamma Regression Score.png

 

 

 

The output of our summary tool reveals that the Average Partial Effect (APE) is 0.046081, or the predicted average cost increases by 0.046081 on average when the number of claims is increased by 1.

Attachments