Variables tab, Statistics sub-tab

Model Related  

Indicates what type of operations can be performed on this variable:

Categorical values can only be tested for equality. A variable with this Op Type is displayed with the Category icon.

Ordinal values, in addition to being tested for equality, have an order defined. A variable with this Op Type is displayed with the Ordinal icon.

Continuous values support arithmetic operators. A variable with this Op Type is displayed with the Numeric icon.

Indicates the type of data represented by the variable, such as numeric (for example, integer, double), string, date, and so on.

Indicates how the variable is used in the model:

Active is used for independent variables, which are usually inputs to the model. A variable with this Usage Type is displayed with the Input icon.

Predicted is used for the dependent variable, which is usually the target of the model. A variable with this Usage Type is displayed with the Target icon.

Supplementary means the variable is not used in the calculation of model. A variable with this Usage Type is displayed with the Extra icon.

Group means the variable is used to group records. For example, grouping items in a transaction. A variable with this Usage Type is displayed with the Group icon.

Order means the variable defines the sequence of records. A variable with this Usage Type is displayed with the Order icon.

The value used to replace a missing value.

If a variable's Usage Type is predicted, then the missing value is replaced when predictive metrics are created, rather than when the model is scored. This is because the dependent variable is only used to train the model, it is not relevant to scoring the model.

Indicates the source of the value used for missing value replacement. This information is for reference only, the value of this statistic does not affect the scoring of the model.

Indicates how values that are identified as outliers are handled. This applies only to continuous variables that are numeric:

asIs means values are used exactly as they are provided.

asExtremeValues means values are replaced by their low or high value extremes (see the information on Low Value and High Value below).

asMissing means these values are treated as missing, and missing value replacement is used.

Lower boundary for outliers. When a value is less than this value and Outlier Value Treatment is asExtremeValues, this value is used instead.

Upper boundary for outliers. When a value is greater than this value and Outlier Value Treatment is asExtremeValues, this value is used instead.

Model Level

For Regression Models

Square of the percent of variance in the dependent, explained by the given of all the independent variables.

R2= SSR/(SSR+SSE)

For Clustering Models

The proportion of variation explained by a particular clustering of the observations.

R2= (TSS-ESS)/TSS

For Time Series Models

R2= 1-SSE/SST

Adjusted R2= 1 - (1 - R2)[(n-1)/(n-k-1)]

where k is the number of independent variables.

Generated from F-Distribution, this p-value gives the probability that a highR2value occurred by chance, which means theR2value is not statistically significant.

G2=-2LL0

Where LL0 is the log likelihood of the model, considering intercepts only. In other words, this is the model not considering any independent variables.

G2=-2LL

Where LL is the log likelihood of the model, considering all of the independent variables.

The p-Value is obtained from a chi-square test with the statistic Likelihood Ratio Test - Degrees of Freedom and the chi-square statistic:

G2=-2(LL0-LL)

Where LLO is the log likelihood of the model, considering intercepts only, and LL is the log likelihood of the model, considering all of the independent variables.

The null hypothesis is H0 = The intercept-only model performs the same as the final model.

The null hypothesis is rejected at α significance level if p-Value < α.

The number of variables in the final model

df = Num_Obs - Min_Obs

where:

Num_Obs is the number of observations.

Min_Obs is the minimum number of observations required to uniquely define the model.

This statistic is sometimes also called the Standard Error of the y estimates.

SEy = √(Σ(yi-y'i)2)/n)

The value of n is dependent on the type of model:

For regression models, n is the number of degrees of freedom for the model.

For time series models, n is the size of the testing dataset.

This statistic is sometimes also called the Sum of Square of Errors.

SSE = Σ(yi-y'i)2

This statistic is abbreviated SSR.

SSR = Σ(y'i-ymean)2

This statistics is abbreviated SST.

SST = Σ(yi-ymean)2

This statistic is similar to Sum of Squares Residual but for Cluster models. This statistic is abbreviated ESS:

ESS = Σ(x

Where x

This statistic is similar to Sum of Squares Residual but for Cluster models. This statistic is abbreviated TSS:

TSS= Σ(x

Where x

The significance of regression is evaluated using F-statistics.

 F = (SSR/dfSSR)/(SSE/dfSSE)

where:

dfSSR is the number of degrees of freedom for the regression sum of squares, which is equal to the number of predicators (g).

dfSSE is the number of degrees of freedom for the error sum of squares which is equal to the number of observations (n) minus the number of predictors (g) minus 1. (n - g - 1)

Optimal is the cumulative sum of records, which are sorted in the ideal order (descending for continuous dependent variables, true values first for categorical dependent variables).

Model is the cumulative sum of records, which are sorted by descending order of the predictions (predictive values for continuous dependent variables, probability for categorical dependent variables).

Random is the straight line from (0%,0%) to (100%,100%).

A matrix with actual outcomes on rows and predicted outcomes on columns, which contains three general regions:

Main diagonal. These are the correct predictions.

Below the main diagonal. These are the false positive or Type I errors, commonly associated with binary results.

Above the main diagonal. These are the false negative or Type II errors, commonly associated with binary results.

This statistic calculates the total distance of all records to their assigned cluster's center.

The constant term of a regression equation. For example, for the slope equation y = mx + b, b is the Intercept. The Intercept is also referred to as the Y-Intercept.

For a full technical definition, refer to the information on Standard Error provided below.

For a full technical definition, refer to the information on RMSE (SEy) provided above.

The total number of values in the data, including missing values, minus the number of missing values.

Total - Missing

Variable Level (Univariate Analysis)

 Also called the expected value.

Mean = Σ(xi)/n

Typically defined for the entire population. It can also be described as the square root of the variance.

Std Dev = √Σ(xi-xmean)2/n

This is also known as the fiftieth percentile value.

Typically only reported for categorical variables. It is possible for a set of values to have more than one mode and, in these cases, the first mode found is listed here.

The smallest number in a set of values

The largest number in a set of values

It is the difference between the twenty-fifth and the seventy-fifth percentiles.

This is also known as the Total Frequency.

This is also known as the Missing Frequency.

Cardinality refers to the uniqueness of data values observed for a variable. The lower the cardinality, the more duplicated values exist. The lowest possible cardinality is one, which means that all the values are the same. Binary indicators and status flags are examples of low cardinality variables. Primary keys and IDs are examples of high cardinality variables. For the purposes of data mining, variables that contain values such as names and area codes are typically considered high cardinality.

Describes the variance between a continuous dependent variable and a categorical independent variable. This includes the p-Value, which represents the probability that the relationship between the variables occurred by chance and is not statistically significant.

Variable Level (Multivariate Analysis)

The calculation depends on the type of model. For additional information on how Importance of a variable is determined, see Determining the importance of a variable.

The standard deviation of a sampling distribution of the coefficient. It can also be described as the square root of the variance.

For the simple linear regression model, y = mx + b:

Std Error = √[Σ((y'i-yi)2/ (n-2))/Σ(xi-xmean)2]

Computed for each coefficient by dividing the coefficient by its standard error.

A multiplicative factor of a term in a regression equation, each independent variable has an associated coefficient that was determined to best describe its relationship with the dependent variable.

Typically, the value is 1.

The p-Value gives the probability of getting a coefficient value in your sample assuming that the coefficient in the population is zero.

For Linear and Exponential Regression models, the p-Value is generated from T distribution.

For Logistic Regression models, the p-Value is generated from ChiSquare distribution, with the square of z-Score as the ChiSquare value.

This p-Value comes from the T-Distribution using the t-Value and is calculated after any variables are eliminated due to variable reduction settings (see Advanced Options).

This p-Value comes from the T-Distribution using the t-Value and is calculated prior to the elimination of any variables due to variable reduction settings (see Advanced Options).

This p-Value comes from the ChiSquare-Distribution using the square of z-Score.

Computed for each coefficient by dividing the coefficient by its standard error. Compare to the t-value (t-critical) from a table based on the degrees of freedom and alpha (usually set at 0.05). Used to assess the significance of individual coefficients, specifically testing the null hypothesis that the regression coefficient is zero. If the observed