Class PartialEdgeStrengthResult

java.lang.Object

edu.cmu.tetrad.sem.PartialEdgeStrengthResult

All Implemented Interfaces:

TetradSerializable, Serializable

public final class PartialEdgeStrengthResult
extends Object
implements TetradSerializable

Result of a partial edge strength computation performed by NNEstimator.computePartialEdgeStrength(String, String, int).

Partial edge strength asks: once Y's other parents are accounted for, does X add anything to held-out prediction of Y? It is computed on the same k folds as NNEstimator.crossValidate(int): on each fold the full model (all parents) and a reduced model (Y's mechanism retrained on the fold's training rows without X) both predict the held-out rows, and the difference in held-out score is reported.

• Continuous child: partialR2 = R²_full − R²_reduced, with R² = 1 − OOS MSE / marginal variance, matching the Cross-Validation table. residualVariance holds the reduced model's held-out MSE.
• Discrete child: partialXentImprovement = xent_reduced − xent_full in nats.

Positive means X carries information about Y that the other parents do not. Near zero means X is redundant given the other parents, which is not the same as X not being a cause; see EdgeStrengthResult for the complementary intervention measure, which stays large for a redundant parent the fitted mechanism actually uses.

See Also:

• Serialized Form

Field Summary

Fields

Modifier and Type	Field	Description
final double	avgMarginalEffect	Signed average marginal effect of the parent on the child: the average, over all held-out rows across the k folds, of the central finite difference (f(x + h) − f(x − h)) / (2h) of the fold's fitted mechanism with respect to the parent, with the other parents held at their observed values and h = 0.05 · SD(X).
final double	avgMarginalEffectStd	Standardized average marginal effect: avgMarginalEffect · SD(X) / SD(Y), with SDs taken from the observed data.
final String	childName	Name of the child variable (head of the edge).
final boolean	discreteChild	true if the child variable is discrete.
final double	fracPositiveDeriv	Fraction of held-out rows on which the local derivative was strictly positive, among rows with a finite derivative.
final int	numFolds	Number of CV folds used.
final String	parentName	Name of the parent variable (tail of the edge).
final double	partialR2	Held-out R² of the full model minus held-out R² of the model with X removed from Y's parents, on the same k folds as the CV table.
final double	partialXentImprovement	OOS cross-entropy improvement of predicting R from X vs.
final double	residualVariance	Held-out MSE of the reduced model (Y predicted from its other parents, without X).

Method Summary

Modifier and Type	Method	Description
boolean	isNonMonotone()	True if the fitted effect changed sign on more than 10% of held-out rows (minority side of fracPositiveDeriv above 0.10), so the average marginal effect mixes regions of opposite sign.
String	toString()
String	toSummaryLine()	Returns a human-readable one-line summary.

Methods inherited from class Object

equals, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Details

parentName

public final String parentName

Name of the parent variable (tail of the edge).

childName

public final String childName

Name of the child variable (head of the edge).

discreteChild

public final boolean discreteChild

true if the child variable is discrete.

partialR2

public final double partialR2

Held-out R² of the full model minus held-out R² of the model with X removed from Y's parents, on the same k folds as the CV table. Positive = X adds predictive information beyond the other parents. NaN for discrete children.

residualVariance

public final double residualVariance

Held-out MSE of the reduced model (Y predicted from its other parents, without X). Kept under its original field name for session compatibility. NaN for discrete children.

partialXentImprovement

public final double partialXentImprovement

OOS cross-entropy improvement of predicting R from X vs. predicting the marginal residual distribution. For discrete children only. NaN for continuous children.

numFolds

public final int numFolds

Number of CV folds used.

avgMarginalEffect

public final double avgMarginalEffect

Signed average marginal effect of the parent on the child: the average, over all held-out rows across the k folds, of the central finite difference (f(x + h) − f(x − h)) / (2h) of the fold's fitted mechanism with respect to the parent, with the other parents held at their observed values and h = 0.05 · SD(X). Uses each fold's full model on that fold's held-out rows, so it is an out-of-sample quantity on the same folds as partialR2. On the parent's original scale (child units per parent unit). NaN unless both parent and child are continuous.

avgMarginalEffectStd

public final double avgMarginalEffectStd

Standardized average marginal effect: avgMarginalEffect · SD(X) / SD(Y), with SDs taken from the observed data. Unit-free, so magnitudes are comparable across parents. NaN unless both parent and child are continuous.

fracPositiveDeriv

public final double fracPositiveDeriv

Fraction of held-out rows on which the local derivative was strictly positive, among rows with a finite derivative. Values near 1 or 0 indicate a monotone fitted effect; intermediate values indicate the effect changes sign over the data range, in which case avgMarginalEffect is an average of opposing effects and should be read with care. NaN unless both parent and child are continuous.

Method Details

isNonMonotone

public boolean isNonMonotone()

True if the fitted effect changed sign on more than 10% of held-out rows (minority side of fracPositiveDeriv above 0.10), so the average marginal effect mixes regions of opposite sign.

Returns:

whether the fitted effect is flagged as non-monotone

toSummaryLine

public String toSummaryLine()

Returns a human-readable one-line summary.

Returns:

a human-readable one-line summary

toString

public String toString()

Overrides:

toString in class Object