What a Component Coefficient Describes
When principal-component (PC) scores are used as predictors, a fitted coefficient describes how the model changes along a component-score direction. That direction combines changes in several original predictors. The coefficient is therefore not, by itself, an effect per unit of any one original variable. Interpreting it correctly requires both the fitted outcome model and the component coefficients used to construct the scores.
The distinction matters in actuarial prediction. A component might combine claim frequency, average severity, and a measure of exposure. A fitted coefficient can summarize how the predicted outcome changes as the combined score changes, but it does not establish that changing one of those original quantities would cause that outcome change.
This definition uses the unit-length eigenvectors of the predictor covariance matrix (or correlation matrix when the inputs are standardized). Some software and reports call these eigenvector entries component loadings. Other conventions call the predictor-component correlations loadings. For standardized PCA, those correlations are \(\sqrt{\lambda_a}p_{ja}\), where \(\lambda_a\) is the eigenvalue of component \(a\) and \(p_{ja}\) is its eigenvector coefficient for predictor \(j\). The coefficient conversion below uses the eigenvector coefficients \(p_{ja}\); if a report gives correlations instead, account for the factor \(\sqrt{\lambda_a}\).
The fitted score coefficient is on the outcome model’s linear-predictor scale. With a Gaussian model and identity link, that is the predicted mean scale. With a logistic model, it is the log-odds scale. With a log link, it is the log-mean scale. The link determines how a change in \(\eta\) translates to the response scale.
Translating Score Coefficients to Original-Predictor Coordinates
If the model uses only the retained scores, substitute \(t=P_k^\mathsf{T}z\) into the linear predictor. This produces coefficients in standardized-predictor coordinates. It is an algebraic re-expression of the same fitted model, not a new fit.
Here \(b_z\) is the vector of coefficients for the standardized original predictors in this re-expression. Its \(j\)th entry is \(b_{z,j}=\sum_{a=1}^{k}p_{ja}\gamma_a\). For a change \(\Delta z\), with the fitted transformation held fixed, the change in the linear predictor is \(\Delta\eta=b_z^\mathsf{T}\Delta z=\gamma^\mathsf{T}P_k^\mathsf{T}\Delta z\). This formula accounts for every retained component’s contribution.
If \(k=p\), the eigenvectors form a complete orthonormal basis, so the conversion expresses the same fitted linear predictor exactly in all original-predictor coordinates. If \(k<p\), it still re-expresses the fitted predictor exactly, but \(b_z\) lies in the span of the retained eigenvectors. The model has no separately estimated effect in omitted component directions. Thus, with dimension reduction, the converted coefficients describe the retained-score model, not an unrestricted model fitted to all original predictors.
For predictors standardized using standard deviations \(s_j\), a coefficient per original unit is \(b_{x,j}=b_{z,j}/s_j\). This follows because \(z_j=(x_j-\mu_j)/s_j\). The model intercept in the re-expression is \(\alpha\); the centered predictors are zero at the training means. If other covariates also appear in the outcome model, this conversion applies to its PC-score terms, with those other terms held fixed.
Why the Conversion Holds
The score is a linear combination of the standardized predictors. Substituting that combination into the fitted model and rearranging the finite sums gives the original-coordinate coefficients. Equivalently, a small or finite predictor change \(\Delta z\) changes each score by \(P_k^\mathsf{T}\Delta z\), so its effect on the linear predictor is the fitted score-coefficient vector multiplied by that score change.
This equality explains both the usefulness and the limit of the conversion. It gives the model-implied change for a specified movement in original-predictor space. It does not turn the PC coefficients into causal effects, nor does it say that the model estimated an independent relationship for each original predictor.
Worked Example: Re-expressing a Two-Component Model
Method: eigenvector matrix multiplication. In a hypothetical Gaussian prediction model with identity link, two standardized predictors have retained component eigenvectors \(p_1=(0.8,0.6)^\mathsf{T}\) and \(p_2=(-0.6,0.8)^\mathsf{T}\). The fitted model is \(\eta=3+2.5t_1-1.0t_2\). Find the coefficients for the two standardized predictors and check the result at \(z=(1,2)^\mathsf{T}\). The common mistake to avoid is assigning the score coefficients directly to the original predictors.
The component matrix and score-coefficient vector are \(P=\begin{pmatrix}0.8&-0.6\\0.6&0.8\end{pmatrix}\) and \(\gamma=(2.5,-1.0)^\mathsf{T}\). Thus, \(b_z=P\gamma\). For predictor 1, \(b_{z,1}=0.8(2.5)+(-0.6)(-1.0)=2.0+0.6=2.6\). For predictor 2, \(b_{z,2}=0.6(2.5)+0.8(-1.0)=1.5-0.8=0.7\). The re-expressed model is \(\eta=3+2.6z_1+0.7z_2\).
Check. At \(z=(1,2)^\mathsf{T}\), the scores are \(t_1=0.8(1)+0.6(2)=2.0\) and \(t_2=-0.6(1)+0.8(2)=1.0\). The score model gives \(\eta=3+2.5(2.0)-1.0(1.0)=7.0\). The re-expressed model gives \(\eta=3+2.6(1)+0.7(2)=7.0\), so the calculations agree. The coefficients 2.6 and 0.7 are a coordinate re-expression; they should not be described as separately estimated or causal effects.
Interpreting a Change in One Original Predictor
A score coefficient describes a one-unit change in a score, with the other included scores fixed. A change to one original predictor alone generally changes several scores at once. To find its model-implied effect, calculate those score changes using the relevant eigenvector coefficients, then combine them with the fitted score coefficients. The result is a change in the linear predictor conditional on the other original predictors being held fixed; it is not automatically an effect of intervening on that predictor.
Worked Example: A One-Component Logistic Model
Method: map the original-predictor change into a score change, then apply the link. A hypothetical logistic model uses one retained score \(t=0.6z_1+0.8z_2\), with fitted linear predictor \(\eta=-1-0.4t\). Both predictors are standardized. Starting from their means, calculate the predicted probability after \(z_2\) increases by one standard deviation while \(z_1\) stays fixed. The common mistake to avoid is treating \(-0.4\) as the change in log-odds for a one-standard-deviation change in either original predictor.
At the starting point, \(z_1=z_2=0\), so \(t=0\) and \(\eta=-1\). The logistic response is \(p=1/(1+\exp(-\eta))\), giving \(p=1/(1+\exp(1))\approx0.2689\).
After the stated change, \(z_1=0\), \(z_2=1\), and \(t=0.6(0)+0.8(1)=0.8\). The new linear predictor is \(\eta=-1-0.4(0.8)=-1.32\), so \(p=1/(1+\exp(1.32))\approx0.2108\). The change in log-odds is \(-0.32\), not \(-0.4\); the odds are multiplied by \(\exp(-0.32)\approx0.7261\). The predicted probability decreases by about \(0.0581\). This is the model’s conditional prediction for the specified predictor change, not evidence that changing \(z_2\) causes the outcome to change.
Returning to Original Measurement Units
Standardization affects the size of the coefficient required for a one-unit change. If an original predictor is measured in dollars, one unit means one dollar; if its standardized version is used, one unit means one training-set standard deviation. Convert before presenting a numerical effect in the original units. For a nonlinear model, report the change on the link scale or calculate the response-scale predictions at the stated covariate values.
Worked Example: Converting to Frequency and Dollar Units
Method: first map score coefficients to standardized-predictor coefficients, then divide by each standard deviation. A hypothetical log-link model predicts a positive outcome using two standardized inputs: claim frequency \(x_1\), with training mean \(0.20\) and standard deviation \(0.05\), and average claim amount \(x_2\), with training mean \(\$1{,}200\) and standard deviation \(\$300\). One retained score has eigenvector \(p_1=(0.6,0.8)^\mathsf{T}\); a second has \(p_2=(-0.8,0.6)^\mathsf{T}\). The fitted score coefficients are \(\gamma=(1.2,-0.5)^\mathsf{T}\). Find the re-expressed coefficients per original unit and the change in \(\eta\) for a frequency increase of \(0.01\), with average claim amount unchanged. The common mistake to avoid is reporting standardized-input coefficients as effects per raw unit.
For standardized predictor 1, \(b_{z,1}=0.6(1.2)+(-0.8)(-0.5)=0.72+0.40=1.12\). For standardized predictor 2, \(b_{z,2}=0.8(1.2)+0.6(-0.5)=0.96-0.30=0.66\). Dividing by the corresponding standard deviations gives \(b_{x,1}=1.12/0.05=22.4\) per unit of frequency and \(b_{x,2}=0.66/300=0.0022\) per dollar of average claim amount.
A frequency increase of \(0.01\) changes the standardized frequency by \(0.01/0.05=0.2\), while the other standardized predictor is unchanged. Therefore \(\Delta\eta=1.12(0.2)=0.224\), equivalently \(22.4(0.01)=0.224\). With a log link, the fitted mean is multiplied by \(\exp(0.224)\approx1.251\), about a 25.1% increase, at the specified fixed value of average claim amount. The calculation is internally consistent: both coefficient routes give the same change in \(\eta\). It is a model-based association, not a causal claim about frequency.
Reporting the Interpretation Carefully
A sound interpretation states the model scale, the score change, and the role of the loadings. For example: “Holding the other included scores fixed, a one-unit increase in PC1 is associated with a 1.2-unit increase in the fitted linear predictor. PC1 combines the standardized input variables in the proportions given by its eigenvector.” If translating to original predictors, state the transformation and units, and clarify that the resulting coefficients re-express the score model.
Avoid saying that a component coefficient measures the effect of “the variables that load positively” without specifying the movement being considered. A one-unit increase in a score means moving in its component direction in predictor space, not increasing every positively weighted original variable by one unit while leaving the rest unchanged. Likewise, if an original predictor alone changes, other component scores may change too; the resulting prediction change must include all those changes.
Check Your Understanding
For each problem, use the stated score definition and model. Interpret changes on the specified scale; answers are provided for checking.
- A Gaussian identity-link model uses one score \(t=0.6z_1+0.8z_2\) with coefficient \(\gamma=1.5\). What are the re-expressed coefficients for \(z_1\) and \(z_2\)? (Answer: \(0.9\) and \(1.2\), respectively.)
- A two-component model has \(P=\begin{pmatrix}0.8&-0.6\\0.6&0.8\end{pmatrix}\) and \(\gamma=(1,-2)^\mathsf{T}\). Calculate \(b_z\). (Answer: \(b_z=(2,-1)^\mathsf{T}\).)
- A logistic model has \(\eta=-0.5+0.5t\), where \(t=0.8z_1+0.6z_2\). Starting at \(z_1=z_2=0\), find the predicted probability after \(z_2\) increases by one and \(z_1\) stays fixed. (Answer: \(t=0.6\), \(\eta=-0.2\), and \(p\approx0.4502\).)
- A one-component model has \(p=(0.6,0.8)^\mathsf{T}\) and \(\gamma=2\). The two original predictors have standard deviations \(0.2\) and \(4\). Find the coefficients per original unit. (Answer: standardized coefficients \((1.2,1.6)\); original-unit coefficients \((6,0.4)\).)
- For standardized PCA, a component has eigenvalue \(\lambda=4\), eigenvector coefficients \(p=(0.5,\sqrt{3}/2)^\mathsf{T}\), and score coefficient \(\gamma=1.2\). If the reported correlation loading for the first predictor is used instead of its eigenvector coefficient, what loading and corresponding multiplier of that loading give the same contribution to \(\eta\)? (Answer: the correlation loading is \(2(0.5)=1\); its multiplier is \(1.2/2=0.6\), giving contribution \(0.6(1)z_1=0.6z_1\), the same as \(1.2(0.5)z_1\).)