Update linear-marker-selection-indices.Rmd

Author: zankrut20

vignettes/linear-marker-selection-indices.Rmd

@@ -85,7 +85,7 @@ library(MASS)
 marker_scores <- matrix(0, nrow = nrow(phen_mat), ncol = ncol(phen_mat))
 colnames(marker_scores) <- colnames(phen_mat)
 
-for (i in 1:ncol(phen_mat)) {
+for (i in seq_len(ncol(phen_mat))) {
     # Fit ridge regression to handle multicollinearity among markers
     # Note: A lambda must be chosen carefully in real scenarios
     fit <- lm.ridge(phen_mat[, i] ~ marker_mat, lambda = 10)
@@ -156,6 +156,53 @@ print(gw_lmsi_res$summary)
 
 The GW-LMSI automatically models all the molecular variance in one step. While it requires computationally heavier matrix inversion (often mitigated by regularization), it tends to predict the net genetic merit more efficiently than standard LMSI when trait heritability is very low or when capturing many small QTL regressions requires immense sample sizes.
 
+## 3. Advanced Concepts and Limitations
+
+### 3.1 Basic Conditions for Constructing the LMSI
+
+According to Lande and Thompson (1990), applying the LMSI correctly requires several primary statistical and biological conditions to be met in the breeding population:
+
+1. **Linkage Disequilibrium:** The markers and the quantitative trait loci (QTL) must be in linkage disequilibrium (LD).
+2. **Additivity:** QTL effects should be combined additively both within and between loci (no epistasis).
+3. **Coupling Mode:** The QTL should be in coupling mode where parental lines hold uniformly positive or negative alleles.
+4. **Genetic Architecture:** Traits should be affected by a few QTLs with large effects rather than exclusively by many small-effect QTLs.
+5. **Heritability:** The traits must inherently have a low heritability ($h^2$). If heritabilities are high, phenotypic selection alone is immensely effective.
+6. **Marker Identification:** Markers correlated with the traits of interest must first be successfully identified.
+
+### 3.2 The "LMSI Paradox" and Efficiency Comparisons
+
+Will a linear marker selection index always dramatically outperform standard phenotypic selection? Not necessarily.
+
+Lande and Thompson (1990) showed that LMSI efficiency relative to basic phenotypic selection ($\lambda_M$) for a single trait is:
+
+$$ \lambda_M=\frac{R_M}{R}=\sqrt{\frac{q}{h^2}+\frac{{\left(1-q\right)}^2}{1-{qh}^2}} $$
+
+Where $q = \frac{\sigma_s^2}{\sigma_g^2}$ is the proportion of additive genetic variance definitively explained by the markers, and $h^2$ is the trait heritability. 
+
+This equation demonstrates the **"LMSI Paradox"** mathematically defined by Knapp (1998): as $h^2$ trends toward zero, $\lambda_M$ approaches infinity. Thus, the massive relative advantage of marker-assisted selection over phenotypic selection occurs explicitly when phenotypic trait heritability is extremely low, and the markers successfully capture a massive fraction of genetic variance. Otherwise, the LMSI offers diminishing returns and can fix QTLs at an overly fast rate.
+
+### 3.3 Statistical Properties of Molecular Indices
+
+Assuming $H$ and the index $I_M$ share a bivariate joint normal distribution, the statistical properties mirror the standard phenotypic LPSI:
+
+1. The variance of the index equals the covariance between the net genetic merit and the index: $\sigma_{I_M}^2=\sigma_{HI_M}$.
+2. The maximized correlation (Accuracy) is $\rho_{HI_M}=\frac{\sigma_{I_M}}{\sigma_H}$.
+3. The Variance of the Predicted Error $Var\left(H-{I}_M\right)=\left(1-{\rho}_{HI_M}^2\right){\sigma}_H^2$ is uniquely minimized.
+
+Because molecular indices incorporate *both* the phenotypic and genotypic marker data, the LMSI accuracy will technically always be higher than the standard LPSI accuracy.
+
+### 3.4 Advanced Matrix Estimation Methods
+
+While the `gw_lmsi()` and `lmsi()` functions in the `selection.index` package abstract away the underlying algebra, molecular index computation routinely suffers from the $p \gg n$ covariance problem (vastly more markers than recorded individuals).
+
+Estimating the marker score variances ($\mathbf{S}$) and the massive $\mathbf{M}$ and $\mathbf{W}$ covariance matrices using least-squares often leads to:
+1. Marker covariance values erroneously overestimating the true genetic covariance ($\mathbf{C}$).
+2. Singular, non-positive definite matrices where classical inverses fail.
+
+The standard solution involves regularization techniques (like the Ridge penalty `lambda` built into `gw_lmsi()`), ensuring matrix invertibility at the cost of slight coefficient bias.
+
 ## Conclusion
 
-The `selection.index` package extends base index selection directly into the molecular realm. The **LMSI** (`lmsi()`) allows breeders to leverage an established pre-calculated marker scorings regression model. The **GW-LMSI** (`gw_lmsi()`) embraces the Genomic Selection paradigm by swallowing all marker and phenotypic data jointly in a single predictive framework, utilizing ridge regularization to bypass classical $p \gg n$ multicollinearity obstacles.
+The `selection.index` package extends base index selection directly into the molecular realm. The **LMSI** (`lmsi()`) allows breeders to leverage an established pre-calculated marker scorings regression model. The **GW-LMSI** (`gw_lmsi()`) embraces the Genomic Selection paradigm by swallowing all marker and phenotypic data jointly in a single predictive framework, utilizing ridge regularization to bypass classical $p \gg n$ multicollinearity obstacles. 
+
+When implemented on traits exhibiting low heritability, these marker-assisted pipelines can drastically outcompete standard phenotypic indices in maximizing Expected Genetic Gain per selection cycle.