RSM 3/26/26

Introduction

  • The lecture begins with an informal greeting. The speaker prepares to discuss covariances in data analysis, especially in the context of experimental setups involving blocks.

Covariance Structure

  • The covariance structure between responses on any two experimental units (peaks) within any block is nonzero.
    • This occurs because experimental units share the same genetics, maternal environment, etc.
    • Effects from one animal may influence others within the same block due to shared characteristics.

Variance-Covariance Matrix

  • The variance-covariance matrix for responses across all blocks is referred to as a block diagram.
  • When observations in different experimental units within a block are equally correlated, this is known as compounding.
    • A 10 x 10 variance-covariance matrix is created for two blocks, each containing five experimental units.
    • Totaling two blocks results in 10 observations in this matrix.

Formation of the F-Ratio

  • The denominator of an F-ratio must be formed using one or more mean squares.
    • This formation is essential such that the expected value of the denominator equals that of the numerator when the null hypothesis is true.
  • The relationship of the F-ratio states:
    • As treatment mean squares increase relative to the error mean squares, the significance of the F value increases, indicating potential effects.

Analysis of Variance (ANOVA)

  • The lecturer notes that in a previous example discussed in class, only one mean square was used, affecting the analysis of mixed effects.
  • For the mixed effects model:
    • The sum of squares comprises the error component, and the sum of squares of the block.
  • Thus, there are two components in analyzing error through the variance-covariance matrix.

Compound Symmetry

  • When experimental units are equally correlated, this relationship exhibits compound symmetry.
    • Compound symmetry refers to the situation where variance and covariance are constant across experimental units within blocks.
  • Different covariance structures exist, revealing that mixed models can more accurately estimate error, particularly with unbalanced data.

Components of Error Estimation

  • The denominator mean square for the error ratio on treatment is calculated as MSE (Mean Square Error).
  • The standard error for an estimated least square mean is derived from the square root of variance:
    • extStandardError=extSquareRootofVarianceoftheMeanext{Standard Error} = ext{Square Root of Variance of the Mean}
  • This approach alone can lead to an incomplete representation of variability as it omits variability due to the block component.

Misestimating Variance

  • The result may be an underestimation of variance components due to not including block variance into error calculations.
  • Software like SAS may wrongly neglect this block component and thus yield incomplete results.

PROC MIXED Advantages

  • The discussion emphasizes using PROC MIXED to better estimate the error component in mixed models.
    • The estimated variance from block influence not only enhances precision but also correctly reflects covariation in the experimental design.

Transition to Chi-Squared Analysis

  • The lecture shifts focus to binary (yes/no) responses, e.g., pregnancy status.
  • Chi-squared tests evaluate whether there is a significant effect based on categorized observations.

Chi-Squared Calculation

  • Chi-square equation given as:
    extChi−squared=(extObserved−extExpected)2extExpectedext{Chi-squared} = \frac{( ext{Observed} - ext{Expected})^2}{ ext{Expected}}
  • Example setup for analysis with 50 ewes divided into 4 pastures. Pregnancy rates are recorded:
    • Pasture 1: 35 pregnant, 15 non-pregnant.
    • Pasture 2: 40 pregnant, 10 non-pregnant.
    • Pasture 3: 25 pregnant, 25 non-pregnant.
    • Pasture 4: 26 pregnant, 24 non-pregnant.
  • Expected values are calculated as total pregnant divided by pastures: (Total Pregnant / Number of Pastures).
    • Concrete numbers: Expected for pasture one = 31.5 pregnant, 18.5 non-pregnant.
  • Differences calculated and squared for the chi-squared formula with results:
    1. For Pasture 1: (3.5)218.5=0.39\frac{(3.5)^2}{18.5} = 0.39
    2. For Pasture 2: (−3.5)218.5=0.39\frac{(-3.5)^2}{18.5} = 0.39
    3. For Pasture 3: (−6.5)231.5=1.34\frac{(-6.5)^2}{31.5} = 1.34
    4. For Pasture 4: (−0.5)231.5=0.08\frac{(-0.5)^2}{31.5} = 0.08
  • Summation yields total chi-squared value of approximately 13.47.

Conclusion on Chi-Squared Significance

  • To determine significance: Compare calculated chi-squared (13.47) to table value (7.81).
  • The total is greater than the critical value at 3 degrees of freedom, indicating that the distribution of pregnancy rates among pastures is statistically significant and they differ markedly.

SAS Application

  • SAS will automatically compute chi-square values, offering more efficiency, but understanding the computation is crucial for comprehension and analysis of results.

Summary

  • Chi-square results suggest that pregnancy rates in different pastures are significantly different, warranting further analysis and interpretations.
  • The consultation concludes with emphasis on building a foundational understanding of mixed models and their applications in data analysis.