10.3 Linear Regression and Multiple Regression Analysis (MRA)

10.3 Linear Regression and Multiple Regression Analysis (MRA)

Governed by the principles of USPAP Standard 1 (for single-property appraisal support) and USPAP Standard 5 (for mass appraisal model development), linear regression is one of the most powerful quantitative methods available to the Certified General Appraiser.

---

1. Simple Linear Regression: The Bivariate Model

Simple linear regression models the mathematical relationship between two variables: one dependent variable ($Y$) and one independent explanatory variable ($X$):

$Y = b_0 + b_1 X + e$

Where:

• $Y$ = Dependent variable (the variable being predicted or explained, such as total sale price, price per square foot, or contract rent per square foot)
• $X$ = Independent variable (the predictor variable, such as gross building area, site size, or building age)
• $b_0$ = $Y$-intercept (the constant; the estimated value of $Y$ when $X = 0$)
• $b_1$ = Slope coefficient (the marginal change in $Y$ associated with a one-unit change in $X$)
• $e$ = Random error term (residual; the difference between actual observed $Y$ and predicted value $\hat{Y}$)

                               SIMPLE LINEAR REGRESSION MODEL
             Sale Price (Y)
                   ^
                   |                                          / (Regression Line: Ŷ = b₀ + b₁X)
                   |                                        /'
                   |                                     .o'  <-- Observed Sale (Y)
                   |                                   .' |   
                   |                                 .'   | Residual (e = Y - Ŷ)
                   |                              .'*-----'   
                   |                            .'   <-- Predicted Value (Ŷ)
                   |                         .'
                   |                      .' 
                   |                   .' 
             b₀ ---+->              .' (Slope b₁ = Contributory Value per Unit of X)
   (Intercept)     |             .'
                   |          .'
                   |       .'
                   +----------------------------------------------------> Property Attribute (X)
                   0                                                     (e.g., Gross Building Area)

The Ordinary Least Squares (OLS) Criterion

Linear regression solves for the line of best fit using Ordinary Least Squares (OLS) estimation. OLS calculates $b_0$ and $b_1$ such that the sum of the squared vertical distances (residuals) between actual data points ($Y_i$) and the predicted regression line ($\hat{Y}_i$) is minimized:

$\text{Minimize } \sum_{i=1}^n e_i^2 = \sum_{i=1}^n (Y_i - \hat{Y}_i)^2$

Squaring the residuals prevents positive and negative errors from canceling each other out and penalizes larger prediction errors more heavily than small ones.

---

2. Multiple Regression Analysis (MRA)

In real estate markets, property prices are never driven by a single attribute. A warehouse's value depends simultaneously on building size, clear ceiling height, dock doors, lot coverage ratio, office buildout percentage, and highway proximity. Multiple Regression Analysis (MRA) extends linear regression to multiple independent variables:

$Y = b_0 + b_1 X_1 + b_2 X_2 + b_3 X_3 + \dots + b_k X_k + e$

Where:

• $Y$ = Dependent variable (e.g., Sale Price in dollars)
• $b_0$ = Intercept term
• $X_1, X_2, \dots, X_k$ = $k$ distinct independent property variables
• $b_1, b_2, \dots, b_k$ = Partial regression slope coefficients (marginal contributory value of each attribute, holding all other variables constant)
• $e$ = Residual error term

Contributory Value of Attributes

The partial regression coefficient ($b_k$) quantifies the Principle of Contribution. If an MRA model for industrial properties produces a coefficient of $b_1 = +145.00$ (+$145.00) for Gross Building Area ($X_1$), holding age, ceiling height, and lot size constant, each additional square foot of warehouse space adds an estimated $145.00 of contributory value to total sale price.

Qualitative Variables: Dummy Variables ($0$ or $1$)

Real estate analysis frequently involves qualitative or categorical features that cannot be measured on a continuous numerical scale (e.g., presence of fire sprinklers, railroad siding, corner location, or zoning classification). These attributes are incorporated into regression models using binary dummy variables (indicator variables):

$\text{Dummy Variable } (D) = \begin{cases} 1 & \text{if the attribute is present} \\ 0 & \text{if the attribute is absent} \end{cases}$

---

3. Statistical Evaluation and Model Diagnostics

A certified appraiser cannot simply accept an MRA model output without auditing its statistical validity. Professional practice requires evaluating four core diagnostic indicators: goodness of fit, prediction precision, variable statistical significance, and multicollinearity.

+---------------------------------------------------------------------------------------------------+
|                             MRA STATISTICAL DIAGNOSTIC SCORECARD                                  |
+---------------------+-----------------------------+-----------------------------------------------+
| DIAGNOSTIC METRIC   | FORMULA / BASIS             | APPRAISAL MEANING & DESIRED THRESHOLD         |
+---------------------+-----------------------------+-----------------------------------------------+
| Coefficient of      | R² = Explained Var /        | Measures % of price variation explained.      |
| Determination (R²)  | Total Variance              | Desired: ≥ 0.80 to 0.90 for commercial MRA.   |
+---------------------+-----------------------------+-----------------------------------------------+
| Adjusted R²         | R²_adj = 1 -                | Penalizes irrelevant variables; prevents      |
|                     | [(1-R²)(n-1)/(n-k-1)]       | artificial inflation of fit.                  |
+---------------------+-----------------------------+-----------------------------------------------+
| Standard Error of   | SEE = √(Σe² / (n - k - 1))  | Standard deviation of residuals; prediction   |
| Estimate (SEE)      |                             | error in dollar terms. Lower is superior.     |
+---------------------+-----------------------------+-----------------------------------------------+
| t-Statistic /       | t = b_k / SE(b_k)           | Tests if attribute has statistically          |
| p-Value             | p-value from t-distribution | significant impact. Desired: |t| ≥ 2.0, p < .05|
+---------------------+-----------------------------+-----------------------------------------------+

1. Coefficient of Determination ($R^2$) vs. Adjusted $R^2$

• $R^2$ (Unadjusted): Represents the proportion of total variation in the dependent variable explained by the regression model ($0.0 \le R^2 \le 1.0$). An $R^2$ of $0.85$ means the model explains 85% of price variation.
• The Flaw of Unadjusted $R^2$: Mathematically, adding any new variable to an OLS model—even completely irrelevant random numbers—will automatically increase (or never decrease) the unadjusted $R^2$. Unethical or naive modelers often "stuff" models with irrelevant variables to show clients an inflated $R^2$.
• Adjusted $R^2$ ($R^2_{\text{adj}}$): Adjusts $R^2$ for degrees of freedom, penalizing the addition of variables that do not contribute meaningful explanatory power: $R^2_{\text{adj}} = 1 - \left[ \frac{(1 - R^2)(n - 1)}{n - k - 1} \right]$ Where $n$ is sample size and $k$ is number of independent variables. If a newly added variable does not improve the model more than would be expected by chance, Adjusted $R^2$ decreases. Appraisers must rely on Adjusted $R^2$ when evaluating model strength.

2. Standard Error of the Estimate (SEE)

The Standard Error of the Estimate measures the dispersion of observed data points around the regression hyperplane. It represents the standard deviation of the residuals:

$\text{SEE} = \sqrt{\frac{\sum_{i=1}^n (Y_i - \hat{Y}_i)^2}{n - k - 1}}$

Unlike $R^2$, which is unitless, SEE is expressed in the exact units of the dependent variable (e.g., ±$45,000 or ±$3.50/SF). A lower SEE indicates greater predictive precision.

3. Hypothesis Testing: $t$-Statistics and $p$-Values

To determine whether an individual property attribute genuinely influences market value, the appraiser conducts a hypothesis test:

• Null Hypothesis ($H_0$): $b_k = 0$ (the attribute has zero contributory value in the market).
• Alternative Hypothesis ($H_1$): $b_k \ne 0$ (the attribute has a statistically significant contributory value).
• The $t$-Statistic: $t = \frac{b_k}{\text{SE}(b_k)}$ Dividing the estimated coefficient by its standard error measures how many standard deviations the coefficient lies away from zero. As a rule of thumb, at a 95% confidence level ($\alpha = 0.05$), a $|t|$-statistic of $2.0$ or greater (and a corresponding $p$-value $< 0.05$) indicates that the attribute is statistically significant.

---

4. Worked MRA Model Interpretation Example

An appraiser calibrates an MRA model to value light industrial flex buildings in a suburban commercial park based on 45 verified transactions ($n = 45$). The model includes three independent variables: Gross Building Area ($X_1$, in SF), Effective Age ($X_2$, in years), and a Dummy Variable for Dock-High Loading ($X_3$: $1 = \text{Dock Loading}, 0 = \text{Grade-Level Only}$):

Regression Output Summary Table

• Model Diagnostics: Sample Size $n = 45$; Number of Predictors $k = 3$; $R^2 = 0.885$; Adjusted $R^2 = 0.877$; SEE = $52,000.

Valuation Application for Subject Property

The subject property is a 25,000 SF flex industrial building with an effective age of 10 years and dock-high loading ($X_3 = 1$):

$\hat{Y} = 450{,}000 + 185(25{,}000) - 12{,}500(10) + 95{,}000(1)$ $\hat{Y} = 450{,}000 + 4{,}625{,}000 - 125{,}000 + 95{,}000 = \mathbf{\$5{,}045{,}000}$

Applying the Standard Error of the Estimate (SEE = $52,000), the 68% confidence interval for this prediction is $5,045,000 ± $52,000 ($4,993,000 to $5,097,000), providing an exceptionally robust statistical foundation for reconciliation.

---

5. Multicollinearity in Real Estate Valuation

Multicollinearity occurs when two or more independent variables in an MRA model are highly correlated with each other. In real estate, multicollinearity is rampant because property characteristics naturally co-vary:

• Larger buildings are almost always situated on larger parcels of land.
• Properties with more square footage have more bathrooms, more parking spaces, and higher utility capacities.

+---------------------------------------------------------------------------------------------------+
|                                MULTICOLLINEARITY DIAGNOSTIC MATRIX                                |
+-------------------------+-------------------------------------------------------------------------+
| SYMPTOM                 | MANIFESTATION IN APPRAISAL REGRESSION                                   |
+-------------------------+-------------------------------------------------------------------------+
| High R² / Low t-Stats   | Model exhibits high R² (e.g., 0.90), but individual variables have      |
|                         | t-statistics below 2.0 and p-values > 0.05 (failure to show significance)|
+-------------------------+-------------------------------------------------------------------------+
| Counter-Intuitive Signs | Coefficients display the wrong economic sign (e.g., a negative dollar   |
|                         | contributory value for additional building square footage).             |
+-------------------------+-------------------------------------------------------------------------+
| Coefficient Instability | Adding or dropping a single sale causes massive, wild swings in the     |
|                         | magnitudes of other property coefficients.                              |
+-------------------------+-------------------------------------------------------------------------+
| High VIF / Correlation  | Bivariate correlation r > 0.80, or Variance Inflation Factor (VIF) > 5  |
|                         | (VIF > 10 indicates severe, fatal multicollinearity).                   |
+-------------------------+-------------------------------------------------------------------------+

Remedying Multicollinearity

When multicollinearity threatens model credibility, appraisers implement recognized modeling remedies:

1. Combine Correlated Variables: Transform building size and lot size into a single ratio metric, such as the Land-to-Building Ratio or Floor Area Ratio (FAR).
2. Drop the Redundant Variable: Remove the variable with lower market importance or higher data error.
3. Stepwise or Ridge Regression: Employ advanced mathematical regularization to constrain coefficient variance.