Transformations and Weighting to Correct Model Inadequacies: In the present study, we

focus on methods and procedures for building regression models when some of the Regression
model assumptions are violated. We place considerable significance on data transformation.
It is not unusual to find that when the response and/or the regressor variables are expressed in
the correct scale of measurement or metric, certain violations of assumptions, such as inequality
of variance, are no longer present. Ideally, the choice of metric should made by the engineer
or scientist with subject-matter knowledge, but there are many situations where this
information is not available. In these cases, a data transformation may be chosen heuristically
or by some analytical procedure.
The method of weighted least squares is also useful in building regression models in
situations where some of the underlying assumptions are violated. We will illustrate how
weighted least squares can be used when the equal variance assumption is not appropriate.
Variance-Stabilizing Transformations: The assumption of constant variance is a basic
requirement of regression analysis. A common reason for the violation of this assumption is
for the response variable y to follow a probability distribution in which the variance is
functionally related to the mean.
For example, if y is a Poisson random variable in a simple linear regression model, then
the variance of y is equal to the mean. Since the mean of y is related to the regressor variable
X, the variance of y will be proportional to X. Variance stabilizing transformations are often
useful in these cases. Thus, if the distribution of y is Poisson, we could regress 𝑦 ′ = √𝑦 against
x since the variance of the square root of a Poisson random variable is independent of the mean.
In another example, if the response variable is proportion 0 ≤ 𝑦𝑖 ≤ 1 and the plot of the
residuals versus 𝑦̂𝑖 has the double bow pattern, then the arcsin transformation 𝑦 ′ = 𝑠𝑖𝑛−1 (√𝑦)
is appropriate.
Several commonly used variance stabilizing transformations are summarized in the following
table

10
 Several commonly used variance - stabilizing transformations are summarized in the
above table. The strength of a transformation depends on the amount of curvature that it
induces. The transformations shown in the above table range from the relatively mild square
root to the relatively strong reciprocal. Generally speaking, a mild transformation applied over
a relatively narrow range of values (e.g., ymax / ymin < 2, 3) has little effect. On the other hand,
a strong transformation over a wide range of values will have a dramatic effect on the analysis.
Sometimes we can use prior experience or theoretical considerations to guide us in
selecting an appropriate transformation. However, in many cases we have no a priori reason
to suspect that the error variance is not constant. Our first indication of the problem is from
inspection of scatter diagrams or residual analysis. In these cases the appropriate
transformation may be selected empirically.
Transformations to Linearize the Model:
The assumption of a linear relationship between y and the regressors is the usual starting point
in regression analysis. Occasionally we find that this assumption is inappropriate. Nonlinearity
may be detected scatter diagrams or residual plots. Sometimes prior experience or theoretical
considerations may indicate that the relationship between y and the regressors is not linear. In
some cases a nonlinear function can be linearized by using a suitable transformation. Such
nonlinear models are called intrinsically or transformably linear.
Several linearizable functions are shown the following figure.

11
 The corresponding nonlinear functions, transformations, and resulting linear forms are
shown in the following table. When the scatter diagram of y against x indicates curvature, we
may be able to match the observed behaviour of the plot to one of the curves in above figure
and use the linearized form of the function to represent the data.

To illustrate a nonlinear model that is intrinsically linear. Consider the exponential function
𝑦 = 𝛽0 𝑒 𝛽1 𝑥 𝜖
This function is intrinsically linear since it can be transformed to a straight line by a logarithmic
transformation
𝑙𝑛𝑦 = 𝑙𝑛𝛽0 + 𝛽1 𝑥 + 𝑙𝑛𝜖
or
𝑦 ′ = 𝛽0′ + 𝛽1 𝑥 + 𝜖 ′
as shown in above table this transformation requires that the transformed error terms ε′=lnε are
normally and independently distributed with mean zero and variance σ2. This implies that the
multiplicative error ε in the original model is log normally distributed. We should look at the
residuals from the transformed model to see if the assumptions are valid. Generally if X and/or
y are in the proper metric, the usual least squares assumptions are more likely to be satisfied,
although it is no unusual to discover at this stage that a nonlinear model is preferable. Various
types of reciprocal transformations are also useful.
1
For example, the model 𝑦 = 𝛽0 + 𝛽1 (𝑥) + 𝜖

can be linearized by using the reciprocal transformation x ′=1 /x. The resulting linearized model
is
𝑦 = 𝛽0 + 𝛽1 𝑥 ′ + 𝜖
Other models that can be linearlized by reciprocal transformations are
1 𝑥
= 𝛽0 + 𝛽1 𝑥 + 𝜖 and 𝑦=𝛽
𝑦 0 +𝛽1 𝑥+𝜖

This last model is illustrated in the above figure g, h. When transformations such as those
described above are employed, the least - squares estimator has least - squares properties with
respect to the transformed data, not the original data.

12