Adv Quant: More on Logistic Regression

In assessing the predictive power of categorical predictors of a binary outcome, should logistic regression be used? In other words, how is the logistic function used to predict categorical outcomes?

Advertisements

Logistic regression is another flavor of multi-variable regression, where one or more independent variables are continuous or categorical which are used to predict a dichotomous/ binary/ categorical dependent variable (Ahlemeyer-Stubbe, & Coleman, 2014; Field, 2013; Gall, Gall, & Borg, 2006; Huck, 2011).  Zheng and Agresti (2000) defines predictive power as a measure that helps compare competing regressions via analyzing the importance of the independent variables.  For linear regression and multiple linear regression, the correlation coefficient and coefficient of determination are adequate for predictive power (Field, 2014; Zheng & Agresti, 2000). The more data that is collected could yield a stronger predictive power (Field, 2014).  Predictive power is used to sell the relationships between variables to management (Ahlemeyer-Stubbe, & Coleman, 2014).

For logistic regression, the predictive power of the independent variables can be evaluated by the concept of the odds ratio for each independent variable (Huck, 2011). Field (2013) and Schumacker (2014) explained that when the logistic regression is calculated, the categorical/binary variables are transformed into ln(odds ratio) and a regression is then performed on this newly scaled variable (scale factor seen in equation 1):

eq1.PNG                                                 (1)

Since the probability of one categorical variable varies between 0à0.999…, the odds ratio value can vary between 0 à 999.999… (Schumacker, 2014). If the value of the odds ratio is greater than 1, it will show that as the independent variable value increases, so do the odds of the dependent variable (Y = n) occurs increases and vice versa (Field, 2013). Thus, the odds ratio measures the constant strength of association between the independent and dependent variables (Huck, 2011; Smith, 2015).  Due to this ln(odds ratio) transformation, logistic regression should be used for binary outcomes.

Field (2013) and Schumacker (2014) further explained that given that this ln(odds ratio) transformation needs to be made on the variables; the way to predict categorical outcomes from the regression formula (2),

  eq2.PNG                                            (2)

is best to explained the probability of the categorical outcome value one is trying to calculate:

eq3                                                (3)

The probability equation (3) can be expressed in multiple ways, through typical algebraic manipulations.  Thus, the probability/likelihood of the dependent variable Y is defined between 0-100% and the odds ratio is used to discuss the strength of these relationships.

References

  • Ahlemeyer-Stubbe, Andrea, Shirley Coleman. (2014). A Practical Guide to Data Mining for Business and Industry, 1st Edition. [VitalSource Bookshelf Online].
  • Gall, M. D., Gall, J. P., Borg, W. R. (2006). Educational Research: An Introduction, 8th Edition. [VitalSource Bookshelf Online].
  • Field, Andy. (2013). Discovering Statistics Using IBM SPSS Statistics, 4th Edition. [VitalSource Bookshelf Online].
  • Huck, Schuyler W. (2011). Reading Statistics and Research, 6th Edition. [VitalSource Bookshelf Online].
  • Schumacker, Randall E. (2014). Learning Statistics Using R, 1st Edition. [VitalSource Bookshelf Online].
  • Smith, M. (2015). Statistical analysis handbook. Retrieved from http://www.statsref.com/HTML/index.html?introduction.html
  • Zheng, B. and Agresti, A. (2000) Summarizing the predictive power of a generalized linear model.  Retrieved from http://www.stat.ufl.edu/~aa/articles/zheng_agresti.pdf

Adv Quant: Logistic Vs Linear Regression

A discussion on the assumptions that must be met for logistic regression and assumptions for regular regression that do not apply in logistic regression and on the types of variables used in logistic regression and regular regression.

To generalize the results of the research the insights gained from a sample of data needs to use the correct mathematical procedures for using probabilities and information, statistical inference (Gall et al., 2006; Smith, 2015).  Gall et al. (2006), stated that statistical inference is what dictates the order of procedures, for instance, a hypothesis and a null hypothesis must be defined before a statistical significance level, which also has to be defined before calculating a z or t statistic value. Essentially, a statistical inference allows for quantitative researchers to make inferences about a population.  A population, where researchers must remember where that data was generated and collected from during quantitative research process.  The orders of procedures are important to apply statistical inferences to regressions, if not the prediction formula will not be generalizable.

Logistic regression is another flavor of multi-variable regression, where one or more independent variables are continuous or categorical which are used to predict a dichotomous/ binary/ categorical dependent variable (Ahlemeyer-Stubbe, & Coleman, 2014; Field, 2013; Gall, Gall, & Borg, 2006; Huck, 2011).  Logistic regression is an alternative to linear regression, which assumes all variables are continuous (Ahlemeyer-Stubbe, & Coleman, 2014). Both the multi-variable linear regression and logistic regression formula are (Field, 2013; Schumacker, 2014):

Y = a + b11 + b2X2 + …                                                       (1)

The main difference between these two regressions is that the variables in the equation (1) represent different types of dependent (Y) and independent variables (Xi).  These different types of variables may have to undergo a transformation before the regression analysis begins (Field, 2013; Schumacker 2014).  Due to the difference in the types of variables between logistic and linear regression the assumptions on when to use either regression are also different (Table 1).

Table 1: Discusses and summarizes the types of assumptions and variables used in both logistic and regular regression, created from Ahlemeyer-Stubbe & Coleman (2014), Field (2013), Gall et al. (2006), Huck (2011) and Schumacker, (2014).

 

Assumptions of Logistic Regression Assumptions for Linear Regression
·         Multicollinearity should be minimized between the independent variables

·         There is no need for linearity between the dependent and independent variables

·         Normality only on the continuous independent variables

·         No need for homogeneity of variance within the categorical variables

·         Error terms a not normally distributed

·         Independent variables don’t have to be continuous

·         There are no missing data (no null values)

·         Variance that is not zero

·         Multicollinearity should be minimized between the multiple independent variables

·         Linearity exists between all variables

·         Additivity (for multi-variable linear regression)

·         Errors in the dependent variable and its predicted values are independent and uncorrelated

·         All variables are continuous

·         Normality on all variables

·         Normality on the error values

·         Homogeneity of variance

·         Homoscedasticity- variance between residuals are constant

·         Variance that is not zero

Variable Types of Logistic Regression Variable Types of Linear Regression
·         2 or more Independent variables

·         Independent variables: continuous, dichotomous, binary, or categorical

·         Dependent variable: dichotomous, binary

·         1 or more Independent variables

·         Independent variables: continuous

·         Dependent variables: continuous

References

  • Ahlemeyer-Stubbe, Andrea, Shirley Coleman. (2014). A Practical Guide to Data Mining for Business and Industry, 1st Edition. [VitalSource Bookshelf Online].
  • Gall, M. D., Gall, J. P., Borg, W. R. (2006). Educational Research: An Introduction, 8th Edition. [VitalSource Bookshelf Online].
  • Field, Andy. (2013). Discovering Statistics Using IBM SPSS Statistics, 4th Edition. [VitalSource Bookshelf Online].
  • Huck, Schuyler W. (2011). Reading Statistics and Research, 6th Edition. [VitalSource Bookshelf Online].
  • Schumacker, Randall E. (2014). Learning Statistics Using R, 1st Edition. [VitalSource Bookshelf Online].
  • Smith, M. (2015). Statistical analysis handbook. Retrieved from http://www.statsref.com/HTML/index.html?introduction.html

Adv Quant: Overfitting & Parsimony

Parsimony in statistical modeling is often discussed in terms of Occam’s razor in the formulation of hypotheses.

Overfitting and Parsimony

Overfitting a regression model is stuffing it with so many variables that have little contributional weight to help predict the dependent variable (Field, 2013; Vandekerckhove, Matzke, & Wagenmakers, 2014).  Thus, to avoid the over-fitting problem, the use of parsimony is important in big data analytics.  Parsimony is describing a dependent variable with the fewest independent variables as possible (Field, 2013; Huck, 2013; Smith, 2015).  The best way to describe this is to use the “Keep It Simple Sweaty,” concept on the regression model.  The concept of parsimony could be attributed to Occam’s Razor, which states “plurality out never be posited without necessity” (Duignan, 2015).  Vandekerckhove et al. (2014) describe parsimony as a way of removing the noise from the signal to create better predictive regression models.

Overfitting in General Least Squares Model (GLM)

For multivariate regressions, a correlation matrix could be conducted on all the variables, to help with identifying parsimony, such that the software will try to maximize the correlation while minimizing the number of variables (Field, 2013).  Smith (2015) stated that the proportion of variation should remain high between the variables and that the correlation between the separate independent variables should be as low as possible. If the correlation coefficient between the independent variables is high (0.8 or higher), then there is a chance that there are extraneous variables (Smith, 2015).   Another technique to achieve parsimony is called the backward stepwise method, which is to run a regression model with all variables, and remove those variables that don’t contribute to the models significantly, or the model could start with one variable and add variables until it has maximized correlation and variance in a forward stepwise method (Field, 2013; Huck, 2015).

Unfortunately, there is still a problem of overfitting when conducting a backward stepwise method, forward stepwise method, or correlation matrix in multivariate linear models.  That is because, computers tend to remove, add or consider variables systematically and mathematically, not based on human knowledge (Field, 2013; Huck, 2015). Thus, it is still important to have a human to evaluate the computational output for logic, consistency, and reliability.  However, if the focus is to reduce overfitting, it should be noted that underfitting should also be avoided.  Underfitting a regression model happens when the model leaves out key independent variables that can help predict the dependent variable from the model (Field, 2013).

Hierarchical regression methods

When the researcher builds a multivariate regression model, they build it in stages, as they tend to add known independent variables first, and add newer independent variables to avoid overfitting in a technique called hierarchical regression (Austin, Goel & van Walraven, 2001; Field, 2013; Huck 2013).  The new and unknown independent variables could be entered in through a stepwise algorithm as abovementioned, or another step could be created where suspected new variables that may have a high contribution to the predictability of the dependent variables are added next (Field, 2013).  Hierarchical regression methods allow the researcher to analyze the differing hierarchical levels by examining not only the correlations between the levels but also the intercepts and slopes, helping drive valid statistical inferences (Austin et al., 2001).

Vandekerckhove et al. (2014) listed these three hierarchical methods for model selection; where each method is balancing between the goodness of fit and parsimony:

  • Akaike’s Information Criterion (AIC) considers how much-observed data influences the belief of one model over the other, but is unreliable with huge amounts of data
  • Bayesian Information Criterion (BIC) considers how much-observed data influences the belief of one model over the other and can handle huge amounts of data, but is known to underfit
  • Minimum Description Length (MDL) considers how much a model can compress the observed data, through identifying regularity within the data values

Vandekerckhove et al, (2014), also stated that the model with the lowest AIC and/or BIC score would be the best to choose.

In conclusion, under parsimony, if adding another variable does not improve the regression formula, then should not be added into the assessment to avoid overfitting (Field, 2013). General Least Squares Models have issues in overfitting because computers systematically and mathematically conduct their analysis and lack the human knowledge to keep removing unneeded variables from the equation.  Hierarchical regression methods can help minimize overfitting through indirect calculation of a parsimony value (Vandekerckhove et al., 2014).

References

  • Austin, P. C., Goel, V., & van Walraven, C. (2001). An introduction to multilevel regression models. Canadian Journal of Public Health92(2), 150.
  • Duignan, B. (2015). Occam’s razor. Encyclopaedia Britannica. Retrieved from https://www.britannica.com/topic/Occams-razor
  • Field, Andy. (2013). Discovering Statistics Using IBM SPSS Statistics, 4th Edition. [VitalSource Bookshelf Online].
  • Huck, S. W. (2013) Reading Statistics and Research (6th ed.). Pearson Learning Solutions. [VitalSource Bookshelf Online].
  • Smith, M. (2015). Statistical analysis handbook. Retrieved from http://www.statsref.com/HTML/index.html?introduction.html
  • Vandekerckhove, J., Matzke, D., & Wagenmakers, E. J. (2014). Model comparison and the principle of parsimony.

Adv Quant: Polynomial Regression in R

For this local polynomial regression, the “oldfaithful.csv” will be used from the open-source data. The eruption times (in minutes) and the waiting time to the next eruption (in minutes) of 272 eruptions are provided for the Old Faithful geyser.

Introduction

For this local polynomial regression, the “oldfaithful.csv” will be used from the open-source data. The eruption times (in minutes) and the waiting time to the next eruption (in minutes) of 272 eruptions are provided for the Old Faithful geyser.

Results

IP2F1.PNG

Figure 1: Density Histograms for eruptions times and eruption waiting times.

IP2F2.PNG

Figure 2: Smoothed density histrogram from local polynomial regresion.

IP2F3.png

Figure 3: Intercomparisson of linear regression (blue), lowess regresion(red), and polynomial regression (green) on the eruption data.

IP2F4

Figure 4: Residual plots for both linear and polynomial regression.

Discussion

The histogram plots (Figure 1) illustrate that both variables, eruption times and eruptions waiting time are both bimodal distributions.  Thus, a linear regression (Figure 3), would not capture the relationship between these two variables.  A polynomial smoothed version of the bimodal curve (Figure 2) show that for low values of the geysers magnitude, there is a low wait time for the next occurrence and vice versa.  The smoothed density curve shows the estimate values of the geyser’s variable distribution better than the bar histogram

LOCFIT (locally fitted regression) and LOWESS (locally weighted scatterplot smoothing regression) are assessed alongside the typical LM (linear regression).  LOCFIT is based on LOWESS, which allows the end user to specify the smoothing parameter and neighborhood size, but LOCFIT affords the end user more control over other the smoothing parameters (Futschik & Crompton, 2004).  Both LOCFIT and LOWESS are methods for regression that uses the nearest-neighbor-based model (Field, 2013; Futschik & Crompton, 2004; Loader, 2013; Smith, 2015).  This analysis will look at all three.

The goal is to see if there is a relationship between the waiting time to the next eruption to the magnitude of the eruption (per eruption time).  Through the linear regression algorithm, the linear model is eruptions = 0.075628 (waiting) – 1.874016.  The Pearson’s correlation coefficient is 0.9008112. Thus 81.14% of the variation could be explained by a linear regression model.  The lowess regression appears not to capture the distribution of data at smaller eruption times, but it is better than the linear regression model since its correlation is 0.9809684, and can explain 0.9622990 of the variation between the variables.

Finally, to evaluate the effectiveness of the linear model and the polynomial model, residuals must be assessed (Figure 4). Both of the residual plots don’t show any discernable pattern. However, the residuals are closer to zero in the polynomial regression, suggesting that it does a better job at explaining the variance between the eruption magnitude and the next eruption wait time.  In conclusion, the best regression for this data set appears to be the polynomial regression.

Code

#

## Use R to analyze the faithful dataset.

## This is a version of the eruption data from the “Old Faithful” geyser in Yellowstone National Park, Wyoming.

##  •     X (primary key)

##  •     eruptions (eruption time [mins])

##  •     waiting (wait time for this eruptions [mins])

#

fateful = read.csv(file=”https://raw.githubusercontent.com/vincentarelbundock/Rdatasets/master/csv/datasets/faithful.csv”, header = TRUE, sep = “,”)

head(fateful)

# Produce density histograms of eruption times and of waiting times.

hist(fateful$eruptions, freq=F, xlab = “eruptions time [mins]”,  main = “Histogram of the eruptions time”)

hist(fateful$waiting, freq=F, xlab = “eruptions waiting time [mins]”,  main = “Histogram of the eruptions waiting time”)

# Produce a smoothed density histogram from local polynomial regression.

install.packages(“locfit”)

library(locfit)

plot(locfit(~lp(fateful$eruptions),data=fateful), xlab = “eruptions time [mins]”,  main = “Histogram of the eruptions time”)

plot(locfit(~lp(fateful$waiting),data=fateful), xlab = “eruptions waiting time [mins]”,  main = “Histogram of the eruptions waiting time”)

# Compare local polynomial regression to regular regression.

lowessRegression = lowess(fateful$waiting, faithful$eruptions, f=2/3)

polynomialRegression = locfit(fateful$eruptions~lp(fateful$waiting))

linearRegression = lm(fateful$eruptions~fateful$waiting)

# Graphing the data

plot(fateful$waiting, fateful$eruptions, main = “Eruption Times”, xlab=”eruption time [min]”, ylab = “Waiting time to next eruption [min]”)

lines(lowessRegression, col=”red”)

abline(linearRegression, col=”blue”)

lines(polynomialRegression, col=”green”)

# summary on the regressions

summary(linearRegression)

# correlations on the regressions

cor(fateful$eruptions,fateful$waiting)

cor(lowessRegression$x, lowessRegression$y)

# Plotting residuals

plot(residuals(linearRegression), main = “residuals for the linear regression”, ylab = “residuals”)

plot(residuals(polynomialRegression), main = “residuals for the polynomial regression”, ylab=”residuals”)

References

Adv Quant: Locally Weighted Scatterplot Smothing (LOWESS) in R

When using the locally weighted scatterplot smoothing (LOWESS) method for multiple regression models in a k-nearest-neighbor-based model, the discussion will revolve on if this is a parametric or nonparametric method as well as some of the advantages and disadvantages of LOWESS from a computational standpoint.

Locally weighted scatterplot smoothing (LOWESS) method for multiple regression models in a k-nearest-neighbor-based model is a regression model with 1+ independent variables, which uses a non-parametric method which creates a smoothed surface/curve (Field, 2013; Smith, 2015).  LOWESS aims not to introduce a parametric model to the data, because doing so, would require much more resources (Cleveland, 1979).  Non-parametric tests have fewer assumptions than parametric tests, such as there are no assumptions on the sampled variable’s population distribution (Field, 2013; Huck, 2013; Schumacker, 2014; Smith, 2015).

Assumptions in the parametric analysis, which are based on the normal distribution, include (1) additivity and linearity; (2) normality; (3) homoscedasticity/homogeneity of variance; and (3) independence (Field, 2013; Huck, 2013). However, the assumption of independence still exists in the non-parametric analysis (Huck, 2013).  Smith (2015) states that these non-parametric analyses are less powerful than parametric analysis.  However, Field (2013) disagrees and says that they are powerful, but admits that there is a loss of information about the magnitude between the observed values.  Huck (2013), states that when using non-parametric analysis correctly, they have the similar power/weight to them as parametric analysis on data that meet the parametric assumptions. Thus, to conduct non-parametric analysis, data values are ranked and arranged, thus higher valued data have higher valued ranks and vice versa (Field, 2013; Huck, 2013; Smith, 2015). Cleveland (1979), describes that only a fraction of the data (local neighbors) are considered at a time, to minimize the weighing function.  Thus, a LOWESS regression is carried out on the ranked data, which help eliminates the effects of outliers, irons out skewed distributions (Field, 2013; Smith, 2015).

Advantages and disadvantages

+ LOWESS doesn’t depend on an underlying population distribution (Field, 2013; Huck, 2013; Schumacker, 2014; Smith, 2015)

+ Looking at the data’s local neighboring data creates a smoothing function, which visually enhances pattern (Cleveland, 1979)

– The LOWESS technique is not a substitute for parametric regression analysis (Huck, 2013).  Thus, to use non-parametric tests, one must reject the null hypothesis: the data follows a defined distribution; with its corresponding alternative hypothesis: the data does not follow a defined distribution (Field, 2013; Huck, 2013).

– LOWESS is computationally heavy, especially depending on the weights chosen (Cleveland, 1979).

– Though the regression formula is easily and visually represented/smoothed, but the regression formula may not be as cleanly written (Cleveland, 1979).

Multiple Regression Analysis

From the R dataset archived website (http://vincentarelbundock.github.io/Rdatasets/), the NOxEmissions.csv file was downloaded, which is the Nox Air Pollution Data and it has 5 variables: primary key, Julian Calendar Day (julday), hourly mean of NOx concentrations in the air in parts per billion (LNOx), hourly sum of NOx car emissions in parts per billion (LNOxEm), and square root of the wind speed in meters per second (sqrtWS).

From this dataset, it is hypothesized, that the wind speed combined with the sum of NOx from car emissions could contribute to the mean Nox concentrations in the atmosphere.  Thus, given that there are multiple independent variables for one dependent variable, then multiple regression analysis is best suited (Field, 2013; Huck, 2013; Schumacker, 2014; Smith, 2015).

IP1.51F1

Figure 1: Histogram of each of the variables in the data set.

IP1.51F2
Figure 2: Simple Linear Regression between each of the independent variables to the dependent variables.  For the image on the right the regression formula is LNOx = -0.86443(sqrtWS) + 5.55885, with a correlation of -0.4300 and for the image on the left the regression formula is LNOx = 0.590066 (LNOxEm) + 0.048645, with a correlation of 0.6399.

IP1.51F3

Figure 3: The summation output of the Linear Multiple Regression, where the regression formula is LNOx= -1.018198 (sqrtWS) + 0.641412 (LNOxEm) + 1.061950, which explains 66.3% of the variation between these variables.

IP1.51F4.png

Figure 4: Normal Quantile-Quantile plot, for the multiple linear regression as described by Figure 3.

The histograms (Figure 1) are not convincing that this could be tested with a normal multiple linear regression analysis, but from the Normal quantile-quantile plot (Figure 4), shows normalcy in the data, justifying the results (Figure 3).  For furthering the understanding of the multiple linear regression, the simple linear regression per independent variable (Figure 2), shows that neither independent variable alone explain the variance between the variables as well as with the multiple regression analysis.

IP1.51F5.png

Figure 5: Multiple LOWESS regression plot with varying smoothing span.

Even though there is normalcy in the data, a LOWESS was still plotted on the data, just to illustrate how the differences between smoothing factors can influence the result.  The smoothing factor describes how small the neighborhood is on the k-nearest neighbor (Cleveland, 1979).  The smaller the smoothing factor, the smaller the neighborhood, and the blue line (f=2/3) is the default value in R (R, n.d.e,).  The larger the smoothing factor, the bigger the neighborhood, over simplifying the result.

Code

NOxData = read.csv(file=”https://raw.githubusercontent.com/vincentarelbundock/Rdatasets/master/csv/robustbase/NOxEmissions.csv”, header = TRUE, sep = “,”)

head(NOxData)

hist(NOxData$LNOx, freq=F, xlab = “hourly mean of NOx concentrations [ppb]”,  main = “Histogram of the hourly mean of NOx concentrations”)

hist(NOxData$LNOxEm, freq=F, xlab = “hourly sum of NOx car emissions [ppb]”,  main = “Histogram of the hourly sum of NOx car emissions”)

hist(NOxData$sqrtWS, freq=F, xlab = “square root of winds [m/s]”, main = “Histogram of the square root of winds”)

# Single Linear Regressions on LNOxEm

## LNOx

plot(NOxData$LNOxEm, NOxData$LNOx)

abline(lm(NOxData$LNOx~NOxData$LNOxEm), col=”red”)

summary(lm(NOxData$LNOx~NOxData$LNOxEm))

cor(NOxData$LNOx,NOxData$LNOxEm)

## sqrtWS

plot(NOxData$sqrtWS, NOxData$LNOx)

abline(lm(NOxData$LNOx~NOxData$sqrtWS), col=”red”)

summary(lm(NOxData$LNOx~NOxData$sqrtWS))

cor(NOxData$LNOx,NOxData$sqrtWS)

# Multiple Linear Regression on both LNOxEM and sqrtWS variables on LNOx

RegressionModel = lm(NOxData$LNOx~ NOxData$LNOxEm + NOxData$sqrtWS)

summary(RegressionModel)     

plot(RegressionModel)

# Pearson’s Correlation between independent variables

cor(NOxData$LNOxEm, NOxData$sqrtWS)

# 95% Confidence Intervals on the regression model

confint(RegressionModel, conf.level=0.95)

# LOWESS MODEL

LowessModel = lowess(NOxData$LNOx~ NOxData$LNOxEm + NOxData$sqrtWS, f=2/3)

LowessModel2 = lowess(NOxData$LNOx~ NOxData$LNOxEm + NOxData$sqrtWS, f=0.01)

LowessModel3 = lowess(NOxData$LNOx~ NOxData$LNOxEm + NOxData$sqrtWS, f=1)

plot(LowessModel,type=”l”,col=”blue”, main=”LOWESS Regression: green is f=1, blue is f=2/3, & red is f=0.01″)

lines(LowessModel2, col=”red”)

lines(LowessModel3, col=”green”)

References