Univariate Normal Distribution; GLM with the Univariate Normal; Least Squares Estimation

Transcription

1 Univariate Normal Distribution; GLM with the Univariate Normal; Least Squares Estimation EPSY 905: Multivariate Analysis Spring 2016 Lecture 3: February 10, 2016

2 Today s Class The building blocks: The basics of mathematical statistics: Ø Random variables: definitions and types Ø Univariate distributions w General terminology w Univariate normal (aka, Gaussian) w Other (continuous) univariate distributions Ø Expected values: means, variances, and the algebra of expectations Ø Linear combinations of random variables The finished product: How the GLM fits within statistics Ø The GLM with the normal distribution Ø The statistical assumptions of the GLM Estimation of GLMs with Least Squares

3 RANDOM VARIABLES AND STATISTICAL DISTRIBUTIONS

4 Random Variables Random: situations in which the certainty of the outcome is unknown and is at least in part due to chance Variable: a value that may change given the scope of a given problem or set of operations = Random Variable: a variable whose outcome depends on chance + (possible values might represent the possible outcomes of a yet-to-beperformed experiment) Today we will denote a random variable with a lower-cased: x

5 Key Features of Random Variables Random variables each are described by a probability density/mass function (PDF) f(x) that indicates relative frequency of occurrence Ø A PDF is a mathematical function that gives a rough picture of the distribution from which a random variable is drawn The type of random variable dictates the name and nature of these functions: Ø Continuous random variables: w f(x) is called a probability density function w Area under curve must equal 1 (found by calculus integration) w Height of curve (the function value f x ): Can be any positive number Reflects relative likelihood of an observation occurring Ø Discrete random variables: w f(x) is called a probability mass function w Sum across all values must equal 1 w The function value f(x) is a probability (so must range from 0 to 1)

6 Uses of Distributions in Data Analysis Statistical models make distributional assumptions on various parameters and/or parts of data These assumptions govern: Ø How models are estimated Ø How inferences are made Ø How missing data may be imputed If data do not follow an assumed distribution, inferences may be inaccurate Ø Sometimes a very big problem, other times not so much Therefore, it can be helpful to check distributional assumptions prior to (or while) running statistical analyses

7 CONTINUOUS UNIVARIATE DISTRIBUTIONS

8 Continuous Univariate Distributions To demonstrate how continuous distributions work and look, we will discuss two: Ø Uniform distribution Ø Normal distribution Each are described a set of parameters, which we will later see are what give us our inferences when we analyze data What we then do is put constraints on those parameters based on hypothesized effects in data

9 Uniform Distribution The uniform distribution is shown to help set up how continuous distributions work For a continuous random variable x that ranges from a, b, the uniform probability density function is: f x = 1 b a The uniform distribution has two parameters: Ø a the lower limit Ø b the upper limit x U a, b In R: the dunif() function gives you the height: dunif(x, min = 0, max = 1, log = FALSE)

10 More on the Uniform Distribution To demonstrate how PDFs work, we will try a few values: x a b f(x) = = = =.1 The uniform PDF has the feature that all values of x are equally likely across the sample space of the distribution Ø Therefore, you do not see x in the PDF f x The mean of the uniform distribution is. (a + b) / The variance of the uniform distribution is. / b a./

11 Univariate Normal Distribution For a continuous random variable x (ranging from to ) the univariate normal distribution function is: 1 f x = exp x μ / < 2πσ 2σ The shape of the distribution is governed by two parameters: Ø The mean μ < Ø The variance σ Ø These parameters are called sufficient statistics (they contain all the information about the distribution) The skewness (lean) and kurtosis (peakedness) are fixed Standard notation for normal distributions is x N(μ <, σ ) Ø Read as: x follows a normal distribution with a mean μ < and a variance σ Linear combinations of random variables following normal distributions result in a random variable that is normally distributed Ø You ll see this later with respect to the GLM

12 Univariate Normal Distribution f(x) f x gives the height of the curve (relative frequency) for any value of x, μ <, and σ

13 More of the Univariate Normal Distribution To demonstrate how the normal distribution works, we will try a few values: x μ x σ2 x f(x) The values from f x were obtained by using Excel Ø Ø The =normdist() function Most statistics packages have a normal distribution function The mean of the normal distribution is μ < The variance of the normal distribution is σ In R the dnorm() function provides the height for a normal distribution: dnorm(x, mean = 0, sd = 1, log = FALSE)

14 EXPECTED VALUES AND THE ALGEBRA OF EXPECTATIONS

15 Expected Values Expected values are statistics taken the sample space of a random variable: they are essentially weighted averages The weights used in computing this average correspond to the densities for a continuous random variable (the heights of the distribution) Notation: the expected value is represented by: E(x) Ø The actual statistic that is being weighted by the PDF is put into the parentheses where x is now Expected values allow us to understand what a statistical model implies about data, for instance: Ø How a GLM specifies the (conditional) mean and variance of a DV

16 Expected Value Calculation For discrete random variables, the expected value is found by: E x = F xp(x = x) < For example, the expected value of a roll of a die is: E x = = 3.5 For continuous random variables, the expected value is found by: E x = L xf x dx < We won t be calculating theoretical expected values with calculus we use them only to see how models imply things about our data

17 Variance and Covariance As Expected Values A distribution s theoretical variance can also be written as an expected value: V x = E x E x / = E x μ < / Ø This formula will help us understand predictions made GLMs and how that corresponds to statistical parameters we interpret For a roll of a die, the theoretical variance is: V x = E x 3.5 / = Ȯ / + Ȯ / + Ȯ ( 3 3.5) / + Ȯ / + Ȯ / + Ȯ / = 2.92 Ø Likewise, the SD is then 2.92 = 1.71 Likewise, for a pair of random variables x and z, the covariance can be found from their joint distributions: Cov x, z = E xz E x E z = E xz μ < μ V

18 LINEAR COMBINATIONS OF RANDOM VARIABLES

19 Linear Combinations of Random Variables A linear combination is an expression constructed from a set of terms by multiplying each term by a constant and then adding the results x = a. v. + a / v / + + a X v X Ø The linear regression equation is a linear combination More generally, linear combinations of random variables have specific implications for the mean, variance, and possibly covariance of the new random variable As such, there are predicable ways in which the means, variances, and covariances change Ø These terms are called the algebra of expectations To guide us through this process, we will use the descriptive statistics from the height/weight/gender example from our 1 st class

20 Descriptive Statistics for Height/Weight Data Variable Mean SD Variance Height Weight , Female Diagonal: Variance Above Diagonal: Covariance Correlation /Covariance Height Weight Female Height Weight.798 3, Female Below Diagonal: Correlation

21 Algebra of Expectations Here are some properties of expected values: x and z are random variables, c and d constants Sums of Constants: E x + c = E x + c V x + c = V x Cov x + c, z = Cov(x, z) Products of Constants: E cx = ce x V cx = c / V x Cov cx, dz = cdcov(x, z) Sums of Random Variables: E cx + dz = ce x + de z V cx + dz = c / V x + d / V z + 2cd Cov x, z

22 Examples for Algebra of Expectations Image you wanted to convert weight from pounds to kilograms (where 1 pound = kg) Weight`a =.453Weight bc The mean (expected value) of weight in kg: E Weight`a = E.453Weight bc =.453E Weight bc =.453Weight bc = = 83.08kg The variance of weight in kg: V Weight`a = V.453Weight bc =.453 / V Weight bc =.453 / 3, = kg / The covariance of weight in kg with height in inches: Cov Weight`a, Height = Cov.453Weight bc, Height =.453Cov Weight bc, Height = = kg inches

23 Don t Take My Word For It R syntax for transforming weight in a DATA step: R syntax and output for marginal descriptive statistics and covariances:

24 Where We Use This The Dreaded ESTIMATE Statement The glht statement in R computes the expected value and standard error (square root of variance) for a new random variable Ø The new random variable is a linear combination of the original model parameters (the fixed effects) Ø The original model parameters are considered random here as their sampling distribution is used (assuming normal errors and a large N) Where: Estimate = 1 β r<srturxvrw + 1 β x/ r<srturxvrw Ø β r<srturxvrw has mean β r<srturxvrw y Ø β x/ r<srturxvrw has mean β x/ r<srturxvrw y Ø There exists a covariance between β r<srturxvrw y w We ll call this Cov β r<srturxvrw y, β x/ r<srturxvrw y and variance se β r<srturxvrw y / and variance se β x/ r<srturxvrw y / and β x/ r<srturxvrw y

25 THE GENERAL LINEAR MODEL WITH WHAT WE HAVE LEARNED TODAY

26 The General Linear Model, Revisited The general linear model for predicting Y from X and Z: Y s = β { + β. X s + β / Z s + β } X s Z s + e s In terms of random variables, under the GLM: e s is considered random: e s N 0,σ r / Y s is dependent on the linear combination of X s, Z s, and e s The GLM provides a model for the conditional distribution of the dependent variable, where the conditioning variables are the independent variables: f(y s X s, Z s ) Ø There are no assumptions made about X s and Z s - they are constants Ø The regression slopes β {, β., β /, β } are constants that are said to be fixed at their values (hence, called fixed effects)

27 Combining the GLM with Expectations Using the algebra of expectations predicting Y from X and Z: The expected value (mean) of f Y p X p, Z p : Y s = E Y s = E β { + β. X s + β / Z s + β } X s Z s + e s Constants Random Variable with E e s = 0 = β { + β. X s + β / Z s + β } X s Z s + E e s = β { + β. X s + β / Z s + β } X s Z s The variance of f Y p X p, Z p : V Y s = V β { + β. X s + β / Z s + β } X s Z s + e s = V e s = σ r /

28 Distribution of f Y s X s, Z s We just found the mean (expected value) and variance implied by the GLM for the conditional distribution of Y s given X s and Z s The next question: what is the distribution of f Y s X s, Z s? Linear combinations of random variables that are normally distributed result in variables that are normally distributed Because e s N 0, σ r / is the only random term in the GLM, the resulting conditional distribution of Y s is normally distributed: Y s X s, Z s N β { + β. X s + β / Z s + β } X s Z s, σ r / Model for the means: from fixed effects; literally gives mean of f Y s X s, Z s Model for the variances: from random effects; gives variance of f Y s X s, Z s

29 Examining What This Means in the Context of Data If you recall from our first lecture, the final model we decided to interpret: Model 5 W s = β { + β. H s H + β / F s + β } H s H F s + e s where e s N 0, σ r / From R:

30 Picturing the GLM with Distributions The distributional assumptions of the GLM are the reason why we do not need to worry if our dependent variable is normally distributed Our dependent variable should be conditionally normal We can check this assumption by checking our assumption about the residuals, e s N 0,σ r / More on this soon

31 More Pictures of the GLM Treating our estimated values of the slopes β {, β., β /, β } and the residual variance σ / r as the true values* we can now see what the theoretical* distribution of f Weight s Height s, Female s looks like for a given set of predictors *Note: these distributions change when sample estimates are used (think standard error of the prediction)

32 Behind the Pictures To emphasize the point that PDFs provide the height of the line, here is the normal PDF (with numbers) that produced those plots: f W s H s, F s = 1 2πσ r / exp = = 1 2πσ r / exp 1 2π 4.73 exp Model for the Variance W s Wˆs 2σ r / W s β { + β. H s H + β / F s + β } H s H F s / 2σ r / W s H s H 82.27F s 1.09 H s H F s The plots were created using the following value for the predictors: H = 67.9 Left plot: H s = 62; F s = 1 Right plot: H s = 76;F s = 0 Model for the Means / /

33 ESTIMATION OF GLMS USING LEAST SQUARES

34 Why Estimation is Important In applied statistics courses, estimation is not discussed very frequently Ø Can be very technical very intimidating Estimation is of critical importance Ø Quality and validity of estimates (and of inferences made from them) depends on how they were obtained Ø New estimation methods appear from time to time and get widespread use without anyone asking whether or not they are any good Consider an absurd example: Ø I say the mean for IQ should be 20 just from what I feel Ø Do you believe me? Do you feel like reporting this result? w Estimators need a basis in reality (in statistical theory)

35 How Estimation Works (More or Less) Most estimation routines do one of three things: 1. Minimize Something: Typically found with names that have least in the title. Forms of least squares include Generalized, Ordinary, Weighted, Diagonally Weighted, WLSMV, and Iteratively Reweighted. Typically the estimator of last resort 2. Maximize Something: Typically found with names that have maximum in the title. Forms include Maximum likelihood, ML, Residual Maximum Likelihood (REML), Robust ML. Typically the gold standard of estimators (and next week we ll see why).

36 Estimation of General Linear Models Recall our GLM (shown here for the prediction of a dependent variable Y s by two independent variables X s and Z s ): Y s = β { + β. X s + β / Z s + β } X s Z s + e s Traditionally (dating to circa 1840), general linear models can be estimated via a process called least squares Least squares attempts to find the GLM parameters (the βs) that minimize the squared residual terms: min {β {, β., β /, β } } F e s / sž.

37 Where We Are Going (and Why We Are Going There) Because the basics of estimation are critical to understanding the validity of the numbers you will use to make inferences from, we will detail the process of estimation Ø Today with Least Squares and then ending with Maximum Likelihood The LS estimation we will discuss is to get you to visualize functions of statistical parameters (the βs here) and data in order to show which estimates we would choose Ø To be repeated: In practice LS estimation for GLMs does not do this (by the magic of calculus and algebra) In the end, we would like for you to understand that not all estimators are created equally and that some can be trusted more than others

38 How Least Squares Estimation Works How Least Squares works is through minimizing the squared error terms but its what goes into error that drives the process: e s = Y s Y s = Y s β { + β. X s + β / Z s + β } X s Z s If you were to do this (and you wouldn t), the process called optimization would go like this: 1. Pick values for regression slopes 2. Calculate Y s and then e s for each person p 3. Calculate OF = e/ sž. s function) (letters OF stand for objective 4. Repeat 1-3 until you find the values of regression slopes

39 Today s Example Data Imagine an employer is looking to hire employees for a job where IQ is important Ø We will only use 5 observations so as to show the math behind the estimation calculations The employer collects two variables: Ø IQ scores Variable Mean SD Ø Job performance IQ Descriptive Statistics: Performance Observation IQ Performance Covariance Matrix IQ Performance

40 Visualizing the Data

41 Let s Play Pick the Parameters This slide is left as a placeholder for the recording we will now do a demonstration in R

42 And The Winner Is

43 Examining the Objective Function Surface

44 LS Estimates of GLMs The process of least squares estimation of GLMs does not need an iterative search Using calculus, a minimum of the objective function can be found Ø This involves taking the first derivative of the objective function with respect to each parameter w Derivative = slope of the tangent line for a given point Ø The first derivative is then set equal to zero w Flat slope = minimum (or maximum or saddle point neither apply here) Ø The equation is then solved for the parameter w Producing the equations you know and love For simple linear regression (one predictor): Ø β = œ š š œ = v žÿtuÿxvr ŸX š v žÿtuÿxvr ŸX = vt st v Ÿtr When we get to matrix algebra, you will know this as β = X X. X y

45 Ramifications of Closed-Form LS Equations Because least squares estimates have a closed form (equations that will provide statistics directly), they will work nearly every time Ø Only fail when collinearity is present (soon you ll know this to mean X X is singular and cannot be inverted) Virtually all other estimators you will encounter in statistics will not have a closed form Ø Even least squares estimators for other types of data (not continuous) Without a closed form, least squares estimates are found by search the objective function for its minimum Ø Like finding the drain of a pool.

46 Why LS is Still in Use Least squares estimates still make up the bulk of GLM analyses because: Ø They are easy to compute Ø They pretty much always give you an answer Ø They have been shown to have good statistical properties The good statistical properties actually come because LS estimates of GLMs match the Maximum Likelihood Estimates Ø We will learn more about maximum likelihood estimation next Ø For now, know that MLEs are the gold standard when it comes to estimates

47 Where LS Fails For all their flexibility, least squares estimates are somewhat limited Ø Only have good properties for basic univariate GLM for continuous data w Normally distributed error terms with homogeneous variance When data are not continuous/do not have normally distributed error terms, least squares estimates are not preferred For multivariate models with continuous data (repeated measures, longitudinal data, scales of any sort), least squares estimates quickly do not work Ø Cannot handle missing outcomes (deletes entire case) Ø Limited in the types of ways of modeling covariance between observations

48 AN INTRODUCTION TO GENERALIZED MODELS

49 A World View of Models Statistical models can be broadly organized as: Ø General (normal outcome) vs. Generalized (not normal outcome) Ø One dimension of sampling (one variance term per outcome) vs. multiple dimensions of sampling (multiple variance terms) w Fixed effects only vs. mixed (fixed and random effects = multilevel) All models have fixed effects, and then: Ø General Linear Models: conditionally normal distribution for data, fixed effects, no random effects Ø General Linear Mixed Models: conditionally normal distribution for data, fixed and random effects Ø Generalized Linear Models: any conditional distribution for data, fixed effects through link functions, no random effects Ø Generalized Linear Mixed Models: any conditional distribution for data, fixed and random effects through link functions Linear means the fixed effects predict the link-transformed DV in a linear combination of (effect*predictor) + (effect*predictor)

50 Observed outcomes: what you collect and evaluate based Unpacking the Big Picture Observed Outcomes (any format) Hypothesized Causal Process Model: Substantive Theory Substantive theory: what guides your study Hypothetical causal process: what the statistical model is testing (attempting to falsify) when estimated

51 The Goal of Generalized Models Generalized models map the substantive theory onto the sample space of the observed outcomes Ø Sample space = type/range/outcomes that are possible The general idea is that the statistical model will not approximate the outcome well if the assumed distribution is not a good fit to the sample space of the outcome Ø If model does not fit the outcome, the findings cannot be believed The key to making all of this work is the use of differing statistical distributions for the outcome Generalized models allow for different distributions for outcomes Ø The mean of the distribution is still modeled by the model for the means

52 What kind of outcome? Generalized vs. General Generalized Linear Models à General Linear Models whose residuals follow some not-normal distribution and in which a linktransformed Y is predicted instead of Y Many kinds of non-normally distributed outcomes have some kind of generalized linear model to go with them: Ø Binary (dichotomous) Ø Unordered categorical (nominal) Ø Ordered categorical (ordinal) Ø Counts (discrete, positive values) These two are often called multinomial inconsistently Ø Censored (piled up and cut off at one end left or right) Ø Zero-inflated (pile of 0 s, then some distribution after) Ø Continuous but skewed data (pile on one end, long tail)

53 Some Links/Distributions (from Wikipedia)

54 3 Parts of a Generalized Linear Model Link Function (main difference from GLM): Ø How a non-normal outcome gets transformed into something we can predict that is more continuous (unbounded) Ø For outcomes that are already normal, general linear models are just a special case with an identity link function (Y * 1) Model for the Means ( Structural Model ): Ø How predictors linearly relate to the link-transformed outcome Ø New link-transformed Y p = β 0 + β 1 X p + β 2 Z p + β 3 X p Z p Model for the Variance ( Sampling/Stochastic Model ): Ø If the errors aren t normally distributed, then what are they? Ø Family of alternative distributions at our disposal that map onto what the distribution of errors could possibly look like

55 A General Linear Model With Binary Outcomes? How can we have a linear relationship between X & Y? Probability of a 1 is bounded between 0 and 1, but predicted probabilities from a linear model aren t bounded Ø Impossible values Linear relationship needs to shut off somehow à made nonlinear Prob (Y=1) ???? Prob (Y=1) X Predictor X Predictor

56 WRAPPING UP

57 Wrapping Up Today discussed estimation, and in the process showed how differing estimators can give you different statistics The key today was to shake your statistical view point: Ø There are many more ways to arrive at statistical results than you may know The take home point is that not all estimators are created equal Ø If ever presented with estimates: ask how the numbers were attained Ø If ever getting estimates: get the best you can with your data Next week your world will further be expanded when we introduce maximum likelihood estimators