1 / 45

Penalized Regression

Penalized Regression. BMTRY 726 3/7/14. Maximum Likelihood. Has been a statistics “staple” (and still is) ML approach -specify a distribution for our data

gamba
Download Presentation

Penalized Regression

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Penalized Regression BMTRY 726 3/7/14

  2. Maximum Likelihood Has been a statistics “staple” (and still is) ML approach -specify a distribution for our data -From this obtain a log-likelihood function, , which specifies how likely the data are for various values of the unknown parameters -Estimate the parameters by maximizing In regression we also use likelihood ratio tests to conduct hypothesis testing about regression parameters and models

  3. Maximum Likelihood However, maximum likelihood estimation can be unsatisfactory -Lack of interpretability: when p is large we may prefer a smaller set of predictors to determine those most strongly associated with y -Large variability: when p is large relative to n, or when columns in our design matrix are highly correlated, variance of parameter estimates is large

  4. Subset Selection We began to address the issue of interpretability of a regression model using the idea of subset selection (1) Examine all possible subsets of predictors from 1 to pand compare models on some criteria (Cp, AIC, …) -Computationally difficult for p > 30 (2) Use a stepwise procedure to select a subset -Forward selection: start with null model and add predictors one at a time according to most significant F- value until no additional predictors have F-value < a -Backward selection: Start with all predictors and remove least significant until all predictors have F-value < a

  5. Problems with Subset Selection The first method is more systematic but still may not provide a clear best subset The second method is done post-hoc and generally can’t be replicated easily Subset selection is also discontinuous -i.e. small changes in the data can results in very different estimates As a result subset selection is often unstable and highly variable, particularly when p is large

  6. Problems with Subset Selection What about collinearity? -difficult to ensure we’ve chosen the “best” subset of predictors -Addition of a predictor to a model when a collinear predictor is already in the model often results in one or both predictors being non-significant due to the variance inflation This results because our estimate of (X’X)-1 is nearly singular due to the near linearity of columns in X- i.e. X is not really full rank

  7. Problems with Subset Selection Models identified by subset selection are reported as if the model was specified apriori This violates statistical principles for estimation and hypothesis testing: -standard errors biased low -p-values falsely small -regression coefficients biased away from zero (as much as 1-2 standard errors off!- Miller, 2002) This brings us to penalized regression -achieve same goals but more stable, continuous, and computationally efficient fashion

  8. Penalized Regression Deals with some of the regression problems by introducing a penalty and changing what we maximize M(q ) is called the objective function and includes the additional terms -The penalty functionp(q ):penalizes “less realistic” values of the b ’s -The regularization (shrinkage) parameterl: controls the tradeoff between bias and variance

  9. Penalized Regression In statistical testing we tend to believe that without tremendous evidence to the contrary, the null hypothesis is true In regression this means that unless you give me a lot of evidence to suggest otherwise, I tend to think most of the model coefficients are closer to zero Thus we are arguing that large coefficients are “less likely” So in penalized regression we want to penalize coefficients the further they get from zero

  10. Penalized Regression We can also think about this in terms of minimizing a loss function rather than maximizing our likelihood. In this case we rewrite our expression as In the regression setting we can write M(q ) in terms of our regression parameters b as follows The penalty function takes the form

  11. Constrained Regression Alternatively, we can think about penalized regression as applying a constraint on the values of q For example we may want to maximize our likelihood but subjected to the constraint that P(q ) <t By the Lagrange multiplier method we arrive at the same regression function

  12. Bayesian Interpretation From a Bayesian prospective, we can think of the penalty arising from a prior distribution on the parameters The objective function M(q) is proportional to the log of the posterior distribution of q given the data By optimizing this objective function, we are finding the mode of the posterior distribution

  13. More on the Penalty Function Let’s consider the impact of the penalty function When q > 1, fq(b ) is continuous and convex It is also convex when q= 1 but not continuous When q< 1 it is not convex and minimization much more difficult

  14. Regularization Parameter Note the regularization parameter l controls the trade-off between the penalty and the fit -If l is too small the tendency is to overfit the data and have a model with large variance -If l is too large, the model may underfit the data yielding a model that is too simple We will discuss how to select an appropriate l a little later

  15. The Intercept In penalized regression we are assuming that the coefficient values near zero are more likely However, there are some considerations in application It does not make sense to apply this to the intercept as this would imply that the mean of y is 0 Thus the intercept isn’t included in the penalty

  16. Other Concerns Another concern is that predictor variable can vary greatly in magnitude (and type) What if the data contain one predictor that takes values between 0 and 1 and a another that takes values between 0 and 100000 In this case, certainly a one unit change is not equivalent in the two predictors q To get around this issue, predictor variables are usually standardized prior to fitting the model to have mean 0 and standard deviation of 1

  17. Other Concerns This can be done without loss of generality as (1) location shifts are absorbed into the intercept (2) Scale shifts can be reversed after the model has been fit If we had to divide xjby a to standardize, we simply divide our transformed coefficient by a to get on our original scale

  18. Centering There are other benefits to centering and scaling our predictors Predictors are now orthoganol to the intercept, thus in the standardized covariate space regardless of the rest of the model This implies that if we center y by subtracting the mean, we don’t really have to estimate

  19. Ridge Regression Now let’s focus on a specific penalized regression Ridge regression penalizes the regression estimated by minimizing In statistics this is also referred to as shrinkage since we are shrinking our coefficients towards 0 We therefore also call the regularization parameter l the shrinkage parameter

  20. Ridge Regression Fortunately the ridge solution is easy to find because our loss function is still a quadratic function of b

  21. Consider an Example What if Xhas perfect collinearity X’X has eigenvalues 10, 6, and 0

  22. Consider an Example In the ridge fix we add lI so X’X has non-zero eigenvalues, for example:

  23. Consider an Example Now suppose we had a response vector Each choice of l results in specific estimates of the coefficients

  24. Ridge Regression Note the solution is very similar to the OLS solution but there is a “ridge” down the diagonal of X’X So how does l impact our solution? This illustrates why we call l the shrinkage parameter In the special case of an orthonormal design matrix we get

  25. Ridge Regression Ridge regression has the most benefit in instances where there is high collinearity. In such a case, the OLS solution for b may not be unique as X may not really be full rank (i.e. X’X is not necessarily invertible) However, in ridge regression, for any design matrix X, the matrix X’X +lIis always invertible! Therefore there is always a unique solution for

  26. Bias-Variance Trade Off In introducing a penalty, when developing our likelihood estimate, we are also introducing bias This is the tradeoff we have to accept for reduced variance Note, the total variance is monotone decreasing with respect to l, while the total squared bias is monotone increasing wrtl.

  27. vs. Existence Theorem: There always exists a l such that the MSE of is less than the MSE of This theorem has huge implications! Even if our model is exactly correct and follows the specified distribution, we can always obtain an estimator with lower MSE by shrinking the parameters towards zero

  28. Choosing Shrinkage Parameter AIC is a one way of choosing models to balance fit and parsimony To apply AIC (or BIC for that matter) to choosing l, we need estimate the degrees of freedom In linear regression we have

  29. Choosing Shrinkage Parameter In ridge regression, this relationship is similar but our H is: What happens if l = 0? What if it is infinity?

  30. Choosing Shrinkage Parameter Now that we can quantify the degrees of freedom in our ridge regression model, we have a was to calculate AIC or BIC We can use this information to choose l, using our estimated degrees of freedom

  31. Choosing Shrinkage Parameter Alternatively we could base our choice of l on how well our ridge model predicts y However, this would be using the data twice (a statistical no-no) In an ideal world we have new data to externally validate our choice of l But wait! We know how to do this in statistics… -We split the data into test and training sets and build the model on the training data and test it on the test data -Even better we could….

  32. K-fold Cross-Validation We split the data into K equally size subsets and construct models on the data leaving out the kth subset (data X-k) We can repeat this for all k subsets at each level of l we want to consider and average the results across the k models .… This procedure is referred to as K-fold cross-validation 5 and 10 are the most common choices for K 1 2 3 4 K

  33. Leave-One-Out CV When K = n we refer to this as leave-one-out cross-validation Similar to leave-one-out residuals (jackknife) in which we look at the effect of leaving an observation out. The main concern here would be poor computational efficiency of repeatedly calculating H However the following approximation, called generalized cross-validation (GCV), may be used instead

  34. Choosing Shrinkage Parameter In general GCV is a popular method for choosing an appropriate shrinkage factor GCV is also the only method readily provided in certain statistical software In this case, though we can do ridge regression in R and SAS, only R provides GCV Lets look at an example

  35. Body Fat Example Recall our regression model > summary(mod13) Call: lm(formula = PBF ~ Age + Wt + Ht + Neck + Chest + Abd + Hip + Thigh + Knee + Ankle + Bicep + Arm + Wrist, data = bodyfat, x = T) Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -18.18849 17.34857 -1.048 0.29551 Age 0.06208 0.03235 1.919 0.05618 . Wt -0.08844 0.05353 -1.652 0.09978 . Ht -0.06959 0.09601 -0.725 0.46925 Neck -0.47060 0.23247 -2.024 0.04405 * Chest -0.02386 0.09915 -0.241 0.81000 Abd 0.95477 0.08645 11.04 < 2e-16 *** Hip -0.20754 0.14591 -1.422 0.15622 Thigh 0.23610 0.14436 1.636 0.10326 Knee 0.01528 0.24198 0.063 0.94970 Ankle 0.17400 0.22147 0.786 0.43285 Bicep 0.18160 0.17113 1.061 0.28966 Arm 0.45202 0.19913 2.270 0.02410 * Wrist -1.62064 0.53495 -3.030 0.00272 ** Residual standard error: 4.305 on 238 degrees of freedom. Multiple R-squared: 0.749, Adjusted R-squared: 0.7353 . F-statistic: 54.65 on 13 and 238 DF, p-value: < 2.2e-16

  36. Body Fat Example Recall our regression model > summary(mod13) Call: lm(formula = PBF ~ Age + Wt + Ht + Neck + Chest + Abd + Hip + Thigh + Knee + Ankle + Bicep + Arm + Wrist, data = bodyfat, x = T) Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -18.18849 17.34857 -1.048 0.29551 Age 0.06208 0.03235 1.919 0.05618 . Wt -0.08844 0.05353 -1.652 0.09978 . Ht -0.06959 0.09601 -0.725 0.46925 Neck -0.47060 0.23247 -2.024 0.04405 * Chest -0.02386 0.09915 -0.241 0.81000 Abd 0.95477 0.08645 11.04 < 2e-16 *** Hip -0.20754 0.14591 -1.422 0.15622 Thigh 0.23610 0.14436 1.636 0.10326 Knee 0.01528 0.24198 0.063 0.94970 Ankle 0.17400 0.22147 0.786 0.43285 Bicep 0.18160 0.17113 1.061 0.28966 Arm 0.45202 0.19913 2.270 0.02410 * Wrist -1.62064 0.53495 -3.030 0.00272 ** Residual standard error: 4.305 on 238 degrees of freedom. Multiple R-squared: 0.749, Adjusted R-squared: 0.7353 . F-statistic: 54.65 on 13 and 238 DF, p-value: < 2.2e-16

  37. Body Fat Example What if we consider ridge regression as an alternative? In R the code is: >library(MASS) >bodyfat2<-scale(bodyfat) >bodyfat2<-as.data.frame(bodyfat2) > lam<-c(seq(0,9.99, by=.01),seq(10,99.9, by=.1),seq(101,10000, by=1)) > ridgemod<-lm.ridge(PBF~., data=bodyfat2, lam=lam) > select(ridgemod) modified HKB estimator is 1.677373 modified L-W estimator is 3.902049 smallest value of GCV at 1.18

  38. Body Fat Example We can plot l versus our estimated coefficients This is called a trace plot There is no direct function in R, we must specify this ourselves- for example > par(mfrow=c(1,2)) > plot(x=0, y=0, xlim=c(0,10000),ylim=c(-0.3, 1.2), type="n", ylab="Coefficients“, xlab="Lambda") > for (i in 1:nrow(ridgemod$coef)) lines(lam, ridgemod$coef[i,]) > for (i in 1:nrow(ridgemod$coef)) lines(lam, ridgemod$coef[i,], col=i) > abline(h=0, lty=2) > abline(v=1.18, lty=3, col="red", lwd=2) > plot(x=0, y=0, xlim=c(0,25),ylim=c(-0.3, 1.2), type="n", ylab="Coefficients“, xlab="Lambda") > for (i in 1:nrow(ridgemod$coef)) lines(lam, ridgemod$coef[i,]) > for (i in 1:nrow(ridgemod$coef)) lines(lam, ridgemod$coef[i,], col=i) > abline(h=0, lty=2) > abline(v=1.18, lty=3, col="red", lwd=2)

  39. Trace Plot

  40. Effect of df(l) Alternatively we can look at the effect of the effective degrees of freedom on lambda

  41. Body Fat Example OLS Ridge Regression > ridgemod<-lm.ridge(PBF~., data=bodyfat, lam=1.18) > ridgemod Int -16.07 Age 0.070 Wt -0.074 Ht -0.084 Neck -0.469 Chest 0.005 Abd 0.901 Hip -0.188 Thigh 0.232 Knee -0.002 Ankle 0.151 Bicep 0.165 Arm 0.440 Wrist -1.659 > summary(mod) Call: lm(formula = PBF ~ ., data = bodyfat2) Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 5.452032 24.760561 0.220 0.82592 Age 0.065291 0.033604 1.943 0.05324 . Wt -0.002408 0.072163 -0.033 0.97341 Ht -0.239064 0.196291 -1.218 0.22451 Neck -0.388059 0.243861 -1.591 0.11291 Chest -0.132094 0.112572 -1.173 0.24184 Abd 0.901663 0.093813 9.611 < 2e-16 *** Hip -0.212855 0.158553 -1.342 0.18076 Thigh 0.202608 0.150403 1.347 0.17928 Knee -0.008189 0.253012 -0.032 0.97421 Ankle -0.008517 0.365308 -0.023 0.98142 Bicep 0.190207 0.187009 1.017 0.31017 Arm 0.191335 0.341657 0.560 0.57601 Wrist -1.605308 0.595433 -2.696 0.00754 ** Residual standard error: 4.28 on 230 degrees of freedom Multiple R-squared: 0.7444, Adjusted R-squared: 0.73 F-statistic: 51.54 on 13 and 230 DF, p-value: < 2.2e-16

  42. Body Fat Example What can we do in SAS? SAS doesn’t estimate GCV which is the preferred method for finding the “best” l However, it does generate trace plots and a plot of the VIF relative to the ridge coefficients Here’s the basic SAS code ods graphics on; proc reg data=bf plots(only)=ridge(unpack VIFaxis=log) outest=b ridge=0 to 10 by .01; model pbf=age wt ht neck chest abd hip thigh knee ankle bicep arm wrist; run; ods graphics off;

  43. SAS Trace Plot

  44. VIF versus Our Ridge Coefficients

More Related