---------------------------------------------------------------------------------------------------------------------------------- help cvlasso lassopack v1.4.2 ---------------------------------------------------------------------------------------------------------------------------------- Title cvlasso -- Program for cross-validation using lasso, square-root lasso, elastic net, adaptive lasso and post-OLS estimators Syntax Full syntax cvlasso depvar regressors [if exp] [in range] [, alpha(numlist) alphacount(int) sqrt adaptive adaloadings(string) adatheta(real) ols lambda(real) lcount(integer) lminratio(real) lmax(real) lopt lse lglmnet notpen(varlist) partial(varlist) psolver(string) ploadings(string) unitloadings prestd fe noftools noconstant tolopt(real) tolzero(real) maxiter(int) nfolds(int) foldvar(varname) savefoldvar(varname) rolling h(int) origin(int) fixedwindow seed(real) plotcv plotopt(string) saveest(string)] Note: the fe option will take advantage of the ftools package (if installed) for the fixed-effects transform; the speed gains using this package can be large. See help ftools or click on ssc install ftools to install. Estimators Description ---------------------------------------------------------------------------------------------------------------------------- alpha(numlist) a scalar elastic net parameter or an ascending list of elastic net parameters. If the number of alpha values is larger than 1, cross-validation is conducted over alpha (and lambda). The default is alpha=1, which corresponds to the lasso estimator. The elastic net parameter controls the degree of L1-norm (lasso-type) to L2-norm (ridge-type) penalization. Each alpha value must be in the interval [0,1]. alphacount(real) number of alpha values used for cross-validation across alpha. By default, cross-validation is only conducted across lambda, but not over alpha. Ignored if alpha() is specified. sqrt square-root lasso estimator. adaptive adaptive lasso estimator. The penalty loading for predictor j is set to 1/abs(beta0(j))^theta where beta0(j) is the OLS estimate or univariate OLS estimate if p>n. Theta is the adaptive exponent, and can be controlled using the adatheta(real) option. adaloadings(string) alternative initial estimates, beta0, used for calculating adaptive loadings. For example, this could be the vector e(b) from an initial lasso2 estimation. The elements of the vector are raised to the power -theta (note the minus). See adaptive option. adatheta(real) exponent for calculating adaptive penalty loadings. See adaptive option. Default=1. ols post-estimation OLS. Note that cross-validation using OLS will in most cases lead to no unique optimal lambda (since MSPE is a step function over lambda). ---------------------------------------------------------------------------------------------------------------------------- See overview of estimation methods. Lambda(s) Description ---------------------------------------------------------------------------------------------------------------------------- lambda(numlist) a scalar lambda value or list of descending lambda values. Each lambda value must be greater than 0. If not specified, the default list is used which is given by exp(rangen(log(lmax),log(lminratio*lmax),lcount)) (see mf_range). lcount(integer)† number of lambda values for which the solution is obtained. Default is 100. lminratio(real)† ratio of minimum to maximum lambda. lminratio must be between 0 and 1. Default is 1/1000. lmax(real)† maximum lambda value. Default is 2*max(X'y), and max(X'y) in the case of the square-root lasso (where X is the pre-standardized regressor matrix and y is the vector of the response variable). lopt after cross-validation, estimate model with lambda that minimized the mean-squared prediction error lse after cross-validation, estimate model with largest lambda that is within one standard deviation from lopt lglmnet use the parameterizations for lambda, alpha, standardization, etc. employed by glmnet by Friedman et al. (2010). ---------------------------------------------------------------------------------------------------------------------------- † Not applicable if lambda() is specified. Loadings & standardization Description ---------------------------------------------------------------------------------------------------------------------------- notpen(varlist) sets penalty loadings to zero for predictors in varlist. Unpenalized predictors are always included in the model. partial(varlist) variables in varlist are partialled out prior to estimation. psolver(string) override default solver used for partialling out (one of: qr, qrxx, lu, luxx, svd, svdxx, chol; default=qrxx) ploadings(matrix) a row-vector of penalty loadings; overrides the default standardization loadings (in the case of the lasso, =sqrt(avg(x^2))). The size of the vector should equal the number of predictors (excluding partialled out variables and excluding the constant). unitloadings penalty loadings set to a vector of ones; overrides the default standardization loadings (in the case of the lasso, =sqrt(avg(x^2)). prestd dependent variable and predictors are standardized prior to estimation rather than standardized "on the fly" using penalty loadings. See here for more details. By default the coefficient estimates are un-standardized (i.e., returned in original units). ---------------------------------------------------------------------------------------------------------------------------- See discussion of standardization in the lasso2 help file. Also see Section Data transformations in cross-validation below. FE & constant Description ---------------------------------------------------------------------------------------------------------------------------- fe within-transformation is applied prior to estimation. Requires data to be xtset. noftools do not use FTOOLS package for fixed-effects transform (slower; rarely used) noconstant suppress constant from estimation. Default behaviour is to partial the constant out (i.e., to center the regressors). ---------------------------------------------------------------------------------------------------------------------------- Optimization Description ---------------------------------------------------------------------------------------------------------------------------- tolopt(real) tolerance for lasso shooting algorithm (default=1e-10) tolzero(real) minimum below which coeffs are rounded down to zero (default=1e-4) maxiter(int) maximum number of iterations for the lasso shooting algorithm (default=10,000) ---------------------------------------------------------------------------------------------------------------------------- Fold variable options Description ---------------------------------------------------------------------------------------------------------------------------- nfolds(integer) the number of folds used for K-fold cross-validation. Default is 10. foldvar(varname) user-specified variable with fold IDs, ranging from 1 to #folds. If not specified, fold IDs are randomly generated such that each fold is of approximately equal size. savefoldvar(varname) saves the fold ID variable. Not supported in combination with rolling. rolling uses rolling h-step ahead cross-validation. Requires the data to be tsset. h(integer)‡ changes the forecasting horizon. Default is 1. origin(integer)‡ controls the number of observations in the first training dataset. fixedwindow‡ ensures that the size of the training dataset is always the same. seed(real) set seed for the generation of a random fold variable. Only relevant if fold variable is randomly generated. ---------------------------------------------------------------------------------------------------------------------------- ‡ Only applicable with rolling option. Plotting options Description ---------------------------------------------------------------------------------------------------------------------------- plotcv plots the estimated mean-squared prediction error as a function of ln(lambda) plotopt(varlist) overwrites the default plotting options. All options are passed on to line. ---------------------------------------------------------------------------------------------------------------------------- Display options Description ---------------------------------------------------------------------------------------------------------------------------- omitgrid suppresses the display of mean-squared prediction errors ---------------------------------------------------------------------------------------------------------------------------- Store lasso2 results Description ---------------------------------------------------------------------------------------------------------------------------- saveest(string) saves lasso2 results from each step of the cross-validation in string1, ..., stringK where K is the number of folds. Intermediate results can be restored using estimates restore. ---------------------------------------------------------------------------------------------------------------------------- cvlasso may be used with time-series or panel data, in which case the data must be tsset or xtset first; see help tsset or xtset. All varlists may contain time-series operators or factor variables; see help varlist. Replay syntax cvlasso [, lopt lse postresults plotcv(method) plotopt(string)] Replay options Description ---------------------------------------------------------------------------------------------------------------------------- lopt show estimation results using the model corresponding to lambda=e(lopt) lse show estimation results using the model corresponding to lambda=e(lse) postresults post lasso2 estimation results (to be used in combination with lse or lopt) plotcv(method) see plotting options above plotopt(string) see plotting options above ---------------------------------------------------------------------------------------------------------------------------- Postestimation: predict [type] newvar [if] [in] [, xb u e ue xbu residuals lopt lse noisily] Predict options Description ---------------------------------------------------------------------------------------------------------------------------- xb compute predicted values (the default) residuals compute residuals e generate overall error component e(it). Only after fe. ue generate combined residuals, i.e., u(i) + e(it). Only after fe. xbu prediction including fixed effect, i.e., a + xb + u(i). Only after fe. u fixed effect, i.e., u(i). Only after fe. lopt use lambda that minimized the mean-squared prediction error lse use the largest lambda that is within one standard deviation from lopt noisily displays beta used for prediction. ---------------------------------------------------------------------------------------------------------------------------- Contents Description Partitioning of folds Data transformations in cross-validation cvlasso vs. Friedman et al.'s glmnet and StataCorp's lasso Examples of usage --General demonstration --Rolling cross-validation with time-series data --Rolling cross-validation with panel data Saved results References Website Installation Acknowledgements Citation of lassopack Description cvlasso implements K-fold cross-validation and h-step ahead rolling cross-validation for the following estimators: lasso, square-root lasso, adaptive lasso, ridge regression, elastic net. See lasso2 for more information about these estimators. The purpose of cross-validation is to assess the out-of-sample prediction performance of the estimator. The steps for K-fold cross-validation over lambda can be summarized as follows: 1. Split the data into K groups, referred to as folds, of approximately equal size. Let n(k) denote the number of observations in the kth data partition with k=1,...,K. 2. The first fold is treated as the validation dataset and the remaining K-1 parts constitute the training dataset. The model is fit to the training data for a given value of lambda. The resulting estimate is denoted as betahat(1,lambda). The mean-squared prediction error for group 1 is computed as MSPE(1,lambda)=1/n(1)*sum([y(i) - x(i)'betahat(1,lambda)]^2) for all i in the first data partition. The procedure is repeated for k=2,...,K. Thus, MSPE(2,lambda), ..., MSPE(K,lambda) are calculated. 3. The K-fold cross-validation estimate of the MSPE, which serves as a measure of prediction performance, is CV(lambda)=1/K*sum(MSPE(k,lambda)). 4. Step 2 and 3 are repeated for a range of lambda values. h-step ahead rolling cross-validation proceeds in a similar way, except that the partitioning of training and validation takes account of the time-series structure. Specifically, the training window is iteratively extended (or moved forward) by one step. See below for more details. Partitioning of folds cvlasso supports K-fold cross-validation and cross-validation using rolling h-step ahead forecasts. K-fold cross-validation is the standard approach and relies on a fold ID variable. Rolling h-step ahead cross-validation is applicable with time-series data, or panels with large time dimension. K-fold cross-validation The fold ID variable marks the observations which are used as validation data. For example, a fold ID variable (with three folds) could have the following structure: +------------------+ | fold y x | |------------------| | 3 y1 x1 | | 2 y2 x2 | | 1 y3 x3 | | 3 y4 x4 | | 1 y5 x5 | | 2 y6 x6 | +------------------+ It is instructive to illustrate the cross-validation process implied by the above fold ID variable. Let T denote a training observation and V denote a validation point. The division of folds can be summarized as follows: Step 1 2 3 +- -+ 1 | T T V | 2 | T V T | 3 | V T T | i 4 | T T V | 5 | V T T | 6 | T V T | +- -+ In the first step, the 3rd and 5th observation are in the validation dataset and remaining data constitute the training dataset. In the second step, the validation dataset includes the 2nd and 6th observation, etc. By default, the fold ID variable is randomly generated such that each fold is of approximately equal size. The default number of folds is equal to 10, but can be changed using the nfolds() option. Rolling h-step ahead cross-validation To allow for time-series data, cvlasso supports cross-validation using rolling h-step forecasts (option rolling); see Hyndman, 2016. To use rolling cross-validation, the data must be tsset or xtset. The options h() and origin() control the forecasting horizon and the starting point of the rolling forecast, respectively. The following matrix illustrates the division between training and validation data over the course of the cross-validation for the case of 1-step ahead forecasting (the default when rolling is specified). Step 1 2 3 4 5 +- -+ 1 | T T T T T | 2 | T T T T T | 3 | T T T T T | t 4 | V T T T T | 5 | . V T T T | 6 | . . V T T | 7 | . . . V T | 8 | . . . . V | +- -+ In the first iteration (illustrated in the first column), the first three observations are in the training dataset, which corresponds to origin(3). The option h() controls the forecasting horizon used for cross-validation (the default is 1). If h(2) is specified, which corresponds to 2-step ahead forecasting, the structure changes to: Step 1 2 3 4 5 +- -+ 1 | T T T T T | 2 | T T T T T | 3 | T T T T T | 4 | . T T T T | t 5 | V . T T T | 6 | . V . T T | 7 | . . V . T | 8 | . . . V . | 9 | . . . . V | +- -+ The fixedwindow option ensures that the size of the training dataset is always the same. In this example (using h(1)), each step uses three data points for training: Step 1 2 3 4 5 +- -+ 1 | T . . . . | 2 | T T . . . | 3 | T T T . . | t 4 | V T T T . | 5 | . V T T T | 6 | . . V T T | 7 | . . . V T | 8 | . . . . V | +- -+ Data transformations in cross-validation An important principle in cross-validation is that the training dataset should not contain information from the validation dataset. This mimics the real-world situation where out-of-sample predictions are made not knowing what the true response is. The principle applies not only to individual observations (the training and validation data do not overlap) but also to data transformations. Specifically, data transformations applied to the training data should not use information from the validation data or full dataset. In particular, standardization using the full sample violates this principle. cvlasso implements this principle for all data transformations supported by lasso2: data standardization, fixed effects and partialling-out. In most applications using the estimators supported by cvlasso, predictors are standardized to have mean zero and unit variance. The above principle means that the standardization applied to the training data is based only on observations in the training data; further, the standardization transformation applied to the validation data will also be based only on the means and variances of the observations in the training data. The same applies to the fixed effects transformation: the group means used to implement the within transformation to both the training data and the validation data are calculated using only the training data. Similarly, the projection coefficients used to "partial out" variables are estimated using only the training data and are applied to both the training dataset and the validation dataset. cvlasso vs. Hastie et al.'s (2010) glmnet and StataCorp's lasso The parameterization used by cvlasso and lasso2 differs from StataCorp's lasso in only one respect: lambda(StataCorp) = (1/2N)*lambda(lasso2). The elastic net parameter alpha is the same in both parameterizations. See the lasso2 help file for examples. The parameterization used by Hastie et al.'s (2010) glmnet uses the same convention as StataCorp for lambda: lambda(glmnet) = (1/2N)*lambda(lasso2). However, the glmnet treatment of the elastic net parameter alpha differs from both cvlasso/lasso2 and StataCorp's lasso. The glmnet objective function is defined such that the dependent variable is assumed already to have been standardized. Because the L2 norm is nonlinear, this affects the interpretation of alpha. Specifically, the default cvlasso/lasso2 and StataCorp's lasso parameterization means that alpha is not invariant changes in the scale of the dependent variable. The glmnet parameterization of alpha, however, is scale-invariant - a useful feature. cvlasso and lasso2 provide an lglmnet option that enables the user to employ the glmnet parameterization for alpha and lambda. See the lasso2 help file for examples of its usage and how to replicate glmnet output. We recommend the use of the lglmnet option in particular with cross-validation over alpha; see below for an example. General introduction using K-fold cross-validation Dataset The dataset is available through Hastie et al. (2015) on the authors' website. The following variables are included in the dataset of 97 men: Predictors lcavol log(cancer volume) lweight log(prostate weight) age patient age lbph log(benign prostatic hyperplasia amount) svi seminal vesicle invasion lcp log(capsular penetration) gleason Gleason score pgg45 percentage Gleason scores 4 or 5 Outcome lpsa log(prostate specific antigen) Load prostate cancer data. . insheet using https://web.stanford.edu/~hastie/ElemStatLearn/datasets/prostate.data, clear tab General demonstration 10-fold cross-validation across lambda. The lambda value that minimizes the mean-squared prediction error is indicated by an asterisk (*). A hat (^) marks the largest lambda at which the MSPE is within one standard error of the minimal MSPE. The former is returned in e(lopt), the latter in e(lse). We use seed(123) throughout this demonstration for replicability of folds. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) . di e(lopt) . di e(lse) Estimate the full model Estimate the the full model with either e(lopt) or e(lse). cvlasso internally calls lasso2 with lambda=lopt or lse, respectively. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, lopt seed(123) . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, lse seed(123) The same as above can be achieved using the replay syntax. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) . cvlasso, lopt . cvlasso, lse If postresults is specified, cvlasso posts the lasso2 estimation results. . cvlasso, lopt postres . ereturn list Cross-validation over lambda and alpha alpha() can be a scalar or list of elastic net parameters. Each alpha value must lie in the interval [0,1]. If alpha() is a list longer than 1, cvlasso cross-validates over lambda and alpha. The table at the end of the output indicates the alpha value that minimizes the empirical MSPE. We recommend using the glmnet parameterization of the elastic net because alpha in this parameterization is invariant to scaling (see above for discussion and the lasso2 help file for illustrative examples). . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, alpha(0 0.1 0.5 1) lc(10) lglmnet seed(123) Alternatively, the alphacount() option can be used to control the number of alpha values used for cross-validation. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, alphac(3) lc(10) lglmnet seed(123) Plotting We can plot the estimated mean-squared prediction error over lambda. Note that the plotting feature is not supported if we cross-validate over alpha. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) plotcv Prediction The predict postestimation command allows to obtain predicted values and residuals for lambda=e(lopt) or lambda=e(lse). . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) . cap drop xbhat1 . predict double xbhat1, lopt . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) . cap drop xbhat2 . predict double xbhat2, lse Store intermediate steps cvlasso calls internally lasso2. To see intermediate estimation results, we can use the saveest(string) option. . cvlasso lpsa lcavol lweight age lbph svi lcp gleason pgg45, seed(123) nfolds(3) saveest(step) . estimates dir . estimates restore step1 . estimates replay step1 Time-series example using rolling h-step ahead cross-validation Load airline passenger data. . webuse air2, clear There are 144 observations in the sample. origin() controls the sample range used for training and validation. In this example, origin(130) implies that data up to and including t=130 are used for training in the first iteration. Data points t=131 to 144 are successively used for validation. The notation `a-b (v)' indicates that data a to b are used for estimation (training), and data point v is used for forecasting (validation). Note that the training dataset starts with t=13 since 12 lags are used as predictors. . cvlasso air L(1/12).air, rolling origin(130) The optimal model includes lags 1, 11 and 12. . cvlasso, lopt The option h() controls the forecasting horizon (default=1). . cvlasso air L(1/12).air, rolling origin(130) h(2) In the above examples, the size of the training dataset increases by one data point each step. To keep the size of the training dataset fixed, specify fixedwindow. . cvlasso air L(1/12).air, rolling origin(130) fixedwindow Cross-validation over alpha with alpha={0, 0.1, 0.5, 1}. . cvlasso air L(1/12).air, rolling origin(130) alpha(0 0.1 0.5 1) Plot mean-squared prediction errors against ln(lambda). . cvlasso air L(1/12).air, rolling origin(130) . cvlasso, plotcv Panel data example using rolling h-step ahead cross-validation Rolling cross-validation can also be applied to panel data. For demonstration, load Grunfeld data. . webuse grunfeld, clear Apply 1-step ahead cross-validation. . cvlasso mvalue L(1/10).mvalue, rolling origin(1950) The model selected by cross-validation: . cvlasso, lopt Same as above with fixed size of training data. . cvlasso mvalue L(1/10).mvalue, rolling origin(1950) fixedwindow Saved results cvlasso saves the following in e(): scalars e(N) sample size e(nfolds) number of folds e(lmax) largest lambda e(lmin) smallest lambda e(lcount) number of lambdas e(sqrt) =1 if sqrt-lasso, 0 otherwise e(adaptive) =1 if adaptive loadings are used, 0 otherwise e(ols) =1 if post-estimation OLS, 0 otherwise e(partial_ct) number of partialled out predictors e(notpen_ct) number of not penalized predictors e(prestd) =1 if pre-standardized, 0 otherwise e(nalpha) number of alphas e(h) forecasting horizon for rolling forecasts (only returned if rolling is specified) e(origin) number of observations in first training dataset (only returned if rolling is specified) e(lopt) optimal lambda (may be missing if no unique minimum MSPE) e(lse) lambda se (may be missing if no unique minimum MSPE) e(mspemin) minimum MSPE macros e(cmd) cvlasso e(method) indicates which estimator is used (e.g. lasso, elastic net) e(cvmethod) indicates whether K-fold or rolling cross-validation is used e(varXmodel) predictors (excluding partialled-out variables) e(varX) predictors e(partial) partialled out predictors e(notpen) not penalized predictors e(depvar) dependent variable matrices e(lambdamat) column vector of lambda values functions e(sample) estimation sample In addition, if cvlasso cross-validates over alpha and lambda: scalars e(alphamin) optimal alpha, i.e., the alpha that minimizes the empirical MSPE macros e(alphalist) list of alpha values matrices e(mspeminmat) minimum MSPE for each alpha In addition, if cvlasso cross-validates over lambda only: scalars e(alpha) elastic net parameter matrices e(mspe) matrix of MSPEs for each fold and lambda where each column corresponds to one lambda value and each row corresponds to one fold. e(mmspe) column vector of MSPEs for each lambda e(cvsd) column vector standard deviation of MSPE (for each lambda) e(cvupper) column vector equal to MSPE + 1 standard deviation e(cvlower) column vector equal to MSPE - 1 standard deviation References Correia, S. 2016. FTOOLS: Stata module to provide alternatives to common Stata commands optimized for large datasets. https://ideas.repec.org/c/boc/bocode/s458213.html Hyndman, Rob J. (2016). Cross-validation for time series. Hyndsight blog, 5 December 2016. https://robjhyndman.com/hyndsight/tscv/ See lasso2 for further references. Website Please check our website https://statalasso.github.io/ for more information. Installation cvlasso is part of the lassopack package. To get the latest stable version of lassopack from our website, check the installation instructions at https://statalasso.github.io/installation/. We update the stable website version more frequently than the SSC version. Earlier versions of lassopack are also available from the website. To verify that lassopack is correctly installed, click on or type whichpkg lassopack (which requires whichpkg to be installed; ssc install whichpkg). Acknowledgements Thanks to Sergio Correia for advice on the use of the FTOOLS package. Citation of cvlasso cvlasso is not an official Stata command. It is a free contribution to the research community, like a paper. Please cite it as such: Ahrens, A., Hansen, C.B., Schaffer, M.E. 2018 (updated 2020). LASSOPACK: Stata module for lasso, square-root lasso, elastic net, ridge, adaptive lasso estimation and cross-validation http://ideas.repec.org/c/boc/bocode/s458458.html Ahrens, A., Hansen, C.B. and M.E. Schaffer. 2020. lassopack: model selection and prediction with regularized regression in Stata. The Stata Journal, 20(1):176-235. https://journals.sagepub.com/doi/abs/10.1177/1536867X20909697. Working paper version: https://arxiv.org/abs/1901.05397. Authors Achim Ahrens, Public Policy Group, ETH Zurich, Switzerland achim.ahrens@gess.ethz.ch Christian B. Hansen, University of Chicago, USA Christian.Hansen@chicagobooth.edu Mark E. Schaffer, Heriot-Watt University, UK m.e.schaffer@hw.ac.uk Also see Help: lasso2, lassologit, rlasso (if installed)
help cvlasso