19
$\begingroup$

I'm getting really interested in the elastic net procedure for predictor shrinkage/selection. It seems very powerful.

But from the scientific point of view I don't know well what to do once I got the coefficients. What question am I answering? These are the variables that most influence that outcome and these are the coefficients which give the best variance/bias ratio during validation?

This is of course a very descriptive/predictive approach compared to the classical p value/confidence intervals approach. Inferential estimation is being studied now by Tibshirani & Co. but is still experimental.

Some people are using the variables chosen by elastic net to perform classical inferential analysis, but that would eliminate the limitation in variance brought by the technique.

Another problem is that since lambda and alpha parameters for elastic net are chosen by cross validation they are subject to random variability. So every time you run (eg.) cv.glmnet() you will select a slightly different subset of predictors with always different coefficients.

I though about solving this considering the right lambda and alpha as random variables and re run the cross validation step n times to get a distribution of these parameters. This way for every predictor I would have the number of occurrences and for every coefficients I would have distribution of results. This should give me more generalizable results with ranges statistics (like sd of the coefficients). It would also be interesting to see whether the lambda and the alpha picked this way approximate to some distribution asymptotically, since that would open up the way for some inference test (but I'm not a statistician so I should not speak about things I don't fully understand).

So finally my question is: Once you get the predictors and the coefficients from an elastic net with cross validation based alpha and lambda, which and how should you present these results? How should you discuss them? what did we learn? Which hypothesis/generalization are we confuting?

$\endgroup$
  • $\begingroup$ I think this is overly broad/unclear to answer appropriately. In some cases I find your statements unclear (eg. what do you mean by "but that would eliminate the limitation in variance brought by the technique.") and on some other cases misled (eg. "every time you run (eg.) cv.glmnet() you will select a slightly different subset of predictors with always different coefficients" - that's not the case every time and even when it happens usually it is not catastrophic given CV was done correctly.) $\endgroup$ – usεr11852 Jan 15 '17 at 1:09
  • $\begingroup$ a motivation I've seen of the elastic net related it to variable clustering (through section 2.3 of the zou, hastie elastic net paper), which is expanded upon in more detail (through a somewhat different method) here: ncbi.nlm.nih.gov/pmc/articles/PMC4011669 $\endgroup$ – user795305 Jan 19 '17 at 4:42
8
$\begingroup$

These methods--the lasso and elastic net--were born out of the problems of both feature selection and prediction. It's through these two lenses that I think an explanation can be found.

Matthew Gunn nicely explains in his reply that these two goals are distinct and often taken up by different people. However, fortunately for us, the methods we're interested in can perform well in both arenas.

Feature Selection

First, let's talk about feature selection. We should first motivate the elastic net from the perspective of the lasso. That is, to quote Hastie and Zou, "If there is a group of variables among which the pairwise correlations are very high, then the lasso tends to select only one variable from the group and does not care which one is selected." This is a problem, for instance, because it means that we're not likely to find an element of the true support using the lasso--just one highly correlated with it. (The paper mentions that this is proven in the LARS paper, which I haven't read yet.) The difficulty of support recovery in the presence of correlation is also pointed out by Wainwright, who showed (in theorem 2a) that the probability of support recovery is bounded above by $0.5$ when there's high correlation between the true support and it's complement.

Now, the l2 penalty in the elastic net encourages features which have coefficients treated as indistinguishable by just the loss and l1 penalty to have equal estimated coefficient. We can loosely see this by noticing that $(a,b) = \arg\min_{a',b': c = |a'| + |b'|} (a')^2 + (b')^2$ satisfies $|a| = |b|$. Due to this, the elastic net makes it so that we're less likely to 'accidentally' make vanish a coefficient estimate which is in the true support. That is, the true support is more likely to be contained within the estimated support. That's good! It does mean there's more false discoveries, but that's a price that most people are willing to pay.

As an aside, it's worth pointing out that the fact that highly correlated features will tend to have very similar coefficient estimates makes it so that we can detect groupings of features within the estimated support which influence the response similarly.

Prediction

Now, we move on to prediction. As Matthew Gunn points out, choosing tuning parameters through cross validation creates an aim to choose a model with minimal prediction error. Since any model selected by the lasso can be selected by the elastic net (by taking $\alpha = 1$), it makes some sense that the elastic net is able to find a model that predicts better than the lasso.

Lederer, Yu, and Gaynanova show, under no assumptions whatsoever on the features, that the lasso and elastic net can both have their l2 prediction error bounded by the same quantity. It's not necessarily true that their bound is tight, but this might be interesting to note since oracle inequalities seem to be a standard way in statistical literature to quantify the predictive performance of estimators--perhaps since the distributions are so complicated! It's also worth noting that Lederer (1)(2) has some papers on lasso predictions in the presence of correlated features.

Summary

In summary, the problems of interest are the true support being within the estimated support and prediction. For support recovery, there's rigorously proven guarantees (through Wainwright) that the lasso selects the correct features to be in the model under assumptions of low correlation between the true support and it's complement. However, in the presence of correlation, we can fall back to the elastic net to be more likely to select the features in the true support to be among all that it selects. (Note that we have to carefully select the tuning parameters here.) And, for prediction when we choose the tuning parameter through cross validation, it makes intuitive sense that the elastic net should perform better than the lasso--especially in the presence of correlation.

Putting aside prediction and some formality, what did we learn? We learned about the true support.

Confidence Intervals

It's worth pointing out that a lot has changed in the past 2 years in regards to valid inference for the lasso. In particular, the work by Lee, Sun, Sun, and Taylor provides exact inference for the coefficients of the lasso conditional on the given model being selected. (Results on inference in lasso for the true coefficients was around at the time of OP's post, and they're well summarized in the linked paper.)

$\endgroup$
  • $\begingroup$ Would it be correct to assume that regularized covariates estimates are probably more similar to those we could find repeating a study? That is, as regularization help minimizing the out of sample prediction error, it could help minimizing the difference from in sample and out of sample estimation? $\endgroup$ – Bakaburg Feb 19 '17 at 18:25
  • 1
    $\begingroup$ @Bakaburg, yeah, that makes sense to say. The regularization creates estimators with lower variance. $\endgroup$ – user795305 Feb 24 '17 at 15:43
9
+25
$\begingroup$

What you're doing with elastic, ridge, or lasso, using cross-validation to choose regularization parameters, is fitting some linear form to optimize prediction. Why these particular regularization parameters? Because they work best for prediction on new data. Shrinking coefficient estimates towards zero, introducing bias, (as is done in either Ridge or Lasso) can reduce overfitting and shrink variance. The idea is for your penalty parameters to strike the right balance in order to optimize prediction on new data.

Imagine the data generating process is:

$$ y_i = f(\mathbf{x}_i, \beta) + \epsilon_i$$

Let $\hat{\beta}$ be our estimate of parameters $\beta$, and let $\hat{y}_j$ be our forecast for observation $j$

How should you present your results? It depends what your underlying research question is! You may want to step back and think deeply about what question you're trying to answer. What does your audience care about? What are you trying to do?

  • Prediction?
  • Estimate coefficients?
  • Variable selection?

It's important to distinguish between two types of research questions:

  1. Questions where you predominantly care about prediction, that is you care about $\hat{y}_j$
  2. Questions where you predominantly care about parameter estimates $\hat{\beta}$.

Off the shelf machine learning techniques can be extremely powerful for the former, $\hat{y}$ prediction problems. As you appear to be recognizing though, standard off the shelf machine learning techniques may be extremely problematic for $\hat{\beta}$, parameter estimate problems:

  • In a high dimensional setting, many different parameterization will give you the same predictions $\hat{y}$. If number of parameters $k$ is high relative to the number of observations $n$, you may not be able to estimate any individual parameter well.
  • Algorithms trained on different folds may have significantly different parameter estimates.
  • The emphasis in machine learning is on prediction, not consistently estimating causal effects. (This contrasts with econometrics where typically the main issue is in consistently estimating causal effects). Prediction, estimating some functional form, is different than estimating causation. Police levels may be a good predictor of crime levels, and this doesn't mean police cause crime.

And as you recognize, there may be issues in interpreting why some machine learning parameterization works. Is your audience comfortable with a prediction black box? Or is how prediction works central to your question?

Lasso and Ridge: classic reasons to use them

  • You can use elastic net for classic machine learning, prediction problems, situations where your main concern is $\hat{y}$. In some sense regularization allows you to include more predictors but still keep overfitting under control.

  • You can use regularization to prevent overfitting. Eg. ridge regression in the context of polynomial curve fitting can work quite nicely.

  • As @Benjamin points out in his answer, Lasso can also be used for variable selection. Under certain regularity conditions, Lasso will consistently select the appropriate model: irrelevant coefficients will be set to zero.

The $L_1$ and $L_2$ penalties, of Lasso and Ridge respectively, bias the coefficient estimates toward zero. If the bias is big, this could be a serious issue if you're trying to interpret coefficient estimates. And to get standard error estimates, you need to do something like bootstrapping; there aren't simple closed form solutions (that I'm aware of). Ridge, lasso, and elastic net have similarities to regular OLS regression, but the regularization and variable selection make inference quite different...

What I keep coming back to is that it's quite difficult to interpret the results of running ridge regression, lasso, or elastic net without some more context of what you're trying to figure out!


Prof. Sendhil Mullainathan gave a talk on machine learning at the January, 2017 AFA meeting which motivated parts of this post.

$\endgroup$
  • 3
    $\begingroup$ This kind of thinking is flawed in my opinion. It is based on the assumption that the underlying phenomenon is simple enough to be comprehended by a human being. High dimensional models are most of the time too complex to be comprehended by humans, but they are very suitable for large scale artificial intelligence. In reality the best predictor is the best interpretation of the phenomenon, whether you can comprehend it or not. $\endgroup$ – Cagdas Ozgenc Jan 19 '17 at 5:56
  • 2
    $\begingroup$ @CagdasOzgenc I think that's a valid point that some functions are hideously complex, difficult to describe to humans but understandable and learnable by machines (eg. chess board evaluation). In these situations, it may be better to throw up your hands, not even try to interpret what the machine learned. On the other hand, there are situations like drug trials where there's a causal effect, some average effectiveness you're trying to estimate in the presence of a multitude of confounders, selection effects etc... These are in some sense different problems and need different techniques. $\endgroup$ – Matthew Gunn Jan 19 '17 at 6:13
  • 1
    $\begingroup$ @Benjamin An underlying problem is that what the OP is most directly asking for, an understandable interpretation of the biased towards zero coefficients from elastic net, may not exist. Imagine you have 10,000 predictors and 5,000 observations. Jointly, your coefficients may do an excellent job at prediction, but individually, each coefficient may be poorly estimated junk. I think it's worth taking a step back and asking what's the underlying research question? What's the objective? Is it finding predictions $\hat{y}$ or estimating some coefficient? Or perhaps something else? $\endgroup$ – Matthew Gunn Jan 19 '17 at 22:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.