What's the bottom
line? How to compare models
Testing the assumptions of linear regression
Additional notes on regression analysis
Spreadsheet with regression formulas
Stepwise and all-possible-regressions
If you are a PC Excel user, you must check this out:
Stepwise and all-possible-regressions
Stepwise regression is a semi-automated process of building a model by successively adding or removing variables based solely on the t-statistics of their estimated coefficients. Properly used, the stepwise regression option in Statgraphics (or other stat packages) puts more power and information at your fingertips than does the ordinary multiple regression option, and it is especially useful for sifting through large numbers of potential independent variables and/or fine-tuning a model by poking variables in or out. Improperly used, it may converge on a poor model while giving you a false sense of security. It's like doing carpentry with a chain saw: you can get a lot of work done quickly, but it leaves rough edges and you may end up cutting off your own foot if you don't read the instructions, remain sober, engage your brain, and keep a firm grip on the controls. It is not a tool for beginners or a substitute for education and experience.
How it works: Suppose you have some set of potential independent variables from which you wish to try to extract the best subset for use in your forecasting model. (These are the variables you will select on the initial input screen.) The stepwise option lets you either begin with no variables in the model and proceed forward (adding one variable at a time), or start with all potential variables in the model and proceed backward (removing one variable at a time). At each step, the program performs the following calculations: for each variable currently in the model, it computes the t-statistic for its estimated coefficient, squares it, and reports this as its "F-to-remove" statistic; for each variable not in the model, it computes the t-statistic that its coefficient would have if it were the next variable added, squares it, and reports this as its "F-to-enter" statistic. At the next step, the program automatically enters the variable with the highest F-to-enter statistic, or removes the variable with the lowest F-to-remove statistic, in accordance with certain control parameters you have specified. So the key relation to remember is: F = t-squared
In the multiple regression procedure in most statistical software packages, you can choose the stepwise variable selection option and then specify the method as "Forward" or "Backward," and also specify threshold values for F-to-enter and F-to-remove. (You can also specify "None" for the method--which is the default setting--in which case it just performs a straight multiple regression using all the variables.) The program then proceeds automatically. Under the forward method, at each step, it enters the variable with the largest F-to-enter statistic, provided that this is greater than the threshold value for F-to-enter. When there are no variables left to enter whose F-to-enter statistics are above the threshold, it checks to see whether the F-to-remove statistics of any variables added previously have fallen below the F-to-remove threshold. If so, it removes the worst of them, and then tries to continue. It finally stops when no variables either in or out of the model have F-statistics on the wrong side of their respective thresholds. The backward method is similar in spirit, except it starts with all variables in the model and successively removes the variable with the smallest F-to-remove statistic, provided that this is less than the threshold value for F-to-remove.
Whenever a variable is entered, its new F-to-remove statistic is initially the same as its old F-to-enter statistic, but the F-to-enter and F-to-remove statistics of the other variables will generally all change. (Similarly, when a variable is removed, its new F-to-enter statistic is initially the same as its old F-to-remove statistic.) Until the F-to-enter and F-to-remove statistics of the other variables are recomputed, it is impossible to tell what the next variable to enter or remove will be. Hence, this process is myopic, looking only one step forward or backward at any point. (Return to top of page)
There is no guarantee that the best model that can be constructed from the available variables (or even a good model) will be found by this one-step-ahead search procedure. Hence, when the procedure terminates, you should study the sequence in which variables were added and deleted (which is usually a part of the output), think about whether the variables that were included or excluded make sense, and ask yourself if perhaps the addition or removal of a few more variables might not lead to improvement. For example, the variable with the lowest F-to-remove or highest F-to-enter may have just missed the threshold value, in which case you may wish to tweak the F-values and see what happens. Sometimes adding a variable with a marginal F-to-enter statistic, or removing one with a marginal F-to-remove statistic, can cause the F-to-enter statistics of other variables not in the model to go up and/or or the F-to-remove statistics of other variables in the model to go down, triggering a new chain of entries or removals leading to a very different model.
While you're studying the sequence of variables entered or removed, you should also watch the value of the adjusted R-squared of the model, which is one of the statistics shown. Usually it should get consistently larger as the stepwise process works its magic, but sometimes it may start getting smaller again. In this case you should make a note of which variables were in the model when adjusted R-squared hit its largest value--you may wish to return to this model later on by manually entering or removing variables. (Return to top of page)
Warning #1: For all the models traversed in the same stepwise run, the same data sample is used, namely the set of observations for which all variables listed on the original input screen have non-missing values, because the stepwise algorithm uses a correlation matrix calculated in advance from the list of all candidate variables. (More about this below.) Therefore, be careful about including variables which have many fewer observations than the other variables, such as seasonal lags or differences, because they will shorten the test period for all models whether they appear in them or not, and regardless of whether "forward" or "backward" mode is used. After selecting your final model, you may wish to return to the original input panel, erase the names of all variables that weren't used in the final model, then re-fit the model to be sure that the longest possible estimation period was used.
Warning #2: If the number of variables that you select for testing is large compared to the number of observations in your data set (say, more than 1 variable for every 10 observations), or if there is excessive multicollinearity (linear dependence) among the variables, then the algorithm may go crazy and end up throwing nearly all the variables into the model, especially if you used a low F-to-enter or F-to-remove threshold.
Warning #3: Sometimes you have a subset of variables that ought to be treated as a group (say, dummy variables for seasons of the year) or which ought to be included for logical reasons. Stepwise regression may blindly throw some of them out, in which case you should manually put them back in later.
Warning #4: Remember that the computer is not necessarily right in its choice of a model during the automatic phase of the search. Don't accept a model just because the computer gave it its blessing. Use your own judgment and intuition about your data to try to fine-tune whatever the computer comes up with.
Warning #5: Automated regression model selection methods only look for the most informative variables from among those you start with, in the limited context of a linear prediction equation, and they cannot make something out of nothing. If you have insufficient quantity or quality of data, or if you omit some important variables or fail to use data transformations when they are needed, or if the assumption of linear or linearizable relationships is simply wrong, no amount of searching or ranking will compensate. The most important steps in statistical analysis are (a) doing your homework before you begin, and (b) collecting and organizing the relevant data.
See this page for more details of the dangers and deficiencies of stepwise regression.
What method should you use: forward or backward? If you have a very large set of potential independent variables from which you wish to extract a few--i.e., if you're on a fishing expedition--you should generally go forward. If, on the other hand, if you have a modest-sized set of potential variables from which you wish to eliminate a few--i.e., if you're fine-tuning some prior selection of variables--you should generally go backward. (If you're on a fishing expedition, you should still be careful not to cast too wide a net, lest you dredge up variables that are only accidentally related to your dependent variable.)
What values should you use for the F-to-enter and F-to-remove thresholds? As noted above, after the computer completes a forward run based on the F-to-enter threshold, it usually takes a backward look based on the F-to-remove threshold, and vice versa. Hence, both thresholds come into play regardless of which method you are using, and the F-to-enter threshold must be greater than or equal to the F-to-remove threshold (to prevent cycling). Usually the two thresholds are set to the same value. Keeping in mind that the F-statistics are squares of corresponding t-statistics, an F-statistic equal to 4 would correspond to a t-statistic equal to 2, which is the usual rule-of-thumb value for "significance at the 5% level." (4 is the default value for both thresholds.) I recommend using a somewhat smaller threshold value than 4 for the automatic phase of the search--for example 3.5 or 3 (but not less than that). Since the automatic stepwise algorithm is myopic, it is usually OK to let it enter a few too many variables in the model, and then you can weed out the marginal ones later on by hand. However, beware of using too low an F threshold if the number of variables is large compared to the number of observations, or if there is a problem with multicollinearity in your data (see warning #2 above). Often this opens the gates to a horde of spurious regressors--and in any case you should manually apply your usual standards of relevance and significance to the variables in the model at the end of the run. Don't just blindly accept the computer's choice. (Return to top of page)
Just in case you're curious about how it's done: At each step in the stepwise process, the program must effectively fit a multiple regression model to the variables in the model in order to obtain their F-to-remove statistics, and it must effectively fit a separate regression model for each of the variables not in the model in order to obtain their F-to-enter statistics. When watching all this happen almost instantaneously on your computer, you may wonder how it is done so fast. The secret is that it doesn't have to fit all these models from scratch, and it doesn't need to reexamine all the observations of each variable. Instead, the stepwise search process can be carried out merely by performing a sequence of simple transformations on the correlation matrix of the variables. The variables are only read in once, and their correlation matrix is then computed (which takes only few seconds even if there are very many variables). After this, the sequence of adding or removing variables and recomputing the F-statistics requires only a simple updating operation on the correlation matrix. This operation is called "sweeping," and it is similar to the "pivoting" operation that is at the heart of the simplex method of linear programming, if that means anything to you. The computational simplicity of the stepwise regression algorithm re-emphasizes the fact that, in fitting a multiple regression model, the only information extracted from the data is the correlation matrix of the variables and their individual means and standard deviations. The same computational trick is used in all-possible-regressions. (Return to top of page)
Stepwise regression often works reasonably well as an automatic variable selection method, but this is not guaranteed. Sometimes it will take a wrong turn and get stuck in a suboptimal region of model space, and sometimes the model it selects will be just one out of a number of almost-equally-good models that ought to be studied together. All-possible-regressions goes beyond stepwise regression and literally tests all possible subsets of the set of potential independent variables. (This is the "Regression Model Selection" procedure in Statgraphics.) If there are K potential independent variables (besides the constant), then there are 2K distinct subsets of them to be tested (including the empty set which corresponds to the mean model). For example, if you have 10 candidate independent variables, the number of subsets to be tested is 210, which is 1024, and if you have 20 candidate variables, the number is 220, which is more than one million. The former analysis would run almost instantly on your computer, the latter might take a few minutes, and with 30 variables it might take hours. All-possible-regressions carries all the caveats of stepwise regression, and more so. This kind of data-mining is not guaranteed to yield the model which is truly best for your data, and it may lead you to get absorbed in top-10 rankings instead of carefully articulating your assumptions, cross-validating your results, and comparing the error measures of different models in real terms.
When using an all-possible-regressions procedure, you are typically given the choice between several numerical criteria on which to rank the models. The two most commonly used are adjusted R-squared and the Mallows "Cp" statistic. The latter statistic is related to adjusted R-squared, but includes a heavier penalty for increasing the number of independent variables. Cp is not measured on a 0-to-1 scale. Rather, its values are typically positive and greater than 1, and lower values are better. The models which yield the best (lowest) values of Cp will tend to be similar to those that yield the best (highest) values of adjusted R-squared, but the exact ranking may be slightly different. Other things being equal, the Cp criterion tends to favor models with fewer parameters, so it is a bit less likely to overfit the data. Generally you look at the plots of R-squared and Cp versus the number of variables to see (a) where the point of diminishing returns is reached in terms of the number of variables, and (b) whether there are one or two models that stand out above the crowd, or whether there are many almost-equally-good models. Then you can look at the the actual rankings of models and try to find the optimum place to make the "cut".
As with any ranking scheme, it's easy to get lost in the trees and lose sight of the forest: the differences in performance among the models near the top of the rankings may not be substantial. (Don't forget that there are dozens, hundreds, or sometimes thousands of models down below!) An improvement in R-squared from, say, 75% to 76% is probably not worth increasing the complexity of the model by adding more independent variables that are not otherwise well-motivated. In fact, it would only reduce the standard deviation of the errors by 2% (as discussed here), which would not noticeably shrink the confidence intervals for forecasts. And there's a very real danger that automated data-mining will lead to the selection of a model which lacks an intuitive explanation and/or performs poorly out-of-sample.
Among the various automatic model-selection methods, I find that I generally prefer stepwise to all-possible regressions. The stepwise approach is much faster, it's less prone to overfit the data, you often learn something by watching the order in which variables are removed or added, and it doesn't tend to drown you in details of rankings data that cause you to lose sight of the big picture.
There are other methods of parameter estimation and variable selection such as ridge regression and lasso regression that are designed to deal with situations in which the candidate independent variables are highly correlated with each and/or their number is large relative to the sample size, but those methods are beyond the scope of this discussion.