The most natural use of multiple regression is when all the variables concerned are continuous. It has been shown by "Monte Carlo" methods (i.e. trying it out with random numbers) that it is not badly affected if either independent or dependent variables are only ordinally measured (e.g. rating scales), so long as the number of categories is not too small.. What happens, though, when we have to deal with nominal measurement, i.e. where the numbers we have are merely labels for categories? If the dependent variable is categorical, we can't use multiple regression at all. However, unordered categories as independent variables can be accommodated.
When an independent variable consists of just two categories, e.g. male
vs female, there is no problem; we have already seen how to include such
dichotomous variables in a regression analysis. Where we have three
or more categories in a variable, however, it would clearly be a serious
error just to include such a variable in a regression analysis directly.
Instead we have to use what are called dummy variables.
Suppose that the variable includes m categories (m>2). What we have to do is to break down this multi-way category into m 2-way categories, each indicating whether or not the observation belongs to a particular one of the m original categories. So suppose a variable called 'faculty' contains codes 1-6 indicating which of the 6 faculties an Exeter student belongs to, assigning the codes in alphabetical order of faculty name. We have to create 6 new variables, one for each faculty, which will indicated whether the student belongs to that faculty. So the values of the variables for the first few cases might look like this:
faculty arts educatio engineer law science socstuds 2 0 1 0 0 0 0 3 0 0 1 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0 0 0 4 0 0 0 1 0 0 6 0 0 0 0 0 1
The new variables are called dummy variables. In effect we have replaced the one 6-way categorical variable 'faculty' with six dichotomous variables, 'arts' to 'socstuds'. We have already seen that we can include dichotomous variables in a regression analysis, so our problem is essentially solved. There are however a couple of loose ends to tie up before we consider how we can make the transformation in practice.
Return to top
Going from a categorical variable to the corresponding set of dummy variables is easy in Minitab: there is a command, INDICATOR, which simply takes a column and produce dummy variables for us. So if we have a 6-way categorical variable stored in C1, we can simply type
MTB> INDICATOR C1 C11-C16
We shall have to give the new variables the appropriate names by hand. Things are not quite so easy in SPSS; here we have to write a short loop in our command file, for example:
do repeat x=arts educatio engineer law science socstuds / i=1 to 6.
if (faculty=i) x=1.
if missing(faculty) x=9.
missing values arts to socstuds (9).
Note that we have to take explicit action to look after missing values.
This is straightforward in either Minitab (use the TALLY command, or TABLE with a single argument) or SPSS (use FREQUENCIES).
In either SPSS or Minitab, the dummy variables can simply be entered in the regression command just like any other variable. Note, though that the original categorical variable should never appear in the regression command.
Testing for the overall effect of the categorical variable is done by entering the set of dummy variables (excluding, of course, the control or modal category) after having entered the remaining independent variables. Suppose we are assessing the effect of A-level score (in points), gender, parental income and faculty on undergraduates' rated satisfaction with their courses at Exeter. Then we might have an SPSS command file like the following (assuming that the dummy variables have been formed up before we saved the system file, and that Arts is the modal faculty):
title satisfaction regression.
variables=ALevels,gender,parincom,satirate,arts to socstuds
/method=enter arts to science.
SPSS will give us an F value for the effects of the group of variables added on each ENTER command, which is exactly what we want.
The principle is the same in Minitab, but it requires more work. We have to be careful to specify the dummy variables last in the list of regressors for the REGRESS command. Then we look at the column headed SEQ SS that comes under the anova table in multiple regression output, and add up the entries for all the m-1 categories. The total can be described as the sum of squares for the categorical variable as a whole. Divide this by m-1, and we shall have the mean square for the categorical variable. This in turn can be divided by the error mean square from the anova table, to give an F statistic which will allow us to test the significance of the entire categorical variable. Remember that F always has two values of degrees of freedom associated with it: in this case, the numerator degrees of freedom are m-1 and the denominator degrees of freedom are the error degrees of freedom from the anova table. Note that Minitab includes the LET command, which allows us to do simple arithmetic, and also constants (called K1, K2 etc to distinguish them from columns) in which we can store the results. This means we can get Minitab to do the arithmetic, by typing lines like
We can then use another Minitab command, CDF (which stands for Cumulative Distribution Function) to look up the significance of the F value we obtain (though, irritatingly, CDF reports the complement of a significance level, i.e. 1-p rather than p). Use Minitab's HELP facility to find out how to use CDF.
Return to top
Up to now, we have assumed that we have a fixed set of independent variables, whose effect on the dependent variable we want to describe and subject to significance tests. However, quite often we have a large set of possible independent variables, and we are interested in finding which subset of them are most useful for predicting the dependent variable. This question divides into two parts: deciding exactly what "most useful" means in this context, and having operationalised that into a precise criterion, finding the set of regressors which best meets this criterion.
A subset of the the possible regressors is usually referred to as a "regression model", because in effect choosing a subset of regressors is specifying a hypothesis, or model, of the variables that are associated with the dependent variable.
If we are choosing between models which all involve the same number of regressors, it is easy to say which is the best: it is the model with the highest value of R2adj. However, more usually, we are choosing between models with different numbers of regressors - for example, we want to know whether it is worth adding an additional regressor or group of regressors to a model which already does a reasonable job at explaining the variation in the data. Here things become more complicated, because, as we already know, adding additional variables means that we are bound to account for more variance, but obviously makes for a less economical model - the issue is whether the gain in prediction is worth the loss in economy. Two criteria are commonly used to assess this issue. We could choose the model with the highest value of R2adj, or we could choose the model with the highest value of F. There are arguments in favour of either policy. The model with the highest R2adj will almost always have more regressors in it than the model with the highest F. Therefore, a reasonable rule is: If you want the most complete description of the data, which is also reasonably efficient, choose the model with the highest R2adj value. If you want the most efficient description, which is also reasonably complete, choose the model with the highest F. Different research projects will call for different strategies.
For purely exploratory analyses, we find the best model by either stepwise or setwise regression. A scenario intermediate between strict hypothesis testing and exploration is that where we know we must consider certain independent variables, but we want to know whether other variables have any effect over and above these "compulsory" regressors. In this case we will use some form of hierarchical regression. Hierarchical and stepwise regression are available in all packages, and are widely used. Setwise regression is in principle more satisfactory, but requires much more computing power, and has only recently come on the scene. It is still not available in SPSS; it is available in Minitab, but restricted to sets of 20 or fewer independent variables.
More formally, what the three procedures do is the following:
If there were no correlation between the regressors, there would never be a problem in deciding which of two regressors was the more important, and therefore in settling on a single best regression model. Usually, though, we are using multiple regression precisely because regressors are correlated, and this leaves us with problems of interpretation. Two models using different sets of regressors can have very similar R2adj and/or F values. Which is really the best model?
Ultimately, if two independent variables are closely correlated (or one is well predicted by two others), no statistical procedure will enable us to say which predicts a dependent variable better - both will do an equally good job. This is called the identification problem. It is a scientific rather than a statistical problem, and it can only be resolved by collecting more data. Usually we will have to change our sampling procedure so as to break down the offending correlation. Suppose we find, in a study of absenteeism among women workers, that parents of pre-school children and part-time workers have high absence rates, but we cannot tell which is the crucial variable, because most of our part-timers have pre-school children and vice versa. We shall have to go out of our way to collect some data from childless part-timers or full-timers who have pre-school children, by oversampling those combinations.
Return to top
Hierarchical regression involves no new techniques. The other two require illustration. You can do stepwise regression in Minitab or SPSS, but SPSS does it better. Setwise regression can only be done in Minitab.
This is very straightforward. You just substitute
for the /method=enter subcommand we have used before. The output will be long, but its interpretation is fairly obvious. One way of testing whether you are falling into local maximum problems is to follow up by doing the same regression but using
This will do backwards stepwise regression, which starts from the model involving all the regressors and then tries to take them out one at a time. If forwards and backwards regression arrive at the same endpoint, it is unlikely to be a local maximum.
SPSS looks for the model with the best F value. However, because it shows you a lot of its intermediate calculations, you can usually identify the model with the best R2adj value, because at each number of regressors, the model with the best F value is also the model with the best R2adj value. It is only when we are comparing models involving different numbers of regressors that the two criteria diverge.
SPSS will report all the usual significance tests for the model it identifies as best. But you should note that choosing a model in this way biases the procedure in favour of producing large R2adj values, and so undermines the logic of significance testing - so it isn't very surprising if the best regression model is reported as significant. Significance testing is only really appropriate when we have specified exact null and alternative hypotheses in advance. So for the purposes of statistical inference, any exploratory regression should be followed up by a test of the chosen regression model on an independent set of data. Some researchers would argue for doing the exploratory regression on a randomly chosen half of the data you have collected, and then following up with a test regression on the other half; others would argue that a completely independent data set should be collected.
This is also straightforward. It uses a new command, BREG (standing for Best Regression). Using BREG is easy. If the dependent variable was in C10, and you had 6 independent variables in C1-C6, you would just type
BREG C10 C1-C6
and Minitab would do the rest. Notice that you don't have to tell BREG
how many regressors there are (though it won't matter if you do).
The output comes in a nice compact form, though you will need to think about it a bit to see how to interpret it. It tells you which regressors are included in the best two models for each possible number of regressors. It also tells you their R2adj values, so you can pick out the best model of all in R2adj terms directly. It doesn't tell you their F values, so if you want the best model in F terms, you will have to use the REGRESS command on the best-R2adj model for each number of regressors, and look to see which one has the best F.
Even if you are interested in the best-R2adj model, you should proceed to use REGRESS on the set of regressors it identifies, so you can find out the values of the coefficients and their significance. You'll also need the F value to assess the significance of the model as a whole, though the same cautionary note applies here as to stepwise regression.
Return to top
It should be clear by now that there is no one best statistics packages
for all purposes. This means that we often need to move a data set between
packages. In general packages will not read each other's private data files:
so Minitab cannot read SPSS system files, and SPSS cannot read Minitab
worksheet files, for example. The usual way to move data from one
package to another is via a text or ascii file.
A text file contains information in a very simple, standard code which can be interpreted by a wide variety of programs; the code most often used is called ASCII, but text files don't use the full list of 256 ASCII codes. The list of symbols allowed in text files varies a bit between programs, but you can rely on being allowed the 26 letters of the English alphabet in both capitals and lower case; the digits 0-9; some but not all punctuation symbols; and some but not all mathematical symbols. In addition you will always be allowed the control code ENTER (used to mark ends of lines). The advantage of text files is that almost any program and almost any computer can use them. So they are used for moving data between one computer and another, as well as between or one program and another. For example, we can prepare a data file in a word processor on a Macintosh, output it as a text file, transfer that to singer, display it on the screen, edit it, or send it to a printer, and read it into Minitab or SPSS for statistical work. Note, though, that both SPSS and Minitab can also produce what are called portable files, which can be used for moving between versions of the same package on different computers: so if you wanted to move data from a Macintosh Minitab worksheet to singer Minitab, you would use a portable worksheet file rather than a text file.
To produce a text file from Minitab, use command WRITE, for example
WRITE 'filename' C1-C10
WRITE without a filename writes to the screen so you can see what the layout looks like. The subcommand FORMAT can be used if you want to write more columns than will comfortably fit on a line, though it is not very easy to use unless you know the Fortran programming language. If you don't provide an extension as part of the filename, WRITE will add .DAT to the name.
The command to produce a text file from SPSS is also called WRITE. It must always be followed by the command EXECUTE; forgetting this is a common and very irritating error. The FORMATS command can be used to vary the output format of each column, and numbers can be given to space the columns out. Life will be much easier at the Minitab end if we use these facilities to make sure that all the variables have some blank space between them:
title writing a file to send to Minitab.
formats index (f3) free1 to law5 (f2).
write outfile=taxasci.DAT / index free1 to law5.
Notice that the output file name is specified by outfile=, not file=; and that the / between the outfile name and the list of variables is essential.
Return to top
Outliers are points which lie far from the main distributions or the main trends of one or more variables. They can be detected by plotting variables against one another (e.g. using Minitab's PLOT command), or using commands like FREQUENCIES in SPSS or TALLY in Minitab to examine the distributions of key variables. Minitab does some checking for you automatically whenever it does a regression, and reports if it finds "unusual observations", which usually are outliers.
Serious outliers should be dealt with as follows:
This refers to the situation where one or more of the independent variables can be predicted almost exactly from the remainder of the set. In this case, the independent variable set is obviously redundant in some sense. If the independent variables are multicollinear, the regression coefficients we calculate will be very unstable - they will vary markedly from sample to sample - so it will be difficult to decide correctly which are the important regressors.
If multicollinearity is extreme, Minitab will refuse to carry out the analysis, but this will only happen with situations way beyond that where we would be wise to drop some variables from the set. SPSS allows us to assess the degree of multicolinearity in the sample, by requesting a measure of tolerance in the statistics subcommand:
/statistics default tol
This gives us 1-R2 for the regression of each independent variable on all the others. If tolerance is low (below 0.1, say) for any independent variable, it should be regarded as a problem.
Another approach to detecting multicollinearity is to run a principal components analysis on the independent variables (first tranforming them to z-scores). If many of the eigenvalues are below 1.0, the variable set is showing serious multicollinearity.
For further discussion of multicollinearity, see Tabachnick & Fidell (1989), p. 130.
Another long word (it means "different variabilities"). Regression assumes that the scatter of the points about the regression line is the same for all values of each independent variable. Quite often, the spread will increase steadily as one of the independent variables increases, so we get a fan-like scattergram if we plot the dependent variable against that independent variable. Another way of detecting heteroscedasticity (and also outlier problems) is to plot the residuals against the fitted values of the dependent variable; a monotonic relationship between residuals and the independent variable (which is what produces the fan-like pattern in the original scattergram) suggests we may have problems with the data. We may be able to deal with this by transforming one or more variables.
For further discussion of heteroscedasticity, see Tabachnik and Fidell (1989), pp. 131-133.
Obviously if you have as many independent variables as you have data points, you can predict the dependent variable value perfectly - but you are not explaining it at all. Economical explanation requires fewer independent variables than cases, but how many fewer? Various rules of thumb have been suggested, and obviously much depends on the amount of noise in the data, the nature of the phenomena being investigated, and the hypotheses being tested. In favourable circumstances (well-defined hypotheses, clean data), five times as many cases as regressors might be enough; even under bad circumstances, a 20:1 ratio ought to be adequate. If in doubt, collect a second sample of data and see whether the results replicate - good scientific advice regardless of the statistical method in use. "Cases" here means fully usable cases, i.e. ones with no missing values on any variable; apparently adequate samples are easily reduced to uselessness by a wide scatter of missing values. For further discussion see Tabachnick and Fidell (1989), pp. 128-192.
Return to top
The data used in examples 1-3 are part of those collected in a questionnaire study of neighbourly help (see Webley & Lea 1993, Human Relations 46, 65-76). After the examples you will find an extract from the questionnaire. People living in different districts were sent different coloured questionnaires, so we knew when the forms came back where they had come from. The corresponding data (including a code for which of 4 districts people came from) are stored in the Singer file /singer1/eps/psybin/stats/neighbor.MTW
NEIGHBOURLINESS SURVEY (Extract, reformatted)
About how long have you lived where you do now?
Less than 6 months / 6-12 months / 1-3 years / 3-10 years / Over 10 years
Where were you living before you moved to your present house?
In the same neighbourhood / Elsewhere in Exeter / Elsewhere in Devon / Elsewhere in Britain / Abroad
How neighbourly do you think the area where you now live is?:
Very unfriendly / Not very friendly / About average / Fairly friendly / Very friendly
Roughly how many people in your street, or in the streets just near you, do you know the names of?
None / 1-5 / 6-20 / More than 20
How many of those people (not counting children) would you call by their first names?
None / 1-5 / 6-20 / More than 20
Male / Female
Under 18 / 18-30 / 31-50 / 51-65 / Over 65
Return to top
Send questions and
comments to the departmental
administrator or to the author
of this page
Goto Home page for this course | previous topic | next topic | FAQ file
Goto home page for: University of Exeter | Department of Psychology | Staff | Students | Research | Teaching | Miscellaneous