SPSS for Windows: Descriptive and Inferential Statistics
This document is the second module of a four module tutorial series. This document describes the use of SPSS to obtain descriptive and inferential statistics. In this module, you will be introduced to procedures used to obtain several descriptive statistics, frequency tables, and crosstabulations in the first section. In the second section, the Chi-square test of independence, independent and paired sample t tests, bivariate and partial correlations, regression, and the general linear model will be covered. If you are not familiar with SPSS or need more information about how to get SPSS to read your data, consult the first module of this four part tutorial, SPSS for Windows: Getting Started. This set of documents uses a sample dataset, Employee data.sav, that SPSS provides. It can be found in the root SPSS directory. If you installed SPSS in the default location, then this file will be located in the following location: C:\Program Files\SPSS\Employee Data.sav.
Some users prefer to use keystrokes to navigate through SPSS. Information on common keystrokes are available in our SPSS 10 for Windows Keystoke Manual.
A common first step in data analysis is to summarize information about variables in your dataset, such as the averages and variances of variables. Several summary or descriptive statistics are available under the Descriptives option available from the Analyze and Descriptive Statistics menus:
After selecting the Descriptives option, the following dialog box will appear:
This dialog box allows you to select the variables for which descriptive statistics are desired. To select variables, first click on a variable name in the box on the left side of the dialog box, then click on the arrow button that will move those variables to the Variable(s) box. For example, the variables salbegin and salary have been selected in this manner in the above example. To view the available descriptive statistics, click on the button labeled Options. This will produce the following dialog box:
Clicking on the boxes next to the statistics' names will result in these statistics being displayed in the output for this procedure. In the above example, only the default statistics have been selected (mean, standard deviation, minimum, and maximum), however, there are several others that could be selected. After selecting all of the statistics you desire, output can be generated by first clicking on the Continue button in the Options dialog box, then clicking on the OK button in the Descriptives dialog box. The statistics that you selected will be printed in the Output Viewer. For example, the selections from the preceding example would produce the following output:
This output contains several pieces of information that can be useful to you in understanding the descriptive qualities of your data. The number of cases in the dataset is recorded under the column labeled N. Information about the range of variables is contained in the Minimum and Maximum columns. For example, beginning salaries ranged from $9000 to $79,980 whereas current salaries range from $15,750 to $135,000. The average salary is contained in the Mean column. Variability can be assessed by examining the values in the Std. column. The standard deviation measures the amount of variability in the distribution of a variable. Thus, the more that the individual data points differ from each other, the larger the standard deviation will be. Conversely, if there is a great deal of similarity between data points, the standard deviation will be quite small. The standard deviation describes the standard amount variables differ from the mean. For example, a starting salary with the value of $24,886.73 is one standard deviation above the mean in the above example in which the variable, salary has a mean of $17,016.09 and a standard deviation of $7,870.64. Examining differences in variability could be useful for anticipating further analyses: in the above example, it is clear that there is much greater variability in the current salaries than beginning salaries. Because equal variances is an assumption of many inferential statistics, this information is important to a data analyst.
While the descriptive statistics procedure described above is useful for summarizing data with an underlying continuous distribution, the Descriptives procedure will not prove helpful for interpreting categorical data. Instead, it is more useful to investigate the numbers of cases that fall into various categories. The Frequencies option allows you to obtain the number of people within each education level in the dataset. The Frequencies procedure is found under the Analyze menu:
Selecting this menu item produces the following dialog box:
Select variables by clicking on them in the left box, then clicking the arrow in between the two boxes. Frequencies will be obtained for all of the variables in the box labeled Variable(s). This is the only step necessary for obtaining frequency tables; however, there are several other descriptive statistics available, many of which are described in the preceding section. The example in the above dialog box would produce the following output:
Clicking on the Statistics button produces a dialog box with several additional descriptive statistics. Clicking on the Charts button produces the following box which allows you to graphically examine their data in several different formats:
Each of the available options provides a visual display of the data. For example, clicking on the Histograms button with its suboption, With normal curve, will provide you with a chart similar to that shown below. This will allow you to assess whether your data are normally distributed, which is an assumption of several inferential statistics. You can also use the Explore procedure, available from the Descriptives menu, to obtain the Kolmogorov-Smirnov test, which is a hypothesis test to determine if your data are normally distributed.
While frequencies show the numbers of cases in each level of a categorical variable, they do not give information about the relationship between categorical variables. For example, frequencies can give you the number of men and women in a company AND the number of people in each employment category, but not the number of men and women IN each employment category. The Crosstabs procedure is useful for investigating this type of information because it can provide information about the intersection of two variables. The number of men and women in each of three employment categories is one example of information that can be crosstabulated. The Crosstabs procedure is found in the Analyze menu in the Data Editor window:
After selecting Crosstabs from the menu, the dialog box shown above will appear on your monitor. The box on the left side of the dialog box contains a list of all of the variables in the working dataset. Variables from this list can be selected for rows, columns, or layers in a crosstabulation. For example, selecting the variable gender for the rows of the table and jobcat for the columns would produce a crosstabulation of gender by job category.
The options available by selecting the Statistics and Cells buttons provide you with several additional output features. Selecting the Cells button will produce a menu that allows you to add additional values to your table. For example, the dialog box shown below illustrates an example in which Expected option in the Counts box and the Row, Column, and Total options in the Percentages box have been selected.
The combination of the two dialog boxes shown above will produce the following output table:
The crosstabulation statistics provide several interesting observations about the data. In the above table, there appears to be an association between gender and employment category as the expected values, which are the values expected by chance, and the actual counts are different from each other. The following section will discuss how to further examine this relationship with inferential statistics.
The Chi-square test for independence is used in situations where you have two categorical variables. A categorical variable is a qualitative variable in which cases are classified in one and only one of the possible levels. A classic example is gender, in which cases are classified in one of two possible levels. The example in the above section, in which Gender and Employment Category are crosstabulated using the SPSS Crosstabs procedure, is an example of data with which you could conduct a Chi-square test of independence testing the null hypothesis that there is no relationship between the two variables.
For instance, you could conduct a test of the hypothesis that there is no relationship between Gender and Employment Category. If this hypothesis were true, you would expect that the proportion of men and women would be the same within each level of Employment Category. In other words, there should be little difference between observed and expected values, where the expected values represent the numbers that would be in each cell when the variables are independent of each other. The difference between observed and expected values is the basis of the Chi-square statistic: it evaluates the likelihood that the differences between the observed and expected values would occur under the null hypothesis that there is no difference between these values. The expected values can be obtained by clicking on the Cells box in the Crosstabs dialog box, as described in the preceding section. Examining the table above, it appears that it is indeed the case that gender and employment category are independent of each other. It appears that there are more women in clerical positions than would be expected by chance, whereas there are more men in custodial and managerial positions than would be expected by chance. Conducting a Chi-square test of independence would tell us if the observed pattern is statistically different from the pattern expected due to chance.
The Chi-square test of independence can be obtained through the Crosstabs dialog boxes that were used above to get a crosstabulation of the data. After opening the Crosstabs dialog box as described in the preceding section, click the Statistics button to get the following dialog box:
By clicking on the box labeled Chi-Square, you will obtain the Chi-square test of independence for the variables you have crosstabulated. This will produce the following table in the Output Viewer:
Inspecting the table in the previous section, it appears that the the two variables, gender and employment category, are related to each other in some way. This finding is implicated by the substantial differences in the observed and expected counts: these differences represent the difference between values expected if gender and employment classification were independent of each other (expected counts) and the actual numbers of cases in each cell (observed counts). For example, if gender and employment classification were unrelated, then it is expected that 38.3 women would be in the manager classification as opposed to the observed number, 10. In this example, the expected value of 38.3 represents the fact that 45.6% of the cases in this dataset are women, so it is expected that 45.6% of the 84 managers in the dataset would also be women if gender and employment classification were independent of each other. The output above provides a statistical hypothesis test for the hypothesis that gender and employment category are independent of each other. The large Chi-Square statistic (79.28) and its small significance level (p < .000) indicates that it is very unlikely that these variables are independent of each other. Thus, you can conclude that there is a relationship between a person's gender and their employment classification.
The t test is a useful technique for comparing mean values of two sets of numbers. The comparison will provide you with a statistic for evaluating whether the difference between two means is statistically significant. T tests can be used either to compare two independent groups (independent-samples t test) or to compare observations from two measurement occasions for the same group (paired-samples t test). To conduct a t test, your data should be a sample drawn from a continuous underlying distribution. If you are using the t test to compare two groups, the groups should be randomly drawn from normally distributed and independent populations. For example, if you were comparing clerical and managerial salaries, the independent populations are clerks and managers, which are two nonoverlapping groups. If you have more than two groups or more than two variables in a single group that you want to compare, you should use one of the General Linear Model procedures in SPSS, which are described below.
There are three types of t tests; the options are all located under the Analyze menu item:
While each of these t tests compares mean values of two sets of numbers, they are designed for distinctly different situations:
To conduct an independent sample t test, first select the menu option shown above, to produce the following dialog box:
To select variables for the analysis, first highlight them by clicking on them in the box on the left. Then move them into the appropriate box on the right by clicking on the arrow button in the center of the box. Your independent variable should go in the Grouping Variable box, which is a variable that defines which groups are being compared. For example, because employment categories are being compared in this analysis, the jobcat variable is selected. However, because jobcat has more than two levels, you will need to click on Define Groups to specify the two levels of jobcat that you want to compare. This will produce another dialog box as is shown below:
Here, the groups to be compared are limited to the groups with the values 2 and 3, which represent the clerical and managerial groups. After selecting the groups to be compared, click the Continue button, and then click the OK button in the main dialog box. The above choices will produce the following output:
The first output table, labeled Group Statistics, displays descriptive statistics. The second output table, labeled Independent Samples Test, contains the statistics that are critical to evaluating the current research question. This table contains two sets of analyses: the first assumes equal variances and the second does not. To assess whether you should use the statistics for equal or unequal variances, use the significance level associated with the value under the heading, Levene's Test for Equality of Variances. It tests the hypothesis that the variances of the two groups are equal. A small value in the column labeled Sig. indicates that this hypothesis is false and that the groups do indeed have unequal variances. In the above case, the small value in that column indicates that the variance of the two groups, clerks and managers, is not equal. Thus, you should use the statistics in the row labeled Equal variances not assumed.
The SPSS output reports a t statistic and degrees of freedom for all t test procedures. Every unique value of the t statistic and its associated degrees of freedom have a significance value. In the above example in which the hypothesis that clerks and managers do not differ in their salaries, the t statistic under the assumption of unequal variances has a value of -16.3, and the degrees of freedom has a value of 89.6 with an associated significance level of .000. The significance level tells us that the probability that there is no difference between clerical and managerial salaries is very small: specifically, less than one time in a thousand would we obtain a mean difference of $33,038 or larger between these groups if there were really no differences in their salaries.
To obtain a paired-samples t test, select the menu items described above and the following dialog box will appear:
The above example illustrates a t test between the variables salbegin and salary which represent employees' beginning salary and their current salary. To set up a paired-samples t test as in the above example, click on the two variables that you want to compare. The variable names will appear in the section of the box labeled Current Selections. When these variable names appear there, click the arrow in the middle of the dialog box and they will appear in the Paired Variables box. Clicking the OK button with the above variables selected will produce output for the paired-samples t test. The following output is an example of the statistics you would obtain from the above example.
As with the independent samples t test, there is a t statistic and degrees of freedom that has a significance level associated with it. The t test in this example tests the hypothesis that there is no difference in clerks' beginning and current salaries. The t statistic, (35.04), and its associated significance level (p < .000) indicate that this in not the case. In fact, the observed mean difference of $17,403.48 between beginning and current salaries would occur fewer than once in a thousand times if there really were no difference between clerks' beginning and current salaries.
Correlation is one of the most common forms of data analysis both because it can provide an analysis that stands on its own, and also because it underlies many other analyses, and can can be a good way to support conclusions after primary analyses have been completed. Correlations are a measure of the linear relationship between two variables. A correlation coefficient has a value ranging from -1 to 1. Values that are closer to the absolute value of 1 indicate that there is a strong relationship between the variables being correlated whereas values closer to 0 indicate that there is little or no linear relationship. The sign of a correlation coefficient describes the type of relationship between the variables being correlated. A positive correlation coefficient indicates that there is a positive linear relationship between the variables: as one variable increases in value, so does the other. An example of two variables that are likely to be positively correlated are the number of days a student attended class and test grades because, as the number of classes attended increases in value, so do test grades. A negative value indicates a negative linear relationship between variables: as one variable increases in value, the other variable decreases in value. The number of days students miss class and their test scores are likely to be negatively correlated because as the number of days of missed classed increases, test scores typically decrease.
To obtain a correlation in SPSS, start at the Analyze menu. Select the Correlate option from this menu. By selecting this menu item, you will see that there are three options for correlating variables: (1) Bivariate, (2) Partial, and (3) Distances. This document will cover the first two types of correlations. The bivariate correlation is for situations where you are interested only in the relationship between two variables. Partial correlations should be used when you are measuring the association between two variables but want to factor out the effect of one or more other variables.
To obtain a bivariate correlation, choose the following menu option:
This will produce the following dialog box:
To obtain correlations, first click on the variable names in the variable list on the left side of the dialog box. Next, click on the arrow between the two white boxes which will move the selected variables into the Variables box. Each variable listed in the Variables box will be correlated with every other variable in the box. For example, with the above selections, we would obtain correlations between Education Level and Current Salary, between Education Level and Previous Experience, and between Current Salary and Previous Experience. We will maintain the default options shown in the above dialog box in this example. The first option to consider is the type of correlation coefficient. Pearson's is appropriate for continuous data as noted in the above example, whereas the other two correlation coefficients, Kendall's tau-b and Spearman's, are designed for ranked data. The choice between a one and two-tailed significance test in the Test of Significance box should be determined by whether the hypothesis you are testing is making a prediction about the direction of effect between the two variables: if you are making a prediction that there is a negative or positive relationship between the variables, then the one-tailed test is appropriate; if you are not making a directional prediction, you should use the two-tailed test if there is not a specific prediction about the direction of the relationship between the variables you are correlating. The selections in the above dialog box will produce the following output:
This output gives us a correlation matrix for the three correlations requested in the above dialog box. Note that despite there being nine cells in the above matrix, there are only three correlation coefficients of interest: (1) the correlation between current salary and educational level, the correlation between previous experience and educational level, and the correlation between current salary and previous experience. The reason only three of the nine correlations are of interest is because the diagonal consists of correlations of each variable with itself, always resulting in a value of 1.00 and the values on each side of the diagonal replicate the values on the opposite side of the diagonal. For example, the three unique correlation coefficients show there is a positive correlation between employees' number of years of education and their current salary. This positive correlation coefficient (.661) indicates that there is a statistically significant (p < .001) linear relationship between these two variables such that the more education a person has, the larger that person's salary is. Also observe that there is a statistically significant (p < .001) negative correlation coefficient (-.252) for the association between education level and previous experience, indicating that the linear relationship between these two variables is one in which the values of one variable decrease as the other increases. The third correlation coefficient (-.097) also indicates a negative association between employee's current salaries and their previous work experience, although this correlation is fairly weak.
The second type of correlation listed under the Correlate menu item is the partial correlation, which measures an association between two variables with the effects of one or more other variables factored out. To obtain a partial correlation, select the following menu item:
This will produce the following dialog box:
Here, we have selected the variables we want to correlate as well as the variable for which we want to control by first clicking on variable names to highlight them on the left side of the box, then moving them to the boxes on the right by clicking on the arrow immediately to the left of either the Variables box or the Controlling for box. In this example, we are correlating current salaries with years of education while controlling for beginning salaries. Thus, we will have a measure of the association between current salaries and years of education, while removing the association between beginning salaries and the two variables we are correlating. The above example will produce the following output:
- - - P A R T I A L C O R R E L A T I O N C O E F F I C I E N T S - - -
Notice that the correlation coefficient is considerably smaller in the output above than in the bivariate correlation example: the correlation between these variables was .661, whereas it is only .281 in the partial correlation. Nevertheless, a statistically significant association (p < .001) exists between these variables. While it may be obvious in the above example that starting and current salaries will be highly correlated, the example illustrates how partial correlations can be used to assess the extent to which variables can be used to explain unique variance by removing the effects of other variables that may be highly correlated with the relationship of interest.
Partial correlations can be especially useful in situations where it is not obvious whether variables possess a unique relationship or whether several variables overlap with each other. For example, if you were attempting to correlate anxiety with job performance and stress with job performance, it would be useful to conduct partial correlations. You could correlate anxiety and a job performance measure while controlling for stress to determine if there were a unique relationship between anxiety and job performance or whether perhaps stress is highly correlated with anxiety--which would result in little remaining variance that could be uniquely attributed to the association between anxiety and job performance.
Regression is a technique that can be used to investigate the effect of one or more predictor variables on an outcome variable. Regression allows you to make statements about how well one or more independent variables will predict the value of a dependent variable. For example, if you were interested in investigating which variables in the employee database were good predictors of employees' current salaries, you could create a regression equation that would use several of the variables in the dataset to predict employees' salaries. By doing this you will be able to make statements about whether knowing something about variables such as employees' number of years of education, their starting salary, or their number of months on the job are good predictors of their current salaries.
To conduct a regression analysis, select the following from the Analyze menu:
This will produce the following dialog box:
This dialog box illustrates an example regression equation. As with other analyses, you select variables from the box on the left by clicking on them, then moving them to the boxes on the right by clicking the arrow next to the box where you want to enter a particular variable. Here, employees' current salary has been entered as the dependent variable. In the Independent(s) box, several predictor variables have been entered, including education level, beginning salary, months since hire, and previous experience.
NOTE: Before you run a regression model, you should consider the method that you use for selecting or rejecting variables in that model. The box labeled Method allows you to select from one of five methods: Enter, Remove,Forward, Backward, and Stepwise. Unfortunately, we cannot offer a comprehensive discussion of the characteristics of each of these methods here, but you have several options regarding the method you use to remove and retain predictor variables in your regression equation. In this example, we will use the SPSS default method, Enter, which is a standard approach in regression models. If you have questions about which method is most appropriate for your data analysis, consult a regression text book, the SPSS help facilities, or contact a consultant.
The following output assumes that only the default options have been requested. If you have selected options from the Statistics, Plots, or Options boxes, then you will have more output than is shown below and some of your tables may contain additional statistics not shown here.
The first table in the output, shown below, includes information about the quantity of variance that is explained by your predictor variables. The first statistic, R, is the multiple correlation coefficient between all of the predictor variables and the dependent variable. In this model, the value is .90, which indicates that there is a great deal of variance shared by the independent variables and the dependent variables. The next value, R Square, is simply the squared value of R. This is frequently used to describe the goodness-of-fit or the amount of variance explained by a given set of predictor variables. In this example, the value is .81, which indicates that 81% of the variance in the dependent variable is explained by the independent variables in the model.
The second table in the output is an ANOVA table that describes the overall variance accounted for in the model. The F statistic represents a test of the null hypothesis that the expected values of the regression coefficients are equal to each other and that they equal zero. Put another way, this F statistic tests whether the R square proportion of variance in the dependent variable accounted for by the predictors is zero. If the null hypothesis were true, then that would indicate that there is not a regression relationship between the dependent variable and the predictor variables. But, instead, it appears that the four predictor variables in the present example are not all equal to each other and could be used to predict the dependent variable, current salary, as is indicated by a large F value and a small significance level.D
The third table in standard regression output provides information about the effects of individual predictor variables. Generally, there are two types of information in the Coefficients table: coefficients and significance tests. The coefficients indicate the increase in the value of the dependent variable for each unit increase in the predictor variable. For example, the unstandardized coefficient for Educational Level in the example is 669.91, which indicates to us that for each year of education, a person's predicted salary will increase by $669.91. A well known problem with the interpretation of unstandardized coefficients is that their values are dependent on the scale of the variable for which they were calculated, which makes it difficult to assess the relative influence of independent variables through a comparison of unstandardized coefficients. For example, comparing the unstandardized coefficient of Education Level, 669.91, with the unstandardized coefficient of the variable Beginning Salary, 1.77, it could appear that Educational Level is a greater predictor of a person's current salary than is Beginning Salary. We can see that this is deceiving, however, if we examine the standardized coefficients, or Beta coefficients. Beta coefficients are based on data expressed in standardized, or z score form. Thus, all variables have a mean of zero and a standard deviation of one and are thus expressed in the same units of measurement. Examining the Beta coefficients for Education Level and Beginning Salary, we can see that when these two variables are expressed in the same scale, Beginning Salary is more obviously the better predictor of Current Salary.
In addition to the coefficients, the table also provides a significance test for each of the independent variables in the model. The significance test evaluates the null hypothesis that the unstandardized regression coefficient for the predictor is zero when all other predictors' coefficients are fixed to zero. This test is presented as a t statistic. For example, examining the t statistic for the variable, Months Since Hire, you can see that it is associated with a significance value of .000, indicating that the null hypothesis, that states that this variable's regression coefficient is zero when all other predictor coefficients are fixed to zero, can be rejected.
The majority of procedures used for conducting analysis of variance (ANOVA) in SPSS can be found under the General Linear Model (GLM) menu item in the Analyze menu. Analysis of variance can be used in many situations to determine whether there are differences between groups on the basis of one or more outcome variables or if a continuous variable is a good predictor of one or more dependent variables. There are three varieties of of the general linear model available in SPSS: univariate, multivariate, and repeated measures. The univariate general linear model is used in situations where you only have a single dependent variable, but may have several independent variables that can be fixed between-subjects factors, random between-subjects factors, or covariates. The multivariate general linear model is used in situations where there is more than one dependent variable and independent variables are either fixed between-subjects factors or covariates. The repeated measures general linear model is used in situations where you have more than one measurement occasion for a dependent variable and have fixed between-subjects factors or covariates as independent variables. Because it is beyond the scope of this document to cover all three varieties of the general linear model in detail, we will focus on the univariate version of the general linear model with some attention given to topics that are unique to the repeated measures general linear model. Several features of the univariate general linear model are useful for understanding other varieties of the model that are provided in SPSS: understanding the univariate model will prove useful for understanding other GLM options.
The univariate general linear model is used to compare differences between group means and estimating the effect of covariates on a single dependent variable. For example, you may want to see if there are differences between men and women's salaries in a sample of employee data. To do this, you would want to demonstrate that the average salary is significantly different between men and women. However, in doing such an analysis, you are likely aware that there are other factors that could affect a person's salary that need to be controlled for in such an analysis. For example, educational background and starting salary are some such variables. By including these variables in our analysis, you will be able to evaluate the differences between men and women's salaries while controlling for the influence of these other variables.
To specify a univariate general linear model in SPSS, go to the analyze menu and select univariate from the general linear model menu:
This will produce the following dialog box:
The above box demonstrates a model with multiple types of independent variables. The variable, gender, has been designated as a fixed factor because it contains all of the levels of interest.
In contrast, random variables are variables that represent a random sample of the possible levels that could be sampled. There are not any true random variables in our dataset; therefore, this input box has been left blank here. However, you could imagine a situation similar to the above example where you sampled data from multiple corporations for our employee database. In that case, you would have introduced a random variable into the model--the corporation to which an employee belongs. Corporation is a random factor because you would only be sampling a few of the many possible corporations to which you would want to generalize your results.
The next input box contains the covariates in your model. A covariate is a quantitative independent variable. Covariates are often entered in models to reduce error variance: by removing the effects of the relationship between the covariate and the dependent variable, you can often get a better estimate of the amount of variance that is being accounted for by the factors in the model. Covariates can also be used to measure the linear association between the covariate and a dependent variable, as is done in regression models. In this situation, a linear relationship indicates that the dependent variable increases or decreases in value as the covariate increases or decreases in value.
The box labeled WLS Weight can contain a variable that is used to weight other variables in a weighted least-squares analysis. This procedure is infrequently used however, and is not discussed in any detail here.
The default model for the SPSS univariate GLM will include main effects for all independent variables and will provide interaction terms for all possible combinations of fixed and random factors. You may not want this default model, or you may want to create interaction terms between your covariates and some of the factors. In fact, if you intend to conduct an analysis of covariance, you should test for interactions between covariates and factors. Doing so will determine whether you have met the homogeneity of regression slopes assumption, which states that the regression slopes for all groups in your analysis are equal. This assumption is important because the means for each group are adjusted by averaging the slopes for each group so that group differences in the covariate are removed from the dependent variable. Thus, it is assumed that the relationship between the covariate and the dependent variable is the same at all levels of the independent variables. To make changes in the default model, click on the Model button which will produce the following dialog box:
The first step for modifying the default model is to click on the button labeled Custom, to activate the grayed out areas of the dialog box. At this point, you can begin to move variables in the Factors & Covariates box into the Model box. First, move all of the main effects into the Model box. The quickest way to do that is to double-click on their names in the Factors & Covariates box. After entering all of the main effects, you can begin building interaction terms. To build the interactions, click on the arrow facing downwards in the Build Term(s) section and select interaction, as shown in the figure above. After you have selected the interaction, you can click on the names of the variables with which you would like to build an interaction, then click on the arrow facing right under the Build Term(s) heading. In the above example, the educ*gender term has already been created. The salbegin*gender and salbegin*educ terms can be created by highlighting two terms at a time as shown above, then clicking on the right-facing arrow. Some of the other options in the Build Terms list that you may find useful are the All n-way options. For example if you highlighted all three variables in the Factors & Covariates box, you could create all of the three possible 2-way interactions by selecting the All 2-way option from the Build Terms(s) drop-down menu, then clicking the right-facing arrow.
If you are testing the homogeneity of regression slopes assumption, you should examine your group by covariate interactions, as well as any covariate by covariate interactions. In order to meet the ANCOVA assumption, these interactions should not be significant. Examining the output from the example above, we expect to see nonsignificant effects for the gender*educ and the gender*salbegin interaction effects:
Examining the group by covariate effects, you can see that both were nonsignificant. The gender*salbegin effect has a small F statistic (.660) and a large significance value (.417), the educ*salbegin effect also has a small F statistic (1.808) and large significance value (.369), and the salbegin*educ effect also has a small F statistic (1.493) and large significance level (.222). Because all of these significance levels are greater than .05, the homogeneity of regression assumption has been met and you can proceed with the ANCOVA.
Knowing that the model does not violate the homogeneity of regression slopes assumption, you can remove the interaction terms from the model by returning to the GLM Univariate dialog box, clicking the Model button, and selecting Full Factorial. This will return the model to its default form in which there are no interactions with covariates. After you have done this, click OK in the GLM Univariate dialog box to produce the following output:D
The default output for the univariate general linear model contains all main effects and interactions between fixed factors. The above output contains no interactions because gender is the only fixed factor. Each factor, covariate, or other source of variance is listed in the left column. For each source of variance, there are several test statistics. To evaluate the influence of each independent variable, look at the F statistic and its associated significance level. Examining the first covariate, education level, the F statistic (37.87) and its associated significance level (.000) indicate that it has a significant linear relationship with the dependent variable. The second covariate, salbegin, also has a significant F statistic (696.12) as can be seen from its associated significance level (.000). In both cases, this indicates that the the values of the dependent variable, salary, increases as the values of education level and beginning salaries increase. The next source of variance, gender, provides us with a test of the null hypothesis that there are no differences between gender groups, or more specifically, there are not differences between men and women's salaries. This test provides a small F statistic (3.87) and a significance level that is not statistically significant (p = .05). In the above model containing education level and beginning salaries as covariates, we are not able to say that there is a statistically significant difference between men and women's salaries. That is, when the model takes into account the variance accounted for by education level and beginning salaries, the variance that can be uniquely attributed to gender is not significantly different from a model in which gender explains no variance.
The repeated measures version of the general linear model has many similarities to the univariate model described above. However, the key difference between the models is that there are multiple measurement occasions of the dependent variable in repeated measures models, whereas the univariate model only permits a single dependent variable. You could conduct a similar model with repeated measurements by using beginning salaries and current salaries as the repeated measurement occasions.
To conduct this analysis, you should select the Repeated Measures option from the General Linear Model submenu of the Analyze menu:
Selecting this option will produce the following dialog box:
This dialog box is used for defining the repeated measures, or within-subjects, dependent variables. You first give the within-subject factor a name in the box labeled Within-Subject Factor Name. This name should be something that describes the dependent variables you are grouping together. For example, in this dialog box, salaries are being analyzed, so the within-subject factor was given the name salaries. Next, specify the number of levels, or number of measurement occasions, in the box labeled Number of Levels. This is the number of times the dependent variable was measured. Thus, in the present example, there are two measurement occasions for salary because you are measuring beginning salaries and current salaries. After you have filled in the Within-Subject Factor Name and the Number of Levels input boxes, click the Add button which will transfer the information in the input boxes into the box below. Repeat this process until you have specified all of your within-subject factors. Then, click on the Define button, and the following dialog box will appear:
When this box initially appears, you will see a slot for each level of the within-subject factor variables that you specified in the previous dialog box. These slots are labeled numerically for each level of the within-subject factor but do not contain variable names. You still need to specify which variable fills each slot of the within-subject factors. To do this, click the variable's name in the variable list on the left side of the dialog box. Next, click on the arrow pointing towards the Within-Subject Variables dialog box to move the variable name from the list to the top slot in the within-subjects box. This process has been completed for salbegin, the first level of the salaries within-subject factor. The same process should be repeated for salary, the variable representing an employee's current salary.
After you have completed the specifications for the within-subjects factors, you can define your independent variables. Between-subject factors, or fixed factors should be moved into the box labeled Between-Subjects Factors(s) by first clicking on the variable name in the variable list, then clicking on the arrow to the left of the Between-Subjects Factor(s) box. In this example, gender has been selected as a between-subjects factor. Covariates, or continuous predictor variables, can be moved into the Covariates box in the same manner as were the between-subjects factors. Above, educ, the variable representing employee's number of years of education, has been specified as a covariate.
This will produce several output tables, but we will focus here on the tables describing between-subject and within-subject effects. However, these tables for univariate analysis of variance may not always be the appropriate. The univariate tests have an additional assumption: the assumption of sphericity. If this assumption is violated, you should use the multivariate output or adjust your results using one of the correction factors in the SPSS output. For a more detailed discussion of this topic, see the usage note, Repeated Measures ANOVA Using SPSS MANOVA in the section, "Within-Subjects Tests: The Univariate versus the Multivariate Approach." This usage note can be found at http://www.utexas.edu/cc/rack/stat.html.
The following output contains the statistics for the effects in the model specified in the above dialog boxes:
This table contains information about the within-subject factor, salaries, and its interactions with the independent variables. The main effect for salaries is a test of the null hypothesis that all levels of within-subjects factor are equal, or, more specifically, it is a test of the hypothesis that beginning and current salaries are equal. The F statistic (18.91) and its associated significance level (p < .000) indicate that you can reject this hypothesis as false. In other words, it appears that there is a statistically significant difference between beginning salaries and salaries after one year of employment. After you have tested this hypothesis, you can then investigate whether the increase in salaries is the same across all values or levels of the other independent variables that are included in the model. The first interaction term in the table tests the hypothesis that the increase in salaries is constant, regardless of educational background. The F statistic (172.10) and its associated significance level (p < .000) allow us to reject this hypothesis as well. The knowledge that this interaction is significant indicates that it is worthwhile to examine characteristics of the interaction. Here, the interaction reflects the fact that employees with higher levels of education received greater pay raises than those with lower levels of education. However, you should always investigate the properties of your interactions through graphical displays, mean comparisons, or statistical tests because a significant interaction can take on many forms. Finally, the third interaction tests the hypothesis that the increase in salaries in the first year of employment differs by gender. The F statistic (25.07) and its significance level (p < .000) indicate that the increase in salaries does vary by gender.
The output for the repeated measures general linear model also provides statistics for between-subject effects. In this example, the model contains two between-subjects factors: employees' education level and their gender. Education level was entered as a covariate in the model, and therefore the statistics associated with it are a measure of the linear relationship between education level and salaries. In contrast, the statistics for the between-subjects factor, gender, represents a comparison between groups across all levels of the within-subjects factors. Specifically, it is a comparison between males and females on differences between their beginning and current salaries. In the above example, both education level and gender are statistically significant. The F statistic (277.96) and significance level (p < .000) associated with education level allows us to reject the null hypothesis that there is not a linear relationship between education and salaries. By rejecting the null hypothesis, you can conclude that there is a positive linear relationship between the two variables indicating that as number of years of education increases, salaries do as well. The F statistic (55.79) for gender and its associated significance level (p < .000) represent a test of the null hypothesis that there are no group differences in salaries. The significant F statistic indicates that you can reject this null hypothesis and conclude that there is a statistically significant difference between men and women's salaries.
To learn more about SPSS, please proceed to the next SPSS tutorial.
14 September 2001
Statistical Support, a division of Research Consulting at ITS
Send us e-mail at firstname.lastname@example.org or submit a feedback form
Copyright 2001, UT Austin
Send questions or comments regarding this page to email@example.com. Send
computing questions to the ITS Help
Desk or call 475-9400.