Overview All statistical procedures have underlying assumptions. In some cases, violation of these assumptions will not change substantive research conclusions. In other cases, violation of assumptions is critical to meaningful research. Throughout this statistical resource, the assumptions of each statistical procedure are clearly indicated in an "Assumptions" section. Establishing that one's data meet the assumptions of the procedure one is using is an expected component of all quantitatively-based journal articles, theses, and dissertations. This module discusses the most common data assumptions found in statistical research, as listed in the table of contents to the right..

Contents Key concepts Additivity Data level Equality of means Homogeneity of variances Homogeneity of variance-covariance matricess Homogeneity of regressions Homoscedasticity Independence Linearity Multicollinearity Multivariate normality Normality, skew, & transformations Normally distributed error Outliers Proper model specification Randomness Sound measurement Sphericity Unidimensionality

Key Concepts and Terms

Assumptions are covered under each statistical topic. See also the separate section on data levels. This section provides information general to all procedures.
• Parametric statistics are those which assume a certain distribution of the data (usually the normal distribution), assume an interval level of measurement, and assume homogeneity of variances when two or more samples are being compared. Most common significance tests (z tests, t-tests, and F tests) are parametric. However, it has long been established that moderate violations of parametric assumptions have little or no effect on substantive conclusions in most instances (ex., Cohen, 1969: 266-267.)
  • Test of near-metricity of ordinal data. Multidimensional scaling provides one method of testing for near-metricity of ordinal data, as discussed in that section.



Assumptions

• SOUND MEASUREMENT
  • Descriptive statistics. All forms of statistical analysis assume sound measurement, relatively free of coding errors. It is good practice to run descriptive statistics on one's data so that one is confident that data are generally as expected in terms of means and standard deviations, and there are no out-of-bounds entries beyond the expected range.

  • Avoiding attenuation. When the range of the data is reduced artificially, as by classifying or dichotomizing a continuous variable, correlation is attenuated, often leading to underestimation of the effect size of that variable.
    

  • Avoiding tautological correlation. When the indicators for latent variable A include one or more of the indicators for latent variable B, definitional overlap confounds the correlation of A and B.

• PROPER MODEL SPECIFICATION
  • Specification of a model refers to not omitting significant variables or including extraneous ones, and also to correctly indicating the direction of arrows connecting the variables in the model. When a misspecification error is corrected by changing the model, all parameter estimates in the model are subject to change, not only in magnitude, but sometimes even in direction. There is no statistical test for misspecification. A good literature review is important in identifying variables which need to be specified.

    As a rule of thumb, the lower the overall effect (ex., R² in multiple regression, goodness of fit in logistic regression), the more likely it is that important variables have been omitted from the model and that existing interpretations of the model will change when the model is correctly specified. The specification problem is lessened when the research task is model comparison, to see which has a better fit to the data, as opposed to when the research task is to justify one model and assess the relative importance of the independent variables.

• DATA LEVEL
  • Data level requirements vary by statistical procedure but most procedures require an interval or ratio level of measurement. It is common in social science, however, to utilize dichotomies and ordinal data, such as Likert scale data. Dichotomies are often included in statistical models (ex., regression models) provided the split is less than 90:10. Ordinal variables are often included in statistical models provided the normality of error terms may be demonstrated, as discussed below. Some researchers require the ordinal scale to have five or greater values. Violations of data level assumptions mean that actual standard error is greater than computed standard error and significance is overestimated (that is, the chance of Type I error is greater than computed).

• UNIDIMENSIONALITY
  • Testing unidimensionality. If the researcher is attempting to measure a construct with multiple indicator variables, then the researcher must demonstrate that the items measure the same thing. There are several methods of doing this, with varying meanings of and stringency for testing for unidimensionality. Some of these methods are:
    1. Cronbach's alpha. Perhaps the most commonly used test, Cronbach's alpha is a measure of the intercorrelation of items. If alpha is greater than or equal to .6, then the items are considered unidimensional and may be combined in an index or scale. Some researchers use the more stringent cutoff of .7. Cronbach's alpha is found in SPSS 13 under Analyze, Scale, Reliability Analysis.

    2. Factor analysis. PCA factor analysis is performed on all the indicators for all the constructs in the study. Indicators should have higher factor loadings on their own constructs than on other constructs. Some researchers also require that loadings be higher than some absolute cutoff value, such as .3. Some researchers also require that indicators not crossload on factors not their own (ex., that all loadings other than their own factor be below some absolute cutoff value, such as .3). Factor analysis is found in SPSS 13 under Analyze, Data Reduction, Factor.

    3. Confirmatory factor analysis in structural equation modeling. The first step of structural equation modeling is confirmatory factor analysis, where the measurement model is assessed separately from the structural model. If goodness of fit measures for the measurement model are acceptable, the researcher concludes that the indicators adequately measure the intended constructs. It is also possible to assess the unidimensionality of a concept by comparing models in structural equation modeling Confirmatory factor analysis can be implemented through AMOS, distributed by SPSS as its structural equation modeling package.

    4. Guttman scaling and other forms of scaling. In Guttman, proximity, Mokken, and certain other types of scales, indicator variables are tested to see if they form a certain relationship to each other, such that the researcher is justified in putting the items in the same scale. The relationships, and hence the meanings of unidimensionality, differ by scale type. Note that the most common form of scale, the Likert scale, does not normally involve testing for unidimensionality. This is discussed in more detail in the section on scales and standard measures. Guttman scaling is found in SPSS 13 under Analyze, Scale, Reliability Analysis.

    Note that a set of items may be considered to be unidimensional using one of the methods above, even when another method would fail to find statistical justification in considering the items to measure a single construct. For instance, it would be quite common to find that items in a large Guttman scale would fail to load on a single factor using factor analysis as a method. Finding satisfactory model fit in SEM would not assure that the Cronbach's alpha criterion was met. The researcher must decide on theoretical grounds what definition and criterion for unidimensionality best serves his or her research purpose. However, some method must always be used before proceeding to use multiple indicators to measure a concept.

• NORMALITY
  • A normal distribution is assumed by many statistical procedures. Various transformations are used to correct non-normally distributed data. Correlation, least-squares regression, factor analysis, and related linear techniques are relatively robust against non-extreme deviations from normality provided errors are not severely asymmetric (Vasu, 1979). Severe asymmetry might arise due to strong outliers. Log-linear analysis, logistic regression, and related techniques using maximum likelihood estimation are even more robust against moderate departures from normality (cf. Steenkamp & van Trijp, 1991: 285). Likewise, Monte Carlo simulations show the t-test is robust against moderate violations of normality (Boneau, 1960).

    Normal distributions take the form of a symmetric bell-shaped curve. The standard normal distribution is one with a mean of 0 and a standard deviation of 1. Standard scores, also called z-scores or standardized data, are scores which have had the mean subtracted and which have been divided by the standard deviation to yield scores which have a mean of 0 and a standard deviation of 1. Normality can be visually assessed by looking at a histogram of frequencies, or by looking at a normal probability plot output by most computer programs.

    The area under the normal curve represents probability: 68.26% of cases will lie within 1 standard deviation of the mean, 95.44% within 2 standard deviations, and 99.14% within 3 standard deviations. Often this is simplified by rounding to say that 1 s.d. corresponds to 2/3 of the cases, 2 s.d. to 95%, and 3 s.d. to 99%. Another way to put this is to say there is less than a .05 chance that a sampled case will lie outside 2 standard deviations of the mean, and less than .01 chance that it will lie outside 3 standard deviations. This statement is analogous to statements pertaining to significance levels of .05 and .01, for two-tailed tests. .

    

  • Skew is the tilt (or lack of it) in a distribution. The more common type is right skew, where the tail points to the right. Less common is left skew, where the tail is points left. A common rule-of-thumb test for normality is to run descriptive statistics to get skewness and kurtosis, then divide these by the standard errors. Skew should be within the +2 to -2 range when the data are normally distributed. Some authors use +1 to -1 as a more stringent criterion when normality is critical. In SPSS, one of the places skew is reported is under Analyze, Descriptive Statistics, Descriptives; click Options; select skew.
    

    Negative skew is right-leaning, positive skew is left-leaning. For each type of skew, the mean, median, and mode diverge, so all three measures of central tendency should be reported for skewed data. Box-Cox transformation may normalize skew. Right-skewed distribution may fit power, lognormal, gamma, Weibull, or chi-square distributions. Left-skewed distributions may be recoded to be right-skewed. (Note: there is confusion in the literature about what is "right" or "left" skew, but the foregoing is the most widely accepted labeling.)

  • Kurtosis is the peakedness of a distribution. A common rule-of-thumb test for normality is to run descriptive statistics to get skewness and kurtosis, then use the criterion that kurtosis should be within the +2 to -2 range when the data are normally distributed (a few authors use the more lenient +3 to -3, while other authors use +1 to -1 as a more stringent criterion when normality is critical). Negative kurtosis indicates too many cases in the tails of the distribution. Positive kurtosis indicates too few cases in the tails. Note that the origin in computing kurtosis for a normal distribution is 3 and a few statistical packages center on 3, but the foregoing discussion assumes that 3 has been subtracted to center on 0, as is done in SPSS and LISREL. The version with the normal distribution centered at 0 is Fisher kurtosis, while the version centered at 3 is Pearson kurtosis. SPSS uses Fisher kurtosis. Leptokurtosis is a peaked distribution with "fat tails", indicated by kurtosis > 0 (for Fisher kurtosis, or > 3 for Pearson kurtosis). Platykurtosis is less peaked "thin tails" distribution, with a kurtosis value < 0 (for Fisher kurtosis, or < 3 for Pearson kurtosis).
    

    Various transformations are used to correct kurtosis: cube roots and sine transforms may correct negative kurtosis. In SPSS, one of the places kurtosis is reported is under Analyze, Descriptive Statistics, Descriptives; click Options; select kurtosis.

  • Dichotomies. By definition, a dichotomy is not normally distributed. Many researchers will use dichotomies for procedures requiring a normal distribution as long as the split is less than 90:10. Dichotomies should not be used as dependents in procedures, such as OLS regression, which assume a normally distributed dependent variable.

  • Shapiro-Wilks W test is a formal test of normality offered in the SPSS EXAMINE module or the SAS UNIVARIATE procedure. This is the standard test for normality. W may be thought of as the correlation between given data and their corresponding normal scores, with W = 1 when the given data are perfectly normal in distribution. When W is significantly smaller than 1, the assumption of normality is not met. That is, a significant W statistic causes the researcher to reject the assumption that the distribution is normal. Shapiro-Wilks W is recommended for small and medium samples up to n = 2000. For larger samples, the Kolmogorov-Smirnov test is recommended by SAS and others. In SPSS, Shapiro-Wilks test is found under Analyze, Descriptive Statistics, Explore; select Both or Plots in the Display group; click Plots and select at least one plot.

  • Kolmogorov-Smirnov D test or K-S Lilliefors test, is an alternative test of normality for large samples, available in SPSS EXAMINE and SAS UNIVARIATE. This is sometimes called the Lilliefors test as a correction to K-S developed by Lilliefors is now normally applied. SPSS (as of Version 9), for instance, automatically applies the Lilliefors correction to the K-S test for normality in the EXAMINE module (but not in the NONPAR module). This test, with the Lilliefors correction, is preferred to the chi-square goodness-of-fit test when data are interval or near-interval. When applied without the Lilliefors correction, K-S is very conservative: that is, there is an elevated likelihood of a finding of non-normality. Note the K-S test can test goodness-of-fit against any theoretical distribution, not just the normal distribution. Be aware that when sample size is large, even unimportant deviations from normality may be technically significant by this and other tests. For this reason it is recommended to use other bases of judgment, such as frequency distributions and stem-and-leaf plots. In SPSS, Shapiro-Wilks test is found under Analyze, Descriptive Statistics, Explore; select Both or Plots in the Display group; click Plots and select at least one plot. The Kolmogorov-Smirnov test in SPSS, is found under Analyze, Descriptive Statistics, Explore; select Both or Plots in the Display group; click Plots and select at least one plot.

  • Boxplot tests of the normality assumption: The SPSS boxplot output option (see the EXAMINE command in SPSS Base) produces charts in which the the Y axis is the interval dependent and categories of the independent are arrayed on the X axis. Inside the graph, for each X category, will be a rectangle indicating the spread of the dependent's values for that category. If these rectangles are roughly at the same Y elevation for all categories, this indicates little difference among groups. Within each rectangle is a horizontal dark line, indicating the mean. If most of the rectangle is on one side or the other of the mean line, this indicates the dependent is skewed (not normal) for that group (category). Further out than the rectangle are the "whiskers," which mark the smallest and largest observations which are not outliers (defined as observations greater than 1.5 inter-quartile ranges [IQR's = boxlengths] from the 1st and 3rd quartiles). Note you can display boxplots for two factors (two independents) together by selecting Clustered Boxplots from the Boxplot item on the SPSS Graphs menu. Example.

  • Graphical methods.
    • A histogram of a variable shows rough normality, and a histogram of residuals, if normally distributed, is often taken as evidence of normality of all the variables.

    • A graph of empirical by theoretical cumulative distribution functions (cdf's) simply shows the empirical distibution as, say, a dotted line, and the hypothetical distribution, say the normal curve, as a solid line.

    • A P-P plot is found in SPSS under Graphs, P-P plots. One may test if the distribution of a given variable is normal (or beta, chi-square, exponential, gamma, half-normal, Laplace, Logistic, Lognormal, Pareto, Student's t, Weibull, or uniform). he P=P plot plots a variable's cumulative proportions against the cumulative proportions of the test distribution.The straighter the line formed by the P-P plot, the more the variable's distribution conforms to the selected test distribution (ex., normal). Options within this SPSS procedure allow data transforms first (natural log, standardization of values, difference, and seasonally difference).

    • A quantile-by-quantile or Q-Q plot forms a 45-degree line when the observed values are in conformity with the hypothetical distribution. Q-Q plots plot the quantiles of a variable's distribution against the quantiles of the test distribution.From the SPSS menu, select Graphs, Q-Q. The SPSS dialog box supports testing the following distributions: beta, chi-square, exponential, gamma, half-normal, Laplace, Logistic, Lognormal, normal, pareto, Student's t, Weibull, and uniform.

  • Resampling is a way of doing significance testing while avoiding parametric assumptions like multivariate normality. The assumption of multivariate normality is violated when dichtomous, dummy, and other discrete variables are used. In such situations, where significance testing is appropriate, researchers may use a resampling method.

  • Normalizing Transformations. Various transformations are used to correct skew:
    1. Square roots, logarithmic, and inverse (1/x) transforms "pull in" outliers and normalize right (positive) skew. Inverse (reciprocal) transforms are stronger than logarithmic, which are stronger than roots.

    2. To correct left (negative) skew, first subtract all values from the highest value plus 1, then apply square root, inverse, or logarithmic transforms.

    3. For power and root transforms, finer adjustments can be obtained by adding a constant, C, where C is some small positive value such as .5, in the transform of X:   X^' = (X + C)^P. When this researcher's data contain zero values, the transform using C is strongly recommended over straight transforms (ex., SQRT(X+.5), not SQRT(X) ), but the use of C is standard practice in any event. Values of P less than one (roots) correct right skew, which is the common situation (using a power of 2/3 is common when attempting to normalize). Values of P greater than 1 (powers) correct left skew. For right skew, decreasing P decreases right skew. Too great reduction of P will overcorrect and cause left skew. When the best P is found, further refinements can be made by adjusting C. For right skew, for instance, subtracting C will decrease skew.

    4. Logs vs. roots: logarithmic transformations are appropriate to achieve symmetry in the central distribution when symmetry of the tails is not important; square root transformations are used when symmetry in the tails is important; when both are important, a fourth root transform may work (fourth roots are used to correct extreme skew)..

    5. Logit and probit transforms. Schumacker & Lomax (2004: 33) recommend probit transforms as a means of dealing with skewness. See Lipsey & Wilson (2001: 56) for discussion of logit and probit transforms as a means of transforming dichotomous data as part of estimating effect sizes.

    6. Percentages may be normalized by an arcsine transformation, which is recommended when percentages are outside the range 30% - 70%. The more observations outside this range or the closer to 0% and/or 100%, the more normality is violated and the stronger the recommendation to use arcsine transformation. However, arcsine transformation is not effective when a substantial number of observations are 0% or 100%, or when sample size is small. The usual arcsine transformation is p^' =   arcsin(SQRT(p)), where p is the percentage or proportion.

    7. Poisson distributions may be normalized by a square root transformation.

    8. Other strategies to correct for skew include collapsing categories and dropping outliers.

    Warnings. Transformations should make theoretical sense. Often, normalizing a dichotomy such as gender will not make theoretical sense. Also note that as the log of zero is undefined and leads to error messages, researchers often add some arbitrary small value such as .001 to all values in order to remove zeros from the dataset. However, the choice of the constant can affect the significance levels of the computed coefficients for the logged variables. If this strategy is pursued, the researcher should employ sensitivity analysis with different constants to note effects which might change conclusions.

    • Transforms in SPSS: Select Transform - Compute - Target Variable (input a new variable name) - Numeric Expression (input transform formula)

  • Box-Cox Transformations of Dependent Variables. Box & Cox proposed a maximum likelihood method in 1964 for determining the optimal power transform for purposes of normalization of data. Power transformations of dependent variables were advanced to remedy model lack of normal distribution, lack of homogeneity of variances, and lack of additivity. In a regression context, Box-Cox transformation addresses the problem of non-normality, indicated by skewed residuals (the transformation pulls in the skew) and/or by lack of homoscedasticity of points about the regression line. In an ANOVA context, the Box-Cox transformation addresses the problem of lack of homogeneity of variances associated with the correlation of variances with means in the groups formed by the independent factors and indicated by skewed distributions within the groups (the transformation reduces the correlation).

    In general, the Box-Cox procedure is to (1) Divide the independent variable into 10 or so regions; (2). Calculate the mean and s.d. for each region; (3). Plot log(s.d.) vs. log(mean) for the set of regions; (4). If the plot is a straight line, note its slope, b, then transform the variable by raising the dependent variable to the power (1 - b), and if b = 1, then take the log of the dependent variable; and (5) if there are multiple independents, repeat steps 1 - 4 for each independent variable and pick a b which is the range of b's you get.

    In practice, computer packages apply an iterative maximum-likelihood algorithm to compute lambda, a Box-Cox parameter used to determine the exact power transformation which will best de-correlate the variances and means of the groups formed by the independent variables. As a rule of thumb, if lambda is 1.0, no transformation is needed. A lambda of +.5 corresponds to a square root transform of the dependent variable; lambda of 0 corresponds to a natural log transform; -.5 corresponds to a reciprocal square root transform; and a lambda of -1.0 corresponds to a reciprocal transform. The Box-Cox transformation is not yet supported in SPSS. However, Michael Speed of the Texas A & M University, has made available SPSS syntax code for Box-Cox transformation at http://www.stat.tamu.edu/ftp/pub/mspeed/stat653/spss/, with tutorials at http://www.stat.tamu.edu/spss.php.

    See Box, G. E. P. and D. R. Cox (1964). An analysis of transformations. Journal of the Royal Statistical Society, B, 26, 211-234; see also Maddala, G. S. (1977). Econometrics. New York: McGraw-Hill. (page 315-317); or Mason, R., L., R. F. Gunst, and J. L. Hess (1989). Statistical design and analysis of experiments with applications to engineering and science. New York: Wiley.

• MULTIVARIATE NORMALITY

  • Multivariate normality is when each variable under consideration is normally distributed with respect to each other variable. Multivariate normal distributions take the form of symmetric three-dimensional bells when the x axis is the values of a given variable, the y axis is the count for each value of the x variable, and the z axis is the values of any other variable under consideration. Structural equation modeling and certain other procedures assume multivariate normality, which is discussed further in the section on SEM.

  • Mardia's statistic is a test for multivariate normality. Based on functions of skewness and kurtosis, Mardia's PK should be less than 3 to assume the assumption of multivariate normality is met. PRELIS (companion software for LISREL) outputs PK. SPSS does not yet support Mardia's PK.

• OUTLIERS can radically alter the outcome of analysis and are also violations of normality. Outliers arise from four different causes, requiring different courses of action:
  1. Errors of data entry: proofread your data for out-of-range entries and other errors.

  2. Not defining missing values: check in SPSS or other statpack to make sure don't know, not home, and other missing values are not being treated as real values.

  3. Unintended sampling: eliminate non-population members from the sample (ex., eliminate unintentionally sampled out-of-town house guests from a sample for the population of city residents).

  4. True non-normal distribution: For a true non-normal distribution with extreme values, the researcher may transform the data to pull in outlier values or may choose to analyse extreme cases separately.

  • Simple outliers are cases with extreme values with respect to a single variable. It is common to define outliers as cases which are more than plus or minus three standard deviations from the mean of the variable.

  • Multivariate outliers are cases with extreme values with respect to multiple variables. Multivariate outliers are operationally defined as cases which have a Cook's Distance greater than some cutoff (some use a cutoff of 1; some use 4/[n - p], where p is the number of parameters in the model; some use 4/[n - k - 1], where n is the number of cases and k is the number of independents.) Leverage is another related way of defining multivariate outliers, with outliers defined as having a leverage value greater than some cutoff (some use .5; others use 2p/n, where p is the number of parameters including the intercept). Mahalanobis distance is a third and very common measure for multivariate outliers. Cases with the highest Mahalanobis D-square values are the most likely candidates to be considered outliers and should be examined.

    In SPSS, select Analyze, Regression, Linear; click the Save button; check Cook's, Mahalanobis, and/or leverage values.

• NORMALLY DISTRIBUTED ERROR

  • A histogram of standardized residuals should show a roughly normal curve when the assumption of regression and certain other techniques is met that error terms are normally distributed. SPSS output shows the normal curve superimposed on such histograms. In SPSS, select Graphs, Histogram. Or select Analyze, Regression, Linear; click Plots; check Histogram.

  • A normal probability plot, also called a P-P Plot, is an alternative method, plotting observed cumulative probabilities of occurrence of the standardized residuals on the Y axis and of expected normal probabilities of occurrence on the X axis, such that a 45-degree line will appear when the observed conforms to the normally expected and the assumption of normally distributed error is met. P-P plotd are found in SPSS under Graphs, P-P plots. Or select Analyze, Regression, Linear; click Plots; check Normal probability plot.

  • Kolmogorov-Smirnov and other normality tests.. In multiple regression one can use the Save option to save residuals to a variable which by default is labeled res_1. Other procedures likewise allow residuals to be saved. One can then use the SPSS Examine procedure to apply tests of normality, such as the Kolmogorov-Smirnov (Lilliefors) and Shapiro-Wilks normality tests, to this variable. These tests are discussed above, in the section on testing for normal distribution.

• HOMOGENEITY OF VARIANCES
  • Levene's test of homogeneity of variance tests the assumption that each group (category) of the independent)(s) has the same variance on an interval dependent. If the Levene statistic is significant at the .05 level or better, the researcher rejects the null hypothesis that the groups have equal variances. The Levene test is robust in the face of departures from normality. Levene's test appears in the SPSS procedures ONEWAY and T-TEST as well as EXAMINE. Usually it is based on deviations from the group mean. However, the EXAMINE module's spread-level plot option also prints versions for Levene's test based on deviations from the group median, deviations from the group median adjusted for degrees of freedom, and the 5% trimmed group mean. The Levene test is more robust in the face of non-normality than more traditional tests like Bartlett's test. In SPSS, select Analyze, Compare Means, One-Way ANOVA; click the Options buttion; check Homogeneity of variance test. Example.

  • Brown & Forsythe's test of homogeneity of variances is a recent test based on criticisms of the Levene test. It tests for equality of group means. The Brown-Forsythe test is more robust than the Levine test when groups are unequal in size and the absolute deviation scores (deviations from the group means) are highly skewed, causing a violation of the normality assumption and the assumption of equal variances. In SPSS, select Analyze, Compare Means, One-Way ANOVA; click Options; select Brown-Forsythe.

  • Bartlett's test of homogeneity of variance is an older, alternative test. Bartlett's test is a chi-square statistic with (k-1) degrees of freedom, where k is the number of categories in the independent variable. The Bartlett's test is dependent on meeting the assumption of normality and therefore Levene's test has now largely replaced it.

  • F-max test arises in the context of ANOVA. If the ratio of the largest to smallest size group in ANOVA is not very unequal (ex., is 4:1 or less), and if F-max (the ratio of the variance in the largest group to the variance in the smallest group) is 10:1 or less, then a rule of thumb is that homogeneity of variances is assumed not to be a problem. If groups are more unequal than 4:1, then the ratio of variance may need to be as low as 3:1.

• SPHERICITY
  In a repeated measures design, the univariate ANOVA tables will not be interpreted properly unless the variance/covariance matrix of the dependent variables is circular in form. Sphericity is an assumption of repeated measures MANOVA, for instance. To conclude that sphericity is not violated, Bartlett's test of sphericity should not be significant. In SPSS, choose Analyze, General Linear Model, Multivariate; click the Options button; check Residual SSCP matrix.

• HOMOGENEITY OF VARIANCE-COVARIANCE MATRICES

  • Box's M test tests the multivariate homogeneity of variances and covariances, as required by MANOVA and some other procedures. When M is not significant, the researcher accepts the null hypothesis that groups do not differ. It has been shown to be a conservative test, failing to reject the null hypothesis too often. It also is highly sensitive to violations of multivariate normality. In SPSS, choose Analyze, General Linear Model, Multivariate; click the Options button; check Homogeneity tests.

• HOMOGENEITY OF REGRESSIONS / TEST OF PARALLELISM
  • In ANCOVA and MANCOVA, the slopes of the regression lines should be the same for each group formed by the categorical variables and measured on the dependents. The more this assumption is violated, the more conservative ANCOVA and MANCOVA become (the more likely to make Type II errors - failing to reject null hypotheses, resulting in false negatives).

  • Violation of the homogeneity of regressions assumption indicates an interaction effect between the covariate(s) and the factor(s). When running an ANCOVA or MANCOVA model in SPSS, include in the model options the interactions between the covariate(s) and each independent factor -- any significant interaction effects indicate that the assumption of homogeneity of regression coefficients has been violated.

  • Homogeneity of regression in SPSS can be tested under the Model button of Analyze, General Linear Model, Univariate; select Custom under the Model button; enter a model with all main effects of the factors and covariates and the interaction of the covariate(s) with the factor(s). These interaction effects should be non-significant if the homogeneity of regressions assumption is met.

  • Tests of parallelism are similar in certain procedures, assuring that the slope of regression lines is the same for each level of a categorical grouping variable. Ordinal regression and the probit response model, for instance, come with a test of parallelism in SPSS and other packages.

• HOMOSCEDASTICITY

  • A standardized scatterplot of the standardized predicted dependent variable (ZPRED in SPSS) by the standardized residuals (ZRESID in SPSS, or one may use studentized deleted residuals, SDRESID; or one might use the observed dependent vs. ZRESID or SDRESID) should show a random pattern across the entire range of ZPRED when, in regression, error is homoscedastic -- that is, when the regression model is equally accurate across the range of the dependent. Sometimes a plot of residuals by the actual dependent values is used to test homoscedasticity also. In SPSS, select Analyze, Regression, Linear; click Plots. If ZPRED, ZRESID, or other needed variables have been saved, you can also use Graphs, Scatterplot.

  • Goldfeld-Quandt test is a formal test of homoscedasticity used when errors take a funnel (fan) shape. Though not directly supported by SPSS, this test involves runninve separate regressions on the upper and lower observations in terms of values on the dependent, then conducting an F-test on the ratio of the error sum of squares (ESS) for the two regressions. To the extent that the ratio of the two ESS coefficients approaches 1.0, the homoscedasticity assumption is met.

  • Glejser test is a formal test of homoscedasticity used when errors tale a bow-tie (bimodal) shape and when sample size is not small. The residuals are regressed on the independent variable (or some function of it) and if the independent has a significant b coefficient, then the researcher concludes that homoscedasticity cannot be assumed.

  • Park test is a related test of homoscedasticity. The squared residuals (or their log) is regressed on the independent variable (or its log), and if the independent has a significant b coefficient, then the researcher concludes that homoscedasticity cannot be assumed.

  • Breusch-Pagan-Godfrey test is another large-sample test of homoscedasticity. The squared residuals are divided by (regression sum of squares (RSS) divided by N), giving the generalized residuals. The generalized residuals are then regressed on all independent variables (m variables) suspected of causing heteroscedasticity. The error sum of squares (ESS) divided by 2, for this regression, follows a chi-square distribution with (m - 1) degrees of freedom. If the null hypothesis is rejected, homoscedasticity cannot be assumed.

  • White’s test involves regressing the squared residuals on all independent variables plus all squares of independent variables plus all crossproducts of independent variables (for a total of t predictor variables in all). The resulting R² is multiplied by N, giving a statistic with a chi-square distribution with t degrees of freedom. If the null hypothesis is rejected, homoscedasticity cannot be assumed.

• LINEARITY
  • Testing for nonlinearity is necessary because correlation, regression, and other members of the general linear model (GLM) assume linearity. A variety of methods are available:

  • Graphical methods. Simple inspection of scatterplots is a common if non-statistical method of determining if nonlinearity exists in a relationship. Even better, a plot of standardized residuals against standardized estimates (fitted values) of the dependent variable should show a random pattern when nonlinearity is absent. In regression, as a rule of thumb, an indicator of possible nonlinearity is when the standard deviation of the residuals exceeds the standard deviation of the dependent. Adding to a model nonlinear terms such as squares or cubes of an independent and then seeing if R² in regression or fit indexes in structural equation modeling improve significantly is another way of testing for nonlinearity.

  • Curve fitting with R-squared difference tests. The SPSS curve fitting module (Analyze, Regression, Curve Fit) will calculater R-squared for a linear model compared to a variety of nonlinear models. One can then employ an F test of R-squared difference between models to see if nonlinear model has a significantly higher R-squared than a linear one. Note that some nonlinear models cannot be fitted if there are missing values, so interpolating using the Missing Values module may be a necessary prerequisite if the researcher wishes to compare these models. Warning: this method involves multiple comparisons which make the nominal significance level more lenient than it appears, so a Bonferroni or other adjustment may be necessary depending on the number of comparisons made.

  • ANOVA test of linearity. One can compute an ANOVA table for the linear and nonlinear components of any pair of variables. If the F significance value for the nonlinear component is below the critical value (ex., < .05), then there is significant nonlinearity. In SPSS, select Analyze, Compare Means, Means; click Options; check Linearity test. Likewise, the Contrasts option in ANOVA can be used as a test for the existence of linear, quadratic, and other polynomial relationships.

  • Eta, the correlation ratio. Eta is a coefficient of nonlinear association. For linear relationships, eta equals the correlation coefficient (Pearson's r). For nonlinear relationships it is greater -- hence the difference between eta and r is a measure of the extent of nonlinearity of relationship. Eta is discussed in a separate section on association. In SPSS, select Analyze, Compare Means, Means; click Options; check Anova table and eta.

  • Adding nonlinear terms to a model. A square term corresponds to a nonlinear pattern with a single curve, upward or downward. A cubic term corresponds to a nonlinear up-then-down or down-then-up pattern. In general, the number of curves is one less than the order of the polynomial term -- for instance, the 4th power will produce three curves. A more shotgun approach is to add all possible power terms to the model, then drop those which are non-significant.

  • Ramsey's RESET test (regression specification error test). Ramsey's general test of specification error of functional form is an F test of differences of R² under linear versus nonlinear assumptions. It is commonly used in time series analysis to test whether power transforms need to be added to the model. For a linear model which is properly specified in functional form, nonlinear transforms of the fitted values should not be useful in predicting the dependent variable. While STATA and some packages label the RESET test as a test to see if there are "no omitted variables," it is a linearity test, not a general specification test. It tests if any nonlinear transforms of the specified independent variables have been omitted. It does not test whether other relevant linear or nonlinear variables have been omitted.
    1. Run the regression to obtain R_o², the original multiple correlation.
    2. Save the predicted values (Y's).
    3. Re-run the regression using power functions of the predicted values (ex., their squares and cubes) as additional independents for the Ramsey RESET test of functional form where testing that none of the independents is nonlinearly related to the dependent. Alternatively, re-run the regression using power functions of the independent variables to test them individually.
    4. Obtain R_n², the new multiple correlation.
    5. Apply the F test, where F = ( R_n² - R_o²)/[(1 - R_n²)/(n-p)], where n is sample size and p is the number of parameters in the new model.
    6. Interpret F: For an adequately specified model, F should be non-significant.

  
  

• MULTICOLLINEARITY
  • Multicollinearity in Regression Models is an unacceptably high level of intercorrelation among the independents, such that the effects of the independents cannot be separated. Under multicollinearity, estimates are unbiased but assessments of the relative strength of the explanatory variables and their joint effect are unreliable. (That is, beta weights and R-squares cannot be interpreted reliably even though predicted values are still the best estimate using the given independents). As a rule of thumb, intercorrelation among the independents above .80 signals a possible problem. Likewise, high multicollinearity is signalled when high R-squared and signficant F tests of the model occur in combination with non-significant t-tests of coefficients.

    That is, whereas perfect multicollinearity leads to infinite standard errors and indeterminant coefficients, the more common situation of high multicollinearity leads to large variances and covariances, large confidence intervals, and insignificant significance coefficients. Power is low (the chance of Type II errors is high - thinking you do not have a relationship when in fact one exists - failure to reject the null hypothesis that the coefficients are not different from zero). R-square is high. The coefficients and their standard errors will be sensitive to changes in just a few observations.

    Methods of handling high multicollinearity are discussed elsewhere.

  • Tolerance is defined as 1 - R-squared, where R-squared is the multiple R of a given independent regressed on all other independent variables. If the tolerance value is less than some cutoff value, usually .20, the independent should be dropped from the analysis due to multicollinearity. This is better than just using simple r > .80 since tolerance looks at the independent variable in relation to all other independents and thus takes interaction effects into account as well as simple correlations. In SPSS, select Analyze, Regression, linear; click Statistics; check Collinearity diagnostics.

  • Variance inflation factor, VIF. Note, the variance-inflation factor, VIF, may be used in lieu of tolerance as VIF is simply the reciprocal of tolerance. The rule of thumb is that VIF > 4.0 when multicollinearity is a problem. Some authors use the more lenient cut-off of VIF >= 5 when multicollinearity is a problem. In SPSS, select Analyze, Regression, linear; click Statistics; check Collinearity diagnostics.

  • Condition indices. Discussed more extensively in the section on regression, condition indices over 15 indicate possible multicollinearity problems and over 30 indicate serious multicollinearity problems. In SPSS, select Analyze, Regression, linear; click Statistics; check Collinearity diagnostics.

  • Multicollinearity in Structural Equation Models (SEM):
    
    • Standardized regression weights: Since all the latent variables in a SEM model have been assigned a metric of 1, all the standardized regression weights should be within the range of plus or minus 1. When there is a multicollinearity problem, a weight close to 1 indicates the two variables are close to being identical. When these two nearly identical latent variables are then used as causes of a third latent variable, the SEM method will have difficulty computing separate regression weights for the two paths from the nearly-equal variables and the third variable. As a result it may well come up with one standardized regression weight greater than +1 and one weight less than -1 for these two paths.

    • Standard errors of the unstandardized regression weights: Likewise, when there are two nearly identical latent variables, and these two are used as causes of a third latent variable, the difficulty in computing separate regression weights may well be reflected in much larger standard errors for these paths than for other paths in the model, reflecting high multicollinearity of the two nearly identical variables.

    • Covariances of the parameter estimates: Likewise, the same difficulty in computing separate regression weights may well be reflected in high covariances of the parameter estimates for these paths - estimates much higher than the covariances of parameter estimates for other paths in the model.

    • Variance estimates: Another effect of the same multicollinearity syndrome may be negative variance estimates. In the example above of two nearly-identical latent variables causing a third latent variable, the variance estimate of this third variable may be negative.

• INDEPENDENCE
  • Durbin-Watson coefficient. Independent observations are assumed by many statistical procedures, including multiple and logistic regression. Examples of non-independent observations are before-after studies, panel studies, and paired comparison data. Independence is tested by the Durbin-Watson coefficient, which uses studentized residuals. The Durbin-Watson statistic should be between 1.5 and 2.5 for independent observations. In SPSS regression output, it is found in the "Model Summary" table. In SPSS, select Analyze, Regression, Linear; click Statistics; check Durbin-Watson.

  • Graphical method. Residuals may be plotted against the case identification number, reflecting the order of data collection. There should be no pattern to this plot if observations are independent. In SPSS, select Graphs, Scatterplot.

• RANDOMNESS
  • Runs Test. The runs test is used to test for randomness in a sample. Note that this is a necessary but not sufficient test for random sampling. A non-random availability sample of, say, students in a class, may be a very biased representation of all students in a university, yet within the class the order of sampling may be random and the sample may pass the runs test. On the other hand, if a purportedly random sample fails the runs test, this indicates that there are unusual, non-random periodicities in the order of the sample inconsistent with random sampling. In SPSS, select Analyze, Nonparametric Tests, Runs.

• ADDITIVITY
  • Tukey's Test for Additivity. Tukey's test estimates the power to which items in a set would need to be raised in order to be additive. In SPSS, select Analyze, Scale, Reliability Analysis; click Statistics; check Tukey's test of additivity.

  • Transforms for additivity. If data do not test additive, sometimes this can be corrected by taking the log of y (the dependent) as a transform.

• EQUALITY OF MEANS
  • Hotelling's T-square is a multivariate test for equality of means among items in a dataset. In SPSS, select Analyze, Scale, Reliability Analysis; click Statistics; check Hotelling's T-square.




Bibliography

• Boneau, C. A. (1960). The effect of violation of assumptions underlying the t-test. Psychological Bulletin, 57: 49-64.
• Cohen, Jacob (1969). Statistical power analysis for the behavioral sciences. NY: Academic Press.
• Hutcheson, Graeme and Nick Sofroniou (1999). The multivariate social scientist: Introductory statistics using generalized linear models. Thousand Oaks, CA: Sage Publications. Chapter two covers data screening for assumptions under GLM.
• Lipsey, Mark W. & Wilson, David B. (2001). Practical meta-analysis. Thousand Oaks, CA: Sage Publications.
• Schumacker, Randall E. and Richard G. Lomax (2004). A beginner's guide to structural equation modeling, Second edition. Mahwah, NJ: Lawrence Erlbaum Associates.
• Steenkamp, J.E.M., & van Trijp, H.C.M. (1991). The use of LISREL in validating marketing constructs. International Journal of Research in Marketing, 8: 283-299.
• Vasu, E. S. (1979). Non-normality in regression analysis: Monte Carlo investigation under the condition of multicollinearity. Working papers in methodology. Number 9. Institute for Research in Social Science, University of North Carolina. Chapel Hill, North Carolina, USA.

Copyright 1998, 2008 by G. David Garson.
Last update 5/20/08.