This document provides an example of conducting a one-way ANOVA in SPSS to compare teacher satisfaction across three school types (public, private, and charter). The ANOVA results show a significant effect of school type on satisfaction. Due to small and unequal sample sizes, the Welch's test was also significant. Post hoc Tukey and Games-Howell tests found charter teachers reported significantly higher satisfaction than public or private teachers, though Games-Howell only marginally significant for public vs. charter. The example illustrates how to conduct follow-up tests after a significant ANOVA.
This document provides an example of conducting a one-way ANOVA in SPSS to compare teacher satisfaction across three school types (public, private, and charter). The ANOVA results show a significant effect of school type on satisfaction. Due to small and unequal sample sizes, the Welch's test was also significant. Post hoc Tukey and Games-Howell tests found charter teachers reported significantly higher satisfaction than public or private teachers, though Games-Howell only marginally significant for public vs. charter. The example illustrates how to conduct follow-up tests after a significant ANOVA.
Original Description:
Statistical analysis of data. With exples for research papers and projects.
This document provides an example of conducting a one-way ANOVA in SPSS to compare teacher satisfaction across three school types (public, private, and charter). The ANOVA results show a significant effect of school type on satisfaction. Due to small and unequal sample sizes, the Welch's test was also significant. Post hoc Tukey and Games-Howell tests found charter teachers reported significantly higher satisfaction than public or private teachers, though Games-Howell only marginally significant for public vs. charter. The example illustrates how to conduct follow-up tests after a significant ANOVA.
This document provides an example of conducting a one-way ANOVA in SPSS to compare teacher satisfaction across three school types (public, private, and charter). The ANOVA results show a significant effect of school type on satisfaction. Due to small and unequal sample sizes, the Welch's test was also significant. Post hoc Tukey and Games-Howell tests found charter teachers reported significantly higher satisfaction than public or private teachers, though Games-Howell only marginally significant for public vs. charter. The example illustrates how to conduct follow-up tests after a significant ANOVA.
Psy 521/621 Univariate Quantitative Methods, Fall 2020 1
ANOVA Example
Below is the output for the SPSS ONEWAY procedure to compare the means of three school types in the hypothetical teacher satisfaction example. 1 This is a pretty small sample size per group and such a small sample is not necessarily recommended. It is certainly legitimate to do an ANOVA with this size sample, but one should be particularly conscious of unequal variances.
Syntax
means tables=satisfaction by school.
oneway satisfaction by school /statistics=welch.
Menus Analyze—Compare Means—One-way ANOVA
Oneway
ANOVA satisfaction
df Mean Square F Sig.
Between Groups 30.000 2 15.000 8.182 .006
Within Groups 22.000 12 1.833
Total 52.000 14
Robust Tests of Equality of Means
satisfaction
Statistic a df1 df2 Sig.
Welch 14.259 2 7.211 .003
a.
Unequal group sizes, very different variances, and small sample sizes can increase Type I error rates (or sometimes increase Type II error rates; Clinch & Keselman, 1982; Tomarken & Serlin, 1986). If the ratio of variances in different groups is less than 4 to 1 and the group sizes are over 5, there is generally a very minor impact on Type I error rates. If data are very non-normal, variances are very unequal, and/or the group sample sizes are small and unequal, the Welch's robust test is an alternative that will generally have a more accurate Type I error rate.
1 It is also possible to do the same one-way ANOVA using several other procedures in SPSS, including the GLM through the menus or GLM, ANOVA, or MANOVA (syntax only) commands. Use of the GLM procedure, glm satisfaction by school., will give the eta-squared (shown as R-squared in output), but the Welch's test is not an option. Newsom Psy 521/621 Univariate Quantitative Methods, Fall 2020 2
Although the group sizes are equal, the small overall sample size and the relatively substantial differences in the standard deviations (and therefore larger differences in the variances) suggest that it may be wise to consider the Welch's test in this case (Welch, 1951). The statistical conclusion does not differ from the standard F-test, however (this could be noted in a write-up, where either or both is reported, if this comes up). With larger Ns, as in most of the applications you will encounter, the correction would likely be unnecessary.
R library(lessR) #do this at the beginning of your run and always run it before you open a data set rm(mydata) library(haven) mydata = read_sav("c:/jason/spsswin/uvclass/school anova.sav")
#note: you may need to add data=mydata after brief=TRUE
Factor Variable: school Levels: 1 2 3 Number of cases (rows) of data: 15 Number of cases retained for analysis: 15
DESCRIPTIVE STATISTICS n mean sd min max 1 5 6.00 2.00 4.00 8.00 2 5 9.00 1.00 8.00 10.00 3 5 6.00 0.71 5.00 7.00 Grand Mean: 7
BASIC ANALYSIS
Summary Table df Sum Sq Mean Sq F-value p-value school 2 30.00 15.00 8.18 0.0057 Residuals 12 22.00 1.83
Association and Effect Size
R Squared: 0.58 R Sq Adjusted: 0.51 Omega Squared: 0.49 Cohen's f: 0.98
Write-up Teacher satisfaction differed significantly among the three groups, F(2,12) = 8.18, p = .006. Because of the substantial differences in variances among groups and the small sample size, the Welch's test was conducted. The results from the Welch's test also indicted a significant difference among the groups, FW (2,12) = 14. 26, p = .003. 2 Teachers were most satisfied in charter schools (M = 9.00, SD = 1.00), and less satisfied in public (M = 6.00, SD=2.00) and private schools (M = 6.00, SD = .71). Approximately 58% of the variance in satisfaction was accounted for by school type, η2 = .58.
References Welch, B. L. (1951). On the comparison of several mean values: An alternative approach. Biometrika, 38, 330-336.
2 The Welch test sometimes may be referred to as W. It is not technically and F, but the F distribution is used as the referent distribution when determining significance. Newsom Psy 521/621 Univariate Quantitative Methods, Fall 2020 3
Post Hoc Examples
The results from the one-way ANOVA do not indicate which of the three groups differ from one another, so, in many cases, it is of interest to follow the analysis with a post hoc test or a planned comparison among particular means. If several comparisons between pairs of means are made, it is a good idea to use a test, such as the Tukey, that controls for alpha inflation.
For the present example, there are not too many possible comparisons to make (just 3), so alpha inflation will not be a tremendous issue. Also, in the case of this artificial example, two of the means are exactly equal and should not require a statistical test. Nevertheless, to illustrate the use of the Tukey post hoc approach. I’ve also included the Games-Howell test in SPSS, because very small sample sizes may be particularly sensitive to the homogeneity of variance assumption.
SPSS
Menus Analyze -> Compare Means -> One-way ANOVA then choose the dependent variable and factor, click Post Hoc and then check Tukey and Games-Howell.
Syntax To obtain the post hoc tests with syntax, the POST HOC subcommand is added.
school school type N 1 2 Tukey HSDa 1.000 public 5 6.0000 3.000 private 5 6.0000 2.000 charter 5 9.0000 Sig. 1.000 1.000 Means for groups in homogeneous subsets are displayed. a. Uses Harmonic Mean Sample Size = 5.000.
R #using the full ANOVA command gives the Tukey follow-up tests ANOVA(satisfaction ~ school)
TUKEY MULTIPLE COMPARISONS OF MEANS
Newsom Psy 521/621 Univariate Quantitative Methods, Fall 2020 4
Write-up The Tukey post hoc tests indicated that teacher satisfaction in the charter schools differed significantly from teacher satisfaction in public, p = .01, and private schools, p = .01. The Games-Howell test, however, suggested that the difference between public and charter schools was only marginally significant, p = .056.
