# SPSS tutorials

BASICS DATA ANALYSIS T-TEST ANOVA CHI-SQUARE TEST

# SPSS Two-Way ANOVA with Interaction Tutorial

Do you think running a two-way ANOVA with an interaction effect is challenging? Then this is the tutorial for you. We'll run the analysis by following a simple flowchart and we'll explain each step in simple language. After reading it, you'll know what to do and you'll understand why.
We'll use depression.sav throughout. The screenshot below shows it in variable view.

## Research Questions

Very basically, 100 participants suffering from depression were divided into 4 groups of 25 each. Each group was given a different medicine. After 4 weeks, participants filled out the BDI, short for Beck's depression inventory. Our main research question is: did our different medicines result in different mean BDI scores? A secondary question is whether the BDI scores are related to gender in any way. In short we'll try to gain insight into 4 (medicines) x 2 (gender) = 8 mean BDI scores.

## Quick Check: Histogram over Scores

Before jumping blindly into statistical tests, let's first see if our BDI scores make any sense in the first place. Before analyzing any metric variable, I always first inspect its histogram. The fastest way to create it is running the syntax below.

*Inspect histogram, see if distribution for bdi scores looks plausible.

frequencies bdi
/format notable
/histogram.

## Result

The scores look fine. We've perhaps one outlier who scores only 15 but we'll leave it in the data. The scores are roughly normally distributed and there's no need to specify any missing values.

## ANOVA Assumptions

When we compare more than 2 means, we usually do so by ANOVA -short for analysis of variance. Doing so requires our data to meet the following assumptions:

• Independent observations (or independent and identically distributed variables). This often holds if each case contains a distinct person and the participants didn't interact.
• Homogeneity: the population variances are all equal over subpopulations. Violation of this assumption is less serious insofar as sample sizes are equal.
• Normality: the test variable must be normally distributed in each subpopulation. This assumption becomes less important insofar as the sample sizes are larger.

## Inspecting Means and Sample Sizes

The first question in our ANOVA flowchart is whether the sample sizes are roughly equal. I like to run a means table for inspecting this because I'm going to need this table anyway for my report. I'll create it with the syntax below.

*Run means table and inspect if sample sizes are roughly equal.

means bdi by gender by medicine
/cells count mean stddev.

*Note: MEANS allows you to choose exactly which statistics you want.

## Result

Note that this table shows the 8 means (2 genders * 4 medicines) that our analysis is all about. Each of these 8 means is based on 10 through 15 observations so the sample sizes are roughly equal.
This means that we don't need to bother about the homogeneity assumption. We can therefore skip Levene's test as shown in our flowchart.

## Running Two-Way ANOVA in SPSS

We'll now run our two-way ANOVA through Analyze General Linear Model Univariate. We'll then follow the screenshots below.

This results in the syntax below. Let's run it and see what happens.

*SPSS Two-Way ANOVA syntax as pasted from screenshots.

UNIANOVA bdi BY gender medicine
/METHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/POSTHOC=medicine(TUKEY)
/PLOT=PROFILE(medicine*gender) TYPE=LINE ERRORBAR=NO MEANREFERENCE=NO YAXIS=AUTO
/PRINT ETASQ HOMOGENEITY
/CRITERIA=ALPHA(.05)
/DESIGN=gender medicine gender*medicine.

## ANOVA Output - Between Subjects Effects

Following our flowchart, we should now find out if the interaction effect is statistically significant. A -somewhat arbitrary- convention is that an effect is statistically significant if “Sig.” < 0.05. According to the table below, our 2 main effects and our interaction are all statistically significant.

The flowchart says we should now rerun our ANOVA with simple effects. For now, we'll ignore the main effects -even if they're statistically significant. But why?! Well, this will become clear if we understand what our interaction effect really means. So let's inspect our profile plots.

## ANOVA Output - Profile Plots

The profile plot shown below basically just shows the 8 means from our means table.If you ran the ANOVA like we just did, the “Estimated Marginal Means” are always the same as the observed means that we saw earlier. Interestingly, it also shows how medicine and gender affect these means.

An interaction effect means that the effect of one factor depends on the other factor and it's shown by the lines in our profile plot not running parallel.
In this case, the effect for medicine interacts with gender. That is, medicine affects females differently than males. Roughly, we see the red line (females) descent quite steeply from “None” to “Pharmaceutical” whereas the blue line (males) is much more horizontal. Since it depends on gender, there's no such thing as the effect of medicine. So that's why we ignore the main effect of medicine -even if it's statistically significant. This main effect “lumps together” the different effects for males and females and this obscures -rather than clarifies- how medicine really affects the BDI scores.

## Interaction Effect? Run Simple Effects.

So what should we do? Well, if medicine affects males and females differently, then we'll analyze male and female participants separately: we'll run a one-way ANOVA for just medicine on either group. This is what's meant by “simple effects” in our flowchart.

## ANOVA with Simple Effects - Split File

How can we analyze 2 (or more) groups of cases separately? Well, SPSS has a terrific solution for this, known as SPLIT FILE. It requires that we first sort our cases so we'll do so as well.

*Interaction statistically significant? Go for simple effects: run one-way ANOVAs for male and female participants separately.

sort cases by gender.

split file separate by gender.

Minor note: SPLIT FILE does not change your data in any way. It merely affects your output as we'll see in a minute. You can simply undo it by running SPLIT FILE OFF. but don't do so yet; we first want to run our one-way ANOVAs for inspecting our simple effects.

## ANOVA with Simple Effects in SPSS

Since we switched on our SPLIT FILE, we can now just run one-way ANOVAs. We'll use Analyze General Linear Model Univariate. The screenshots below guide you through the next steps.

This results in the syntax below. Let's run it.

*Simple effects: SPLIT FILE by gender and run one-way ANOVAs for males and females separately.

UNIANOVA bdi BY medicine
/METHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/POSTHOC=medicine(TUKEY)
/PLOT=PROFILE(medicine) TYPE=LINE ERRORBAR=NO MEANREFERENCE=NO YAXIS=AUTO
/PRINT ETASQ HOMOGENEITY
/CRITERIA=ALPHA(.05)
/DESIGN=medicine.

## ANOVA Output - Between Subjects Effects

First off, note that the output window now contains all ANOVA results for male participants and then a similar set of results for females. According to our flowchart we should now inspect the main effect.

The effect for medicine is statistically significant. However, this just means it's probably not zero. But it's not very strong either as indicated by its partial eta squared of 0.208. This should't come as a surprise. The 4 medicines don't differ much for males as we saw in our profile plots.

## ANOVA Output - Post Hoc Tests

Our main effect suggests that our 4 medicines did not all perform similarly. But which one(s) really differ from which other one(s)? This question is addressed by our post hoc tests, in this case Tukey's HSD (“honestly significant differences”) comparisons.

This table compares each medicine with each other medicine (twice). The only comparison yielding p < 0.05 is “None” versus “Pharmaceutical”. Our profile plots show that these are the worst and best performing medicines for male participants. The means for all other medicines are too close to differ statistically significantly.

## Female Participants

According to our flowchart, we're now done -at least for males. Interpreting the results for female participants is left as an exercise for the reader. However, I do want to point out the following:

• our profile plots show a much steeper line for females than for males.
• the main effect of medicine has a much higher partial eta squared of 0.63 for females.
• the effect of medicine has p = 0.000 for females and p = 0.019 for males.
• for females, all post hoc comparisons are statistically significant except for “Homeopathic” versus “Placebo” (p = 0.997).

All these findings indicate a much stronger effect of medicine for females than for males; there's a substantial interaction effect between medicine and gender on BDI scores. And this -again- is the reason why we need to analyze these groups separately and rerun our ANOVA with simple effects -like we just did.

## Final Notes

Do you still think running a two-way ANOVA with an interaction effect is challenging? I hope this tutorial helped you understand the main line of thinking. And -hopefully!- things start to sink in for you -perhaps after a second reading.

# Let me know what you think!

*Required field. Your comment will show up after approval from a moderator.

# This tutorial has 19 comments

• ### By Ruben Geert van den Berg on November 8th, 2019

Hi Deven, good comment!

I only ran the initial histogram to see if the frequency distribution looks plausible in the first place and if any extreme values should be excluded. That wasn't the case: the data look credible.

I don't really care about the normality assumption for the ANOVA: I don't need normality because my sample sizes are reasonable.

I don't see any reason for not starting off with the factorial ANOVA. You may guess that there may be an interaction but you don't know if it's statistically significant and you don't know its effect size -partial eta squared- either.

The factorial ANOVA tells us that -and more- and thus provides for a better informed decision whether to (not) proceed with main effects or simple effects.

Does that make any sense to you? Or did I miss something?

Thanks!!

SPSS tutorials

• ### By DJR on November 7th, 2019

You ran the histogram with all subjects included and then concluded that it is roughly normal, but the means plots and the large interaction effect led you down the path of separating by gender.

Should you not have then re-run the histogram(s) based on that split? My guess based on the means plot is that the original histogram is actually bi-modal rather than normally distributed with different mean and sigma for each gender. In which case I probably woudn't have run the two factor ANOVA at all, but would have stuck with the single factor ANOVA split by gender.

• ### By Lester Gilbert on May 16th, 2019

Nice tutorial, but sadly doesn't take us all the way.
(1) The error term for the one-way ANOVAs should be set equal to the overall error as reported by the two-way. We assumed equal variances, and we get a larger error df.
(2) A second set of one-way ANOVAs is needed for the second factor. In the example, the gender effect needs to be considered for each of the four medicines.
(3) These are indeed "simple main effects". Given a significant two-way interaction, we need to look at the simple main effects of the first factor at each of the categories of the second, and then at the simple main effects of the second factor at each of the first factor categories. Consider a three-way ANOVA. A significant three factor ABC interaction needs us to consider the simple interaction effects AB at each C, AC at each B, and BC at each A. If the simple interaction of AB is significant at, say, C2, then we need the simple simple (sic) main effects of A at C2 and each level of B, and of B at C2 and at each level of A...

• ### By Rossel on May 15th, 2019

I have a design with two factors, each of 3 levels, and I need to obtain the simple effects. How do you proceed?

Thanks