Interactions
Using SPSS to Test Interaction Effects: A Factorial ANOVA and Moderation Analysis
Introduction
Customer behaviour can have important consequences for employees who interact directly with customers. But does an apology repair the damage caused by a difficult customer, and does its effectiveness depend on whether the customer’s behaviour was an honest mistake or a deliberate act? This was another quantitative research methods exercise I completed during the first trimester of my PhD in 2026. The analysis provided an opportunity to move beyond simple comparisons of group means and examine interaction and moderation effects using factorial analysis of variance in IBM SPSS Statistics.
The exercise draws on experimental data associated with research by Robertson, Rotman, McQuilken and Ringer on customer failures, apologies and frontline service employee well-being. The data is unpublished and unavailable outside of class use.
Research Problem
Participants were randomly assigned to one of four customer-service scenarios. The customer either:
- made an honest pricing mistake and apologised;
- deliberately switched a price and apologised;
- made an honest pricing mistake without apologising; or
- deliberately switched a price and did not apologise.
The central question was whether the effect of a customer’s apology on employee job satisfaction depended on whether the original behaviour represented an honest mistake or intentional misconduct.
The analysis also examined whether having personally experienced a similar customer incident further moderated this relationship, and whether controlling for gender changed the findings.
Analytical Approach
Two new binary variables were created in SPSS from the original four-condition variable:
Apology
- 1 = Customer apologised
- 0 = Customer did not apologise
Customer behaviour
- 1 = Honest mistake
- 0 = Intentional price switching
The primary analysis used a 2 × 2 factorial ANOVA, with job satisfaction entered as the dependent variable and apology and customer behaviour entered as fixed factors.
The critical statistical test was the Apology × Customer Behaviour interaction. Simple main effects were subsequently examined to determine precisely where any significant interaction occurred.
Prior experience with a similar customer situation was then added as a third factor to test a three-way interaction. Gender was subsequently included as a categorical control variable.
Method
The dataset contained 518 participants, distributed almost evenly across the four experimental conditions.
ConditionnHonest mistake + apology130Intentional behaviour + apology132Honest mistake + no apology126Intentional behaviour + no apology130
Participants ranged from 19 to 68 years of age (M = 34.50, SD = 9.75). The sample included 296 men (57.1%), 216 women (41.7%), and six participants (1.2%) identifying as another gender or preferring not to specify.
Of the 516 participants who answered the prior-experience question, 301 reported that a similar situation had previously happened to them at work, while 215 reported that it had not.
Job satisfaction was assessed using three seven-point items concerning participants’ anticipated feelings about their job, organisation, and work following the customer encounter. The three-item measure demonstrated high internal consistency, Cronbach’s α = .90, and the items were combined into an overall job-satisfaction score.
Results
Apology × Customer Behaviour Interaction
Mean job satisfaction differed across the four experimental conditions.
Table 1
Job Satisfaction by Experimental Condition
The factorial ANOVA revealed a statistically significant Apology × Customer Behaviour interaction, F(1, 514) = 4.74, p = .030, partial η² = .009.
Get Chimdi Chukwukere’s stories in your inbox
Join Medium for free to get updates from this writer.
There was also a significant main effect of apology, F(1, 514) = 7.90, p = .005, partial η² = .015, and a significant main effect of customer behaviour, F(1, 514) = 18.23, p < .001, partial η² = .034. Because these main effects were qualified by the interaction, however, the simple effects provide the more meaningful interpretation.
Levene’s test was significant, F(3, 514) = 5.37, p = .001, indicating that the assumption of homogeneity of variance was violated. The experimental groups were nevertheless very similar in size, ranging from 126 to 132 participants, although the variance violation should still be considered when interpreting the results.
Simple Main Effects
When the customer had made an honest mistake, participants reported significantly greater job satisfaction when the customer apologised (M = 5.16) than when the customer did not apologise (M = 4.51), F(1, 514) = 12.29, p < .001.
By contrast, when the customer had intentionally switched the price, an apology produced virtually no difference. Job satisfaction was similar when the customer apologised (M = 4.32) and when the customer did not apologise (M = 4.24), F(1, 514) = 0.20, p = .653.
The implication is clear: an apology improved job satisfaction following an honest mistake, but it did not meaningfully repair the effect of deliberate customer misconduct.
Figure 1
Interaction of Customer Apology and Behaviour on Job Satisfaction
Did Previous Experience Further Moderate the Effect?
Prior experience was then introduced as a third factor. Two participants had missing values on this measure, resulting in an analytical sample of 516 participants for this model.
The Apology × Customer Behaviour × Prior Experience interaction was statistically significant, F(1, 508) = 4.07, p = .044, partial η² = .008.
Among participants who had not previously experienced a similar workplace situation, the Apology × Customer Behaviour interaction was significant, F(1, 211) = 8.18, p = .005, partial η² = .037. For these participants, an apology substantially improved job satisfaction following an honest mistake but provided virtually no benefit following intentional misconduct.
Among participants who had previously experienced a similar situation, the corresponding interaction was not significant, F(1, 297) = 0.03, p = .865.
Figure 2
Prior Experience and the Apology × Customer Behaviour Interaction
What Happens When Gender Is Controlled?
Gender was then included as a categorical control variable.
Gender itself was significantly associated with job satisfaction, F(2, 512) = 7.82, p < .001, partial η² = .030. Men reported higher overall job satisfaction (M = 4.76) than women (M = 4.26), although the very small number of participants in the third gender category means comparisons involving that group should be treated cautiously.
Importantly, controlling for gender did not eliminate the primary Apology × Customer Behaviour interaction. The interaction remained statistically significant, F(1, 512) = 5.97, p = .015, partial η² = .012.
Gender therefore did not explain away the central finding that the effectiveness of a customer apology depended on whether the original behaviour was accidental or intentional.
The more complex three-way interaction was less robust. After controlling for gender, the Apology × Customer Behaviour × Prior Experience interaction declined to F(1, 506) = 3.81, p = .051, partial η² = .007.
It therefore fell just outside the conventional threshold of statistical significance.
Conclusions and Implications
The analysis demonstrates why interaction effects are important: examining only the overall effect of an apology would have hidden an important difference between types of customer behaviour.
Overall, an apology was associated with higher anticipated job satisfaction, but its effectiveness depended on what the customer had done. When the incident resulted from an honest mistake, an apology meaningfully improved job satisfaction. When the customer had deliberately behaved dishonestly by switching a price, apologising afterwards made little difference.
The findings therefore suggest that employees may distinguish between accidental mistakes and intentional misconduct when evaluating the restorative value of an apology. An apology may help repair an honest error but may be insufficient when the original behaviour was perceived as deliberately dishonest.
Previous experience initially appeared to further moderate the relationship. However, the three-way interaction became just non-significant after gender was controlled. This secondary result should therefore be interpreted more cautiously than the primary two-way interaction.
Because participants were randomly assigned to the four experimental conditions, the study provides stronger grounds for causal inference regarding the experimentally manipulated customer scenarios than would be possible in a purely observational study. Within the context of the experiment, the combination of customer behaviour and apology influenced anticipated job satisfaction.
However, prior experience and gender were measured rather than experimentally manipulated, so their relationships with job satisfaction should not be interpreted causally. The dependent variable also measured anticipated responses to a hypothetical scenario rather than actual long-term job satisfaction or real workplace behaviour.
Overall, the exercise demonstrates a central principle of moderation analysis: the effect of one variable can depend substantially on the level of another variable, and examining interaction effects can reveal relationships that simple main-effect analyses may overlook.
Reference
Robertson, N., Rotman, J., McQuilken, L., & Ringer, A. (2023). The customer is often wrong: Investigating the influence of customer failures and apologies on frontline service employee well-being. Psychology & Marketing, 40(4), 825–844.
