If you were a researcher

What you care about

As a researcher you want a result you can trust and a clear idea of how far it reaches. You care about validity, whether the study measures what it claims to, and reliability, whether the measurement is consistent. You also care about whether others can repeat the study and get the same result.

What worries you about the original study

• Scorers who could guess the condition. The male model scored every session. The authors admit that children who had seen the aggressive model could be "readily identified" by their behaviour.
• One model of each sex. Any difference between "the male model" and "the female model" could be due to those two people's size, voice or style.
• Small groups. Six children per group gives little power to detect real differences and makes chance differences look meaningful.
• Learning and doing are mixed up. The test counted only what children chose to do, not what they had learned.
• A measure that may not be aggression. Hitting a toy built to be hit.
• No delayed test. Nothing was measured after the session.

Your redesign

1. Get ethics approval first. Obtain parental consent and each child's assent, and drop or replace the deceptive frustration step, as described on the participant page. A design an ethics committee would reject is not a usable redesign.
2. Video-record every session and have it scored by two trained coders who do not know the child's condition.
3. Use several models. At least three men and three women, each filmed performing the identical script, with children randomly assigned to models. Now differences between male and female models cannot be blamed on one person.
4. Plan the sample size in advance. Use a power analysis, a calculation of how many children are needed to detect an effect of a given size, and register the hypotheses and analysis plan publicly before collecting data. This is pre-registration.
5. Separate learning from performance. After free play, offer a small reward for each act the child can show "the grown-up did", as Bandura did in 1965.
6. Add a measure involving a person. For example, a game in which the child can choose to take stickers from, or share stickers with, another child who is not present. This gives a measure of harm or kindness towards a person, without any real person being hurt.
7. Test again after one week, using the same room and toys.
8. Run it in several places, including at least one site outside the United States, with the same written protocol.
9. Share the data and materials so others can check and repeat the work.

Why

Each change answers a specific weakness. Blind coding deals with observer bias. Several models separate the sex of a model from the individual. A planned, larger sample guards against chance results. Psychology has learned through the replication crisis of the 2010s that many famous findings were weaker than they first seemed, often because of small samples and flexible analysis. Pre-registration and open data are now standard protections.

The incentive test answers the most important theoretical question the 1961 study left open. Bandura's 1965 study found that when children were offered rewards to show what the model did, differences between the groups largely disappeared. Children who had copied little in free play had still learned the acts. A modern study should measure both.

What your redesign gives up

• It stays in the laboratory. Blind coding, scripted models and fixed rooms keep the artificial setting, so ecological validity remains limited.
• It is slower and more expensive, with several models, many more children, video equipment and several sites.
• The person-based measure is still a game. Taking stickers from an absent child is closer to real harm than hitting a toy, but it is still not the same as real aggression.