Piliavin (1969) Subway Samaritan Study

The Piliavin, Rodin, and Piliavin (1969) subway Samaritan study tested the bystander effect by staging a stranger’s collapse in public.

It was a covert field experiment. That means passengers on the New York underground did not know they were part of a study.

Passengers helped an apparently ill victim on 62 of 65 trials, far more often and faster than a drunk victim, and larger crowds did not reduce helping.

The study also produced the arousal:cost–reward model: distress arouses a bystander, and a quick weighing of costs and rewards decides the response.

Aim

This study investigated how bystanders react when a stranger collapses on a train.

The researchers tested four factors thought to influence helping:

• Would an ill person get more help than a drunk person (the type of victim)?
• Would people help others of the same race before helping those of different races?
• If a model person started helping the victim, would that encourage others to also help?
• Would the number of bystanders who saw the victim influence how much help was given?

Background: Bystander Studies Before 1969

The subway study belongs to a wave of research triggered by the 1964 murder of Kitty Genovese in New York. Newspapers reported, inaccurately, that dozens of neighbours had watched and done nothing.

John Darley and Bibb Latané responded with controlled laboratory experiments. They found that the more bystanders were present, the less likely, and the slower, any one person was to help.

Would that hold in real life?

Laboratory Evidence for Inhibited Helping

In Darley and Latané’s (1968) staged seizure study, participants believed they were the only witness to another student’s apparent epileptic fit. They helped almost every time.

Participants who believed four other bystanders were present helped far less often, and far more slowly.

This is diffusion of responsibility: the tendency to feel less personally responsible for helping as more bystanders are present, because responsibility is shared.

A companion study found that passive others also stopped people reacting to smoke pouring into a room (Latané & Darley, 1968). They called this group inhibition, or pluralistic ignorance. Each person reads everyone else’s calm as a sign that nothing is wrong.

A third study bridged the two designs.

In Latané and Rodin’s (1969) “lady in distress” experiment, a woman fell and cried out in an adjoining room. Bystanders could see and hear each other, but not the victim.

Larger groups again helped less often and more slowly. Strangers inhibited one another more than friends did. Friends could read each other’s concern, rather than assuming a stranger’s calm meant nothing was wrong.

Why Test Helping on a Subway?

Piliavin et al. (1969) asked whether these laboratory effects would survive a real, unavoidable, face-to-face emergency.

The subway offered three conditions the laboratory studies could not fully reproduce:

• Enclosed Space: strangers could not easily leave.
• Visible Victim: the collapse happened directly in front of them.
• Clear Emergency: there was no ambiguity about what was wrong.

The design also went one step beyond Latané and Rodin (1969). On the trains, bystanders could see both the victim and each other.

None of the earlier laboratory studies had tested that combination.

The researchers also varied whether the victim seemed ill or drunk, the victim’s race, and the behaviour of a planted model helper. Group size varied naturally from trial to trial, so helping could be compared across crowds of different sizes.

Procedure

This study was a field experiment on a 7 ½ minute non-stop journey on a New York underground train, using various coaches along the train. The journey was the express run on the Eighth Avenue line between 59th Street and 125th Street.

Participants were passengers who were on board.

A team of 4 university students (a male victim, a male model, and 2 female observers) staged the emergency to observe how passengers would react.

About 70 seconds in, the ‘victim’ staggered forward and collapsed. This staged ‘emergency’ happened in the designated ‘critical area’, the part of the carriage where passengers’ helping was recorded.

After collapsing, the victim lay on his back on the floor. If not helped earlier in the journey by a participant or model, the model assisted the victim at the end of the journey.

Participants’ reactions were then watched by covert observers.

Participants

The victims were males aged 26–35. Three were white and one was black. All were identically dressed in a US army-style jacket, old trousers, and no tie.

The ‘drunk’ smelled of alcohol and carried a spirits bottle wrapped in a brown paper bag (38 trials). The ‘ill’ victim appeared sober and carried a black cane (65 trials).

The models were males aged 24–29. They wore casual, non-identical clothes. They helped by raising the victim to a sitting position and staying with him.

Findings

• Spontaneous Help: models were rarely needed, as passengers usually helped quickly on their own.
• Ill vs Drunk: the ill victim was helped on 62 of 65 trials, the drunk victim on only 19 of 38.
• Speed of Help: median latency was about 5 seconds for the ill victim and roughly 109 seconds for the drunk victim.
• Little Diffusion: group size made no meaningful difference to whether, or how quickly, the victim was helped. If anything, more bystanders sometimes meant slightly faster help.
• Gender: males are more likely to help than females, making up 60% of travelers but 90% of first helpers.
• Race: race had little overall effect, although a drunk victim was less likely to receive opposite-race help.
• Avoidance: the longer no help was offered, the less important the model became and the more likely bystanders were to leave the area, especially with the drunk victim.
• Spontaneous Comments: these were more common in the drunk condition.

Help for the ill victim usually arrived within seconds, well before the model’s 70-second cue. That left little time for any group process of diffusion to develop.

The Arousal:Cost–Reward Model

Piliavin et al. (1969) proposed the arousal:cost–reward model. It is a major alternative to Latané and Darley’s decision model. They called it a ‘fine tuning’ of that earlier account. Like the decision model, it has two stages before a bystander acts.

Stage 1: Emotional Arousal

Seeing someone in distress produces an unpleasant emotional response. This physiological arousal is the model’s basic motivational engine.

Without arousal, a bystander has little reason to act. The greater the arousal, the more likely a bystander is to help.

Acting is simply the fastest way to switch the discomfort off. Several factors intensify this arousal.

Empathy with the victim raises it further. Physical closeness to the emergency matters too. So does how long the crisis drags on.

Piliavin et al. (1969) grounded this first stage in Schachter’s (1964) theory of emotion. Raw physiological arousal does not arrive with a label. The situation tells a bystander how to read it: as fear, disgust, sympathy, or a mixture.

On the subway, the same jolt could read as sympathy for the sober victim with the cane. It could read as wary disgust for the man who smelled of drink.

The physiological jolt is much the same either way. That may help explain why the two victims drew such different responses.

Arousal is the engine. It is not yet the steering wheel.

Stage 2: The Cost-Reward Calculation

Arousal supplies the motivation to act. A second, cognitive stage decides exactly what to do.

The bystander weighs the costs and rewards of each option. This often happens fast, and not fully consciously.

They pick whichever option cuts the arousal fastest, for the fewest costs.

• Costs of helping: effort, time, lost resources, risk of harm, and unpleasant emotions such as disgust or embarrassment.
• Rewards of helping: gratitude from the victim and onlookers, social approval, and the self-satisfaction of having acted.
• Costs of not helping: guilt, the disapproval of others, and damage to one’s self-esteem.

These weights differ between people. So the model predicts patterns, not certainties.

Help was more likely for the ill victim than the drunk one. The drunk victim raised the cost, and lowered the sympathy, of helping.

Men offered help more often too. Their social role reduces self-blame, even though their perceived risk is higher.

Same-race bystanders were also likelier to help a drunk victim. Perceived risk is higher then, and same-race trust may matter.

A Worked Example

Picture a commuter watching each victim collapse in turn. The commuter is trapped in the carriage, so walking away will not switch the arousal off. The same model, with the costs re-weighted, predicts both outcomes.

	Ill victim (cane)	Drunk victim (bottle)
Cost of helping	Low: modest effort, little personal risk	Higher: the smell and bottle add disgust and the risk of an unpredictable reaction
Cost of not helping	High: guilt, and disapproving looks in an enclosed space	Lower: he seems partly responsible for his own state, which lowers sympathy
Cheapest way to cut arousal	Step forward and help	Look away, or reinterpret him as “just a drunk”
Result	Helped on 62 of 65 trials	Helped on 19 of 38 trials, and more slowly

Later Development of the Model

Piliavin et al.’s original two-stage sketch was elaborated over the following decade.

In a further subway experiment, Piliavin, Piliavin, and Rodin (1975) gave the victim a disfiguring facial birthmark. This raised the perceived cost of approaching him.

Help fell to about 61% of trials.

Diffusion of responsibility, which the 1969 study had failed to detect, began to reappear too. It returned once the costs of helping rose.

The model was elaborated most fully in the book Emergency Intervention (Piliavin, Dovidio, Gaertner, & Clark, 1981).

It added many interacting factors: the bystander’s own state, and the victim’s characteristics.

Together, these shape both how aroused a bystander becomes, and how they weigh the costs and rewards of helping.

Dovidio et al. (1991) reviewed two decades of evidence.

They concluded that emotional reactions to another person’s distress play an important role in motivating helping. This supports the model’s core logic.

Comparison with the Decision Model

Latané and Darley’s (1970) decision model frames helping as a sequence of five decisions. A bystander can fall off at any step and stay passive.

1. Notice the event.
2. Interpret it as an emergency.
3. Take personal responsibility.
4. Decide how to help.
5. Act.

In that tradition, other people cause failure through diffusion of responsibility and pluralistic ignorance, where each person reads everyone else’s calm as a sign that nothing is wrong.

Piliavin et al. complicate this story in two ways.

First, in a real emergency the crowd did not inhibit helping. Diffusion of responsibility largely failed to appear, because the situation was unambiguous and inescapable.

Second, their model is motivational rather than purely cognitive. It foregrounds arousal and the economics of helping, whereas the decision model foregrounds perception and attribution.

The two models are complementary rather than rival, and Piliavin et al. called theirs a ‘fine tuning’. Whether a crowd helps or freezes depends on the structure of the situation: ambiguity, escapability, and cost. There is no blanket law of “apathy in numbers”.

The study also corrects the myth built around the Kitty Genovese case. In a clear emergency, real bystanders frequently do act.

Critical Evaluation

Piliavin et al.’s study has real strengths and real limitations. Later research has also tested how well its central finding has held up. These are summarised below.

1. Ecological Validity: a genuine field emergency, covertly observed, means the high helping rates reflect how people really behave, not how they think they should behave in a lab.
2. Reliability and Rich Data: two independent observers recording about 4,450 passengers across 103 trials gave statistical and qualitative data and allowed an inter-rater reliability check (agreement between observers).
3. Ethical Issues: passengers gave no informed consent, could not withdraw from a staged collapse, and could never be debriefed afterwards.
4. Loss of Control and Generalisability: carriage composition, time of day, and passenger mood could not be controlled, and the sample was drawn from one city, line, and era.
5. The Confound of Victim Characteristics: the drunk and ill victims differed on more than sobriety alone, making it hard to isolate which single feature reduced helping.
6. An “Overly Calculating” Model?: critics argue the model makes helping sound too rational, and that arousal and helping are only correlated, not proven to cause one another.
7. Contemporary Research Supports the Finding: a meta-analysis and CCTV analysis of real disputes suggest crowds inhibit helping far less in clear, serious emergencies, echoing Piliavin et al.

Ecological Validity

The study’s biggest strength is its ecological validity. Real subway passengers faced what looked like a genuine collapse.

They were not volunteers who had signed up for an experiment.

The observers recorded behaviour covertly, from behind newspapers. None of the passengers knew they were being watched.

So there were no demand characteristics steering their behaviour.

This is why the very high helping rates, 62 of 65 trials for the ill victim, likely reflect real emergency behaviour. They are not rehearsed lab politeness.

The trade-off is a loss of control. Carriage composition, time of day, and passengers’ moods varied freely.

None of that could be held constant. A field experiment like this one buys realism at some cost to the tightness of its causal claims.

Reliability and Rich Data

About 4,450 passengers took part in the study.

That is a large sample, for a field experiment.

It generated two kinds of evidence: qualitative and quantitative.

The quantitative data, whether and how quickly someone helped, let the researchers run statistical comparisons across victim type, race, and group size.

The qualitative data came from spontaneous comments passengers made as the scene unfolded.

Together they build a fuller picture.

These add insight into why people acted, or held back. A frequency count alone cannot show that.

Two independent observers recorded each trial, not one. This let the researchers check inter-rater reliability.

That means how consistent different observers are when recording the same event.

Scale, two data types, and a reliability check together make this an unusually well-evidenced field study.

Ethical Issues

This raises some of psychology’s clearest ethical issues.

Passengers boarded expecting an ordinary commute. The study broke several ethical principles:

• No Informed Consent: passengers gave no informed consent, with no idea that an emergency, and their reactions to it, were staged and recorded.
• No Right to Withdraw: they could not leave once the ‘victim’ collapsed, because walking away from a moving train is not realistic.
• No Debriefing: most seriously, trials ran continuously on anonymous commuters who dispersed at the next stop, so they could never be found afterwards.
• Possible Distress: some may have carried genuine distress or lingering guilt over not helping, and the researchers had no chance to check they were unharmed.

Against this, the study caused no physical harm. It produced knowledge about real helping that a consenting, debriefed lab sample could not have provided.

Loss of Control and Generalisability

As a field experiment, the study sacrificed control over many factors. These vary from trial to trial: carriage crowding, time of day, passengers’ moods.

None of these extraneous variables could be held constant.

So it is hard to be certain that victim type and crowd size, rather than a background factor, drove any single trial’s outcome.

The sample, though large at around 4,450 passengers, came from one city, one train line, and one culture, all at a single point in time.

New York subway commuters of that era are not necessarily representative of bystanders elsewhere.

The researchers themselves noted that results on an open street, where people can simply walk away, might well differ.

Helping norms are also shaped by culture and history. The study’s figures should not be read as universal constants.

The Confound of Victim Characteristics

The victims did not differ cleanly.

The drunk victim smelled of alcohol and carried a bottle in a paper bag. He also invited a moral judgement that the sober, cane-carrying victim did not.

Several features changed at once.

Sobriety, smell, the prop carried, and the blame a bystander assigns all shifted together.

So it is hard to isolate which single feature reduced helping.

Disgust, risk, or blame could each be responsible.

A further limitation: the same small pool of student confederates played every victim across all 103 trials.

Actors differ too.

Individual differences between them could also have shaped how passengers responded.

Realism is gained in a field study like this one. But something is lost too: the ability to isolate exactly which variable is doing the causal work.

An “Overly Calculating” Model?

Critics argue the arousal:cost–reward model makes bystanders sound too calculating.

In a real emergency, people do not consciously tot up the pros and cons of helping this deliberately.

There is also a causal-inference problem. Arousal and helping are typically only correlated in this kind of field data.

Yet the model treats arousal as the actual cause of helping.

Defenders reply that the weighing is quick and largely unconscious. Dovidio et al. (1991) reviewed two decades of follow-up evidence and concluded that emotional reactions to distress do motivate helping.

So the “accountancy” language is a loose metaphor. It stands for a real, fast psychological process, not a literal ledger.

Contemporary Research Supports the Finding

Two modern lines of evidence support Piliavin et al.’s central finding. One explains why the subway result differed from the laboratory studies, and the other tests it on real, unstaged conflicts.

Why the Laboratory and the Subway Disagreed

Fischer et al. (2011) meta-analysed the bystander literature from the 1960s to 2010. Their data came from over 7,700 participants and 105 independent effect sizes.

Overall, more onlookers did mean less individual helping (g = −0.35, a small-to-moderate effect). But this bystander effect shrank in three kinds of situation:

• Danger: the situation was perceived as dangerous rather than non-dangerous.
• Perpetrator: a perpetrator was present.
• Physical Cost: the costs of intervening were physical rather than non-physical.

In certain dangerous emergencies the effect was no longer negative at all. Bystanders could even act as physical back-up for one another.

The authors noted that this pattern fits the arousal:cost–reward model. Dangerous emergencies are recognised faster and more clearly as real, which raises arousal and so increases helping.

That is the model Piliavin et al. proposed. The subway collapse was a clear, visible emergency that passengers could not escape. So the 1969 result looks like part of a wider pattern, not a one-off.

Real Public Conflicts on CCTV

Philpot et al. (2020) analysed 219 real public disputes captured on CCTV in the UK, the Netherlands, and South Africa. At least one bystander intervened in about 90% of incidents.

The average event drew nearly four interveners. Echoing the 1969 subway result, the larger the group present, the more likely it was that someone helped.

Unlike a laboratory study, these were genuine, unstaged conflicts. Together, the meta-analysis and the CCTV evidence support Piliavin et al.’s original, surprising finding: intervention, not apathy, is the norm in a clear, unavoidable emergency.

Half a century on, the subway study’s most striking result still holds up against real-world data. Groups help. They do not simply freeze.

References

Darley, J. M., & Latané, B. (1968). Bystander intervention in emergencies: Diffusion of responsibility. Journal of Personality and Social Psychology, 8(4, Pt.1), 377–383. https://doi.org/10.1037/h0025589

Dovidio, J. F., Piliavin, J. A., Gaertner, S. L., Schroeder, D. A., & Clark, R. D. (1991). The arousal: Cost–reward model and the process of intervention: A review of the evidence. In M. S. Clark (Ed.), Prosocial behavior (pp. 86–118). Sage.

Fischer, P., Krueger, J. I., Greitemeyer, T., Vogrincic, C., Kastenmüller, A., Frey, D., Heene, M., Wicher, M., & Kainbacher, M. (2011). The bystander-effect: A meta-analytic review on bystander intervention in dangerous and non-dangerous emergencies. Psychological Bulletin, 137(4), 517–537. https://doi.org/10.1037/a0023304

Latané, B., & Darley, J. M. (1968). Group inhibition of bystander intervention in emergencies. Journal of Personality and Social Psychology, 10(3), 215–221. https://doi.org/10.1037/h0026570

Latané, B., & Darley, J. M. (1970). The unresponsive bystander: Why doesn’t he help? Appleton-Century-Crofts.

Latané, B., & Rodin, J. (1969). A lady in distress: Inhibiting effects of friends and strangers on bystander intervention. Journal of Experimental Social Psychology, 5(2), 189–202. https://doi.org/10.1016/0022-1031(69)90046-8

Philpot, R., Liebst, L. S., Levine, M., Bernasco, W., & Lindegaard, M. R. (2020). Would I be helped? Cross-national CCTV footage shows that intervention is the norm in public conflicts. American Psychologist, 75(1), 66–75. https://doi.org/10.1037/amp0000469

Piliavin, I. M., Piliavin, J. A., & Rodin, J. (1975). Costs, diffusion, and the stigmatized victim. Journal of Personality and Social Psychology, 32(3), 429–438. https://doi.org/10.1037/h0077092

Piliavin, I. M., Rodin, J., & Piliavin, J. A. (1969). Good Samaritanism: An underground phenomenon? Journal of Personality and Social Psychology, 13(4), 289–299.

Piliavin, J. A., Dovidio, J. F., Gaertner, S. L., & Clark, R. D. (1981). Emergency intervention. Academic Press.

Schachter, S. (1964). The interaction of cognitive and physiological determinants of emotional state. In L. Berkowitz (Ed.), Advances in experimental social psychology (Vol. 1, pp. 49–80). Academic Press. https://doi.org/10.1016/S0065-2601(08)60048-9