Statistically, it's impossible to predict domestic violence homicide. That doesn't mean we should give up

Shouldn't we have seen it coming? Whenever a woman is killed by her partner in Australia, domestic violence experts quickly line up to call out the "red flags" that had apparently been overlooked: the perpetrator had access to guns; the victim had previously reported being strangled by him; he'd made specific, detailed threats to kill her.

Had police, or bail magistrates, or even family violence workers been paying more attention to these well-known "risk factors", the argument goes, had they taken the victim's concerns seriously and stepped in, her death could have been prevented.

The idea that domestic homicide is predictable is so entrenched that a reporter's recent well-intended claim that "statistically, the most reliable way to predict a domestic violence homicide is to believe the victim" barely raised an eyebrow.

Victims must always be believed. The uncomfortable truth, however, is that statistically, it is almost impossible to predict lethal family violence because it is so rare.

At least that is the conclusion of a new study, by researchers from Swinburne University's Centre for Forensic Behavioural Science, which examined whether 12 factors commonly believed to be high-risk indicators of domestic homicide could accurately predict lethal or near-lethal family violence.

"It's very difficult to predict what we call 'low base rate' events, or events that are very, very infrequent," said Benjamin Spivak, senior lecturer in forensic psychology at Swinburne University and one of the study's authors. "But what we kept seeing throughout the literature … is a lot of claims that we can predict it — that we can zero in on homicide and target interventions in a precise way."

It sounds sensible in theory, Spivak said, but often those interventions are "quite intensive or intrusive" — for instance, laws that make it harder for alleged domestic violence offenders deemed "high risk" to get bail. "If you're trying to avoid something horrific like a homicide … you may want to take extreme steps like that to stop it. But if you're not taking into account the fact that the overwhelming majority of people who you're targeting will never go on to have that kind of outcome, is that justifiable?"

Some might argue that remanding thousands of people is justifiable if it saves even one life. But locking people up is costly, not just to taxpayers but individuals and families — it disrupts employment, housing and social connections, and can have serious mental health consequences, potentially exacerbating family violence.

The prediction 'problem'

For their study, soon to be published in the Journal of Family Violence, Spivak and his colleagues collected nearly 40,000 family violence reports recorded by Victoria Police between September 2019 and March 2020 involving an alleged victim and an alleged perpetrator. Each party was followed up for 365 days after the initial report to establish whether they had been charged with lethal or near-lethal family violence, including homicide or attempted homicide, or intentionally or recklessly causing serious injury involving gross violence. All up, there were 55 cases, equivalent to 140 per 100,000 people (0.14 per cent).

The researchers also examined the data police recorded at the time of the initial report on a form called the L17, essentially a series of 39 questions that screen for numerous risk factors, including 12 that are widely considered "high risk" indicators for domestic homicide – things like access to firearms, threats to kill, strangulation, alcohol or substance use, and mental health problems. 

They then used a statistical model to determine whether there was a predictive relationship between any of the risk factors — in combinations or on their own — and lethal or near-lethal family violence. By weighting all of the risk factors, the model generated a range of probability estimates, and threshold scores, indicating how likely lethal or near-lethal violence was.

The results showed that few of the risk factors were associated with lethal or near-lethal violence, and the vast majority of cases the model identified as high-risk did not go on to be violent in that way. The reason for this, the researchers explain, is that lethal or near-lethal violence occurred in only 14 of every 10,000 family violence cases over 12 months, but the risk factors involved were common to thousands — sometimes tens of thousands — of individuals.

Take strangulation, for instance, which is often described as a "red flag" because a significant number of female domestic homicide victims had experienced it. A history of strangulation was recorded in more than 3,000 of the Victorian family violence incidents the researchers examined, but fewer than six of those cases went on to involve lethal or near-lethal violence.

Just one risk factor — a victim experiencing violence during pregnancy — did significantly predict lethal or near-lethal family violence, but the absolute change in probability it caused was minute, the researchers found, and it would make "no meaningful difference to the accuracy of risk assessment if applied in practice".

Of course, family violence is notoriously under-reported, and the researchers only examined incidents that came to the attention of police — doesn't that muddy the results? Because lethal and near-lethal violence is likely not under-reported, Spivak said, expanding the sample of family violence reports would only make the prediction "problem" worse — it would mean lethal violence is even rarer. In other words, they've given prediction its best shot.

Search for the haystack, not the needle

What is the upshot? The study highlights just how rare domestic homicide is; although it spiked in 2023-24, Australia's intimate partner homicide rate has fallen since 1989 — not that experts are exactly sure why.

And so for co-author Troy McEwan, professor of clinical and forensic psychology at Swinburne University, it is crucial that family violence workers, and the community more broadly, recognise that predicting which people will kill their partner is simply not possible — it's like trying to find a needle in a haystack.

Instead, McEwan said, the focus should be on finding the haystack: on identifying people who are at increased risk of repeat and severe family violence — something police forces are equipped to do — and intervening in those cases. In doing so, if those interventions are effective, we will likely reduce harm overall, and hopefully also prevent some domestic homicides.

"I'd also really like to see more focus on what in the field are called dynamic or acute indicators … I'd stop calling them risk factors and start calling them acute indicators, because they're not about risk — they're acute indicators of concern," said McEwan, who led the development of Victoria Police's family violence risk assessment tool. "So, 'How concerned are we about this family right now given what's been happening the last few weeks and months?' And, 'Let's help that family now.'

"Sometimes that means detaining someone because we're so worried that we need to actually create a barrier," McEwan added. "But a lot of the time it means … taking steps to provide adequate services in the community that can protect victims … but also actually respond to the person who is causing harm and helping them cause less harm. That would be a far more useful long-term expenditure than locking people up for a period of time, then letting them out when nothing has been done in the interim."

That is not to suggest that risk assessment is pointless, or frontline workers are off the hook. Instances of "poor practice", McEwan said — of police officers not taking reports of abuse seriously, of criminal justice system actors or family violence workers overlooking critical information — can have catastrophic consequences. In that way, attention to policing failures or cultural problems is important, especially if it leads to practice improvements.

"Some people who read our paper may not like it, partly because they might think that what we're saying is, 'Throw out the baby with the bath water, just don't bother. Don't worry if someone says they've been strangled … it's not relevant'," McEwan said. "But that is not what we're saying.

"When people are in fear for their life, when they're being stalked, when they've been strangled, when people threaten to kill them, we obviously need to take that very, very seriously and we need to respond. What we shouldn't be doing, though, is saying that means a homicide will happen."

Editor's note, Dec 2: An earlier version of this article incorrectly stated the percentage of cases of lethal/non-lethal family violence in the study.