- Reanalysis of gateway studies shows their models predict smoking no better than chance when tested properly.
- Most papers never reported how accurate their models were, and many authors refused to share their results.
- Using better methods made the models only slightly more accurate – and vaping turned out to be one of the weakest predictors of smoking.
- The findings match real-world data: teen smoking keeps falling, suggesting a “diversion effect” rather than a gateway.
A new analysis has cast serious doubt on the so-called gateway theory – the claim that vaping leads teenagers into smoking.
The study, led by scientist and data analyst Floe Floxon with psychologist Dr Ray Niaura, found that the statistical models behind gateway claims are not reliable and fail to predict which young people will take up cigarettes.
The findings were presented earlier this month at the 2025 annual meeting of the Society for Research on Nicotine and Tobacco (SRNT-E) in Cluj-Napoca, Romania.
“The key point here is that if a model cannot accurately predict which adolescents go on to smoke and which do not, then any inference made from such a model is not going to be reliable,” Floxon said.
Models that don’t work
Gateway papers usually rely on odds ratios from large surveys, claiming that teens who try vapes are more likely to smoke later. But, Floxon explained, odds ratios alone don’t prove a cause – especially when smoking is already rare.
According to the CDC’s 2024 National Youth Tobacco Survey, only 1.4 per cent of adolescents smoked cigarettes.
That rarity creates a problem. Models can look “accurate” simply by predicting that almost no one smokes. “It doesn’t matter how large the odds ratio is or how small the p value is if the model is getting its predictions wrong,” said Floxon.
To deal with this, researchers use balanced accuracy, which measures how well a model predicts both smokers and non-smokers. A balanced accuracy of 50 per cent means the model is no better than flipping a coin.
When Floxon and Niaura reanalysed published gateway studies, they found exactly that. “For the studies we were able to analyse, we found that the balanced accuracies are extremely low, close to 0.5, which is no better than chance prediction.”
Missing information
The team also discovered that most of the existing papers never reported their model accuracy at all. “We couldn’t find a single paper that reported its model’s accuracy,” Floxon said. When they contacted study authors directly, “most did not share this information.”
Those that did share data showed poor results. And because many studies used the same datasets, the authors argue it is reasonable to assume the unshared models perform just as badly.
Trying to improve the models
The team then tested whether they could do better with modern methods. They adjusted for the imbalance in the data using balanced class weights, tested performance using cross-validation, and tried out machine learning models such as support vector machines and neural networks.
These steps raised the balanced accuracy by 10 to 20 percentage points – an improvement, but still modest. More importantly, the improved models showed that vaping was not a strong factor in predicting smoking.
“Ecigarette use at baseline had negligible effect on how well the model predicted smoking at follow-up,” Floxon explained. Instead, classic risk factors – wanting to smoke, having friends who smoke, and other substance use – were far more powerful predictors.
No sign of a gateway
Taken together, the findings undermine the idea of a gateway from vaping to smoking. “We see no evidence of a net gateway in population level trends, which represent the reality,” Floxon said. Smoking among teenagers has continued to decline, even as vaping has become more common.
This suggests a “diversion effect” instead. This means that some young people who might once have tried cigarettes are trying vapes instead, but the overall number of smokers keeps falling.
A call for rigour
Floxon stressed that none of this means adolescent vaping is harmless. “None of this is to say that it’s okay for adolescence to use ecigarettes. But at the very least, there is no apparent gateway effect.”
The bigger issue, he argued, is how the gateway theory has been built on models that don’t pass basic checks. “The odds ratios are wrong because the models are wrong,” he said. “A model which cannot accurately predict smoking initiation cannot accurately tell us what is associated with smoking initiation.”
The team’s message to researchers and policymakers is that models must be tested properly, and focus on the predictors that actually matter. Floxon said: “You cannot just throw data into a logistic regression model and hope for the best. You have to rigorously evaluate the performance of the model before you can even begin to worry about the size and statistical significance of an association.”
