The finding comes from a study tracking the accessibility of scientific data over time, conducted at the University of British Columbia.
Researchers attempted to collect original research data from a random set of 516 studies published between 1991 and 2011.
While all data sets were available two years after publication, the odds of obtaining the underlying data dropped by 17 per cent per year after that, they reported.
"Publicly funded science generates an extraordinary amount of data each year," UBC visiting scholar Tim Vines said. "Much of these data are unique to a time and place, and is thus irreplaceable, and many other data sets are expensive to regenerate.
"The current system of leaving data with authors means that almost all of it is lost over time, unavailable for validation of the original results or to use for entirely new purposes," the researchers reported in the journal Current Biology.
Vines is calling on scientific journals to require authors to upload data onto public archives as a condition for publication.
"Losing data is a waste of research funds and it limits how we can do science," he said. "Concerted action is needed to ensure it is saved for future research."