You're looking at the "Safe WRs for a Variety of Lazy Portfolios" right? Essentially it looks like this is comparing just domestic US stock/bond portfolio to portfolios that also contain other asset classes. Because some of these asset classes are not correlated with the performance of stocks OR bonds in the dataset, that means these portfolios didn't drop nearly as much in bad years. Since the worst years define what the safe withdrawal rate of a given portfolio is, that means these portfolios have higher SWRs.
Some of the most common additional asset classes used in alternative portfolios are international/developing world stocks or bonds, and gold. Just to give you both sides, below is my attempt to summarize why these asset classes don't get used in a lot of "conventional" simulations.
I really wish we had better historical data on international stocks or bonds going back as far as the USA dataset (1870s-present). Unfortunately a lot of the international stock/bond data (whether country specific or global index) in the public sector doesn't go back far at all.*
Adding gold to backtested portfolios tends to increase the SWR
substantially. However, a big driver of this is that in the middle of "stagflation" in the mid 1970s, when inflation was destroying the value of fixed rate bonds, and the stock market dropped like a rock and then stayed down for more than a decade, the USA also went off the Bretton Woods system (end of the gold standard, read up on the "Nixon shock"), and the price of an oz of gold shot from a fixed rate of $35/oz to $450/oz by the early 80s (not adjusting for inflation, which was admittedly substantial). Since we're not back on the gold standard, we cannot go off the gold standard again in a future economic crisis, so the question is whether gold prices would respond in a similar fashion to economic conditions like the 1970s in the future.**
So basically the amount of historical data for these asset classes is shorter and, arguably, may not be representative. Of course excluding things because they make our models messy runs the risk of turning into
a spherical cow in a frictionless vacuum. Either way, no easy answer.
*Typically not even to world war II, although I know there are much better datasets that go back further for the folks who do this for a living and subscribe to various financial databases.
**Reasonable people can disagree on this point.