Did anyone run IRV models/simulations on 2016 US election?

I'm aware of four such efforts, three of which assume a national popular vote, and one which looks at the implementation of IRV in individual states.

Firstly, an article entitled Dancing with Donald: Polarity in the 2016 Presidential Election, by Robert Chuchro, Kyle D'Souza, & Darren Mei. In their paper, they use polling data (n~750) obtained through Amazon Mechanical Turk, and weighted to match the demographics of voters in the 2016 election. They then proceed to run a number of simulations using various alternative voting mechanisms, including Simple Plurality, Plurality with Runoff, and IRV, which they call Ranked Choice.

The full results are below, sorted by rank in the IRV simulation, in which Bernie Sanders was elected.

                  Ranked Choice   Plurality   w/ Runoff   Approval   Coombs   Copeland   Borda   Moderation   Polarization
Candidates                                                                                                                
Bernie Sanders                1           2           1          1        1          1       1            1              3
Donald Trump                  2           1           2          3        8          8       5            7              8
Hillary Clinton               3           3           3          2        7          3       3            5              7
Gary Johnson                  4           4           6          4        2          2       2            2              1
John Kasich                   5           5           4          7        5          7       7            4              5
Marco Rubio                   6           6           5          5        4          5       4            6              2
Ted Cruz                      7           7           7          6        6          6       8            3              6
Jill Stein                    8           8           8          8        3          4       6            8              4

The second effort is contained in Comparing Voting Methods: 2016 US Presidential Election, by Herrade Igersheim, François Durand, Aaron Hamlin, & Jean-François Laslier. They too, perform their own polling to obtain preference-based voting intention, but in addition, split respondents into two groups. One group faced a choice of the four main presidential candidates - Trump, Clinton, Johnson, & Stein - while the other had a larger selection, including Sanders, Cruz, McMullin, Bloomberg, and Castle. Notably, their polling took place in the week leading up to the election.

Once data was collected, they simulated the election results under approval voting, range voting, and IRV. Under IRV, Hillary Clinton was elected in both simulations. The percentage vote after each round is below. A full analysis of vote transfers is included in the article.

Short Candidate List
          0   1   2    3
Clinton  46  48  54  100
Trump    39  39  46    0
Johnson  11  13   0    0
Stein     4   0   0    0

Long Candidate List
             0    1   2   3   4   5   6   7    8
Clinton     30   30  30  30  31  32  33  54  100
Trump       28   28  28  28  29  31  39  46    0
Sanders     21   21  21  21  22  25  28   0    0
Cruz        10   10  10  11  12  13   0   0    0
Johnson      4    5   5   5   6   0   0   0    0
Bloomberg    4    4   4   5   0   0   0   0    0
McMullin     2    2   2   0   0   0   0   0    0
Stein       <1   <1   0   0   0   0   0   0    0
Castle      <1    0   0   0   0   0   0   0    0

Finally, and perhaps slightly less academically rigorously, Vox has published this analysis by David Shor, 'a senior data scientist at Civis Analytics, a Democratic data and polling firm'. It surveyed just over a thousand registered voters who said that they had voted in 2016.

Bear in mind that this analysis was performed in the days immediately after the election, so detailed demographic information was not available - survey weighting was performed to reflect 'the election outcome and Civis’s best guess as to the demographics of the 2016 electorate.'

The results are presented below, and show Hillary Clinton being elected president.

The only study I could find which simulates the effect of IRV at a state-level was published by an undergraduate student, Nicholas Joyner, supervised by Michele Joyner an Associate Professor at East Tennessee State University. Entitled Utilization of Machine Learning to Simulate the Implementation of Instant Runoff Voting, it uses data from the Cooperative Congressional Election Study to guess which of the two main candidates voters who supported third-party candidates might rank as their second choice, allowing a pseudo-IRV system to be simulated.

Initially, the simulation is run at the state-level in the states where neither of the main candidates achieved an absolute majority of votes - where IRV would have made a difference. After 1,000 simulations, 5% of the time, there was no difference to the 2016 overall electoral vote count. 83% of the time, Trump gained more electoral votes, and in the remaining 12% of cases, Clinton received more electoral votes. However, Trump still retained a majority of the electoral votes, and was therefore elected president, in every simulation.

The study does point out, however, that if IRV had been implemented from the start, the public might have been more likely to support third-party candidates from the start. To test the results under this assumption, a 'modified vote count' is used, which arbitrarily increases the votes cast for third-party candidates by 10%, taken equally from Trump & Clinton. Running the simulation under these conditions for every state now without an absolute majority yielded the same results as before.