As I look at the MORE list above, I notice the weakest rank for 6 of the 9 Midwest teams is given by the NPI. Maybe you’d be interested to know those 6 compete with only 15% of their opponents not in the Midwest while the other 3 play 42% that aren’t? i.e. The teams not rated weakest by the NPI are almost 3 times more likely to play opponents not from the Midwest compared to the teams the NPI rates the weakest.
From the 19 on this list not Midwest, there is just 1 team the NPI gives its weakest rank. (New Paltz) This means the NPI gives a team on this list its worst rank 67% when it’s from the Midwest (6 of 9), compared to the worst rank when it’s not, just 5% of the time. (1 of 19) i.e. IT IS MORE THAN 12 TIMES LIKELY THAT THE NPI OFFERS A TEAM ON THIS LIST ITS WORSE RANK WHEN IT’S FROM THE MIDWEST! (.666 divided by .053…)
Look at the MORE’s Top 16 teams above. The NPI was the weakest rank just 3 times. Every single time it was a CCIW team! What do you think the chances of that would be if it really were a coincidence?
Digging Deeper into the NPI’s Problem with Midwest Volleyball Teams:
The plot directly below shows the distribution of 128 teams across the landscape being ranked by all 4 experts above. (Not the AVCA coaches who only choose 20.) These 512 data points are plotted 4 at a time, vertically placed above every volleyball program’s median z-score. (The median z-score might be the only strength order better than the MORE seen above because continuous metrics producing discrete ordinals offer more precision.)
The graph below is very much like a visual representation of “MORE.” There are 4 different colored points aligned vertically across its spectrum, just like the MORE above shows those same 4 ranks from every expert horizontally in rows. Instead of the median of the ranks dictating the order like in the table above, I plot the 4 ranks as a function of the median Z-Score of the expert’s metrics. This shows how well these models align with one another and offers clarity on the variability each has to its own. Looking closely, you can see the 4 models are almost indistinguishable, visually laying on top of each other, with the exception of Massey’s for the weakest third of teams. i.e. The thin blue curve a “hair” above the other 3 seen at the upper left.
Think of any model as the bullseye of a target and the distance away from it as the rings around that bullseye. If you had 128 red-feathered arrows, a quarter of them manufactured in the Midwest, and shot them at the target with your bow, then…
Wouldn’t you ask why those particular Midwest produced arrows had the propensity to consistently miss above the bullseye farther away than most others? More than 90% of the time should be concerning enough to take pause when using the same bow settings and routine every time.
All 4 expert models produce metrics which are normally distributed. I calculated the Z-Score average for every team from the 3 non-essential experts – Massey, Inside Hitter, & T100. They are the x values of points shown below. The y value of every point is the Z-Score for the same team according to the latest NPI metrics. I plotted Midwest Teams in red and non-Midwest Teams in blue. You may notice a second y-axis with different values to the right labeled “NPI Points Equivalent.” Here is the explanation for that:
The mean of NPI Metrics on 3/30/25 was 49.05 and its standard deviation 8.98. This makes a Z-score of 0.00 aligned with the NPI score of 49.05 & the Z-Score of 1.00 exactly 8.98 points above that at 58.03. Add another 8.98 to arrive at a Z-Score of 2.00 being 67.01, just a hair higher than Southern Va. as seen by the 2nd highest blue point plotted on the graph. (Note – I use a 25.0, 24.0, & 23.0 as NPI metric placeholders for 3 weak teams the T100 tracks that the NPI doesn’t report, so if you were to look at the NPI site you might compute its mean to be 49.57 and the standard deviation at 8.21 for teams it lists.)
I spent an hour of my life I’ll never get back looking up schedules of the 31 Midwest teams to account for their foes, 21of them plotted above. It occurred to me that if deflation is related to Midwest teams because of algorithmic bias in the NPI, then any red points higher above the red line would have been “tugged” from blue line influence. This should indicate Midwest teams who played more matches against non-Midwest teams were benefited. Points closest to being on the red line should be teams who played against less than that, and so it would follow the red points below the red line would be hypothesized to have played the least proportion of non-Midwest teams.
The 6 red points displaced highest above the Midwest model were teams playing nearly 40% of its opponents not from the Midwest. (This number was bolstered by Baldwin Wallace, Mount Union, and Wittenberg combining to play more than half their schedule not against the MW, and when combined with Carthage, Aurora, and Westminster, each closer to 20%, brought the total just shy of 40%.) The Midwest teams indicated by the 7 points closest to being on the red line played about 15% of their matches against non-Midwest teams, and the remaining 8 red points displaced below the red line were teams with just 13% of matches against those not in their region. An increasing percent of each rate stated above from Cal-Lu & UCSC when the percent of non-Midwest opponents declined. (Note: The remaining Midwest teams not graphed above and from the bottom half played only 10% of opponents not from the Midwest, nearly half of these against the two West Coast teams. They are generally deflated between 2.5 & 4 NPI points but with no potential to be injured beyond their pride. This the reason for only plotting teams from the upper half. i.e. The top 10 drive at-large bids and the weakest of all the auto-bids needed to be seeded is almost always among the top 55 teams. i.e. z = 0.2 for about 130 teams)
Does the nature of the graph above look different comparing the NPI to only one expert at a time rather than the average of all 3? Take a look below to see the answer to this question:
The intention for showing every relationship individually by an expert model to the NPI is two-fold:
[1] To demonstrate the overall pattern, regardless of which expert is chosen, continues to be true.
[2] Scanning from top to bottom, notice the gaps between the lines of best fit get a little thinner each time going from Massey to Inside Hitter to T100, though it remains noticeable & significant.
Massey and Inside Hitter might think the T100 is biased against the Midwest, but there would be insufficient evidence to prove it because it isn’t so. In the past, I’ve attempted to prove theirs has algorithmic bias in favor of the Midwest to support my notion the T100 was properly calibrated to perform its function well. Came close one time a couple years ago, but sample sizes came up too short. I walked away from those efforts believing the truth absolutely must lie somewhere between, shaded in the T100’s direction, though. LOL – Over time I have seen all three converging reasonably closer to where I no longer even think that way. Partly due to D3 volleyball team’s improving and more parity in the middle 50% of the distribution than before, and also because my experience has produced confidence in that calibration. It didn’t hurt to later find out the FIVB ranks international teams with a similar construct to the T100.
Now comes a new model, the NPI, with its higher variability than any of these 3 which tilts almost twice as far in the opposite direction from the T100 as the other two do. Some like to eloquently say models intended to predict shouldn’t be the same as those intended to reward. Prediction is based on strength, and strength is evidenced by what comes before. A rewarding model is also based on what comes before, but if it rewards in a way to give more credit for having defeated weaker teams, then maybe its rewards are fallacious. The March Madness selection committee uses both predictive and rewards-based models, too. A real cool one called WAB (Wins Above the Bubble) which I think could serve D3 better than the NPI, at least for Men’s Volleyball.