The AP is not estimating the popular vote. It is projecting the number of state delegates (multiplied by 100).
For example in Taylor County, there were 30 delegates elected to the county convention, 21 for Clinton, and 9 for Sanders. The IDP shows Clinton with 70% of the "vote", which is 21/30. That is, she won 70% of the delegates to the county convention.
Taylor County elects 2 state delegates. AP and the Atlas show this as 200 units, and divides it as 140 60.
I don't know how the IDP estimates the state delegates. They could conceivably do it just like the AP and the Atlas, but not multiply by 100. So it would be 1.4 to 0.6, and then they would round at the state total.
This would have the advantage that it would would work with partial county results. Let's imagine that Bedford was still out, which would make it Clinton 16, Sanders 7, and 7 delegates not yet chosen.
Clinton could be projected as 16/30 * 2 = 1.07 state delegates (or 107 units)
Sanders would be 7/30 * 2 = 0.47 (or 47)
Still out 7/30 * 2 = 0.47 (or 47).
The alternative would be to attempt to project what will happen at the county convention. If we assume that all the county delegates will show up and vote for their candidate, then a 21:9 split translates to 1.4 and 0.6 state delegates, which rounds to 1 for Clinton and 1 for Sanders.
How I got to a 21:9 split for county delegates in Taylor County.
County delegates are allocated to each precinct in proportion to the Obama(P2012) + Hatch(G2014) vote:
Blockton 92
Bedford 403
Clearfield 161
Gravity 127
Legion 183
Lenox 649
New Market 174
The precinct results are:
Blockton: 100%:0%. which means 1:0 or some integer multiple.
Bedford 71.4%:28.6%, 5:2 or some multiple.
Clearfield 100%:0%, 1:0 or some multiple
Gravity 50%:50%, 1:1 or some multiple.
Legion 66.7%:33.3%, 2:1 or some multiple.
Lenox 63.6%:36.4%, 7:4 or some multiple.
New Market 66.7%: 33.3%, 2:1 or some multiple.
If the multiple in every precinct is 1, then it would be 19:9, which would give Clinton 67.9%, which is not the 70.0% shown on the IDP website.
Sw we calculate the Obama+Hatch vote divided by the minimum number of delegates for a precinct.
Blockton 92/1 = 92.0
Bedford 403/7 = 57.5
Clearfield 161/1 = 161.0
Gravity 127/2 = 63.5
Legion 183/3 = 61.0
Lenox 649/11 = 59.0
New Market 174/3 = 58.0
The ratio of Obama+Hatch to delegates appears to be around 60, with Blockton and Clearfield needing more delegates.
Clearfield 161/2 = 80.5, and 100%:0% = 2:0 county delegates.
Blockton 92/2 = 46.0, and 100%:0% = 2:0 county delegates.
With these adjustments the county convention would be 21:9, which matches the 70%:30% split on the IDP website.
As a further check, we divide the 1789 Obama+Hatch votes by 30 = quota of 59.6 votes/delegate allocated.
Blockton 92/Q = 1.54 (rounds to 2)
Bedford 403/Q = 6.76 (7)
Clearfield 161/Q = 2.70 (round to 3)
Gravity 127/Q = 2.13 (rounds to 2)
Legion 183/Q = 3.07 (rounds to 3)
Lenox 649/Q = 10.88 (rounds to 11)
New Market 174/Q = 2.92 (rounds to 3)
This totals to 31. It appears that for some reason, county conventions have delegates that are a multiple of 5. Since the number of county delegates is fixed, we drop a delegate from Blockton, which has the smallest fraction over 0.50.
This would mean that it was actually:
Blockton 1:0
Clearfield 3:0
And still 21:9 for the county.
But this still doesn't tell us the popular vote. We don't know know how many attended the caucus and Legion, and we know that, Clinton had more support than Sanders (but it could be one person, or even a lucky coin flip), and Sanders had 1/6 of the delegates (this is the threshold when electing 3 delegates). So Clinton had between 50% and 83.3% of an unknown number of voters in Legion.