7,649
edits
Changes
Jump to navigation
Jump to search
Urban Start-up Agglomeration and Venture Capital Investment (view source)
Revision as of 11:09, 9 September 2019
, 11:09, 9 September 2019→Maximum R-squared
[[File:Portland3HullsLowestHighest.png|right|300px]] Two further options are to find the lowest-highest and the first-after-peak. These often coincide. The lowest-highest finds the highest layer with x hulls and then works back down the layers taking the lowest in the continuous chain of x hulled layers. The first-after-peak finds the first layer with x hulls in or after the layers where there is the peak number of hulls. This last approach is a little problematic because sometimes there isn't a peak - its just flat - and it is inconceivable that there could be two or more peaks.
Computing the Lowest-Highest, running a PCA (see below) and using three factors in a regression to store R2, N, and adjusted R2, gives the results in the '''R2 On Hulls''' sheet of '''Images Review.xlsx'''. The first 20 placeIDs were checked (i.e., 10% of the sample). In most cases (16/20), the maximum R2 and R2-Adjusted coincided when considering only cases when N>=10. Note that R2=1 and R2 adjusted is missing for all cases where N <=4, and there there is very high volatility in both measures for N<10.
There were some draws out of scope (i.e., with N=9), and we should take the lowest(?) layer in the event of a draw (they will have the same N with prob ~1). There are some cases to look at:
*10 - Austin (3885 layers): Came up 3 when it could have been anything up to 23 (with N>=10). Looking at a map, this might be ok. Austin's has some fairly homogenous big clusters in 2017.
*9 - Atlanta (1414 layers) came up 8/8(N>=10) on R2 and 1 on adjusted R2, which was the biggest spread. All other discrepancies were of a single layer. It is hard to see either answer in the 2017 image, which looks like it has 2 clusters, or maybe 3 or 4. The r2 does jump at level 8 (to 0.50 from 0.28 in the layer before, with the previous highest being 0.39 at level 1), but level 8 has N=13 and is the second highest R2adj at 0.327 as compared with 0.332 in level 1.
*13 - Bellevue (1181 layers) had R2 and R2adj in agreement at level 2, but there were 7 levels with N>=10 and 13 levels over all. 2 looks ok on the map in 2017 (3 or 4 would likely be better). I expect that this is a case of a place that had dramatic growth...
*6 - Alpharetta (580 layers) comes it at 1 for both R2 and R2adj even with N<10 but has 6 layers, 4 of which have N>=10. It looks like a pretty homogeneous place in Georgia.
We also need to check some of the really big places:
*San Fran (160 - 12,946) is 52 on both (N=12, >=10). Picking any N>=10 up to N=15 will always find the largest hull count. At N=16 it finds the smallest large hull count (41 to 47 are all N=16).
*Boston (22 - 3,506) is 17 with N=10 and 11 (N=17) with N>10 on both measures.
*New York (122 - 11,466) is 27 with both measures with N>10 (N=18). It is 67 and then 66 at N=10 on both! N=12's 65 hulls is in fourth place.
*Palo Alto (134 - 3,492) is 1 with both measures. Even allowing N<10 there isn't an issue until N=5.
===Image Analysis===
