This assumes each failure is independent. Often a single failure will cause many bad pixels on the same screen. So you might have 98 good screens 1 with one bad pixel and one with 5,000 bad pixels when the per pixel odds say you are going to end up with 1 good screen if you are lucky and 99 bad ones.

I simulated 100 screens with different numbers of pixels where the probability the first pixel is bad in a screen is p0 = 1e-6 and the presence of each bad pixel increases the probability of the next one by the same amount (ie probability for a second bad pixel is 2e-6 and for last pixel is p0*number_already_bad_px). A screen was accepted if it had 0 bad pixels. The results look pretty much the same:

  num_pixels frac_accepted
       1e+01          1.00
       1e+02          1.00
       1e+03          1.00
       1e+04          0.98
       1e+05          0.89
       1e+06          0.32
       1e+07          0.00

Of course this was just a convenient scheme to simulate. The greatest number of bad pixels was 86k for the 10 million px screen (so chance the last pixel was bad was ~9%). It does show that the independent pixel model isn't necessarily very far off though.

You messed up the math. You need to count (# number of pixels that where bad) / (total number of pixels produced) not simply call the odds for the first bad pixels as the rate of bad pixels.

You are saying to adjust p0 so I get that the frequency of bad pixels = 1e-6? That makes sense but it would probably now depend on what size screens are being considered... not sure if I'll go back to it.

I didn't actually calculate the probabilities, it was just my off-the-cuff estimates. 37% could be 'about half' :P