At 1st glance, dataset is amazing.
Index cases: infected people who took train. Contacts - potential 2ndary cases - are passengers who sat w/in 3 rows or 5 columns of index case. Metric/Response variable: did contact get sick in next 14d (attack rate in %).
Index cases: infected people who took train. Contacts - potential 2ndary cases - are passengers who sat w/in 3 rows or 5 columns of index case. Metric/Response variable: did contact get sick in next 14d (attack rate in %).
As in all contact tracing studies, index vs contact (primary vs secondary case) is based on person w/ earlier date of symptom onset. Obviously not ideal since incubation period is 2-14d (& can get infector/infectee backwards) but unless one has more info, it's what we have.
Paper presents fraction of contacts infected w/ trip duration & distance in both x-y directions (seat rows vs columns). At 1st I thought this study could be amazingly informative for understanding risk in space/time. But 1 piece of missing data & 1 fig suggests something else.
Crucial question is: Is there any evidence transmission actually occurred on train. 1st glance suggests yes. Strong relationship b/w travel time & attack rate, as we'd expect if transmission occurred on train. But... 
Key piece of missing data is relationship b/w passengers. Data were anonymous so no info on whether adjacent passengers were more likely in same household & patterns just reflect prob that contact is in same household (higher for adjacent seats & nearby rows).
Authors show supplemental figure that I think makes it clear that study is measuring relationships among people, not transmission on trains. Authors' brilliant idea was to look at patterns for co-seated passengers that do or don't start/end at same destinations vs those that do.
If all transmission is occurring on train, pattern should be same for both. If substantial transmission occurs on train but some away from train, pattern should be similar but diluted. But if no transmission on train then pattern should be absent for people w/ diff destinations.
Here's the comparison. There is no effect of co-travel time on attack rate for passengers that don't share start/end destinations (right), but strong pattern for those that do (left). Would like to seating fig presented same way. I'd guess it is same.
So then how can we explain cool patterns shown in paper? Sadly, very simply. If I want to take train w/ spouse, where would we sit? As close together as possible, & adjacent rather than in diff rows. If can't sit together, sit 1 column seat apart in same row.
If can't get that seat, then same column, 1 row apart. I've had many experiences on airplanes when traveling alone where family asks me to swap seat so they can be together & their seat is 1-2 row or column seats away.
How do we explain increasing attack rate w/ travel time? Also pretty simple. What are odds that strangers sit together for short trip (commuting, short business trip)? Pretty high.
What are odds for 8 hr trip? Much lower (or vice versa - more likely to take long trip w/ fam)
What are odds for 8 hr trip? Much lower (or vice versa - more likely to take long trip w/ fam)
Thus, while I think the dataset is amazing & authors did fantastic job w/ careful analysis, I think results simply tell us about where family members try to sit in relation to each other, & not about virus transmission on trains.
Put another way, I think it's possible there were zero transmission events from these index passengers to contacts ON TRAIN & 100% of transmission could have occurred off the train in households/other settings & I don't see how we can distinguish these possibilities.
Would love to hear counter-arguments that show that at least some transmission occurred on the train, especially from authors. Most convincing, noted above, would be fig of attack rate vs distance (row, column) for two types of passengers - those sharing destinations & not.
If spatial pattern (higher risk for closer seats & especially in adjacent seats) is robust for passengers that don't share to-from destinations then evidence for transmission strong. If no pattern (like travel duration) then no evidence for transmission on train.
Addition: @mlipsitch pointed me to Table S3 (which I somehow missed) which shows same data as Fig S3 above, except for all trip durations. Also shows that high attack rate for adjacent seat & 5-8 hr trips drops to 0(!) when not including passengers w/ same To/From destinations.
In making plot in tweet above (shown again here) I realized diff in scales in 2 panels Fig S3 above (which I didn't notice at first). Not only is there no increase w/ trip duration but attack rate is ~0 for all durations for different itineraries/strangers.
Would love comment from authors but this is pretty strong evidence to me that ~ZERO transmission is occurring on train & attack rate patterns on train just reflect seating choices of people from same household/friend/work circle & off-train is where transmission is happening.
