(It's even entirely possible that this particular fortnight of data-collection were so disrupted by conditions on the ground that exact timings were impossible to establish for otherwise confirmed counts submitted at intervals less frequently than daily, so the compilers made best-guesses that intentionally or unintentionally smoothed the contrast across these few days'-worth of cumulated detail. Not an issue, if it still builds up to the right total in the larger trend until some nitpick went selectively looking for "something" by p-hacking the data, and found this artefact that nonetheless represents.)
Probably should say they're doing that, if that's what happened. Those limitations are noted for other periods.
Maybe they did say that. Maybe there were a whole lot of caveats about how they have to deal with minor difficulties of compiling the data, all understandable and none of them changing the long-term accuracy. The graph-makers may have decided to not pass on the caveats (
especially if it was contrary to their intended interpretation...).
You presented a graph that did not have any kind of caveat (nor "limitations noted for other periods"), seemingly aware of common 'graphing tricks' like non-linear axes that might mean either an objection is raised or that you whould raise one yourself (although, in the right circumstances, something like a log-normal or log-log layout showing a linear trend is
exactly what you'd expect from perfect data) but missing all kinds of other issues that were
at least a problem with the view being presented.
Looking at the new graph, you can see that the decision of the (...briefly checks link...) March 7th article to highlight a two week period of data originating from nigh-on half a year ago was clearly being selective for a 'notable' artefact, as I thought. All else aside (yes, the new graph quotes various missing/vague periods[1], but need not be exhaustive of all passing issues), that doesn't look like a
particularly manipulated set of data anywhere like the original complaint alleged (and carefully contrived its window on the data to 'prove').
Clusters (or the
lack of clusters) can always be found if you go looking into data with an eye for supporting a presupposed agenda.
In reality, I'd expect a less massaged (<= still need
not be 'manipulated', in a meaningful way, just a degree of being nicified for viewing) set of data to be very ragged (
e.g.[2][3][4][5]). But the seperation of presentation from source is always a consideration. And observer bias can spin the stories told all kinds of ways. Direct stories or meta ones.
("Lies, damn lies and statistics", indeed...)
[1] The pre-truce flatness and the post-truce jump aren't explained by the given caveats. A critical eye on the trend certainly ought to consider these features. If you want me to speculate (here from my armchair, and not going out of my way to check for supporting evidence), it just looks like deaths had been happening at
almost the same rough rate as before, just unrecorded for some reasons still not noted. Removing a number of truce-defered combat deaths that did not happen, you can imagine the underlying death-rate (perhaps from humanitarian failures, or those attacked by pre-truce actions who then succumbed (or were finally found, dead all along, in the rubble of buildings) days later) continuing during the hidden gaps and delays in the data-gathering process. Moreover, if the reporting methods involved up to a week of delay from (recorded) death up to finding itself on graphs (maybe becase of the required cross-checking for accuracy) then the pre-truce disruptions (the "storm before the calm") cause the flattening in that region. The 31/Nov to 2/Nov period was mostly 'catch up' while they tried to return to their original thoroughness. - Noting that I'm
not saying that I trust the underlying data unconditionally, but (given not unreasonable assumptions) a real-world situation could quite easily translate to such an ultimately truthful trend,
despite the fog of war and other issues.
[2] Here 'nicified' for viewing by a rolling average, instead.
[3] By this time, the infrastructure of the territory is even more disrupted, naturally. Which means less desk-jockeys getting in the way of the (sporadic) reporting procedures.
[4] And, of course, it's actually data leading up to publication, rather than cherry-picked from history.
[5] Also, to forestall misunderstanding, that's of courze a 'reported per day' bar-graph, rather than a 'cummulated total' one.