Quick note on the ‘bunching’ hypothesis. While that particular post and suggestion is mostly an artefact of the US tax code and would lead to years that look like 20%/​0%/​20%/​0%/​etc., there’s a similar-looking thing that can happen for non-US GWWC members, namely that their tax year often won’t align with the calendar year (e.g. UK is 6th April − 5th April, Australia is 1st July − 30th June I believe).
In these cases I would expect compliant pledge takers to focus on hitting 10% in their local tax year, and when the EA survey asks about calendar years the effect will be that the average for that group is around 10% but the actual percentage given will range anywhere from 0 − 20% (if ~10% is being given), but often look like 13% one calendar year, 8% the next, 11% the year after that, etc. In other words, they will appear to be meeting the pledge around 50% of the time in your data. Yet the pledge is being kept by all such members continuously through that period. Eyeballing your 2017 graph of the actual distributions of percentages given, there are a lot of people in the 8-10% range, who are the main candidates for this.
Since both most US members and most non-US members have good reasons to not hit 10% in every calendar year, the number I find most compelling is the one in the bunching section that averages 2015 and 2016 donations (and finds 69% compliance when doing so). But that number suffers from not knowing if those people were actually GWWC members in 2015. It just knows they were members when they took the survey in 2017. GWWC had large growth around that time, so that’s a thorny issue. Then the 2018 survey solves the ‘when did they join’ problem but can’t handle any level of donations not exactly aligning with the 2017 calendar year.
My best guess thinking over all this would be that 73% of the GWWC members in this EA survey sample are compliant with the pledge, with extremely wide error bars (90% confidence interval 45% − 88%). I like Jeff’s suggestion below as a way to start to reduce those error bars.
Quick note on the ‘bunching’ hypothesis. While that particular post and suggestion is mostly an artefact of the US tax code and would lead to years that look like 20%/​0%/​20%/​0%/​etc., there’s a similar-looking thing that can happen for non-US GWWC members, namely that their tax year often won’t align with the calendar year (e.g. UK is 6th April − 5th April, Australia is 1st July − 30th June I believe).
In these cases I would expect compliant pledge takers to focus on hitting 10% in their local tax year, and when the EA survey asks about calendar years the effect will be that the average for that group is around 10% but the actual percentage given will range anywhere from 0 − 20% (if ~10% is being given), but often look like 13% one calendar year, 8% the next, 11% the year after that, etc. In other words, they will appear to be meeting the pledge around 50% of the time in your data. Yet the pledge is being kept by all such members continuously through that period. Eyeballing your 2017 graph of the actual distributions of percentages given, there are a lot of people in the 8-10% range, who are the main candidates for this.
Since both most US members and most non-US members have good reasons to not hit 10% in every calendar year, the number I find most compelling is the one in the bunching section that averages 2015 and 2016 donations (and finds 69% compliance when doing so). But that number suffers from not knowing if those people were actually GWWC members in 2015. It just knows they were members when they took the survey in 2017. GWWC had large growth around that time, so that’s a thorny issue. Then the 2018 survey solves the ‘when did they join’ problem but can’t handle any level of donations not exactly aligning with the 2017 calendar year.
My best guess thinking over all this would be that 73% of the GWWC members in this EA survey sample are compliant with the pledge, with extremely wide error bars (90% confidence interval 45% − 88%). I like Jeff’s suggestion below as a way to start to reduce those error bars.