Starting on 29th September, 160 German citizens will participate in a series of meetings to discuss national food and nutrition policy. These 160 people will form a Citizens' Council; the council was instigated by Germany's national parliament, the Bundestag, and the Council has been asked to present recommendations for action by February 29, 2024.
The Citizens Council is tasked with understanding how Germany can eat more healthily and more sustainably. They will cover topics such as the labeling of food with a view to animal welfare, environmental standards, food waste, and the price and taxation of food.
Sortition Foundation worked with the Nexus Institute to recruit these 160 people. We describe the process in detail below. The final part of the process was a live lottery, in which Bärbel Bas, the President of the German Bundestag, drew the winning assembly "out of a hat" (actually 3 transparent containers). This lottery was live-streamed from the Bundestag (see video above or photo).
Stage 1: Postal Invitations
We started by selecting 84 municipalities by lottery out of the approximately 11,000 municipalities all across Germany. These 84 lottery-selected municipalities were chosen in such a way as to ensure a population-weighted distribution across ever Germany State, and also ensured different size municipalities were selected (German municipalities can range in population from less than 2000 to more than 3.5 million). This was necessary as the resident registers in Germany are held at the municipal level.
We then wrote to these 84 municipalities to obtain a lottery-selected list of their residents from each municipalities complete resident list, such that the final list included approximately 20,000 people (in the end we had 19,300 people). We wrote to these people to invite them to express an interest in participating in the Citizens' Council via a webpage or by calling a free phone number. When expressing their interest people gave some demographic and other data (for reasons explained below).
The registration period lasted for approximately a month, at the end of which 2220 people had expressed an interest in being a part of the assembly.
Stage 2: Stratified Selection
As part of the registration process, all potential participants were required to share some basic information about themselves. We asked them to confirm their name and address (this allowed us to confirm that they had received an invitation), their date of birth, their gender and their level of education. We also asked them whether they were vegan, vegetarian or neither; this question was asked to ensure that there was people with a range of eating habits in the Citizens' Council as this was considered highly relevant to the subject that the assembly was tasked to consider.
We then used this information as input into our "sortition algorithm"; this process extended our standard technique (more below) by selecting 1000 possible assemblies, each assembly comprised of 160 people selected by lottery from the pool of 2220 people who had registered their interest in such a way that every assembly is a representative sample (so, for instance, the age profile of the assembly members is broadly similar to the age profile of the German population as a whole). Details of the specific algorithm we used, including information about the fairness of the algorithm, can be found here.
For this event we stratified across six categories, as follows. We provide translations where needed to understand the pie charts that follow. Further information about how to understand the pie charts is below.
- Gender ("geschlecht") into man ("herr"), woman ("frau") and other ("divers");
- Age ("altersgruppen");
- Education ("bildungsabschluss") into high ("hoch"), middle ("mittel") and low ("niedrig");
- State ("wohrnort im bundesland");
- Municipality size ("gemeindegrossenklassen") into large ("gross"), middle ("mittel") and small ("klein");
- Diet ("ernahrung"): into vegan ("vegan"), vegetarian ("vegetarisch") and neither ("nicht").
The way to understand these pie charts is as follows:
- Column 1 (Target): These pie charts give information about the German population as a whole, using various publicly available statistics. As an example, in the second row, you can see that 26% of the population in population is aged 65 or over ("65 Jahre").
- Column 2 (Respondents): These pie charts summarise the information that was provided to us by the 2220 people who signed up as potential participants. There is some skewing in statistics here compared with our target. The most notable skewing is for education ("bildungsabschluss") where the percentage of respondents with high ("hoch") qualifications is much higher than we would expect from a random sample of the population.
- Column 3 (Confirmed Selected): These pie charts summarise information about a typical assembly selected using our algorithm. Notice that, thanks to our use of a sortition algorithm, the pie charts in this column are very similar to the target charts in the first column. As part of the recruitment process all of these people will be directly contacted to confirm that they were still willing and able to participate -- in the event that this is not the case, we can use the sortition algorithm to replace people who dropped out with others who share similar characteristics. There will typically be some small variation in these pie charts, due to the random selection process, but all will be broadly similar.
Stage 3: Live Lottery
Ordinarily, when we support democratic lotteries, at the end of Stage 2 we select by lottery one group of people that fit all the targets and these people form the selected Citizens' Council. However, for increased transparency, the Bundestag wanted us to do a bit more than this! Although all of the code we use to select people is open-source and publicly available for inspection or download, there was a desire to integrate a live physical lottery into the process to further increase the transparency and (hopefully) also generate some interest and excitement in the Citizens' Council by demonstrating clearly that the process is a lottery.
To these ends, we didn't just select one group of 160 people, we selected 1000 groups of 160 people. Every one of these 1000 possible councils satisfied the targets described at Stage 2. What is more this set of 1000 groups had some special properties which we detail below.
The 1000 possible councils were each assigned a number from 000 through to 999. All of the 2220 respondents appeared on multiple possible councils and so, after we selected these 1000 groups, our partners at Nexus Institut contacted the 2220 respondents to tell them which possible councils they were on. All of the respondents were then invited to tune into the live lottery (see the video at the top).
This live lottery was the final stage of the recruitment process. As you will see if you watch the video, the president of the Bundestag selected Council numbers 1, 8 and 7. This means that the people on Council 187 were selected to form the German Citizens' Council on Food Policy. The pie charts in the third column above give the actual figures for this finally-selected council.
Properties of the 1000 Citizens' Councils
Another reason for a live lottery was to ensure that each of the 2220 respondents had an obvious and transparent chance of participating in the Citizens' Council. This was certainly achieved: every one of those 2220 respondents appeared on at least one of the 1000 possible councils. However, our algorithm chose the 1000 councils very carefully to make sure that we did better than this. An explanation follows but, be warned, it is a little technical!
In a perfectly fair world, we would want every respondent to appear on the same number of the 1000 possible councils. This would mean that every respondent was on 160*1000/2220 = 72.1 councils and that each respondent has about a 7 in 100 chance of being selected.
However, the fact that we want a representative assembly -- one that meets the targets described above -- means that it is not possible for every respondent to have an equal chance of being selected for the assembly. For instance, we saw above that in the education category, many more highly qualified people responded, compared with lower qualified people. Indeed only 70 of the respondents had lower qualifications and our target asks for 33 people in the assembly to have this level of education. This means that a little less than half of those 70 respondents will be in the final assembly and so these 70 people are going to be in around half of the 1000 assemblies, which is many more than the 72 we expect if everything is perfectly fair and the pool of respondents matched the German population exactly. Likewise, people with high levels of education are likely to be in a smaller number of the assemblies than the 72 we might expect.
So, having accepted that perfect fairness is not possible if we want to match our demographic targets, our algorithm first calculates what the maximum possible fairness is, and seeks to achieve this. In mathematical terms, the algorithm seeks to maximise the minimum number of assemblies that any respondent appears in. In this particular case, the algorithm was able to make sure that every respondent was in at least 18 assemblies. Such a person would have a roughly 2 in 100 chance of being selected, which is less than the 7 in 100 we would like but is the best the maths will allow.
The graph below shows how many assemblies each respondent was in. You can see that more than three quarters of the respondents were in between 20 and 40 assemblies, meaning they had between 2 in 100 and 4 in 100 chance of being selected. There are a small number of people (44 in total) who are in more than 520 assemblies, meaning they had more than 1 in 2 chance of being selected. These will likely have been people with lower educational qualifications, as described above.
It is important to emphasise that, although we said that the algorithm "chose the 1000 councils very carefully", it is still a random algorithm -- however the randomness is constrained to ensure that all of the targets are met and we try and maximise fairness, as described above. One consequence of the requirement that we maximise fairness is that some of the 1000 assemblies were, in fact, duplicates. The algorithm produces repeats because, for instance, there may be particular respondents who do not easily fit into many assemblies and so, to ensure they don't appear in too few assemblies, some assemblies are duplicated.
The graph below shows the level of duplication among the 1000 assemblies. There were 892 distinct assemblies in total, of which 789 appeared exactly once, 99 appeared twice, 3 appeared three times and 1 appeared four times.
What happened next?
The live lottery got significant coverage in the Germany press, noticeably from Tagesschau, ZDF, Welt and RTL. The official Bundestag report of the recruitment process is here. You can also explore the demographics of the assembly by clicking through a webapp created by Philipp.
We will provide details about the recommendations of the Citizens' Council, once they are made public in February 2024.