Friday, 19 August 2016

Some Thoughts on CCMG's Parallel Vote Tabulation (PVT) Exercise

The Christian Council Monitoring Group (CCMG), an alliance of faith-based organizations in Zambia, carried out a Parallel Vote Tabulation (PVT) exercise during last week's presidential elections. PVTs are a type of vote count verification mechanism and can be useful in checking the validity of final election results. A PVT was actually critical in proving to Dr. Kenneth Kaunda that he'd lost the 1991 elections. Dr. Kaunda is obviously a gracious man and would likely have accepted the results anyway. It's not improbable to imagine, however, that he would have protested the extent of his loss in the absence of a PVT (he only scooped 25% of the vote that year).

The CCMG has subsequently released a statement on the accuracy of the count from last week's presidential elections and concluded thus:

Now that the Electoral Commission of Zambia (ECZ) has declared results for the 2016 presidential elections, CCMG affirms that its PVT estimates for the presidential election are consistent with the ECZ’s official results. All stakeholders, particularly political parties, that participated in the election should have confidence in the ECZ’s presidential results. 

What is a PVT and how is it carried out in practice? And what does it mean for PVT estimates to be consistent with ECZ's official results?

In theory, a PVT works like this: a trusted entity (like CCMG) places agents at polling stations across the country. These agents observe all facets of the voting process from the voting itself to the counting of ballots afterwards. And the main and important idea is this: as the Electoral Commission collects and aggregates counted votes at polling stations across the country, the CCMG also does the same (hence the phrase "parallel tabulation"). At the end of the process, the Electoral Commission announces their final tally and so does the CCMG. The two totals should ideally be the same.

In practise, however, the CCMG cannot manage to send agents to everyone of the country's polling stations (which numbered 7,700 last week). They have to pick a "representative sample" of the 7,700 polling stations. Alternatively, a PVT can be conducted across all polling stations (called a Comprehensive PVT). But quite apart from it being prohibitively costly, a Comprehensive PVT, surprisingly, is more prone to systematic errors than simply sampling polling stations (read more here).

In picking its sample of polling stations last week, the CCMG used simple random sampling (SRS) techniques. But to guard against the possibility that the SRS only picked polling stations in one part of the country, the CCMG employed a version of Clustered SRS so that eventually polling stations were sampled from every province, district and constituency. The final sample contained 1,001 polling stations whose distribution largely mirrored the actual distribution of polling stations across the country. For example, since the Copperbelt holds 14% of all polling stations in the country, the CCMG's sample also had 14% of the polling stations coming from the Copperbelt. Using random sampling ensures that the characteristics of the polling stations (skill set of polling agents, availability of electricity etc...) in the sample are representative of characteristics across all the 7,700 polling stations. We don't want to only pick polling stations where there is electricity because this might not be representative of all polling stations in the country.

Since the CCMG worked with a sample of 1,001 stations and not the entire population of 7,700 stations, the CCMG's pronouncement on the election (for example what percentage of the vote went to candidate X) is called an estimate. And this estimate will come with a margin of error precisely because we are working with a sample and not the actual thing. (Technical note: The margin of error comes from the idea that if CCMG repeated this process of randomly picking 1,001 polling stations many times, they'd obtain a different estimate each time. The collection of these different estimates would form a distribution from which we could work out the margin of error). Ideally you want the margin of error to be as small as possible.

Table 1 below provides estimates of vote shares for each candidate from CCMG's PVT in addition to ECZ's official vote shares from last week's elections. The table also provides margins of error.      

                                  Table 1

An ECZ vote share is consistent with CCMG's vote share if the ECZ share is contained within CCMG's share plus the margin of error. For instance, Edgar Lungu of the Patriotic Front was declared by the ECZ to have obtained 50.4% of valid votes cast (see the second column). The CCMG's PVT, based on 1,001 polling stations, estimated Mr. Lungu's share at 50.2% with a margin of error of +/-2.5%. This implies a lower bound estimate of 47.7% and an upper bound of 52.7% of the vote share. ECZ's announced share for Mr. Lungu falls right within this range and is therefore consistent with CCMG's estimate.  

Now, there is an important proviso to this process that I have not seen anyone else raise in the discussion of the CCMG's PVT this past week. And this involves the issue of measurement error which is different from the margin of error. The margin of error is a given in a process where you use a sample to make statements about a population. In making such a statement, there is, however, an assumption that what you are measuring (eg the vote count) is done properly. In practice, you are always going to commit some type of measurement error but the idea is that such errors should be small and not systematic. An illustration: suppose there are two groups of people A and B that are engaged in trying to figure out the average weight of Lusaka residents. Group A is going to use a random sample and Group B will visit each one of Lusaka's residents and take their weights. Both groups are each given a scale to measure the weight. But it turns out that both scales were manufactured with the same type of defect - they add 2kilos to the actual weight. If the sampling is done right (i.e. margin of error is minimized), the estimate of the average from Group A and the one from Group B will be consistent with one another, but might be 2kilos more because of a systematic pattern in measurement error. So if there is measurement error and if it's systematic, the PVT process will not be informative of the voting process. Just as the average weights from our defective scales will not be informative of the actual average weight in Lusaka.

Lastly, according to CCMG's PVT, Mr. Lungu's vote share could range anywhere from 47.7% to 52.7% of valid votes cast. This PVT tells us that there is likely a range of vote shares where Mr. Lungu did not get the majority needed to prevent a run-off (from 47.7% to 50%) and there is likely a range where he did (50% + 1 to 52.7%). Note that this conclusion holds even when there is zero measurement error, which the CCMG assumes here. I don't know what to make of this last point or if it's even important at all.


Monday, 1 August 2016

Where to Voter Turnout?

The 2016 General Elections in Zambia are less than a fortnight away - in 10 days time to be precise. At this late stage in the game, discussion seems to have shifted from talking about the popularity of the PF/UPND, the two clear front-runners for the presidency, to talking about what voter turnout will be like on August 11.

Focusing on voter turnout, at this stage, is quite understandable: there's little that the PF/UPND can do now to convince a *significant* fraction of PF/UPND voters to vote for them. Minds are already largely made up. The question now is will voters turn out in large numbers to vote for their preferred candidates?    

So what's been the historical pattern of voter turnout in Zambia's General Elections since the reintroduction of multiparty democracy in 1990?

Figure 1 below plots voter turnout against date of election. The first thing noticeable is that getting a voter turnout of at least 70% has been historically difficult -- the only time this happened was in September 2006 when the turnout was 71%. The average turnout for the 7 elections that have involved choosing a head of state since 1991 is 53% (59% if you exclude the two presidential by-elections in 2008 and 2015).

                               Figure 1: Voter turnout, 1991 to 2015

                                Source: ECZ

The second noticeable pattern from Figure 1 is that in "normal times", relatively low voter turnout has coincided with a change of government (for this conclusion, we need to exclude the extraordinary presidential by elections of 2008 and 2015 -- voter turnout in these elections is a function of many things, not least of which is the fact that the elections are not preceded by an updating of the voters' register given their suddenness). 

Figure 2 below excludes the two presidential by elections of 2008 and 2015, and there it becomes clear that the two lowest voter turnouts (1991 and 2011) coincided with a change of government (I was actually surprised to learn that the momentous elections of 1991 had such low voter turnout!). 

                            Figure 2: Turnout excluding 2008/2015 by elections

                                Source: ECZ

So what could be going on here? At an intuitive level it does make sense why low voter turnout might coincide with a change of government: the many supporters of the once popular ruling party might choose to stay away, because of disillusionment, instead of voting for the opposition party (a protest vote of sorts). At the sametime the relatively fewer supporters of the opposition might come out in full force.

So will this pattern play out on August 11? Only time will tell. And the election on August 11 is different to those that have come before precisely because there’s a rule change. It’s no longer “first-past-the goal post”. The winner has to command a majority to win the election otherwise there’s a second round of voting. And a second round might bring about a change of heart from the ruling party’s protest voters – they might think that the ruling party’s defeat (or lack of clear victory) in round 1 was punishment enough and perhaps the party has learnt its lesson. Or the opposition's near victory in round 1 might scare the ruling party's protest voters into coming out in full force for round 2 -- the thought of the opposition forming government might be too much to bear even though the ruling party has done little for them. 

But all this could be armchair theorizing on my part. So don’t listen to me but vote wisely next week.