The Most Corrupt Counties in Kenya Ranked

This is the second part of my corruption series, where I try to figure out whether we can tell which counties are corrupt and whether some counties are more corrupt than others. If you haven’t already, please read the first article here.

To recap, though, there’s a good reason why we should know whether counties are corrupt and by how much. Previous evidence has shown a strong correlation between corruption and poor economic growth, and we need to investigate and see if that relationship exists locally. In the meantime, it’s worth noting that Kenya ranks as one of the most corrupt countries in the world, with a Corruption Perception Index (CPI) score of 32 out of 100.

The problem in Kenya and other African countries is not just the embezzlement of funds, bribery, and other openly corrupt behaviors; it is also the cultures of corruption, which are behaviors that, while not explicitly illegal or corrupt, can only happen under the auspices of a highly corrupt environment. In my calculations of a sub-national corruption index, I consider these factors and develop an index that captures not only explicitly corrupt behavior but also behaviors and uses of funds that can be regarded as suspiciously corrupt. These include county expenditures on domestic and foreign travel and monthly sitting allowances for MCAs.

In addition, I also look at other explicit evidence of corruption from the EACC. For example, I analyzed all the tweets by EACC from 17th March 2024 to 1st February 2023. I then looked at evidence of corruption in the counties, ranging from whether any county executives were being investigated for graft or whether the EACC had impounded any other county officials for corruption. I also looked at the recently published report by the EACC titled “National Ethics and Corruption Survey1.” The survey provides evidence of corruption in the counties across several metrics, such as the prevalence of corruption, the likelihood of bribery, the impact of bribery, the size of bribery by county, and bribery in the counties as a share of the national bribe.

This word cloud shows the most common words from the EACC Twitter (X) page. It’s pretty informative because it shows us the most common forms of corruption, the individuals involved, and which counties are mainly implicated.

Our calculations will, therefore, be based on the following nine measures:

1. Counties that were adversely mentioned or are being investigated by EACC for corruption.
2. Likelihood of bribery demand2 in each county (EACC survey).
3. Prevalence of bribery3 by county (EACC survey).
4. Impact of bribes4 in each county (EACC survey).
5. The average size of a bribe in each county (EACC survey).
6. Share of national bribe5 by county (EACC survey).
7. Average Monthly Sitting Allowance for MCAs FY22/236
8. Expenditure on domestic travel by County FY22/23
9. Expenditure on foreign travel by County FY22/23

Using these variables, we can come up with a corruption score that we can use to rank the counties from the most corrupt to the least. As discussed in the previous post, it requires normalizing all these variables into a standardized scale, multiplying them by certain weights, and then summing the figures into one overall score.

Step 1: Getting z-Scores

In the table below, we can see both the formula for calculating z-scores and the corresponding z-scores. Essentially, we need the mean and the standard deviation of the values. The mean of our variable (Graft_Mention_EACC) was 3.06, and the SD was 5.495. For a county like Nairobi with 33 mentions, you do (33–3.06)/5.495 = 5.449. Do the same for all the other eight measures we are looking at. In the last column, we can see our z-scores for each county.

Step 2: Weighting

Someone might ask, how do we know that whichever variables we use are reasonable measures of corruption, and if indeed they are suitable measures, do they all measure corruption equally? The answer to the second question is no. I believe the best evidence of bribery is successfully arresting and charging a corrupt person or, at least, successfully seizing stolen property or recovering proceeds of corruption. That is the best evidence of corruption we can have. In such a case, data that entails corruption cases that were successfully prosecuted would be weighted the highest since it has a robust correlation with actual corruption.

In Kenya, this depends on who the EACC pursues and who they choose to charge formally, and because the EACC does not prosecute, the rest is left to the ODPP and the courts. This explains why we are using EACC mentions. It is the closest we can get to the smoking gun, considering that once cases are forwarded to the ODPP, he might choose to litigate, shelf, or ask for more evidence, a behavior that takes us closer to politics than science. So, in this case, we stick with EACC mentions as the best evidence of actual corruption.

If some variables are better predictors of corruption than others, then the only way to scale the information they give is to weigh them differently. There are many ways to do this. One can, for example, run a correlation or regression and see how much variance is explained by each variable or how strongly the variables correlate with actual corruption. In some cases, however, weighting can be more arbitrary, assigning weights based on expert judgment. I used the Principle Component Analysis (PCA) for our case.

Step 3: Principle Component Analysis

PCA gives us two critical pieces of information for this post. First, it tells us our most essential variables when measuring corruption. We have a list of nine measures, but PCA can exclude some and leave us with the most important. From these, we can choose to discard all other variables and only remain with the ones suggested by PCA. Second, it tells us how much these variables matter. It is from these that we derive the weightings for our data. But first things first, PCA gives us this first table.

Some might be interested in whether conducting a PCA was the best thing we could have done. The 0.632 KMO test tells us precisely that, and the 0.000 is the significance level of that result. I am writing a separate article explaining why I am moving away from these methods of looking at data. In the meantime, we follow doctrine, and 0.632 is reasonable enough for our analysis. Others might say 0.7 is the cut-off point or 0.8, but you know what? All these are arbitrary benchmarks plucked from thin air.

We are interested in the table below. It shows us eigenvalues from the PCA of our variables. We see that our first three components explain 73.9% of the variance. It also suggests that we could exclude measures that weakly correlate with our variable and remain with the first three only. The first component explaining 36.8% of the variance is the EACC mentions (3.309) we discussed earlier, followed by the likelihood of bribery (2.192) in each county. The third is the prevalence of bribery (1.147) in the counties. The latter two come from the survey done by the EACC, and coincidently, the best explanation of corruption comes from our analysis, where we extracted counties mentioned in corruption-related activities on EACC’s Twitter (X) page.

Only the eigenvalues are relevant for this discussion, and I will use them as weights. We have nine variables from which we have extracted z-scores. Each one of them will be multiplied by its corresponding eigenvalue. And because we are also interested in cultures of corruption, we shall not exclude any of the variables this time round. The table below shows what we get when we multiply the z-scores for each variable and the corresponding weight. We now have standardized values that are weighted. Let’s call them composite scores.

Step 4: Sum them up

Our next step is to sum up the weighted composite scores for each county across the nine variables. Since each county had different original values and thus different z-scores, the weighted sums will give us the overall score for the county. We can use these scores to rank the counties from the most corrupt to the least, as seen in the table below. Remember that our measures are weighted differently. Suppose County A was mentioned by the EACC 6 times, and County B spent a conspicuous amount of money on domestic travel. Our analysis shows that ranking will be biased (weighted) to incriminate the former as more corrupt. The reason is that while counties might misuse their budgets for domestic travel, thus symbolizing corrupt behavior, that measure is not a stronger predictor of corruption than being caught by the EACC engaging in corruption.

Step 5: Ranking Counties by Corruption

On the left, we have Kenya’s top 15 most corrupt counties; on the right, we have the least corrupt counties. What do you think of the list? Has your experience of corruption in any of these counties been different? Do you believe our analysis has been unfair to Nairobi? Which county do you think needed to be higher up or lower? I’ve left out counties in the middle, but if you are interested in the raw dataset, view it here.

Business Daily recently published an article based on the Auditor General’s report about six counties earmarked for poor management of funds. From their list, three counties (Nairobi, Kiambu, and Baringo) are on our top 15 list. Only Narok was on the list but appears in position 31 of our list. The others were Nyamira and Tana River, ranked 27th and 28th on our list.

Looking to publish these ranks every year to see how counties change positions over the years and whether Nairobi redeems itself in the long run.

Notes

1 National Ethics and Corruption Survey, 2023: Evidence from Households in Kenya.

2 “The likelihood of bribery indicator represents the number of respondents from whom bribes were demanded or expected as proportion of the total number of respondents who reported seeking public services or visiting an institution or county office, respectively.”

3 The prevalence of bribery indicator captured the portion of respondents who paid a bribe. This indicator represents the number of respondents who paid bribes as a proportion of the total number of respondents who reported seeking public services or visiting an institution or county office, respectively.

4 The impact indicator represents the proportion of respondents who reported having accessed a particular service, institution, or county only after paying a bribe.

5 The share of national bribe indicator measures the proportion of actual bribes paid as a percentage of all bribes reported to have been paid for a service in an institution or a given county.

6 County Governments Budget Implementation Review Report FY 2022/2023. Office of the Controller of Budget.