Ask how many people work in Uganda's dairy sector and you will get answers ranging from one hundred thousand to nearly five million, depending on who you ask and what they are counting. Both numbers can be defended. Neither is wrong. That gap, and what sits inside it, is the whole problem of employment modelling in an agricultural value chain where most of the work is informal, seasonal, and invisible to any register.
I spent part of this year building a bottom-up estimate of employment across Uganda's dairy value chain, as part of a skills demand and supply study run through the FAO Investment Centre in support of a World Bank operation. This post is about how that kind of estimate is actually built, where it gets hard, and what separates a number a reviewer will trust from one they will quietly discard.
Why the easy answer is the wrong one
The tempting approach is top-down. Take a sector GDP contribution, apply a labour intensity ratio borrowed from somewhere plausible, and report a round number. It is fast and it is nearly always indefensible, because the ratio is doing all the work and nobody can see it.
Uganda's dairy sector makes the case for a different method. Roughly 80 percent of marketed milk moves through informal channels: raw milk sold by roadside vendors, carried on the back of a boda-boda motorcycle, boiled and resold in a trading centre, or turned into ghee in a household kitchen. None of these actors files a return. None appears in a processor's payroll. A top-down number cannot see them, and they are most of the sector.
A number you can interrogate is worth more than a number you have to take on faith. Exposure is the point, not a weakness.
The approach: build it in pieces, show every piece
The model breaks the value chain into six nodes and estimates each separately. The logic in each node is deliberately simple: for each type of actor you establish a count of units and a figure for jobs per unit, and you multiply.
Three disciplines make the difference between this being rigorous and being arithmetic dressed up as analysis.
Separate what is sourced from what is assumed. In the working model every input cell is colour-coded. A figure traceable to the UBOS livestock census looks different from a figure estimated from field observation, which looks different again from a placeholder waiting for better data. Hiding the judgement does not make it disappear. It just means the reviewer finds it later, and trusts you less for having buried it.
Count persons engaged and full-time equivalents separately. In a sector built on part-time household labour, these two numbers are wildly different. A farmer milking three cows before walking to a second job is engaged in dairy, but not full-time. You need both, side by side, or the number misleads whoever reads it.
Build a cross-check from an independent direction. Any single method can drift. Alongside the actor-by-actor build, I estimated informal marketing employment a second way, using job-per-litre coefficients from FAO's dairy work in Kenya. When two methods aimed at the same quantity from different starting points land close together, confidence rises. When they diverge, you have found something worth understanding.
The hard part is not the maths. It is the boundary.
Here is the decision that moved my headline figure by 2.4 million people, and it has nothing to do with any multiplier.
The question is which households count. Uganda's National Livestock Census 2021 records about 2.3 million cattle-keeping households. A separate figure puts dairy farming households at 1.2 million. The difference is not an error in either source. They are counting different populations.
Anchor the model to 2.3 million and you get a total north of five million persons engaged. Anchor it to 1.2 million and you get under three million. Same model, same multipliers, one definitional choice, and the answer roughly doubles.
The single most important sentence in an employment model is the one that states which population it counts. A reader who does not know that will compare your number against someone else's, see a gap of two million people, and conclude that one of you cannot count.
Where the estimates fought back
Independent sources disagree, and reconciling them is the work. An independent sector factsheet put total dairy employment at about 5.03 million. My bottom-up model produced 5.2 million. Reassuring, until you look closely and notice the factsheet counted 145 milk collection centres where the government figure is 729. Agreement at the top can hide disagreement in the parts. Only a component-by-component comparison catches it.
The commercial tier is where a skills study lives, and it is the easiest to undercount. The census found that 18.6 percent of cattle-keeping households kept exotic or crossbred animals, implying something like 428,000 commercially oriented farms. An early version of my model carried closer to 79,000 in its intensive categories. For the headline total this barely matters. For a skills study it matters enormously, because the commercial tier is precisely where hired labour, formal qualifications, and identifiable training demand concentrate.
A figure you cannot source is a liability, however plausible. One widely repeated figure in my model — a national count of milk bars — traced back to a study I could never locate. Rather than quote it as fact, I flagged it as an order-of-magnitude estimate awaiting a source. Reviewers forgive the gap. They do not forgive discovering the guess.
What this is really about
Numbers like these do not exist to be impressive. They exist to be used: to size a training programme, to justify an investment, to set a policy target, to tell a funder where the jobs actually are. A number that cannot survive scrutiny fails at the one job it has.
The craft is not producing a figure. It is producing a figure whose every assumption is visible, whose boundary is stated, whose weak points are flagged rather than hidden, and which a sceptical reviewer can pull apart and put back together with their trust intact.