Writing a statement of work for an analytics study

The statement of work decides the study before anyone bids

A one-time analytics study is bought on paper. The firms that read your statement of work cannot see the meeting where the question came up, the data system everyone already knows is messy, or the deputy director who needs an answer by budget season. They see the document. If the document is vague, careful firms price for the risk they cannot see, less careful firms price as if there were none, and you end up comparing quotes that describe different studies.

A good study SOW does three jobs at once. It lets a stranger price the work. It lets an evaluator tell a strong approach from a thin one. And it gives the program office something to enforce when the draft report arrives. Most weak SOWs do the first job passably and miss the other two.

What follows is the section-by-section structure we would expect to see in a study SOW that draws strong quotes, with the mistakes that most often undo each one.

Background: what a stranger needs to know

The background section is for the reader who has never heard of your program. Two or three paragraphs is usually right. Describe the program or policy being studied, who it serves, how long it has run, and why the question is being asked now.

Then describe what is already known. If an earlier evaluation exists, say so and say where it fell short. If staff have run a preliminary analysis, describe what it showed. Quoters use this to judge how much of the work is new, and the honest answer changes their hours more than any other fact in the document.

Leave out the mission statement and the agency history. Neither helps anyone price or plan.

Objectives framed as the decisions the study informs

This is the section most SOWs get wrong. Objectives are written as activities: "analyze program data," "assess outcomes," "evaluate effectiveness." Those are things a firm does. They do not tell anyone what a finished study has to answer.

Frame each objective as a decision and the question that informs it. For example: "Leadership will decide in the next budget cycle whether to expand the program to additional counties. The study must estimate the change in the outcome attributable to the program for participants enrolled in the last three years, and how that estimate varies by county size." A firm reading that knows what kind of design is required, what comparison it needs, and what precision will matter.

List the research questions explicitly and number them. Numbered questions become the backbone of the final report and the first acceptance criterion: did the report answer each one, including saying plainly when the data could not support an answer.

That outcome is more common than a late study. A study that cannot tell you which decision it serves will answer a question nobody asked, on time and within budget. Spending an extra week on the objectives section is the cheapest insurance in the whole procurement.

Tasks: what has to happen, not every keystroke

Tasks describe the work in a sequence the government can follow. A typical study has five or six: a kickoff and work plan, data acquisition and preparation, analysis, interim findings, a draft report, and a final report with a briefing.

Say what must happen in each task and leave the method to the quoter unless the method itself is a requirement. If the agency needs a specific approach for consistency with prior years, say so and explain why. If not, ask quoters to propose the method and make method quality part of the evaluation. Prescribing every step turns the quote into a price for your design, which removes the reason to hire experts in the first place.

The work plan task deserves special attention. Require a written design memo early in the period, usually within the first month, that restates the research questions, specifies the method, lists the data elements needed, and names the risks. Government approval of that memo is the point where both sides agree what the study is. Every later disagreement gets resolved against it.

Deliverables with formats and acceptance criteria

A deliverable has a name, a due date or trigger, a format, required content, a government review period, and acceptance criteria. When any of those six is missing, the deliverable becomes a negotiation.

Acceptance criteria for a study describe completeness and quality, never agreement with the findings. They answer questions such as: does the report address every numbered research question, does it state data sources and limitations, does it include uncertainty for each estimate, can the delivered code reproduce the tables, and does the document meet federal accessibility requirements.

The difference between a vague deliverable and an enforceable one is easier to see than to describe.

Always split the report into a draft and a final. Give the government a fixed review window for the draft, require consolidated comments from one point of contact, and require the firm to answer each comment in the final or explain in writing why it did not. That gives the deliverable a defined number of turns.

Data access and government-furnished information

Data access is where analytics studies lose the most time, and it is almost never described well. "The government will provide all necessary data" is a promise that nobody can price.

Describe each data set the study will use: what it contains, roughly how many records, what years, who owns it, where it lives, and what form it will arrive in. Say whether the firm will receive extracts or work inside a government environment. Say what agreements are needed before data can move, such as a data use agreement with a state partner, and who is responsible for getting them signed.

Then state the government's obligations with dates. "Within 15 business days after award, the government will provide extracts of the following tables" is a commitment both sides can plan around. If the government cannot meet it, the schedule moves with it, and the SOW should say so in one sentence.

If the data includes personal information, involve the privacy office before release. The contracting officer may need to add privacy and security clauses, and the firm may need to meet specific handling requirements. Quoters price those requirements, so they belong in the SOW, not in a kickoff surprise.

Schedule, meetings and travel

Write the schedule as a table of deliverables with due dates expressed relative to award or to a government action. Dates tied to government actions, such as "10 business days after receipt of data," protect both sides from delays neither controls.

Set the meeting cadence explicitly: a kickoff within two weeks of award, a short standing check-in every two weeks, and a briefing at each major deliverable. Say whether meetings are virtual. A firm pricing a study with weekly in-person meetings prices a different study than one with biweekly calls.

If any travel is needed, describe it: how many trips, to where, for how many people and days. Ask for travel on its own line with a not-to-exceed amount rather than buried in labor rates. Trips that never happen then cost nothing.

Qualifications for key roles

A study is mostly the judgment of a few people. Name the roles that matter, usually a study lead and a senior analyst or methodologist, and say what each must be able to do.

Write qualifications as capabilities tied to the work, not as years of experience copied from a labor category. "Has led at least two quasi-experimental evaluations using administrative data, including the final report" tells an evaluator what to look for. "Fifteen years of experience" only filters out capable people.

Designate those roles as key personnel and require that substitutions be approved in advance by the contracting officer with someone of equal or better qualifications. That one requirement does more to protect the study than any amount of evaluation language, because it keeps the people you selected on the work.

When people are part of the study: human subjects and the Paperwork Reduction Act

Two sets of rules can add months to a study that involves surveys, interviews or data about individuals, and both need to be in the SOW from the start.

Human subjects protections. Research involving human subjects that a federal agency conducts or supports generally requires review by an institutional review board or a determination that it is exempt. Many program evaluations qualify for an exemption or fall outside the definition of research, but that determination is made by the agency's designated office, not assumed by the program. The SOW should say who will make it and allow time for it.

The Paperwork Reduction Act. When an agency, or a contractor on its behalf, collects information by posing identical questions to ten or more members of the public, the collection generally needs approval from the Office of Management and Budget. The process includes public comment periods published in the Federal Register and commonly takes several months. A survey planned for month three of a twelve-month study, with no clearance started, will not happen in month three.

The practical answer is to identify any collection before the SOW goes out, ask whether an existing approval or a generic clearance covers it, and if not, either start the clearance early or design the study around administrative data. Then write the dependency into the schedule.

Data rights and delivered code

The point of a study is a result the agency can defend and build on. That requires the work behind the result, not only the report.

List the code, the cleaned analytic files, the data dictionary and the documentation as deliverables, with the same acceptance treatment as the report. Require that the code run end to end from the delivered inputs and reproduce the main tables and figures. Ask the contracting officer to confirm that the data rights terms in the contract give the government the rights it needs to run, modify and share these materials with other contractors later.

Ask quoters to identify any existing proprietary tools they intend to use and how the government will be able to reproduce results without them. A firm that relies on a licensed model or platform can still be the right choice, but the agency should know before award that rerunning the analysis will need that license.

The section checklist

Before a study SOW leaves the program office, it should contain each of these.

• Background that a stranger can read in five minutes, including what is already known
• Objectives framed as decisions, with numbered research questions
• Tasks in sequence, with a design memo approved early in the period
• Deliverables with format, content, due date, review period and acceptance criteria
• Data and government-furnished information, described set by set, with government delivery dates
• Schedule and meeting cadence, with dates tied to award or government actions
• Key roles and qualifications, written as capabilities, with substitution approval
• Travel, described trip by trip, on its own priced line
• Human subjects and information collection, identified with an owner and time allowed
• Code, data and documentation delivered, with a reproduction test and confirmed data rights
• Place of performance and security requirements, including where data may be stored and processed

Common mistakes and the fix for each

Most problems in a study SOW repeat across agencies. These are the ones we see most often in the requests we read.

Mistake	What it causes	The fix
Objectives written as activities	Quotes describe effort, not answers, and cannot be compared	Name the decision and number the research questions
"All necessary data will be provided"	Careful firms price high; the schedule fails in month one	Describe each data set, its owner, its form and the delivery date
Deliverable titles with no acceptance criteria	Endless revision cycles and disputes at the final report	Format, content, review window and a completeness test for each
Every step of the method prescribed	You buy your own design and lose the expertise you were paying for	State required methods only where consistency demands them
Years of experience as the only qualification	Filters out capable people and admits weak ones with long resumes	Describe what the person must have done, tied to this study
A survey with no clearance plan	The collection cannot start when the schedule says it will	Identify collections early and put the approval time in the schedule
Only a report delivered	The analysis cannot be rerun or extended after the firm leaves	Deliver code, data list and README, and test that it reproduces

None of these fixes require legal drafting. They require the program office to write down what it already knows, in a form a stranger can use. That is most of the job.

Frequently asked questions