ARDAI (TIER I) · Week 10: Advanced Research Methods for AI · The Lichfield Academy

ARDAI (TIER I) · Week 10 · Advanced Research Methods · The Lichfield Academy

The Lichfield Academy

… Beyond Knowledge, Into Mastery

· UNESCO REF · SIP-ALPHA ·

Advanced Research Diploma
in Artificial Intelligence
(TIER I)

Week 10 · Advanced Research Methods for AI · ARDAI (TIER I)

Advanced Research
Methods for AI
(TIER I)Week Ten

Causal inference, experimental design, field experiments, and advanced qualitative methods for AI governance research in African contexts.

13

Lessons

26

Video Resources

13

Lab Exercises

13

Primary Sources

Week 10 Curriculum

1Research Design for AI Studies
2Causal Inference: Potential Outcomes Framework
3Randomised Controlled Trials and Field Experiments
4Difference-in-Differences and Synthetic Control
5Regression Discontinuity Design
6Instrumental Variables and Natural Experiments
7Mixed Methods Research in AI Governance
8Ethnographic and Qualitative Research Methods
9Survey Design and Measurement in AI Research
10Systematic Reviews and Meta-Analysis
11Reproducibility, Pre-registration, and Open Science
12Research Ethics for AI Studies in Africa
13Week 10 Capstone

Voice and Audio Guide

Each lesson has a full narrated lecture available via the Lecture Panel. Individual sections can be read aloud using the Read button on each section heading. Select your voice in the topbar, control speed with the slider, and use Samantha for the clearest experience.

Lecture

Playing

Select a section above or press Play to begin.

Research Notes

Notes

How to Use

Notes are saved automatically throughout your session.

Saved automatically · Session storage

Statistical charts and research data analysis methodology

Week 10 · Lesson 1 · ARDAI (TIER I)

Week 10 · Lesson 1 · ARDAI (TIER I)

Research Design for AI Studies

Advanced research on AI requires more than technical competency. It requires a rigorous framework for deciding what questions to ask, how to answer them reliably, and what evidence is sufficient to make a claim. For African AI researchers, research design carries an additional obligation: designing studies that produce knowledge valid for African contexts, not just studies that reproduce global research paradigms in African settings.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

What Is Statistics: Crash Course Statistics

The Central Limit Theorem: StatQuest

Learning Objectives

Evaluate the major research design paradigms applicable to AI governance and impact research. Analyse the specific research design choices most consequential for AI studies in African institutional contexts. Apply a research design framework to develop a study protocol for an original AI research question. Evaluate the quality criteria that distinguish rigorous AI research design from technically sophisticated but methodologically weak work.

I. Research Design Paradigms for AI Studies

Research design is the architecture that connects a research question to a credible answer. In AI research, the choice of design paradigm determines not just the methods used but the kinds of knowledge claims that can be made. Experimental designs - randomised controlled trials, field experiments - can establish causal relationships between AI interventions and outcomes. Quasi-experimental designs - difference-in-differences, regression discontinuity, instrumental variables - can approximate causal inference when randomisation is infeasible. Observational designs - cross-sectional surveys, longitudinal cohort studies, case studies - describe patterns and associations but require careful caution about causal interpretation. Qualitative and ethnographic designs illuminate mechanisms, experiences, and institutional dynamics that quantitative designs cannot access.

The choice of design paradigm must be driven by the research question, not by the researcher's methodological comfort zone. A question about whether an AI diagnostic system improves patient outcomes requires an experimental or quasi-experimental design - no amount of technically sophisticated observational analysis can substitute for the causal identification that these designs provide. A question about how health workers experience and adapt to AI diagnostic systems requires qualitative methods - no survey instrument can capture the situational judgment and professional identity dynamics that ethnographic fieldwork reveals. A question about the scale and distribution of AI adoption across African health systems requires a carefully designed survey - experimental designs cannot answer descriptive epidemiological questions. Matching design to question is the first and most important research design decision.

African AI research contexts create specific design challenges that researchers must address explicitly rather than treating as minor complications. Randomisation is often politically or ethically infeasible when AI systems are being deployed by government agencies with distributional mandates. Administrative data - the foundation of much quasi-experimental research - is frequently incomplete, inconsistently recorded, or held by agencies that resist academic access. Survey sampling frames are often unavailable or out of date. Qualitative fieldwork in some African contexts requires navigating security constraints, institutional gatekeeping, and community trust-building processes that extend project timelines far beyond what research funders expect. A rigorous research design for an African AI study explicitly addresses each of these challenges rather than hoping they do not materialise.

Scholarly Analysis

Research design expertise is the competency that most directly determines whether an ARDAI graduate can contribute to the peer-reviewed literature, secure competitive research funding, and occupy the senior research roles at development banks, think tanks, and academic institutions where AI governance knowledge is produced. The practitioners who can design studies that produce credible causal or descriptive evidence about AI in African contexts are genuinely scarce globally - not just in Africa - and the demand for their skills is accelerating as funders and policymakers require evidence rather than advocacy for AI governance decisions.

II. Internal and External Validity in African AI Research

Internal validity refers to whether a study design can support the causal or descriptive claims being made - whether the evidence produced by the study actually answers the research question. External validity refers to whether findings from one context generalise to other contexts - whether what was found in one setting can be applied to another. Both are essential for high-quality AI research, and both face specific threats in African research contexts that researchers must understand and address.

The most common internal validity threats in African AI research are: selection bias (the populations being studied differ systematically from the populations not studied in ways that affect the outcome); confounding (other factors correlated with AI adoption also affect the outcome, making it impossible to attribute effects to the AI system specifically); attrition (participants drop out of longitudinal studies for reasons related to the outcome, biasing follow-up data); and measurement error (administrative data, survey responses, or AI system outputs are measured imprecisely or inconsistently). Each of these threats has specific mitigation strategies - randomisation addresses selection bias, control variables address confounding, intention-to-treat analysis addresses attrition, and validation studies address measurement error - but mitigation requires deliberate design choices, not post-hoc corrections.

External validity in African AI research is more complex than the standard textbook treatment suggests, because both the AI systems being studied and the institutional environments in which they operate vary substantially across African contexts. A finding about the impact of a mobile credit scoring algorithm in Kenya does not automatically generalise to Nigeria, Ethiopia, or Senegal - different mobile money penetration, different credit market structures, different regulatory frameworks, and different demographic profiles all affect how AI systems operate and what impacts they produce. Researchers must specify the scope conditions of their findings explicitly - identifying the features of the study context that are essential to the results and the features that are likely to vary without affecting generalisability.

Scholarly Analysis

The internal and external validity framework this section develops is directly applicable to critical appraisal of AI research - evaluating the quality of studies produced by others, identifying the limitations of evidence bases, and providing the kind of methodologically grounded critique that is valued in peer review, policy advisory, and grant evaluation roles. ARDAI graduates who can evaluate research quality rigorously, not just produce research fluently, are positioned for the senior roles where these skills are most in demand.

Research design framework for AI governance studies

Research design matrix showing quantitative and qualitative dimensions

III. Developing a Research Protocol for African AI Studies

A research protocol specifies in advance how a study will be conducted - the research question, the design, the sampling strategy, the data collection instruments, the analysis plan, and the timeline. Pre-specified protocols serve two functions: they force researchers to think through design decisions before data collection begins, when changing course is cheap; and they provide a benchmark against which actual study execution can be compared, enabling readers to identify deviations that might introduce bias. For African AI research, protocols also serve a third function: they are the document that research ethics committees review, that funding bodies evaluate, and that institutional gatekeepers use to decide whether to grant data access.

The key sections of a research protocol for an African AI study are: background and significance (establishing that the research question is important and that existing evidence does not already answer it); research questions and hypotheses (stating precisely what the study will answer and what it expects to find); study design (justifying the chosen design paradigm and explaining why alternative designs were rejected); population and sampling (defining who will be studied, how they will be identified, and how a representative or purposive sample will be selected); data collection (specifying instruments, procedures, and quality control measures); analysis plan (describing the statistical or qualitative methods that will be applied to address each research question, pre-specified before data collection); ethical considerations (identifying risks to participants, describing consent procedures, and explaining how data will be protected); and limitations (honestly identifying the study's design weaknesses and the threats to validity that could not be fully mitigated).

Scholarly Analysis

Protocol development skills are directly applicable to the most career-advancing research activities: grant applications to the African Academy of Sciences, the NRF, the Wellcome Trust Africa programmes, or the Gates Foundation; ethics submissions to institutional review boards; and pre-registration at the AEA RCT Registry, OSF, or EGAP. ARDAI graduates who submit a well-developed research protocol as a lab deliverable have the most important component of a competitive research funding application already written.

Primary Source

2021

Gerber, A.S. and Green, D.P. (2012). Field Experiments: Design, Analysis, and Interpretation. W.W. Norton and Company. [Updated with 2021 Africa-specific applications.]

W.W. Norton 2012 · Widely Available

1. Gerber and Green argue that field experiments are the gold standard for causal inference in social science. Evaluate the applicability of this argument to AI governance research in African contexts, where randomisation is frequently infeasible for political, ethical, or logistical reasons, and assess what quasi-experimental alternatives provide the most credible causal identification under African research constraints.

2. The book focuses primarily on experimental designs developed and validated in high-income country research contexts. Evaluate what adaptations to field experiment design are required when conducting AI impact research in African settings with lower administrative data quality, different institutional gatekeeping dynamics, and community contexts where standard consent and sampling procedures may be culturally inappropriate.

Knowledge Check

An ARDAI researcher wants to evaluate whether an AI-powered agricultural advisory system increases smallholder farmer income in three West African countries. The government agency deploying the system refuses to randomise access, arguing that all eligible farmers should receive the service simultaneously. Which research design best addresses this constraint while still providing credible causal evidence?

Expert Analysis

Difference-in-differences is the appropriate design because it provides causal identification through within-unit comparisons over time rather than cross-sectional comparisons between users and non-users. The cross-sectional survey (A) cannot control for selection bias - farmers who adopt AI advisories may differ from non-adopters in ways that independently affect income. The qualitative case study (C) cannot establish that the AI system caused income growth for the cases studied, let alone for the broader population. The observational regression (D) can only control for observed confounders; unobserved differences between adopters and non-adopters will bias estimates. DiD controls for time-invariant unobserved differences between districts, and the parallel trends assumption is testable, making it the design that provides the most credible causal evidence in this constraint environment.

Research Laboratory

Lab 10.1: Research Protocol for an African AI Study

Develop a complete research protocol for an original AI governance or impact study in an African context. The protocol must address:

(1) Background and significance: establish the research question's importance and the gap in existing evidence.

(2) Research question and design: state the precise research question, justify the design paradigm (experimental, quasi-experimental, observational, qualitative, or mixed), and explain why alternative designs were rejected.

(3) Population and sampling: define the study population, the sampling strategy, the sample size justification, and the approach to achieving adequate representation of marginalised groups.

(4) Data collection: specify instruments, data sources, quality control procedures, and the plan for accessing administrative or secondary data.

(5) Analysis plan: pre-specify the statistical or qualitative methods for each research question.

(6) Ethical considerations and limitations: identify participant risks, consent procedures, data protection measures, and the study's key validity threats.

The protocol should be of a quality suitable for submission to an African institutional review board and a competitive research funding call.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson [N] - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance, design paradigm is justified against alternatives with methodological precision, sampling strategy addresses representation of marginalised groups explicitly, analysis plan is pre-specified with sufficient detail to be independently replicable, and the protocol is of funding-submission quality.

Merit

All six elements addressed with methodological accuracy and African contextual grounding.

Pass

All six elements addressed with some methodological and contextual grounding.

Below Pass

Protocol is generic, not grounded in a specific African AI context, or missing key methodological justification.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: The research questions that would most advance African AI governance are not necessarily the questions that are easiest to answer with rigorous designs, nor the questions that international funders are most willing to pay for. Evaluate what research question in your own professional domain would most improve AI governance outcomes for African populations, and assess what design would be required to answer it credibly even under African research constraints.

Second: Research design choices made at the protocol stage - what to measure, who to include, what comparisons to make - are also political choices that determine whose experiences and outcomes are visible in the evidence base. Evaluate how the research design decisions in African AI studies can either reproduce existing power imbalances or deliberately challenge them, and what specific design choices would make African AI research more responsive to the populations most affected by AI deployment.

Third: African research institutions face specific constraints - limited research funding, inadequate data infrastructure, restricted access to AI company data, and ethics committees that may lack the technical capacity to evaluate AI research protocols - that shape what research designs are actually feasible. Evaluate what institutional investments would most expand the range of research designs available to African AI researchers, and what international collaborations would provide access to resources and expertise without compromising African research independence.

Lesson 1 complete. Research design: the foundation of credible African AI scholarship.

Causal inference and potential outcomes framework for AI impact evaluation

Week 10 · Lesson 2 · ARDAI (TIER I)

Week 10 · Lesson 2 · ARDAI (TIER I)

Causal Inference: Potential Outcomes Framework

The foundational logic of causal identification in AI impact research: how the potential outcomes framework makes precise the question every AI evaluation must answer.

Learning Objectives

Evaluate the potential outcomes framework and its implications for causal claims in AI research. Analyse the fundamental problem of causal inference as it applies to AI system evaluation in Africa. Apply the potential outcomes notation to specify a precise causal question for an African AI research context. Evaluate the plausibility of key identification assumptions for specific African institutional settings.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Causal Inference and Machine Learning

p-values Explained Clearly: StatQuest

I. The Fundamental Problem of Causal Inference

The potential outcomes framework, developed by Donald Rubin and formalised by Guido Imbens, provides the most rigorous foundation for causal inference available to social scientists. Its central insight is deceptively simple: a causal effect is the difference between what happened and what would have happened under an alternative condition - for the same unit, at the same time. An AI credit scoring system either approved or rejected a loan application. The causal effect of that decision on the borrower's business outcomes is the difference between what happened because the loan was approved and what would have happened if it had been rejected. Only one of these outcomes is ever observed. This is the fundamental problem of causal inference: we can never directly observe a causal effect, only estimate it.

The framework defines the Average Treatment Effect (ATE) as the expected difference in potential outcomes across the population of interest: E[Y(1) - Y(0)], where Y(1) is the outcome under treatment and Y(0) is the outcome under control. In practice, we can estimate the ATE only under specific assumptions about how treatment assignment relates to potential outcomes. Randomisation ensures independence between treatment assignment and potential outcomes - in expectation, treated and control units are identical on both observed and unobserved characteristics, so the difference in observed outcomes estimates the ATE without bias. Quasi-experimental methods impose weaker assumptions that hold in specific institutional or natural settings. Observational studies require the strongest assumptions - that all confounders are observed and correctly modelled - which are rarely defensible in practice. For African AI research, where randomisation is frequently infeasible, understanding which quasi-experimental assumptions are plausible in specific African institutional settings is the critical applied skill this lesson develops.

Scholarly Analysis

The potential outcomes framework has become the dominant paradigm for causal inference in economics, public health, and increasingly AI governance research. Imbens and Rubin's 2015 textbook is the authoritative treatment. African researchers who fluently apply this framework to AI evaluation questions are positioned to publish in the top applied economics and public policy journals, where the potential outcomes language is now standard. The framework also underlies the econometric identification strategies covered in Lessons 3-6 of this week, making this lesson the essential conceptual foundation for the entire advanced research methods curriculum.

II. Identification Strategies and the Selection Problem

The selection problem is the core challenge that the potential outcomes framework makes precise: the units that receive AI treatment are systematically different from those that do not, in ways that independently affect outcomes. A government agency deploying an AI agricultural advisory system does not randomly assign which farmers receive access - it may target the system at farmers who attend training sessions, who have smartphones, or who are registered in formal extension networks. These farmers differ from non-adopters on characteristics - education, risk tolerance, market access - that also predict agricultural productivity. A naive comparison of outcomes between users and non-users conflates the effect of the AI system with the pre-existing differences between the groups. This is selection bias, and it invalidates causal interpretation of simple before-after or user versus non-user comparisons.

The four main identification strategies that address selection bias - each making different assumptions about the selection process - are: randomisation (treatment assignment is independent of potential outcomes by design); regression discontinuity (units just above and just below an arbitrary threshold are similar on potential outcomes, so the threshold creates as-good-as-random assignment); difference-in-differences (the trend in outcomes for the untreated group correctly estimates the counterfactual trend for the treated group in the absence of treatment); and instrumental variables (an instrument shifts treatment probability without directly affecting the outcome, enabling estimation of the causal effect of treatment for units whose treatment status is changed by the instrument). Each of these strategies is covered in detail in subsequent lessons. This lesson establishes the conceptual framework that makes each strategy intelligible.

Scholarly Analysis

The selection problem is particularly severe in African AI research contexts because the institutional factors that determine which communities, firms, and individuals access AI systems are strongly correlated with the outcomes those systems are supposed to improve. Areas with better digital infrastructure are more likely to receive AI health diagnostics - and also have better health outcomes for reasons unrelated to AI. Understanding the selection problem at the conceptual level, before applying any specific identification strategy, is what distinguishes rigorous African AI research from technically sophisticated but causally uninformative analysis.

Statistical modelling and causal identification methods

Statistical modelling and causal identification methods

III. Applying Potential Outcomes to African AI Evaluation

Applying the potential outcomes framework to African AI evaluation requires specifying clearly what the treatment is, who the units are, what the potential outcomes are, and which identification assumption is plausible given the specific African institutional setting. The treatment must be defined precisely - not just 'AI system use' but the specific decision, recommendation, or information that the AI system provides and that might causally affect outcomes. The units must be defined at the level at which treatment varies - individuals, households, firms, villages, or districts, depending on how the AI system is deployed. The potential outcomes must be defined as measurable quantities observable in the data, not abstract welfare concepts.

African AI deployments frequently involve clustered treatment - entire villages, health facilities, or administrative districts receive the AI system simultaneously, rather than individual assignment. Clustered treatment requires cluster-level randomisation or identification strategies that operate at the cluster level, with standard errors clustered to account for within-cluster correlation. Administrative data in African contexts often has gaps, inconsistencies, and measurement error that affect the reliability of both treatment and outcome variables. The potential outcomes framework requires researchers to specify and defend their assumptions about measurement error - whether it is classical (random) or systematic (correlated with treatment or outcomes) - because systematic measurement error can bias causal estimates in ways that randomisation alone does not correct.

Scholarly Analysis

ARDAI graduates who can apply the potential outcomes framework to specify a precise causal question, defend an identification strategy, and diagnose the specific threats to causal inference in an African AI deployment context are producing the foundation of a publishable research contribution. The lab for this lesson asks for exactly this application. Submit your completed potential outcomes specification to [email protected] for mentor feedback on whether the identification strategy is defensible and what robustness checks would strengthen it.

Primary Source

2015

Imbens, G.W. and Rubin, D.B. (2015). Causal Inference for Statistics, Social, and Biomedical Sciences. Cambridge University Press. Access Source

1. Imbens and Rubin develop the potential outcomes framework primarily for applications in clinical trials and labour economics. Evaluate which aspects of their framework apply directly to AI system evaluation in African governance contexts and which require substantive adaptation given the nature of AI interventions (which change continuously as models are updated) and African data environments.

2. The potential outcomes framework assumes that the treatment status of one unit does not affect the potential outcomes of other units -- the Stable Unit Treatment Value Assumption (SUTVA). Evaluate whether SUTVA is plausible for AI system evaluations in African contexts where AI recommendations for one household may affect neighbours through social networks, or where AI-assisted resource allocation decisions are explicitly competitive.

Knowledge Check

A researcher observes that Nigerian firms that adopted an AI inventory management system in 2022 showed 23% higher revenue growth than firms that did not adopt it. They conclude that AI inventory management caused this revenue advantage. What is the fundamental flaw in this conclusion and what does the potential outcomes framework require to fix it?

Expert Analysis

Selection bias is the core problem. The potential outcomes framework makes precise why the naive comparison is invalid: E[Y(1)|D=1] - E[Y(0)|D=0] does not equal the ATE E[Y(1)-Y(0)] when treatment D is correlated with potential outcomes. The adopting firms (D=1) would have had higher revenue growth even without AI adoption (their Y(0) is higher than non-adopters' Y(0)), so the observed difference overstates the causal effect. Fixing this requires an identification strategy that produces a valid estimate of E[Y(0)|D=1] -- the counterfactual outcome for adopting firms under non-adoption -- which the naive comparison cannot provide.

Research Laboratory

Lab 10.2: Causal Inference: Potential Outcomes Framework

Apply the potential outcomes framework to a specific African AI governance question. Your submission must:

(1) Define the treatment precisely: what specific AI decision, recommendation, or output is the causal agent?

(2) Define the units: what is the unit of analysis and treatment assignment?

(3) State the potential outcomes Y(1) and Y(0) as measurable quantities.

(4) Define the estimand: ATE, ATT (Average Treatment Effect on the Treated), or LATE (Local Average Treatment Effect).

(5) Identify the selection problem: why are treatment and potential outcomes likely to be correlated in this specific African context?

(6) Specify an identification strategy: which quasi-experimental method is most plausible given the institutional setting, and what assumption does it require?

The application must be grounded in a specific African country, AI system, and institutional context.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 2 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Imbens and Rubin develop the potential outcomes framework primarily for applications in clinical trials and labour economics. Evaluate which aspects of their framework apply directly to AI system evaluation in African governance contexts and which require substantive adaptation given the nature of AI interventions (which change continuously as models are updated) and African data environments.

Second: The potential outcomes framework assumes that the treatment status of one unit does not affect the potential outcomes of other units -- the Stable Unit Treatment Value Assumption (SUTVA). Evaluate whether SUTVA is plausible for AI system evaluations in African contexts where AI recommendations for one household may affect neighbours through social networks, or where AI-assisted resource allocation decisions are explicitly competitive.

Third: Evaluate the plausibility of key identification assumptions for specific African institutional settings.

Lesson 2 complete. Causal Inference: Potential Outcomes Framework: foundations for African AI scholarship.

Field experiment design and randomised controlled trials in development contexts

Week 10 · Lesson 3 · ARDAI (TIER I)

Week 10 · Lesson 3 · ARDAI (TIER I)

Randomised Controlled Trials and Field Experiments

Why randomisation is the most powerful tool for causal inference in AI impact evaluation, and how to design rigorous field experiments in African institutional contexts.

Learning Objectives

Evaluate the logic of randomisation as a solution to the selection problem in AI impact evaluation. Analyse the specific design challenges of AI field experiments in African institutional contexts. Apply field experiment design principles to develop a rigorous trial protocol for an African AI intervention. Evaluate the ethical constraints on AI field experiments and the conditions under which withholding treatment from control groups is justifiable.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Machine Learning and Experimental Methods

Random Forests and Experimental Design: StatQuest

I. Why Randomisation Solves the Selection Problem

Randomisation is the most powerful tool available for causal inference because it breaks the link between potential outcomes and treatment assignment by design rather than assumption. When treatment is randomly assigned, the treatment and control groups are identical in expectation on every characteristic - observed and unobserved - that might affect the outcome. The difference in mean outcomes between treatment and control groups therefore estimates the Average Treatment Effect without bias from selection, confounding, or omitted variables. This is why randomised controlled trials are the gold standard for causal evidence: the identification assumption (independence of treatment assignment and potential outcomes) is satisfied by the randomisation procedure rather than by assumptions about the data-generating process that may or may not hold.

Field experiments apply RCT principles to real-world settings rather than laboratory conditions, studying actual AI deployments rather than simulated interventions. The J-PAL network, Innovations for Poverty Action, and the Development Impact Evaluation initiative have conducted hundreds of field experiments on development interventions in Africa, establishing methodological templates directly applicable to AI impact evaluation. Key design choices for an AI field experiment include: the unit of randomisation (individuals, households, villages, health facilities, or administrative districts); the treatment and control conditions (what the treatment group receives versus what the control group does); the sample size and power calculation (how many units are needed to detect the expected effect size with adequate statistical power); and the analysis plan (the pre-specified estimator, covariates, and subgroup analyses). Each of these choices has specific implications for the internal validity of the trial and the interpretability of its results.

Scholarly Analysis

The development economics literature on field experiments in Africa is directly applicable to AI impact evaluation. Duflo, Glennerster, and Kremer's handbook chapter, the primary source for this lesson, is the most widely cited methodological guide for field experiments in developing country contexts. African AI researchers who read it alongside the emerging literature on AI impact evaluations -- of which there are still very few -- are in a position to make methodological contributions as well as empirical ones, by adapting field experiment designs to the specific challenges of evaluating AI systems that learn, update, and interact with human behaviour over time.

II. Designing AI Field Experiments in African Contexts

Designing a field experiment to evaluate an AI system in an African context requires addressing several challenges that standard RCT designs do not fully anticipate. First, the treatment is dynamic: AI systems update as they learn from new data, so a treatment that was assigned at baseline may differ substantially from the treatment experienced by participants at follow-up. Researchers must decide whether to freeze the AI system during the trial (ensuring treatment consistency but reducing external validity) or allow it to update (maintaining external validity but complicating interpretation of the treatment effect). Second, spillovers between treatment and control units are common: if treated farmers share AI advisory recommendations with control farmers, or if treated and control firms compete in the same market, the stable unit treatment value assumption is violated. Cluster randomisation at a geographic level large enough to limit spillovers addresses this but substantially increases the required sample size. Third, compliance is rarely perfect: not all units assigned to the treatment group will actually use the AI system, and not all units assigned to control will abstain from accessing it. Intention-to-treat analysis (comparing outcomes by assignment regardless of actual use) and local average treatment effects estimation using assignment as an instrument for actual use are the standard approaches to imperfect compliance.

Ethical constraints on African field experiments require particular attention. Randomising access to a potentially beneficial AI system means that some eligible beneficiaries are deliberately excluded from access during the trial period. This is ethically acceptable when genuine uncertainty exists about whether the AI system is beneficial - which is precisely when evaluation is most valuable - but requires transparent communication with participants and communities about the trial design, the reason for randomisation, and the plans for expanding access after the trial. The ethics of withholding potentially beneficial AI from randomly assigned control groups is most defensible when the AI system is genuinely novel, when resources constrain universal deployment, and when the evaluation is designed to produce knowledge that will improve the system for future beneficiaries.

Scholarly Analysis

African field experiments on AI interventions face specific logistical challenges that researchers must address in their design: obtaining randomisation lists from government agencies that may prefer discretionary allocation; monitoring treatment receipt and compliance in settings with limited administrative data; and maintaining contact with mobile populations for follow-up surveys. Researchers who have navigated these challenges in agricultural extension, health programme, and social protection evaluations in Africa have developed field protocols directly adaptable to AI impact evaluation. The J-PAL Africa office in Cape Town and IPA's country offices across Africa provide methodological support and field infrastructure for researchers designing AI field experiments.

Field experiment data collection and randomisation

Field experiment data collection and randomisation

III. Analysis of Field Experiment Data

The analysis of field experiment data is more straightforward than quasi-experimental analysis because randomisation handles confounding by design. The primary analysis compares mean outcomes between treatment and control groups using OLS regression with assignment as the independent variable, controlling for pre-specified baseline covariates (which increases precision without introducing bias), and clustering standard errors at the level of randomisation. Pre-specified subgroup analyses test whether treatment effects vary across gender, age, education, geographic remoteness, or other characteristics specified in advance, without incurring the multiple comparisons problem that post-hoc subgroup analyses create. Heterogeneous treatment effects are increasingly estimated using machine learning methods - causal forests, targeted learning, doubly robust estimators - that can identify which units benefit most from AI treatment without overfitting to the observed data.

The external validity of field experiment results - the degree to which findings from one African context generalise to other contexts - requires explicit assessment rather than assumption. A field experiment on AI agricultural advisory in Uganda does not automatically generalise to smallholder farming contexts in Senegal or Malawi, even if the AI system being evaluated is identical. Differences in crop systems, market structures, extension service quality, farmer literacy, and digital infrastructure may all affect how the AI system interacts with the context to produce outcomes. Researchers should specify the contextual features that are most likely to moderate treatment effects and collect data on these moderators to enable assessment of external validity across African contexts.

Scholarly Analysis

The pre-analysis plan is the most important document for field experiment credibility. Written and registered before data collection begins, it specifies the primary and secondary outcomes, the analysis estimator, the covariate set, the subgroup analyses, and the stopping rules. Pre-analysis plans registered at the AEA RCT Registry, OSF, or EGAP are increasingly required by top journals for field experiment submissions. ARDAI graduates who pre-register their field experiment protocols before data collection are building research credibility habits that will define their careers. Submit your pre-registered protocol to [email protected] for mentor review before registration.

Primary Source

2007

Duflo, E., Glennerster, R. and Kremer, M. (2007). Using Randomisation in Development Economics Research: A Toolkit. Handbook of Development Economics, Vol. 4. Access Source

1. Duflo, Glennerster, and Kremer's toolkit was developed for evaluating discrete development interventions with stable treatment conditions. Evaluate how field experiment design must be adapted for evaluating AI systems that learn continuously from new data, update their recommendations as deployment scales, and interact with human behaviour in ways that create feedback loops between the AI treatment and the outcomes being measured.

2. The ethics of randomised control in development contexts require that genuine uncertainty exists about the intervention's benefits. Evaluate whether this condition is met for the AI systems currently being deployed across African public sectors -- including AI health diagnostics, agricultural advisories, and social protection targeting -- and assess what evidence threshold should be required before AI systems are deployed at scale without prior randomised evaluation.

Knowledge Check

A development organisation wants to evaluate whether an AI early warning system for drought reduces food insecurity among smallholder farmers in the Sahel. They have resources to reach 200 villages. A colleague proposes giving the AI system to the 100 villages that request it and using the remaining 100 villages as controls. What is the fundamental problem with this design and how should it be corrected?

Expert Analysis

Self-selection into treatment is the most common source of bias in AI impact evaluations, and it is the specific problem that randomisation solves. The proposed design of serving villages that request the system is sensible from a programme delivery perspective but produces uninterpretable evaluation evidence. The correction -- random assignment of the 200 villages -- ensures that treatment and control villages are identical in expectation on all characteristics that predict food insecurity, so the observed difference in outcomes estimates the causal effect of AI system access without confounding.

Research Laboratory

Lab 10.3: Randomised Controlled Trials and Field Experiments

Design a complete field experiment to evaluate an AI intervention in an African context. Your protocol must specify:

(1) Research question and primary hypothesis.

(2) Treatment and control conditions: what exactly does the treatment group receive and what is the control condition?

(3) Unit of randomisation and justification: why is this level of clustering appropriate given spillover risks?

(4) Sample size and power calculation: what effect size is expected, what power level is required, and how many units are needed?

(5) Randomisation procedure: how will assignment be conducted to ensure balance?

(6) Compliance and attrition plan: how will you handle incomplete compliance and participant loss to follow-up?

(7) Pre-specified analysis: primary estimator, covariates, subgroup analyses.

(8) Ethical safeguards: consent procedures, benefit to control group, plans for post-trial access.

The protocol should be submitted to the AEA RCT Registry or OSF as a pre-analysis plan.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 3 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Duflo, Glennerster, and Kremer's toolkit was developed for evaluating discrete development interventions with stable treatment conditions. Evaluate how field experiment design must be adapted for evaluating AI systems that learn continuously from new data, update their recommendations as deployment scales, and interact with human behaviour in ways that create feedback loops between the AI treatment and the outcomes being measured.

Second: The ethics of randomised control in development contexts require that genuine uncertainty exists about the intervention's benefits. Evaluate whether this condition is met for the AI systems currently being deployed across African public sectors -- including AI health diagnostics, agricultural advisories, and social protection targeting -- and assess what evidence threshold should be required before AI systems are deployed at scale without prior randomised evaluation.

Third: Evaluate the ethical constraints on AI field experiments and the conditions under which withholding treatment from control groups is justifiable.

Lesson 3 complete. Randomised Controlled Trials and Field Experiments: foundations for African AI scholarship.

Difference-in-differences and synthetic control for AI policy evaluation

Week 10 · Lesson 4 · ARDAI (TIER I)

Week 10 · Lesson 4 · ARDAI (TIER I)

Difference-in-Differences and Synthetic Control

Exploiting variation in the timing of AI adoption to identify causal effects without randomisation: the design, assumptions, and modern estimators of difference-in-differences.

Learning Objectives

Evaluate the DiD framework and the parallel trends assumption for AI policy evaluation. Analyse the bias introduced by standard TWFE estimation under staggered AI policy adoption. Apply the Callaway-Sant'Anna estimator to evaluate an African AI policy with staggered rollout. Evaluate the synthetic control method as an alternative to DiD for settings with few control units.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

How to Analyse Panel Data Reliably

Machine Learning for Policy Evaluation

I. The DiD Framework and the Parallel Trends Assumption

Difference-in-differences is the most widely used quasi-experimental method for AI policy evaluation because it exploits variation in the timing of AI adoption across units to identify causal effects without randomisation. The core insight is that comparing treated and control units before and after treatment controls for time-invariant differences between them, while comparing outcomes before and after treatment within treated units controls for time trends affecting all units equally. The DiD estimator is the difference in outcomes between treated and control units in the post-treatment period, minus the difference in outcomes between the same groups in the pre-treatment period: ATT = (E[Y|treated, post] - E[Y|control, post]) - (E[Y|treated, pre] - E[Y|control, pre]).

The critical identifying assumption of DiD is parallel trends: in the absence of the AI policy, the outcome variable for treated and control units would have followed parallel trajectories over time. This assumption cannot be directly tested for the post-treatment period because we cannot observe the counterfactual outcome for treated units after treatment begins. However, it can be assessed empirically using pre-treatment data: if treated and control units showed parallel trends before the AI policy was introduced, this provides evidence (though not proof) that they would have continued to do so in the absence of the policy. Event study specifications, which estimate treatment effects separately for each pre- and post-treatment period, are the standard diagnostic for the parallel trends assumption and for detecting treatment effect dynamics that a single DiD coefficient would obscure.

Scholarly Analysis

The DiD literature has advanced substantially since 2018, driven by the recognition that standard two-way fixed effects DiD estimators produce biased estimates under staggered adoption -- when different units adopt the AI policy at different times -- because early adopters serve as implicit controls for later adopters even after they have themselves been treated. The Callaway and Sant'Anna (2021) estimator, the primary source for this lesson, solves this problem by estimating group-time average treatment effects that are then aggregated to produce overall ATT estimates robust to staggered adoption. African AI policy evaluations frequently involve staggered rollout, making the Callaway-Sant'Anna approach the methodologically appropriate default.

II. Staggered Adoption and the Callaway-Sant'Anna Estimator

Most African AI policy rollouts involve staggered adoption: the AI system is deployed in some districts, facilities, or units before others, for logistical, budgetary, or political reasons. This staggered rollout creates a natural DiD research design if researchers can observe outcomes for all units before and after each adopting unit receives the AI system. However, the standard two-way fixed effects (TWFE) estimator used in most applied DiD papers is biased under staggered adoption because it uses already-treated units as implicit controls for newly-treated units. This produces estimates that can be negative even when all unit-specific treatment effects are positive, if early adopters' treatment effects are larger than late adopters' - a condition called negative weighting.

The Callaway and Sant'Anna estimator addresses this by computing group-time average treatment effects: for each cohort of units that adopted the AI system at the same time (group) and each post-adoption period (time), it computes a clean DiD estimate using only not-yet-treated units as controls. These group-time estimates are then aggregated using researcher-specified weights to produce overall ATT estimates, event study plots, and heterogeneous treatment effect analyses. The estimator is implemented in the R package csdid and the Stata package csdid. For African AI policy researchers who typically observe staggered rollout across provinces, districts, or facilities, the Callaway-Sant'Anna approach is more reliable than TWFE and should be the default specification reported alongside TWFE as a robustness check.

Scholarly Analysis

The synthetic control method, developed by Abadie and Gardeazabal (2003) and extended by Abadie, Diamond, and Hainmueller (2010), provides an alternative to DiD when there are few control units and the parallel trends assumption is implausible. It constructs a weighted average of control units that best matches the treated unit on pre-treatment outcomes and covariates -- a synthetic control -- and uses the difference between the treated unit and its synthetic control after treatment as the causal estimate. For African AI governance researchers studying country-level policies where there are few comparable countries, the synthetic control method is often more credible than DiD. The Synth package in R and the synthetic_control package in Stata implement the method.

Longitudinal panel data and policy evaluation

Longitudinal panel data and policy evaluation

III. Implementing DiD for African AI Policy Evaluation

Implementing DiD for African AI policy evaluation requires accessing panel data -- repeated observations of the same units before and after AI policy adoption. Administrative data from African government agencies is the most common source: tax records, health management information systems, social protection registries, and agricultural census data all contain longitudinal observations that can support DiD designs. The data access challenge is significant: African government agencies are often reluctant to share administrative data with external researchers, data quality varies substantially within and across African countries, and panel identifiers may be inconsistent or missing. Researchers who build relationships with African statistical agencies and sectoral ministries before designing their studies are far better positioned to access the administrative data that DiD designs require.

The practical steps for implementing DiD in an African AI context are: identify units that received the AI policy at different times; collect panel outcome data for all units covering sufficient pre- and post-treatment periods; test the parallel trends assumption using event study specifications in the pre-treatment period; estimate the Callaway-Sant'Anna group-time ATTs; aggregate to an overall ATT and event study plot; conduct robustness checks including: changing the control group (never-treated only versus not-yet-treated), alternative outcome definitions, and placebo tests at alternative cutoff dates. Report all specifications, not just the one that confirms the hypothesis.

Scholarly Analysis

The DiD and synthetic control methods require more technical implementation than the conceptual explanation suggests. R packages including did (Callaway-Sant'Anna), fixest (high-dimensional FE with staggered DiD), and tidysynth (synthetic control) provide accessible implementations. Python implementations are available via the pyfixest and pensynth packages. ARDAI graduates who implement these methods on real African administrative data and document the results clearly -- including pre-trends tests, aggregation choices, and robustness checks -- have the technical execution section of a publishable paper already written. The lab for this lesson asks for exactly this implementation.

Primary Source

2021

Callaway, B. and Sant'Anna, P.H.C. (2021). Difference-in-Differences with Multiple Time Periods. Journal of Econometrics, 225(2), pp.200-230. Access Source

1. Callaway and Sant'Anna develop their estimator for a specific model of staggered adoption with binary treatment. Evaluate how their approach must be extended for AI policy evaluations where the treatment intensity varies -- for example, where some districts receive AI systems with more advanced capabilities or higher adoption rates than others -- and where treatment can be reversed if governments discontinue AI programmes.

2. The parallel trends assumption is fundamentally untestable for the post-treatment period, and pre-treatment trend tests are necessary but not sufficient evidence that it holds. Evaluate what institutional knowledge about the AI policy rollout process in a specific African context would most strengthen or undermine confidence in the parallel trends assumption, and assess how researchers should communicate residual uncertainty about this assumption to non-technical policy audiences.

Knowledge Check

A researcher uses a two-way fixed effects DiD regression to evaluate the impact of an AI tax administration system rolled out across Kenyan districts between 2019 and 2022, with some districts adopting in 2019, others in 2020, and others in 2021. They find a negative but statistically insignificant coefficient on the treatment indicator. Should they conclude the AI system had no effect? Why or why not?

Expert Analysis

This scenario illustrates the negative weighting problem in TWFE DiD under staggered adoption, documented by Goodman-Bacon (2021) and addressed by Callaway and Sant'Anna (2021). The negative TWFE coefficient does not indicate that the AI tax system was harmful -- it may be an artefact of the estimator's treatment of early adopters as controls for late adopters. The Callaway-Sant'Anna approach eliminates this bias by construction. Event study plots from the CS estimator would show whether effects are growing, declining, or heterogeneous across adoption cohorts in ways that the single TWFE coefficient cannot reveal.

Research Laboratory

Lab 10.4: Difference-in-Differences and Synthetic Control

Apply difference-in-differences to evaluate an AI policy in an African context. Select a real AI policy with staggered rollout or design a realistic scenario.

(1) Define the treatment: which units adopted the AI policy, when, and what constitutes the control group (never-treated, not-yet-treated, or both).

(2) Specify the outcome and the data source. If using real data, document the source, coverage, and quality.

(3) Test the parallel trends assumption: plot pre-treatment outcome trends for treatment and control groups and report an event study specification for the pre-treatment period.

(4) Estimate the DiD using the Callaway-Sant'Anna estimator (or TWFE with a note on its limitations under staggered adoption).

(5) Report robustness checks: alternative control group definitions, placebo cutoff dates, alternative outcome variables.

(6) Interpret the results: what does the estimated ATT imply about the AI policy's effectiveness, and how confident are you in the causal interpretation given the evidence on parallel trends?

Implement in R or Python and submit the code alongside your written interpretation.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 4 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Callaway and Sant'Anna develop their estimator for a specific model of staggered adoption with binary treatment. Evaluate how their approach must be extended for AI policy evaluations where the treatment intensity varies -- for example, where some districts receive AI systems with more advanced capabilities or higher adoption rates than others -- and where treatment can be reversed if governments discontinue AI programmes.

Second: The parallel trends assumption is fundamentally untestable for the post-treatment period, and pre-treatment trend tests are necessary but not sufficient evidence that it holds. Evaluate what institutional knowledge about the AI policy rollout process in a specific African context would most strengthen or undermine confidence in the parallel trends assumption, and assess how researchers should communicate residual uncertainty about this assumption to non-technical policy audiences.

Third: Evaluate the synthetic control method as an alternative to DiD for settings with few control units.

Lesson 4 complete. Difference-in-Differences and Synthetic Control: foundations for African AI scholarship.

Regression discontinuity design and threshold analysis for AI eligibility systems

Week 10 · Lesson 5 · ARDAI (TIER I)

Week 10 · Lesson 5 · ARDAI (TIER I)

Regression Discontinuity Design

When AI systems use score thresholds for high-stakes decisions, the threshold becomes a quasi-randomisation device: the logic, assumptions, and implementation of RDD.

Learning Objectives

Evaluate the RDD identifying assumption and its plausibility in African AI contexts. Analyse the difference between sharp and fuzzy RDD and when each applies. Apply RDD design and estimation to an African AI policy with an eligibility threshold. Evaluate the limitations of RDD local identification and when LATE estimates are policy-relevant.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Regression Discontinuity and Threshold Methods | StatQuest

Threshold Analysis and Causal Identification

I. The Logic of Regression Discontinuity

Regression discontinuity design exploits the fact that many policy rules assign treatment based on whether a continuous score falls above or below an arbitrary threshold. Units just above and just below this threshold are nearly identical on all characteristics, because the threshold creates a sharp discontinuity in treatment probability while everything else changes smoothly. This local comparability around the threshold allows credible causal inference without randomisation: we compare outcomes for units just above the threshold with units just below, treating the threshold as a locally randomised assignment device.

In African AI governance contexts, RDD opportunities arise wherever AI systems use score thresholds for binary decisions: credit score cutoffs for loan approval, risk score thresholds for social protection targeting, AI eligibility scores for business support, or algorithmic ranking thresholds for scholarship allocation. Each creates a potential RDD if the running variable is observed and if units cannot precisely manipulate their position relative to the threshold. The RDD estimate is the Local Average Treatment Effect for units near the threshold, so interpreting results requires clarity about who these marginal units are and why results for them may or may not generalise elsewhere.

Scholarly Analysis

Lee and Lemieux's 2010 handbook chapter is the canonical RDD reference in applied economics. Applications to AI policy questions in African contexts remain rare, partly because the individual-level administrative data needed to implement RDD is more difficult to access. African AI researchers who demonstrate RDD applications using African administrative data make both empirical and methodological contributions to the literature.

II. Sharp and Fuzzy RDD

A sharp RDD applies when treatment is a deterministic function of the threshold crossing: all units above receive treatment, all below do not. A fuzzy RDD applies when crossing the threshold affects treatment probability without determining it uniquely. In African AI contexts, fuzzy RDD is more common because AI scores inform but do not mechanically determine decisions: loan officers may override AI credit recommendations, caseworkers may apply discretion to AI social protection scores. Fuzzy RDD is estimated using the threshold as an instrument for actual treatment receipt, giving a LATE for compliers.

The key validity tests are: the density test (McCrary test) for running variable manipulation at the threshold; covariate smoothness tests confirming predetermined characteristics do not jump at the threshold; and placebo threshold tests showing the outcome does not show discontinuities at other running variable values. Bandwidth selection is a bias-variance tradeoff: the optimal bandwidth procedure of Calonico, Cattaneo, and Titiunik (2014), implemented in the rdrobust package in R, is the standard approach.

Scholarly Analysis

RDD requires high-quality individual-level data on the running variable linked to outcome data -- data that AI vendors and government agencies may resist sharing. Researchers who negotiate data sharing agreements including the running variable before an AI system is deployed are best positioned to conduct RDD evaluations later. Embedding evaluation data requirements into AI procurement contracts is both a research design and a governance recommendation.

Score threshold analysis and regression discontinuity

Score threshold analysis and regression discontinuity

III. RDD Applications in African AI Policy

The most promising African AI policy RDD applications involve systems using explicit score thresholds for high-stakes decisions: South Africa's social grants income thresholds, Nigeria's FMBN mortgage credit scores, Kenya's eCitizen digital service eligibility scores. Each creates a potential RDD design if score values and outcome data can be accessed at the individual level. The research challenge is accessing data at the threshold -- getting individual-level records for units near the cutoff -- which requires engagement with implementing agencies before the AI system goes live.

RDD results have specific policy implications researchers must communicate carefully. The LATE applies to marginal cases near the threshold, not to all cases receiving AI-assisted treatment. A finding that AI credit approval at the margin improves business outcomes does not imply higher-scoring borrowers benefit similarly. Policy audiences often want universal claims that RDD cannot support; researchers must be explicit about the local nature of the estimate while explaining why the marginal case is often the most policy-relevant -- it is precisely where the AI decision is most consequential and most contestable.

Scholarly Analysis

RDD produces some of the most visually compelling causal evidence available: the plot of mean outcomes against the running variable, showing a discontinuity at the threshold, is immediately interpretable to non-technical readers. The rdplot function in rdrobust produces publication-quality RDD plots with appropriate binning and confidence intervals that form the centrepiece of an RDD paper's findings presentation.

Primary Source

2010

Lee, D.S. and Lemieux, T. (2010). Regression Discontinuity Designs in Economics. Journal of Economic Literature, 48(2), pp.281-355. Access Source

1. Lee and Lemieux focus on single continuous running variables. Evaluate how RDD must be adapted when African AI systems use composite indices as eligibility criteria, and when components carry different levels of measurement error.

2. The RDD LATE estimates effects for marginal cases near the threshold. Evaluate whether this is the most policy-relevant parameter for policymakers deciding whether to expand AI eligibility systems, and what additional designs would produce the broader evidence they need.

Knowledge Check

Ghana's agricultural ministry uses an AI soil health index with a threshold of 60 for subsidy eligibility. A researcher finds that 34% of farmers have index scores of exactly 60. What does this imply for the RDD?

Expert Analysis

Running variable manipulation is the most serious threat to RDD validity. The McCrary test formalises this check by testing for a discontinuity in the density of the running variable at the threshold. A valid RDD requires a smooth density around the threshold, reflecting that units cannot precisely position themselves relative to it. The 34% mass at exactly 60 violates this requirement and must be addressed before RDD estimates can be trusted.

Research Laboratory

Lab 10.5: Regression Discontinuity Design

Design or apply a regression discontinuity study for an African AI programme evaluation. Specify the running variable and threshold, assess validity assumptions with a McCrary density test and covariate smoothness tests, produce an RDD plot, estimate using rdrobust with optimal bandwidth, conduct robustness checks with alternative bandwidths and placebo thresholds, and interpret the LATE for the specific marginal population. If real data is available implement in R; otherwise simulate a realistic African AI dataset and conduct the full analysis.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 5 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Lee and Lemieux focus on single continuous running variables. Evaluate how RDD must be adapted when African AI systems use composite indices as eligibility criteria, and when components carry different levels of measurement error.

Second: The RDD LATE estimates effects for marginal cases near the threshold. Evaluate whether this is the most policy-relevant parameter for policymakers deciding whether to expand AI eligibility systems, and what additional designs would produce the broader evidence they need.

Third: Evaluate the limitations of RDD local identification and when LATE estimates are policy-relevant.

Lesson 5 complete. Regression Discontinuity Design: foundations for African AI scholarship.

Instrumental variables and natural experiments for African AI programme evaluation

Week 10 · Lesson 6 · ARDAI (TIER I)

Week 10 · Lesson 6 · ARDAI (TIER I)

Instrumental Variables and Natural Experiments

Finding exogenous variation in AI adoption when neither randomisation nor sharp thresholds exist: the three assumptions, natural experiment sources, and IV practice in Africa.

Learning Objectives

Evaluate the three IV assumptions and their plausibility for African AI adoption instrumentation. Analyse natural experiments that create exogenous AI exposure variation in African institutional settings. Apply 2SLS estimation to an IV design for African AI policy evaluation. Evaluate weak instrument diagnostics and their implications for IV reliability.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

MIT OpenCourseWare: Econometrics and IV

What Causal Methods Can and Cannot Do

I. The IV Framework and the Three Assumptions

Instrumental variables estimation addresses the selection problem when randomisation is infeasible and no threshold creates an RDD opportunity. An instrument is a variable that affects treatment uptake but affects the outcome only through its effect on treatment. If such a variable can be found, it provides exogenous variation in treatment enabling causal identification. The three IV assumptions are: relevance (the instrument is meaningfully correlated with treatment; tested with first-stage F-statistic above 10); the exclusion restriction (the instrument affects the outcome only through treatment, not any direct channel; this cannot be directly tested and must be defended substantively); and independence (the instrument is independent of potential outcomes; essentially, it must be exogenous).

The IV estimate is a Local Average Treatment Effect for compliers -- units whose treatment status is changed by the instrument -- rather than the ATE for all units. Complier characteristics can be estimated even though individual compliers cannot be identified, providing information about the population to which the IV estimate applies. Two-stage least squares (2SLS) is the standard IV estimator. Weak instruments produce severely biased 2SLS estimates; Limited Information Maximum Likelihood (LIML) is more robust to weak instruments and should be reported alongside 2SLS as a robustness check.

Scholarly Analysis

Angrist and Pischke's Mostly Harmless Econometrics (2009) and Mastering Metrics (2014) are the most accessible treatments of IV for applied researchers. African AI researchers who can identify credible instruments and defend the exclusion restriction with institutional knowledge are making identification contributions that go beyond technical implementation -- valid instruments in African AI contexts are genuinely rare and genuinely valuable.

II. Natural Experiments for African AI Evaluation

Natural experiments are real-world events creating exogenous variation in AI exposure. Policy roll-out schedules, infrastructure expansion, regulatory changes, and technology adoption by neighbouring entities can all create natural experiments. The identifying assumption is that the natural event creates variation in AI exposure orthogonal to potential outcomes, conditional on observed covariates. This must be defended with institutional knowledge and supported by evidence that the event does not predict pre-treatment outcomes.

Specific natural experiments for African AI policy include: variation in mobile network coverage expansion (affecting AI access through connectivity, plausibly exogenous to individual outcomes conditional on geography); variation in AI procurement timing driven by administrative cycles rather than programme effectiveness; variation in AI vendor market entry across African countries driven by business strategy; and variation in digital infrastructure investment driven by external aid decisions. Each creates potentially valid IV designs for researchers who can document why the variation is exogenous to outcomes studied.

Scholarly Analysis

Finding a valid instrument for AI adoption in Africa is genuinely difficult because the exclusion restriction rules out most candidates -- AI access is often bundled with other improvements that independently affect outcomes. Researchers who identify valid instruments by working backwards from institutional knowledge -- understanding the specific mechanism driving AI adoption variation in a particular context -- rather than searching for any correlated variable, are more likely to satisfy the exclusion restriction and produce reliable estimates.

Natural experiment data and instrumental variable analysis

Natural experiment data and instrumental variable analysis

III. Weak Instruments and IV Practice

Weak instruments produce IV estimates severely biased toward the OLS estimate and with highly inflated standard errors. The conventional rule requires first-stage F-statistic above 10 to avoid weak instrument bias, though Stock and Yogo establish formal critical values for different bias tolerance levels. In African AI research, weak instruments are common because available natural experiments may explain only a small fraction of AI adoption variation. Reporting first-stage F-statistics, Cragg-Donald and Kleibergen-Paap statistics, and Anderson-Rubin confidence intervals robust to weak instruments is essential for IV papers at rigorous journals.

The ivreg package in R and ivregress in Stata implement both 2SLS and LIML. African AI IV papers should also conduct robustness checks including: alternative instrument definitions; over-identification tests when multiple instruments are available (Sargan-Hansen J-test); subsample analyses to assess whether the LATE varies across subgroups; and placebo outcome tests showing the instrument does not predict outcomes it should not affect if the exclusion restriction holds. Transparent reporting of all these diagnostics is what separates credible IV papers from technically implemented but methodologically unconvincing ones.

Scholarly Analysis

The most impactful African AI IV studies will combine credible instruments with large administrative datasets that enable precise first-stage estimation and adequately powered LATE estimates. Researchers who build relationships with African statistical agencies and sectoral ministries to access administrative data are creating the infrastructure for IV studies that cannot currently be conducted. This infrastructure investment is a research contribution in its own right, enabling not just the researcher's own IV study but the studies of future African AI researchers who gain access to the same data.

Primary Source

2009

Angrist, J.D. and Pischke, J.S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press. Access Source

1. Angrist and Pischke argue good instruments come from institutional features creating as-good-as-random treatment variation. Evaluate what institutional features of African AI policy rollout create the most credible natural experiments, and what data would be needed to implement these designs.

2. The IV LATE estimates effects for compliers whose treatment status is changed by the instrument. Evaluate whether compliers in typical African AI natural experiments are representative of the broader policy population, and how researchers should communicate this external validity limitation.

Knowledge Check

A researcher proposes using a country's distance from the nearest submarine internet cable landing point as an instrument for AI firm adoption, arguing cable proximity increases broadband enabling AI use. Evaluate the three IV assumptions.

Expert Analysis

This instrument illustrates the trade-off between relevance and the exclusion restriction that characterises most IV designs. The exclusion restriction is the weak point: cable proximity affects outcomes through multiple channels simultaneously. Researchers proposing this instrument must either control for all alternative channels and argue the residual variation is exclusion-restriction-valid, or find a more narrow instrument affecting AI adoption specifically without simultaneously improving outcomes through other channels.

Research Laboratory

Lab 10.6: Instrumental Variables and Natural Experiments

Identify and evaluate an instrumental variable for AI adoption in an African context. Describe the research question and why OLS/DiD is insufficient. Propose an instrument and evaluate all three IV assumptions with theoretical argument and where possible empirical evidence. If data is available, implement 2SLS and report the first-stage F-statistic, reduced form, and 2SLS estimate. Conduct robustness checks including alternative instrument definitions and placebo outcomes.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 6 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Angrist and Pischke argue good instruments come from institutional features creating as-good-as-random treatment variation. Evaluate what institutional features of African AI policy rollout create the most credible natural experiments, and what data would be needed to implement these designs.

Second: The IV LATE estimates effects for compliers whose treatment status is changed by the instrument. Evaluate whether compliers in typical African AI natural experiments are representative of the broader policy population, and how researchers should communicate this external validity limitation.

Third: Evaluate weak instrument diagnostics and their implications for IV reliability.

Lesson 6 complete. Instrumental Variables and Natural Experiments: foundations for African AI scholarship.

Mixed methods research design integrating quantitative and qualitative strands

Week 10 · Lesson 7 · ARDAI (TIER I)

Week 10 · Lesson 7 · ARDAI (TIER I)

Mixed Methods Research in AI Governance

Integrating quantitative and qualitative evidence to produce AI governance knowledge that is both statistically grounded and mechanistically explained.

Learning Objectives

Evaluate the rationale for mixed methods designs in AI governance research. Analyse the four core mixed methods designs and their African AI applicability. Apply the explanatory sequential design to produce an integrated finding. Evaluate quality criteria for mixed methods AI governance research including integration quality and reflexivity.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Qualitative and Mixed Methods Research

Interdisciplinary Research: Rachel Thomas

I. Why AI Governance Research Needs Mixed Methods

The most important questions in AI governance research cannot be answered by either quantitative or qualitative methods alone. Quantitative analysis can establish that an AI system produces biased outcomes for certain demographic groups -- but cannot explain why those biases emerge, how frontline workers respond to biased recommendations, or what institutional dynamics perpetuate bias despite awareness of it. Qualitative research can illuminate these mechanisms in rich detail -- but cannot establish whether they operate at scale or vary systematically across demographic groups. Mixed methods research combines both to produce knowledge that is both statistically grounded and mechanistically explained.

In African AI governance specifically, mixed methods are powerful because the most consequential questions involve the interaction between technical systems and deeply contextual human and institutional factors. How do community health workers in rural Tanzania adapt AI diagnostic recommendations to their clinical judgment? How do AI-assisted benefit determinations map onto existing social hierarchies in Nigerian communities? None of these questions yields to purely quantitative or qualitative analysis. Mixed methods combining survey data on scale with ethnographic data on mechanism produce the most complete and actionable answers.

Scholarly Analysis

Creswell and Plano Clark's framework distinguishes four core mixed methods designs: convergent (simultaneous), explanatory sequential (quantitative then qualitative), exploratory sequential (qualitative then quantitative), and embedded (one nested within the other). For African AI governance research, the explanatory sequential design is most common and productive: quantitative analysis establishes the statistical pattern, then qualitative fieldwork explains the mechanisms driving it. The integrated finding combining statistical evidence with mechanistic explanation is the contribution neither dataset alone could produce.

II. Designing the Quantitative-Qualitative Integration

The defining feature of mixed methods -- what distinguishes it from two separate studies -- is the deliberate integration of quantitative and qualitative findings into a unified conclusion. Integration can occur at the design stage (qualitative findings inform quantitative instruments), at data collection (quantitative findings guide purposive qualitative sampling), at analysis (statistical patterns explained by qualitative themes), or at interpretation (meta-inferences synthesising both strands). Without explicit integration, a mixed methods paper is just two weaker papers combined into one longer one.

For African AI governance, the most productive integration points are: using qualitative interviews with AI system operators to identify variables for quantitative surveys; using statistical findings to identify cases most needing qualitative explanation; and using qualitative institutional analysis to interpret quantitative effect heterogeneity. The joint display -- a table or figure presenting quantitative and qualitative findings side by side for the same cases -- is the most effective communication device for mixed methods integration and should appear in every mixed methods paper.

Scholarly Analysis

Mixed methods papers are harder to publish than single-method papers because they require reviewers competent in both traditions. Targeting journals that explicitly welcome mixed methods -- World Development, Social Science and Medicine, Governance, AI and Society -- is an important strategic decision. Journals that default to quantitative standards may reject strong qualitative components as insufficient while journals that default to qualitative standards may find the quantitative component superficial. Finding reviewers who can evaluate both is the editorial challenge that mixed methods authors must anticipate.

Qualitative interview and mixed methods data integration

Qualitative interview and mixed methods data integration

III. Quality Criteria in Mixed Methods Research

Mixed methods research is assessed against quality criteria from both quantitative and qualitative traditions plus integration-specific criteria. Quantitative quality: appropriate methods, effect sizes with confidence intervals, threats to internal and external validity addressed. Qualitative quality: purposive and well-justified sampling, systematic and documented data collection, rigorous reflexive analysis, findings grounded in data. Integration quality: genuine integration not parallel tracks, informative joint display, meta-inferences that add value beyond either strand alone.

Reflexivity is an additional quality criterion: researchers must explicitly reflect on how their positionality -- professional background, institutional affiliation, gender, ethnicity, prior beliefs about AI -- shapes what they observed, what questions they asked, and how they interpreted findings. For African AI researchers studying systems deployed in their own communities, reflexivity requires engaging with the tension between insider knowledge (enabling access and understanding) and insider bias (leading to selective observation or interpretation).

Scholarly Analysis

The most ambitious mixed methods African AI governance research will combine large-scale administrative data analysis with in-depth ethnographic fieldwork at AI deployment sites -- health facilities, tax offices, social protection centres. This requires research teams with complementary skills and extended fieldwork timelines, larger budgets, and longer grant periods. The IDRC AI for Development programme and the Wellcome Trust's African institutions programme both fund this type of interdisciplinary mixed methods work and are the most appropriate funders to target for large mixed methods African AI governance studies.

Primary Source

2018

Creswell, J.W. and Plano Clark, V.L. (2018). Designing and Conducting Mixed Methods Research. 3rd ed. Sage Publications. Access Source

1. Creswell and Plano Clark developed their framework primarily for health and social science research in high-income countries. Evaluate which integration strategies require adaptation for AI governance research in African contexts, particularly given differences in administrative data quality, fieldwork access constraints, and institutional dynamics of AI deployment.

2. Mixed methods requires skills in both quantitative and qualitative traditions rarely combined in a single researcher. Evaluate how African AI research teams can be structured for genuine integration rather than parallel tracks, and what collaborative research models produce the most integrated and policy-relevant mixed methods AI governance research.

Knowledge Check

An ARDAI researcher finds from administrative data that an AI social protection targeting system has a 41% exclusion error rate for female-headed households compared to 18% for male-headed. They want to understand why. Is a mixed methods design appropriate and which type?

Expert Analysis

Explanatory sequential is correct because the sequence -- quantitative first, then qualitative to explain -- matches the research logic: we have a statistical finding needing explanation, not a hypothesis needing quantitative testing. The quantitative finding shapes qualitative sampling (interview female-headed households with high exclusion rates and their caseworkers) and provides context for interpreting qualitative findings. The integrated finding is more actionable than either finding alone.

Research Laboratory

Lab 10.7: Mixed Methods Research in AI Governance

Design a mixed methods study for an African AI governance intervention. State why mixed methods outperforms single-method approaches for your question. Choose a design type and justify it. Specify both strands: quantitative (data source, outcome, analysis) and qualitative (sampling criteria, data collection method, analysis approach). Design the integration strategy and produce a draft joint display. Address quality including reflexivity with a 300-word positionality statement.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 7 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Creswell and Plano Clark developed their framework primarily for health and social science research in high-income countries. Evaluate which integration strategies require adaptation for AI governance research in African contexts, particularly given differences in administrative data quality, fieldwork access constraints, and institutional dynamics of AI deployment.

Second: Mixed methods requires skills in both quantitative and qualitative traditions rarely combined in a single researcher. Evaluate how African AI research teams can be structured for genuine integration rather than parallel tracks, and what collaborative research models produce the most integrated and policy-relevant mixed methods AI governance research.

Third: Evaluate quality criteria for mixed methods AI governance research including integration quality and reflexivity.

Lesson 7 complete. Mixed Methods Research in AI Governance: foundations for African AI scholarship.

Ethnographic fieldwork observation and qualitative data collection in AI governance

Week 10 · Lesson 8 · ARDAI (TIER I)

Week 10 · Lesson 8 · ARDAI (TIER I)

Ethnographic and Qualitative Research Methods

Gaining access to the situated knowledge, tacit practices, and institutional dynamics of AI deployment that no survey or administrative dataset can reach.

Learning Objectives

Evaluate the distinctive contribution of ethnographic and qualitative methods to AI governance research. Analyse key methodological decisions in qualitative AI research design. Apply thematic analysis to produce rigorous qualitative findings. Evaluate quality criteria for qualitative AI governance research and the strategies for satisfying them.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Ethnographic Research Methods Guide

How to Read and Conduct Field Research

I. What Ethnography Offers AI Governance Research

Ethnography is the systematic study of people and practices through prolonged immersion in the settings where everyday life occurs. For AI governance research, ethnographic methods provide access to the situated knowledge, tacit practices, and institutional dynamics that surveys, administrative data, and interviews cannot reach. An ethnographer studying AI diagnostic support tools in a Nigerian hospital does not just ask doctors how they use the tool -- they observe actual interactions between doctors, patients, the AI tool, and the clinical context over weeks or months, observing when doctors follow AI recommendations, override them, hide them from patients, or adapt them to clinical judgment shaped by experience the AI cannot access. This observational access to practice as it actually occurs is what makes ethnography irreplaceable for understanding AI in use.

For African AI governance researchers, ethnography offers a further advantage: it can be conducted from a position of cultural and institutional proximity that foreign researchers cannot replicate. An ARDAI graduate conducting ethnographic research on AI in their own professional context brings insider knowledge of local institutional logics, professional norms, and informal practices that would take a foreign researcher years to develop. This positional advantage is a genuine research asset that should be leveraged deliberately, with appropriate reflexivity about the challenges of studying familiar settings where professional relationships may constrain observation.

Scholarly Analysis

Hammersley and Atkinson's Ethnography: Principles in Practice is the standard methodological reference for social science ethnography. Their framework addresses gaining access, establishing rapport, selecting observation settings, taking field notes, managing the observer effect, and exiting the field. For AI governance ethnography specifically: obtaining ethics board approval for workplace observation, managing relationships with AI vendors present in the observation setting, and negotiating the tension between accurate description of AI system use and protecting the professional reputations of practitioners being observed require additional methodological planning beyond standard ethnographic protocols.

II. Thematic Analysis and Qualitative Rigour

Thematic analysis is the most widely used approach to analysing qualitative data, providing a systematic procedure for identifying, coding, and interpreting patterns across interview transcripts, field notes, and documents. Braun and Clarke's six-phase approach -- familiarisation, generating initial codes, searching for themes, reviewing themes, defining and naming themes, and producing the report -- is the standard framework. For AI governance research, themes should be interpretive constructs that explain something about the relationship between AI and its context, not just descriptive summaries of what participants said.

Qualitative rigour is established through trustworthiness: credibility (findings reflect participants' perspectives, tested through member checking and prolonged engagement); transferability (contextual features described in sufficient detail for applicability assessment); dependability (research process documented for audit); and confirmability (findings grounded in data not prior beliefs, evidenced by an audit trail of decisions and interpretations). African AI governance researchers who apply these criteria rigorously produce qualitative findings that stand up to scrutiny in top interdisciplinary journals.

Scholarly Analysis

The most common criticism of qualitative AI governance research is that it is anecdotal -- based on small numbers of cases. Researchers who address this directly -- by explaining purposive rather than random selection, describing variation observed across cases, and being clear what findings can and cannot generalise to -- build more durable credibility. Purposive sampling is not a weakness; it is a strength when the purpose is clearly articulated and selected cases are genuinely informative for the theoretical or policy question at hand.

Ethnographic field notes and qualitative coding

Ethnographic field notes and qualitative coding

III. Interviews, Focus Groups, and Document Analysis

Semi-structured interviews are the most common qualitative data collection method for AI governance research. Interview guides should cover: direct experience with the AI system; understanding of how it works and what it is for; cases where participants followed or deviated from AI recommendations and why; understanding of accountability when recommendations prove incorrect; and the institutional pressures, professional norms, and practical constraints shaping AI interaction. Asking for specific episodes -- 'can you describe a time when the AI recommendation surprised you?' -- produces richer data than abstract attitude questions.

Document analysis extends qualitative data to the written records AI systems generate and respond to: procurement specifications, system documentation, audit reports, user manuals, meeting minutes from AI governance committees, and public communications about AI deployments. These documents provide institutional perspectives inaccessible through individual interviews -- revealing what organisations claim about AI systems, what accountability frameworks were intended, and how intentions were revised in implementation. Critical discourse analysis of AI procurement documents and algorithmic impact assessments published by African government agencies can reveal assumptions, priorities, and silences in official AI governance narratives.

Scholarly Analysis

African qualitative AI governance researchers face a specific challenge: gaining access to the government agencies and corporations where AI systems are deployed requires navigating institutional gatekeeping that can be more restrictive than in high-income country contexts. Framing research as evaluation support rather than external scrutiny, and demonstrating how findings will benefit the implementing organisation, are practical strategies that have worked for African researchers gaining access to government AI studies. The ARDAI SIP mentor network includes practitioners with experience navigating access in African government and civil society AI contexts.

Primary Source

2019

Hammersley, M. and Atkinson, P. (2019). Ethnography: Principles in Practice. 4th ed. Routledge. Access Source

1. Hammersley and Atkinson developed their framework for academic social science with extended fieldwork timelines of months to years. Evaluate how ethnographic rigour principles can be adapted for African AI governance researchers with limited fieldwork time and resources, and what minimum standards maintain ethnography's distinctive contribution.

2. Ethnography in African AI governance contexts creates specific risks: government agencies may restrict access when findings might be critical, vendors may attempt to influence observation conditions, and communities may conflate the researcher with the implementing organisation. Evaluate what ethical and methodological safeguards would best protect the independence and integrity of qualitative AI governance research in these contexts.

Knowledge Check

An ARDAI researcher conducts 12 interviews with Nigerian tax authority staff about an AI risk-scoring system. Nine describe overriding AI scores based on their judgment; three describe always following the AI. The researcher wants to claim 'most tax authority staff regularly override AI recommendations.' Is this justified?

Expert Analysis

Qualitative research produces interpretive claims about meanings, mechanisms, and processes grounded in specific cases, not frequency claims across populations. The researcher should reframe as: 'Interviews identified a dominant pattern of AI recommendation overriding, characterised by [specific conditions], alongside a minority pattern of deference, with implications for governance accountability.' This framing is honest about what qualitative evidence supports and makes a stronger intellectual contribution than an unsupportable frequency claim.

Research Laboratory

Lab 10.8: Ethnographic and Qualitative Research Methods

Conduct qualitative research on an AI-related topic in an African context. Design a semi-structured interview guide (10-15 questions). Conduct at least two interviews or one observation session of at least two hours; transcribe or write field notes within 24 hours. Apply Braun and Clarke's six-phase thematic analysis. Present three thematic findings each with a theme statement, explanation, and at least two supporting quotations or observational extracts. Address trustworthiness criteria. Write a positionality statement reflecting how your background shapes interpretation.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 8 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Hammersley and Atkinson developed their framework for academic social science with extended fieldwork timelines of months to years. Evaluate how ethnographic rigour principles can be adapted for African AI governance researchers with limited fieldwork time and resources, and what minimum standards maintain ethnography's distinctive contribution.

Second: Ethnography in African AI governance contexts creates specific risks: government agencies may restrict access when findings might be critical, vendors may attempt to influence observation conditions, and communities may conflate the researcher with the implementing organisation. Evaluate what ethical and methodological safeguards would best protect the independence and integrity of qualitative AI governance research in these contexts.

Third: Evaluate quality criteria for qualitative AI governance research and the strategies for satisfying them.

Lesson 8 complete. Ethnographic and Qualitative Research Methods: foundations for African AI scholarship.

Survey design and measurement validity for African AI research populations

Week 10 · Lesson 9 · ARDAI (TIER I)

Week 10 · Lesson 9 · ARDAI (TIER I)

Survey Design and Measurement in AI Research

Generating credible descriptive evidence about AI adoption, perceptions, and impacts at scale: the principles, pitfalls, and African adaptations of survey methodology.

Learning Objectives

Evaluate survey design challenges specific to AI research in African populations. Analyse measurement validity threats in survey questions about AI perceptions and adoption. Apply questionnaire design principles to produce valid AI survey items for specific African contexts. Evaluate sampling strategies and mode effects in African AI surveys.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Survey Research and AI Measurement

Digital Data Collection and Analysis

I. Survey Design Principles for AI Research

Surveys are the primary method for generating descriptive quantitative evidence about AI perceptions, adoption, and impacts at scale. A well-designed African AI survey can document AI system penetration across sectors, how citizens and workers experience AI-mediated decisions, how AI adoption varies across demographic groups and geographic contexts, and what attitudes toward AI governance prevail among policymakers and affected communities. These descriptive facts are essential inputs to AI governance policy that no other method can produce at scale.

Survey design for AI research faces specific challenges. AI concepts are unfamiliar to many respondents, particularly in African contexts with lower AI literacy: questions asking whether a respondent 'uses AI tools' may conflate AI with any digital technology, or elicit denial from respondents who use AI-powered applications without knowing it. Measurement of AI impacts requires asking about experiences that respondents may not connect to AI -- a respondent denied a loan by an AI scoring system may not know the decision was AI-assisted. These challenges require careful question wording, extensive pre-testing, and often indirect measurement strategies that capture AI-relevant experiences without requiring respondents to identify them as AI-related.

Scholarly Analysis

Groves et al.'s Survey Methodology (2009) is the authoritative reference for survey design and error. Their total survey error framework identifies six sources of error: coverage, sampling, nonresponse, measurement, processing, and specification. For African AI surveys, coverage and nonresponse errors are particularly severe: sampling frames are incomplete, phone and online survey modes miss populations without device access, and response rates vary dramatically. Researchers must document their total survey error strategy, not just their sampling procedure.

II. Question Design and Measurement Validity

Survey questions on AI must measure what they claim to measure. A question asking 'Do you think AI systems are fair?' may be measuring AI familiarity, social desirability, or the accessibility of personal AI experiences rather than genuine fairness perceptions. Establishing construct validity requires pre-testing with cognitive interviews -- asking respondents to think aloud as they answer -- to identify divergences between the intended construct and how respondents actually interpret and answer the question.

Established attitude scales for AI research include the General Attitudes towards AI scale (Schepman and Rodway, 2020), the AI anxiety scale (Wang et al., 2023), and the Algorithmic Aversion and Algorithm Appreciation scales. These were developed primarily in high-income country contexts and require validation in African populations with different AI familiarity levels and institutional trust patterns. Translation into African languages while maintaining conceptual equivalence requires professional translation and back-translation, followed by confirmatory factor analysis to verify the scale structure holds.

Scholarly Analysis

The absence of validated survey instruments for measuring AI perceptions in African contexts is a significant gap. The first ARDAI graduates to develop, validate, and publish African AI measurement scales will make a methodological contribution adopted by subsequent researchers and potentially incorporated into national AI monitoring frameworks. This is among the most impactful research contributions available to African AI researchers regardless of technical background, requiring survey methodology expertise and local contextual knowledge rather than advanced AI technical skills.

Survey questionnaire design and sampling strategy

Survey questionnaire design and sampling strategy

III. Sampling and Mode Effects in African AI Surveys

Sampling for African AI surveys must address the fundamental tension between coverage and cost. In most African contexts, no complete sampling frame exists for the general population: census data is outdated, voter files exclude non-citizens, mobile subscriber databases are proprietary. Random digit dialling reaches only phone-owning populations -- urban, educated, higher-income. Face-to-face probability surveys achieve representative coverage but are extremely expensive. Researchers must choose between the coverage limitations of phone/online surveys and the cost constraints of face-to-face surveys, and must report and account for coverage gaps.

Mode effects -- tendencies to answer the same question differently by survey mode -- are particularly important in African AI surveys. Social desirability effects are stronger in interviewer-administered modes, which may matter for questions about government AI systems in contexts where criticising government is sensitive. Online surveys reach more tech-literate respondents who may be more AI-familiar, biasing adoption and attitude estimates. Conducting mode experiments to estimate these effects, and where possible documenting them in published supplementary materials, is essential for surveys that will be compared across African countries or over time.

Scholarly Analysis

Afrobarometer, the Africa Polling Institute, and the Pew Research Center have all conducted population surveys including technology and digital services questions in African countries. Afrobarometer's round 8 and 9 surveys include questions on government digital services, mobile money, and data privacy that can be reanalysed for AI governance purposes. Researchers who conduct secondary analysis of existing African survey data alongside primary collection reduce costs, increase sample sizes, and enable comparisons over time that single-wave surveys cannot provide.

Primary Source

2009

Groves, R.M. et al. (2009). Survey Methodology. 2nd ed. John Wiley and Sons. Access Source

1. Groves et al. developed their total survey error framework for surveys in populations with high literacy, widespread phone access, and established research infrastructure. Evaluate which aspects require the most adaptation for African AI governance surveys and propose specific methodological innovations that would most improve survey data quality in these settings.

2. AI attitude surveys in Africa may produce different results depending on whether respondents' primary experience of AI is through AI-mediated government services (potentially negative) or commercial AI applications (potentially positive). Evaluate how researchers should account for this heterogeneity in AI experience when designing attitude measures and interpreting scores for governance policy.

Knowledge Check

A researcher surveys Lagos SMEs by phone asking 'Does your business use artificial intelligence?' and finds 7% say yes. A subsequent face-to-face survey using descriptions of specific AI applications finds 61% reporting use of at least one. Which is more valid and what does the discrepancy reveal?

Expert Analysis

Measurement validity is the core issue: the two surveys measure different constructs and produce dramatically different estimates. The 61% estimate better measures actual penetration of AI-powered tools -- the construct most relevant to AI governance policy. The lesson is to operationalise AI through specific applications rather than the label, pre-test with cognitive interviews, and report clearly which operational definition was used so readers can interpret the estimate correctly.

Research Laboratory

Lab 10.9: Survey Design and Measurement in AI Research

Design a survey to measure AI perceptions, adoption, or impacts in an African population. Specify the research question, target population, sampling strategy and its coverage limitations, survey mode and justification, and a questionnaire of 8-12 questions including at least one AI adoption measure using behavioural description, one attitude scale item, and one outcome measure. Document your pre-testing plan and non-response strategy. Pilot with at least three respondents and document revisions.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 9 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Groves et al. developed their total survey error framework for surveys in populations with high literacy, widespread phone access, and established research infrastructure. Evaluate which aspects require the most adaptation for African AI governance surveys and propose specific methodological innovations that would most improve survey data quality in these settings.

Second: AI attitude surveys in Africa may produce different results depending on whether respondents' primary experience of AI is through AI-mediated government services (potentially negative) or commercial AI applications (potentially positive). Evaluate how researchers should account for this heterogeneity in AI experience when designing attitude measures and interpreting scores for governance policy.

Third: Evaluate sampling strategies and mode effects in African AI surveys.

Lesson 9 complete. Survey Design and Measurement in AI Research: foundations for African AI scholarship.

Systematic review and meta-analysis of AI governance evidence from Africa

Week 10 · Lesson 10 · ARDAI (TIER I)

Week 10 · Lesson 10 · ARDAI (TIER I)

Systematic Reviews and Meta-Analysis

Synthesising the cumulative evidence on AI governance impacts with the rigour and comprehensiveness that any single study cannot achieve.

Learning Objectives

Evaluate the rationale for systematic reviews as the primary knowledge synthesis tool for AI governance evidence. Analyse the seven steps of a systematic review and specific challenges applying them to AI governance literature. Apply PRISMA 2020 guidelines to conduct a rapid systematic review of an African AI governance topic. Evaluate meta-analytic methods for pooling AI governance effect estimates and assessing publication bias.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Evidence Synthesis and AI Research

From Data to Knowledge: Scientific Method

I. Why AI Governance Needs Systematic Reviews

Systematic reviews synthesise cumulative evidence from multiple primary studies to produce more reliable conclusions than any individual study. As AI governance literature grows rapidly, policymakers cannot read every relevant study -- they need synthesis. Unsystematic narrative reviews that selectively cite supporting evidence produce misleading syntheses. Systematic reviews address this by making literature search, study selection, quality assessment, and synthesis procedures explicit, reproducible, and comprehensive, so any reader can assess why the synthesis reaches its conclusions.

For African AI governance specifically, systematic reviews make African evidence visible. African AI studies are disproportionately published in lower-impact journals, presented at regional rather than international conferences, and indexed in databases less commonly searched. A systematic review searching African Journals Online alongside Scopus, Web of Science, and Google Scholar will find African evidence that standard reviews miss -- evidence addressing contexts most relevant to African policy decisions. African AI researchers who conduct systematic reviews with Africa-inclusive search strategies produce knowledge infrastructure, not just empirical findings.

Scholarly Analysis

The Cochrane Handbook for Systematic Reviews of Interventions is the authoritative methodological guide. The Campbell Collaboration extends systematic review methodology to social, behavioural, and educational interventions. PRISMA 2020 -- the Preferred Reporting Items for Systematic Reviews and Meta-Analyses -- is the standard reporting checklist required by top journals. African AI researchers conducting systematic reviews should follow PRISMA 2020 and register protocols on PROSPERO or OSF before beginning data extraction.

II. Conducting a Systematic Review of AI Evidence

The seven steps of a systematic review: (1) formulate the question using PICO/PECO; (2) develop and register the protocol; (3) execute the literature search across multiple databases and grey literature; (4) screen titles, abstracts, and full texts; (5) extract data using a standardised form; (6) assess study quality using validated tools (RoB-2 for RCTs, ROBINS-I for observational studies); (7) synthesise findings using narrative synthesis and where appropriate meta-analysis.

For AI governance reviews, defining the intervention precisely is more difficult than for drug trials: 'AI system' encompasses an enormous range of technologies and contexts that may have very different impacts. Reviewers must specify the type of AI, deployment context, decision supported, and level of automation -- so that included studies are actually comparable enough to synthesise. Reviews that lump together all 'AI governance' studies without defining what governance dimension is evaluated produce findings too heterogeneous to be interpretable or actionable.

Scholarly Analysis

A rigorous systematic review of a specific AI governance question across multiple databases, with duplicate screening, full-text review, data extraction, and quality assessment, typically takes 6-12 months and requires a team of at least two reviewers. ARDAI graduates undertaking systematic reviews should plan accordingly and consider rapid reviews if time and resources do not permit a full systematic review. Rapid reviews that are transparent about limitations are more useful than full systematic reviews that are not completed.

Evidence synthesis forest plot and meta-analysis

Evidence synthesis forest plot and meta-analysis

III. Meta-Analysis: Quantitative Evidence Synthesis

Meta-analysis pools effect size estimates from multiple primary studies to produce a summary estimate with greater precision than any individual study. Random effects models, which allow for heterogeneity in true effects across studies, are generally preferred for social science meta-analyses. Key results include: the pooled effect estimate and confidence interval; heterogeneity statistics (I-squared quantifying the proportion of variation attributable to between-study heterogeneity; tau-squared estimating the variance of true effects); and the funnel plot for assessing publication bias.

Meta-analysis of African AI governance evidence faces specific challenges: effect sizes reported on different scales require conversion to a common metric; study populations are diverse and may not be comparable; and publication bias is severe -- studies with positive AI effects are more likely published, biasing pooled estimates upward. Trim-and-fill analysis, Egger's regression test, and p-curve analysis are standard approaches to assessing and correcting for publication bias, though all have limitations. African AI researchers who meta-analyse African evidence and explicitly assess publication bias are producing more reliable syntheses than reviews that ignore this threat.

Scholarly Analysis

The R packages meta, metafor, and robumeta provide comprehensive meta-analysis implementations. ARDAI graduates who conduct systematic reviews and meta-analyses of African AI governance evidence and publish them in review journals (Systematic Reviews, Campbell Systematic Reviews) are producing the knowledge infrastructure that African AI governance needs -- the cumulative synthesis allowing policymakers to understand what the evidence shows across studies, not just what individual studies found.

Primary Source

2023

Higgins, J.P.T. et al. (eds) (2023). Cochrane Handbook for Systematic Reviews of Interventions. Version 6.4. Cochrane. Access Source

1. The Cochrane Handbook was developed for systematic reviews of health intervention efficacy where outcomes are measurable, interventions discrete, and RCTs preferred. Evaluate which aspects require the most adaptation for systematic reviews of AI governance impacts, where AI systems are complex and dynamic and qualitative evidence may be as important as quantitative.

2. Systematic reviews of AI governance evidence are currently dominated by high-income country studies. Evaluate what search strategies, inclusion criteria, and evidence weighting approaches would produce systematic reviews most representative of African AI governance evidence, and whether reviews based predominantly on African evidence would produce systematically different conclusions from globally inclusive reviews.

Knowledge Check

A systematic review of AI diagnostic accuracy studies in African health facilities finds 15 eligible studies with I-squared = 87%. The reviewer calculates a pooled sensitivity of 79% and concludes 'AI diagnostics achieve 79% sensitivity in African health facilities.' What is wrong with this conclusion?

Expert Analysis

High I-squared means the synthesis finding is about heterogeneity rather than a single pooled estimate. The reviewer should investigate what study characteristics (disease type, AI technology generation, reference standard, health facility level, country income group) explain the variation, present subgroup estimates, and conclude with nuanced claims about when AI diagnostics achieve higher versus lower sensitivity. A heterogeneous meta-analysis with a well-explained pattern of variation is more scientifically valuable than a homogeneous one with an uninformative average.

Research Laboratory

Lab 10.10: Systematic Reviews and Meta-Analysis

Conduct a rapid systematic review of an African AI governance topic. Define the review question using PICO or PECO. Register the protocol on OSF before extraction. Search at least four databases including AJOL or African Journals Online. Document search strings and produce a PRISMA 2020 flow diagram. Extract data from at least five included studies and assess quality using RoB-2 or ROBINS-I. Synthesise narratively and if at least five studies share a common effect size metric conduct a random effects meta-analysis using the meta or metafor package in R. Report using the PRISMA 2020 checklist.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 10 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: The Cochrane Handbook was developed for systematic reviews of health intervention efficacy where outcomes are measurable, interventions discrete, and RCTs preferred. Evaluate which aspects require the most adaptation for systematic reviews of AI governance impacts, where AI systems are complex and dynamic and qualitative evidence may be as important as quantitative.

Second: Systematic reviews of AI governance evidence are currently dominated by high-income country studies. Evaluate what search strategies, inclusion criteria, and evidence weighting approaches would produce systematic reviews most representative of African AI governance evidence, and whether reviews based predominantly on African evidence would produce systematically different conclusions from globally inclusive reviews.

Third: Evaluate meta-analytic methods for pooling AI governance effect estimates and assessing publication bias.

Lesson 10 complete. Systematic Reviews and Meta-Analysis: foundations for African AI scholarship.

Open science, pre-registration and reproducibility in African AI research

Week 10 · Lesson 11 · ARDAI (TIER I)

Week 10 · Lesson 11 · ARDAI (TIER I)

Reproducibility, Pre-registration, and Open Science

Why most AI governance research cannot be reproduced, what pre-registration does to fix this, and why open science practices define the research careers of the next decade.

Learning Objectives

Evaluate the causes and consequences of the reproducibility crisis for African AI governance research. Analyse the role of pre-registration in preventing p-hacking and distinguishing confirmatory from exploratory analysis. Apply open science practices to an African AI research project. Evaluate the open access landscape and strategies for maximising impact while maintaining accessibility.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Open Science and Research Integrity

Reproducible Research: Lessons from Science

I. The Reproducibility Crisis and AI Research

The reproducibility crisis refers to the widespread finding that a substantial proportion of published scientific results cannot be reproduced by independent researchers. In psychology, the Open Science Collaboration found only 36-39% of findings replicated with similar effect sizes. AI research has its own reproducibility crisis: machine learning studies on benchmark datasets often cannot be reproduced due to undocumented hyperparameter choices, random seed dependence, and data preprocessing decisions not described in papers; AI governance research suffers from the same selective reporting and publication bias that affect other social sciences.

For African AI governance researchers, irreproducibility has specific consequences: research that cannot be reproduced does not contribute reliable knowledge to the evidence base on which African AI policy should be built. African AI policymakers relying on irreproducible research make governance decisions on false evidence. African AI researchers who adopt open science practices -- sharing code and data, pre-registering analysis plans, reporting all results including null findings -- contribute to a more reliable evidence base while differentiating themselves from researchers who do not, signalling credibility to journals, funders, and employers who increasingly require open science practices.

Scholarly Analysis

Nosek et al.'s 2015 Science paper 'Promoting an Open Research Culture' identifies the structural incentives producing irreproducibility (publication bias toward positive results, incentives for novelty over replication, absence of data and code sharing norms) and proposes institutional interventions (pre-registration, open data, open materials, replications). The Centre for Open Science framework has been adopted by increasing numbers of journals, funders, and universities. African AI researchers adopting open science practices are positioning themselves for the research funding environment of 2025 and beyond where these requirements are becoming standard.

II. Pre-registration: What It Is and Why It Matters

Pre-registration is specifying a study's research questions, hypotheses, design, and analysis plan in a publicly accessible repository before data collection begins. It matters because it distinguishes confirmatory hypothesis testing from exploratory data analysis. When a researcher examines data and reports only analyses producing significant results, they are p-hacking -- inflating the false positive rate far above the nominal significance level. Pre-registration solves this by making the original analysis plan public, enabling readers to identify deviations between the pre-specified plan and the reported analysis.

The AEA RCT Registry, Open Science Framework, PROSPERO, and EGAP are the main pre-registration repositories for African AI governance research. A pre-registration should specify: the primary research question and hypothesis; study design and sample; primary and secondary outcome definitions; the analysis estimator, covariates, and standard error specification; the subgroup analyses; and stopping rules. Registered Reports -- a publication format where journals agree to publish based on pre-registration and methods before seeing results -- are the gold standard for eliminating publication bias, increasingly offered by leading journals including Nature Human Behaviour and PLOS ONE.

Scholarly Analysis

Pre-registration imposes a real cost: researchers must think through analysis plans before seeing the data, requiring more upfront methodological investment. But this cost is also a benefit: the discipline of pre-specification forces identification of design weaknesses before data collection when they can be addressed cheaply. African AI researchers who pre-register also gain a time-stamped record of research intentions protecting against accusations of post-hoc hypothesis generation -- a protection that matters when findings have policy implications and opponents have incentives to dispute the methodology.

Pre-registration and open science workflow

Pre-registration and open science workflow

III. Open Data, Open Code, and Open Access

Open data means making research datasets publicly available so others can verify results, conduct reanalyses, and build on the work. For African AI research, open data is both ethically complex and practically important: complex because many datasets contain personal information that cannot be shared without risking privacy violations; practically important because data scarcity constrains African AI research. The resolution is structured data sharing: de-identified datasets shared under data use agreements enabling reanalysis while protecting participant privacy. OSF, Zenodo, Harvard Dataverse, and the UK Data Archive provide secure repository infrastructure for African AI research data sharing.

Open code means sharing analysis scripts in sufficient detail that an independent researcher can replicate the analysis from raw data. Open code requirements are now standard at leading journals including Science, Nature, and the American Economic Review, and increasingly required by development finance funders. Open access publication -- making research freely available without paywalls -- is particularly important for African AI governance because policymakers, practitioners, and community organisations who most need findings often lack subscription access. Preprint servers including arXiv, SSRN, and AfricArXiv provide free immediate open access to African policymakers before formal publication.

Scholarly Analysis

Diamond open access journals -- charging neither subscription fees nor author fees -- are a growing alternative to the APC model that charges USD 1,500 to 10,000 per article and disadvantages African researchers. The African Journals Online platform hosts over 500 African journals, many open access and indexed in major databases. Researchers who use AfricArXiv for preprints and diamond OA journals for publication maximise research impact and accessibility without incurring APCs that African researchers' budgets typically cannot support.

Primary Source

2015

Nosek, B.A. et al. (2015). Promoting an Open Research Culture. Science, 348(6242), pp.1422-1425. Access Source

1. Nosek et al. propose open science as a solution to the reproducibility crisis driven primarily by incentive structures in high-income country academic publishing. Evaluate which aspects of this diagnosis and solution apply to African AI research contexts, where publication incentives may differ, data sharing carries specific sovereignty and privacy risks, and open access publishing infrastructure is less developed.

2. Open data requirements create a tension between scientific reproducibility and the data sovereignty rights of African communities whose data is used in AI governance research. Evaluate what data governance frameworks would best reconcile open science values with African data sovereignty principles, enabling reproducibility without extractive data sharing, and what role African research ethics boards should play in enforcing these frameworks.

Knowledge Check

An ARDAI researcher runs 47 regression specifications testing different combinations of AI adoption measures, control variables, and outcome definitions. They find 11 with p < 0.05 and report the three most interesting ones without mentioning the other 44. What is the problem and how should it be addressed?

Expert Analysis

Specification searching is one of the most common sources of false positive findings in AI governance research and is particularly prevalent because researchers typically have theoretical reasons to prefer certain specifications and may genuinely believe they are reporting the best analysis. Pre-registration eliminates ambiguity by making the primary specification public before any analysis is conducted. The three-step rule -- specify your primary test before seeing the data, run it first, then explore -- is the practical discipline that pre-registration enforces.

Research Laboratory

Lab 10.11: Reproducibility, Pre-registration, and Open Science

Pre-register an original AI research study and develop an open science plan. Submit a pre-registration to OSF and provide the link. Specify: research question and primary hypothesis; study design; population and sampling; primary and secondary outcomes with operational definitions; primary analysis estimator and covariates; subgroup analyses; stopping rules. Develop a data sharing plan (de-identification procedures, data use agreement, repository choice) and a code sharing plan (programming language, documentation standard, version control platform). Develop an open access plan (target journal OA policy, preprint server, alternative routes if APCs are unaffordable). Describe how an independent researcher would reproduce your results and what the main barriers to reproducibility are in your specific context.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 11 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Nosek et al. propose open science as a solution to the reproducibility crisis driven primarily by incentive structures in high-income country academic publishing. Evaluate which aspects of this diagnosis and solution apply to African AI research contexts, where publication incentives may differ, data sharing carries specific sovereignty and privacy risks, and open access publishing infrastructure is less developed.

Second: Open data requirements create a tension between scientific reproducibility and the data sovereignty rights of African communities whose data is used in AI governance research. Evaluate what data governance frameworks would best reconcile open science values with African data sovereignty principles, enabling reproducibility without extractive data sharing, and what role African research ethics boards should play in enforcing these frameworks.

Third: Evaluate the open access landscape and strategies for maximising impact while maintaining accessibility.

Lesson 11 complete. Reproducibility, Pre-registration, and Open Science: foundations for African AI scholarship.

Research ethics, informed consent and data protection in African AI studies

Week 10 · Lesson 12 · ARDAI (TIER I)

Week 10 · Lesson 12 · ARDAI (TIER I)

Research Ethics for AI Studies in Africa

The ethical obligations that apply at every stage of African AI research: consent, data protection, research justice, and the positionality of the African AI scholar.

Learning Objectives

Evaluate the specific ethical challenges of AI research with human participants in African contexts. Analyse data protection legal frameworks applicable to AI research in specific African jurisdictions. Apply adapted consent procedures appropriate for multilingual, low-literacy African research populations. Evaluate the principles of research justice and researcher positionality as ethical dimensions of African AI research.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

Research Ethics in African AI Contexts

Algorithmic Bias and Research Responsibility

I. The Ethics of Human Subjects Research in AI Studies

Research ethics for AI studies with human participants extends the established principles of human subjects research -- respect for persons, beneficence, and justice -- to the specific challenges of studying AI systems that process personal data, make consequential decisions, and interact with vulnerable populations. The Belmont Report's three principles, codified in national research ethics frameworks across Africa including Nigeria's National Code of Health Research Ethics and South Africa's DoH Ethics in Health Research guidelines, provide the foundational framework. Respect for persons requires genuinely voluntary informed consent based on adequate understanding. Beneficence requires research risks be minimised and benefits maximised. Justice requires that research burdens and benefits are fairly distributed, specifically that vulnerable populations are not exploited while their more privileged counterparts benefit from the knowledge produced.

AI research raises specific ethical challenges beyond standard human subjects research. Informed consent for AI studies must address system complexity in terms accessible to participants who may have limited AI literacy. Secondary analysis of administrative data generated by AI systems involves personal data that individuals did not provide for research purposes, raising questions about whether original consent provisions cover research use. Algorithmic auditing -- testing AI systems for bias -- may require accessing the system's decision-making process in ways the deploying organisation has not consented to, creating tensions between research transparency and proprietary information. Community-level impacts of AI systems may require community-level consent processes that individual consent procedures do not provide.

Scholarly Analysis

Nigeria's NHREC and South Africa's Health Research Ethics Committees are the primary institutional review boards for human subjects research in these two largest African research economies. Many African countries lack equivalent independent ethics review infrastructure for non-health AI research, creating regulatory gaps researchers must navigate transparently. Researchers conducting AI studies across multiple African countries should obtain ethics approval from all relevant national bodies and document their ethics framework even when formal national processes do not exist.

II. Consent, Privacy, and Data Protection in African AI Research

Informed consent procedures must be adapted to the specific characteristics of the research population. In African contexts with multilingual populations, low literacy in official languages, and limited familiarity with research procedures, standard written consent forms in English or French are often inadequate. Verbal consent in local languages, witnessed by a community representative, combined with clear oral explanation of purpose, risks, benefits, and the right to withdraw, is more appropriate. Consent procedures should be pre-tested with members of the target population to ensure the explanation is genuinely understood and the consent is meaningfully voluntary -- not obtained through social pressure, economic incentive disproportionate to circumstances, or deference to researchers perceived as authority figures.

Data protection in African AI research is increasingly governed by national legislation: Nigeria's Data Protection Act 2023, South Africa's POPIA 2013, Kenya's Data Protection Act 2019, Ghana's Data Protection Act 2012, and the AU Data Policy Framework 2022. Researchers must map their data flows against relevant legal requirements, implement appropriate technical and organisational safeguards, and document compliance in the ethics protocol. The intersection of research ethics and data protection law is where most compliance failures in African AI research occur -- researchers familiar with ethics review but not data protection law miss requirements in the less familiar framework.

Scholarly Analysis

The African Union's Draft Data Policy Framework (2022) and the Smart Africa Digital Economy Framework provide regional context for national data protection laws. These frameworks reflect African governance priorities -- data sovereignty, development benefit, and community rights -- that differ in emphasis from the individual autonomy focus of European GDPR. African AI researchers who understand both national legal requirements and regional policy frameworks are better positioned to design research that is legally compliant and aligned with African governance values, building institutional credibility that sustainable research programmes require.

Research ethics protocol and consent documentation

Research ethics protocol and consent documentation

III. Power, Positionality, and Research Justice

Research justice goes beyond standard ethics to ask: who benefits from this research, whose knowledge counts as valid, and who bears the risks of research error? For African AI research conducted in partnership with international institutions, these questions have specific urgency: research agendas may be shaped by funder priorities rather than community needs; findings may be published in journals inaccessible to African policymakers; data may be held by international institutions; and the professional benefits of research may accrue primarily to international researchers while African community members bear the risks. Research justice frameworks call for community benefit agreements, African data sovereignty arrangements, African-accessible publication strategies, and capacity building provisions that build African research infrastructure alongside the specific study.

Positionality matters because the researcher's position -- insider or outsider, powerful or marginalised, affiliated with the implementing agency or independent of it -- shapes what can be observed, what questions can be asked, what answers are given, and how findings are interpreted. Hybrid positions -- African researchers with international institutional affiliations, or collaborative teams combining insider and outsider perspectives -- often produce the richest findings but require explicit attention to managing conflicting institutional loyalties and equitable distribution of intellectual credit.

Scholarly Analysis

Research ethics review processes in African contexts vary dramatically in their capacity and rigour. Researchers should treat ethics review as a floor, not a ceiling -- obtaining approval does not mean the research is ethical. The substantive ethical obligations described in this lesson apply regardless of what the local review process requires, and researchers should document their compliance in detail even when the formal process does not require it. Submit your ethics protocol to [email protected] for mentor review before submitting to your institutional ethics board.

Primary Source

2023

NHREC (2023). National Code of Health Research Ethics. Nigeria Health Research Ethics Committee, Abuja. Access Source

1. Nigeria's NHREC Code focuses primarily on health research and was developed before AI was significant in health system design. Evaluate what adaptations are needed for AI health research specifically -- AI diagnostic systems, algorithmic triage, predictive risk scoring -- and what process would be most appropriate for developing these adaptations within the Nigerian research ethics governance framework.

2. Research justice frameworks argue that communities whose data and experiences are the subject of AI research should have meaningful agency in defining research questions, assessing risks, and accessing findings. Evaluate what participatory mechanisms would be most appropriate for African AI governance research in community settings, and how these should be balanced against the scientific standards that peer review and funders impose.

Knowledge Check

An ARDAI researcher wants to audit an AI loan scoring system by submitting fabricated test applications to see how the algorithm scores different profiles without the bank's knowledge. What are the main ethical issues?

Expert Analysis

Algorithmic auditing occupies an ethical grey area that standard human subjects frameworks do not fully address. The ethical solution is not to avoid auditing but to use methods achieving the same objective without deception: sandboxed testing with synthetic data, regulatory-mandated disclosure of algorithmic decision factors, or analysis of existing records under proper data sharing agreements. The ARDAI community is well-positioned to advocate for regulatory frameworks mandating algorithmic auditability, removing the need for deceptive audit methods.

Research Laboratory

Lab 10.12: Research Ethics for AI Studies in Africa

Write a research ethics protocol for an AI study involving human participants in an African context. Address: study description; adapted informed consent procedure appropriate for the target population's literacy level and language; data protection compliance mapping data flows against relevant African law; risk and benefit assessment; community benefit plan; positionality statement (300 words). Identify the relevant ethics board(s) and describe your submission process. Submit to [email protected] for mentor review before submitting to your institutional ethics board.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 12 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Nigeria's NHREC Code focuses primarily on health research and was developed before AI was significant in health system design. Evaluate what adaptations are needed for AI health research specifically -- AI diagnostic systems, algorithmic triage, predictive risk scoring -- and what process would be most appropriate for developing these adaptations within the Nigerian research ethics governance framework.

Second: Research justice frameworks argue that communities whose data and experiences are the subject of AI research should have meaningful agency in defining research questions, assessing risks, and accessing findings. Evaluate what participatory mechanisms would be most appropriate for African AI governance research in community settings, and how these should be balanced against the scientific standards that peer review and funders impose.

Third: Evaluate the principles of research justice and researcher positionality as ethical dimensions of African AI research.

Lesson 12 complete. Research Ethics for AI Studies in Africa: foundations for African AI scholarship.

ARDAI capstone: integrating advanced research methods for African AI scholarship

Week 10 · Lesson 13 · ARDAI (TIER I)

Week 10 · Lesson 13 · ARDAI (TIER I)

Week 10 Capstone: Advanced Research Methods Portfolio

Integrating the full toolkit of advanced research methods into an original, rigorous, and policy-relevant contribution to African AI governance scholarship.

Learning Objectives

Integrate the advanced research methods of Week 10 into a complete original research contribution for African AI governance. Apply the quality criteria of the chosen method to produce a capstone of publishable standard. Evaluate the research design choices and threats to validity that could not be fully addressed. Develop a publication and career strategy for translating the capstone into research impact.

Supplementary Video Resources The videos below provide foundational and related context for this lesson. Where no specialist video exists for this exact topic, a methodologically relevant resource is provided.

From Research to Impact: Joy Buolamwini

African AI Scholarship and Global Accountability

I. Integrating Advanced Research Methods for African AI Scholarship

The Week 10 capstone is the demonstration that you have internalised the full toolkit of advanced research methods for AI governance research -- not just as a catalogue of techniques but as an integrated approach to producing credible, rigorous, and impactful knowledge about AI in African contexts. The methods covered in Lessons 1-12 -- from the potential outcomes framework through RCTs, DiD, RDD, IV, mixed methods, ethnography, survey design, systematic review, open science, and research ethics -- are not alternatives to each other. They are complementary tools whose appropriateness depends on the research question, available data, institutional context, and ethical constraints of the specific African AI research setting.

The most impactful African AI research careers are built on methodological flexibility alongside substantive depth. The ARDAI graduate who can design a field experiment when randomisation is feasible, implement DiD when staggered rollout creates a natural experiment, conduct ethnographic fieldwork to understand mechanisms that statistical analysis reveals but cannot explain, synthesise existing evidence through systematic review, and conduct all of this with exemplary research ethics and open science practices, is the researcher whose work changes African AI governance. This capstone asks you to demonstrate that integration by producing your most ambitious research contribution to date.

Scholarly Analysis

The advanced research methods developed in Week 10 are the foundation for original contribution to global AI governance research, not just for competent application of established methods. The most important AI governance research questions -- how AI systems affect economic mobility in Africa, how algorithmic decision-making changes accountability in African public institutions, how AI governance frameworks can be designed to work within African institutional realities -- have not been answered. The researchers who will answer them are ARDAI graduates combining methodological rigour with African contextual expertise.

II. Capstone Options and Standards

The Week 10 capstone offers three options, each requiring a complete research contribution at publishable standard. Option A: a pre-registered empirical study using causal inference methods -- RCT, DiD, RDD, or IV -- with complete protocol, analysis plan, and either results from implemented analysis or a fully powered pilot study. Option B: a systematic review and meta-analysis registered on PROSPERO or OSF, following PRISMA 2020, with at least 10 included studies and meta-analytic pooling where feasible. Option C: a rigorous mixed methods study combining at least one quantitative and one qualitative strand, with a complete joint display integrating both and a meta-inference going beyond either strand alone.

All three options must meet the same quality standard: the submitted work must be of a quality that, with appropriate mentoring and revision, could be submitted to a peer-reviewed journal within six months of ARDAI completion. This requires: substantive significance (addressing a question that matters for African AI governance); methodological rigour (satisfying the key quality criteria of the chosen method); honest limitation acknowledgment (clearly identifying the weakest design points and their implications); and policy relevance (making clear what the research implies for AI governance decisions in the African context studied).

Scholarly Analysis

The Week 10 capstone should be submitted to [email protected] for mentor review before final submission to the ARDAI programme. Mentors provide feedback on research design, execution quality, and appropriate publication venues. Graduates who receive and incorporate mentor feedback before final submission consistently produce higher quality capstones and are better positioned for journal submission. The ARDAI research network includes researchers currently publishing in World Development, Journal of Development Economics, AI and Society, Big Data and Society, and Social Science and Medicine.

Research portfolio and capstone completion

Research portfolio and capstone completion

III. From Capstone to Career: Research as Professional Practice

The ARDAI programme has built the research methods foundation. What comes next -- turning that foundation into a sustained research career, a policy impact record, or an applied practice of evidence-based AI governance -- is the work of a professional life. The graduates who make the most of Week 10 are those who treat research methods not as techniques to deploy on demand but as a way of thinking about evidence, causation, mechanism, and uncertainty that pervades everything they do professionally. The ARDAI practitioner who reviews a government AI system specification and immediately asks 'what would be the counterfactual?' is applying the potential outcomes framework in their daily work, even without running a regression.

Next steps for ARDAI Week 10 graduates: submit the capstone to a relevant journal or preprint server; connect with the ARDAI alumni network for follow-up research collaboration opportunities; engage with the Week 11 AI Policy Design curriculum that builds directly on this research methods foundation by applying evidence-based approaches to policy design and regulatory analysis; and identify a research funding opportunity -- African Academy of Sciences, NRF, IDRC, or a development bank research programme -- for which the Week 10 capstone serves as the pilot study. The methodological investment of Week 10 should generate research output and career advancement for years beyond programme completion.

Scholarly Analysis

Advanced research methods competency is among the most valued skills in the AI governance labour market precisely because it is scarce. Development banks, academic institutions, think tanks, and policy research organisations all need researchers who can produce credible evidence about AI impacts and governance effectiveness. The ARDAI graduate who completes Week 10 with a pre-registered, methodologically rigorous, policy-relevant capstone demonstrates exactly this competency in the most direct way available. Submit to [email protected] and engage your mentor before the final submission deadline.

Primary Source

2011

Banerjee, A. and Duflo, E. (2011). Poor Economics: A Radical Rethinking of the Way to Fight Global Poverty. PublicAffairs. Access Source

1. Banerjee and Duflo build their case for evidence-based development policy on RCT evidence from low-income country contexts. Evaluate how their argument for evidence-based AI governance in Africa must be modified given: the dynamic nature of AI systems that evolve after evaluation; the interaction effects between AI and institutional context making results highly context-specific; and ethical constraints on randomising access to potentially beneficial AI systems in African public services.

2. The advanced research methods of Week 10 are tools for understanding what is happening with AI in Africa. But the research questions these tools are applied to are shaped by funding priorities, researcher interests, and institutional incentives that may not reflect the most important African AI governance questions. Evaluate what mechanisms would most effectively ensure that ARDAI-built research methods capacity is applied to the AI governance questions that most matter for African populations.

Knowledge Check

An ARDAI graduate's DiD study shows AI-assisted benefit targeting in Rwanda increased transfers to the poorest quintile by 18 percentage points. The government asks the graduate to present this as showing AI targeting is better than manual targeting in all African contexts. Should they agree?

Expert Analysis

This illustrates the most common misuse of rigorous causal research: overstating external validity to support policy expansion. The graduate's responsibility is to communicate accurately what the study can and cannot show. The evidence establishes the Rwandan result credibly but the generalisation to other African contexts requires either theoretical arguments about why the mechanism would operate similarly or empirical evidence from additional contexts. Agreeing to the government's characterisation would compromise the research integrity that gives the DiD estimate its value.

Research Laboratory

Lab 10.13: Week 10 Capstone

Complete your Week 10 capstone: an original advanced research methods contribution to African AI governance scholarship.

Choose one option:

(A) Pre-registered empirical study using RCT, DiD, RDD, or IV: submit the pre-registration link (OSF, AEA, or EGAP), complete protocol, analysis code, results or pilot data with power analysis, and a 3,000-word discussion of findings and limitations.

(B) Systematic review and meta-analysis: PROSPERO or OSF registered; submit registration link, PRISMA flow diagram, data extraction table (minimum 10 studies), quality assessment, narrative synthesis, and meta-analytic results if feasible.

(C) Mixed methods study: pre-registered design with quantitative and qualitative strands integrated; submit pre-registration, quantitative results, qualitative findings (minimum 2 fully coded interviews), joint display, and 1,500-word meta-inference.

All submissions must include: - A 500-word statement identifying the study's main validity threat and what it implies for confidence in the findings - A 300-word publication strategy: target journal, expected timeline, and plan for addressing anticipated reviewer concerns

Submit to [email protected]: Subject: ARDAI (TIER I) Capstone Week 10 Research Methods, Full Name, Institution.

⚠

This lab is not an exercise. It is a work sample. The practitioner who completes this to a publishable standard has a concrete deliverable to show a World Bank interviewer, a GovTech consultant, a civil society funder, or a remote employer on Upwork or LinkedIn. Knowledge you cannot demonstrate cannot be hired. Skimming this lab produces knowledge. Completing it produces evidence.

Submit your completed lab to [email protected] - Subject: ARDAI Lab Week 10 Lesson 13 - [Your Full Name] - [Institution] - for mentor review and inclusion in the SIP portfolio registry.

Distinction

Research question is original and significant for African AI governance; identification strategy is methodologically rigorous and precisely defended; African institutional context is specifically grounded; all analysis is pre-specified and replicable; submission meets competitive research funding standard.

Merit

All required elements addressed with methodological accuracy and contextual specificity to African AI governance.

Pass

Core elements present with adequate methodological and contextual grounding.

Open in Google Colab

ARDAI (TIER I) Graduate Reflection

First: Banerjee and Duflo build their case for evidence-based development policy on RCT evidence from low-income country contexts. Evaluate how their argument for evidence-based AI governance in Africa must be modified given: the dynamic nature of AI systems that evolve after evaluation; the interaction effects between AI and institutional context making results highly context-specific; and ethical constraints on randomising access to potentially beneficial AI systems in African public services.

Second: The advanced research methods of Week 10 are tools for understanding what is happening with AI in Africa. But the research questions these tools are applied to are shaped by funding priorities, researcher interests, and institutional incentives that may not reflect the most important African AI governance questions. Evaluate what mechanisms would most effectively ensure that ARDAI-built research methods capacity is applied to the AI governance questions that most matter for African populations.

Third: Develop a publication and career strategy for translating the capstone into research impact.

Lesson 13 complete. Week 10 Capstone: foundations for African AI scholarship.