Reflections

Fourth year was a rollercoaster… It challenged me in ways I did not anticipate. It encouraged me to skill-up. It put me out of my comfort zone. It somehow increased my imposter syndrome in some settings but completely emboldened my confidence in other ways. Many of my reflections about research from third year carried over into fourth year, but I would say that this year was shaped by Adaptation and AI.

Adaptation

One of the biggest skills a PhD teaches you is how to adapt to failure and change. I was balancing multiple projects: a solo-authored RCT in India, a co-authored development project in Uganda, a co-authored project with my professor, and a solo-authored labor project that used administrative data. I had initially entered fourth year with a detailed Gantt chart detailing which projects I would prioritize each month. I tried to predict and stagger my fieldwork, but timelines constantly shifted and evolved.

I encountered a few big roadblocks on two of my development economics projects: (1) For my solo-authored RCT in India, my implementing partner was no longer able to continue with me. While they supported the project, in the time it took to get the funding, their priorities had shifted and resources would not be available for my research. I was luckily able to find another implementing partner, but the process delayed timelines quite substantially. (2) For my co-authored development project in Uganda, timelines were delayed due to us changing the overall design. We had submitted a pre-analysis plan to a journal and received valuable feedback worth incorporating.

Adapting to these challenges required communication, creativity, perseverance, and project management. I was able to be flexible with my time management to prioritize projects involving administrative data during low-fieldwork periods (and vice versa) while constantly keeping my partner organizations engaged.

In this way, I felt that my PhD forced me to wear different hats:

• I was the researcher - designing the empirical strategy, running the analysis, and thinking of additional extensions/checks
• I was the interpreter - distilling complex design choices to non-technical partners and packaging the story of my research to make others care
• I was the project manager - balancing timelines, budgets, clients, while maintaining weekly communication

This year was the first year I felt that being a researcher was akin to a professional career.

I think that a common misconception by PhD students is that we are researchers with specific training in specific skills. While it is true that we learn valuable technical skills and grow expertise, I think we need to acknowledge the different hats we wear a bit more and learn to “sell” ourselves as working professionals.

AI

AI is everywhere. I don’t live in the Bay Area, so I don’t hear AI chatter as much, but being geographically close to the epicenter of AI innovation certainly brings about a certain buzz. I had read numerous blogs, LinkedIn posts, and seen YouTube videos expressing how revolutionary AI can be for research. But it all still seemed a bit overwhelming given the content from economists in Academia, software developers, and big tech companies.

I wanted to “upskill” for many reasons - increased efficiency in work, better career prospects, and curiosity. However, I had a few fears: (i) I didn’t want to lose my critical thinking skills and constantly default to an LLM. (ii) I wanted coding assistance, but also wanted to be able to explain, test, and verify what I was doing. (iii) Above all, I wanted to ensure data quality/accuracy. While models have been getting better from the early days in which there were reports about made up citations, LLMs can still pretty convincingly make mistakes. But regardless of these fears, I knew that I didn’t want to be left behind.

The wave of economics PhD students right now remembers a time without AI in which we read textbooks, did research more “manually”, and learned to code. We are still navigating our workflows and can adapt to the changing times. I am of the opinion that if we don’t use AI and figure out our complementary niche, we will be left behind.

So I sat down to “upskill.” The first blog I read was Scott Cunningham’s blog on how Claude Code transformed his workflow. Then, numerous followed - LinkedIn posts, substacks, youtube videos. My current workflow is pretty basic, but works for me. I use VS Code as my IDE, connect Claude Code to it, and integrate all projects with GitHub. I led an internal Professional Development session with my cohort to describe this workflow and teach everyone a basic workflow. I know many folks like myself were intimidated about knowing where to start and perhaps just using AI via writing prompts independently with repeated context each time. Now that more folks are integrating AI into their workflows, we have a bigger group of folks to discuss tools with together.

I follow the following principles when using AI:

• Teach myself aloud - If I ever ask an LLM to explain a paper, argument, proof, or piece of code I think is foundational, I try to summarize it myself out loud and in my own words - almost as if I am teaching. I don’t ever want to be in a situation in which I relied on AI so heavily I can’t respond to questions people might ask me about my own research. I find that this helps me catch mistakes and frame my research in my voice.
• Keep writing - I am not at a stage in which I feel comfortable asking an LLM to write. I find that it distorts my voice. The process of writing is an important skill I do not ever want to lose. I write all of my own work and even if an AI tool helps frame something, the process of re-writing and re-structuring helps me catch mistakes, think of extensions, and only treat AI as a soundboard.
• Code with AI - I use Claude Code and it has been a game changer. I was at first very reluctant to use AI to produce code, and in my fourth year, I still worked on a few projects in which I coded everything myself. But I have begun to integrate agentic coding more and more. I started with more “low-stakes” work like data visualization or more standard exercises I have experience with (ex: power calculations). But I have recently started using it to produce sections of cleaning/analysis code. I still verify everything myself and prompt using my coding background. I leverage local .md files and use/create different skills.
• Constantly learn - It is hard to keep up with all of the developments. A quick google search on AI for Economists will yield many important GitHub repositories, blogs, and LinkedIn posts. I have found some tips extremely helpful and have subsequently set-up my projects to manage different agents, re-iterate and check work in loops, define committees of different perspectives, preserve skepticism etc. But I find that fully copying someone else’s published workflow and trying to download/copy all skills is a bit overwhelming. I try my best to instead “shop around” - introduce tools at a slower pace when I need them and discuss tools with peers.

Conclusion

As I have been doing for the past four years, I noted my activities in a running list. I’m proud of the progress I have made and confidence as a researcher I have built!

Fourth Year Video

Here is my one-second-a-day video of the moments of joy that defined my fourth year: