My analysis isn't working and I'm stuck
Step back to what the analysis needs to show, rebuild from a simple version and get another person to look at it with you.
In short
Start here, today. Write, in plain language, what this analysis needs to show for your thesis, in one or two sentences. Then sketch what the main table, figure or theme would look like if it worked. This takes about 15 minutes.
After that, work through these steps. Choose one to jump to it.
Step by step
Tick each step as you finish it. Open a step for the how.
0 of 6 steps done
Restate what the analysis must show
Go back to your research question and the claim this chapter or paper has to support. Analyses often drift into answering something more ambitious than you need. Narrowing the target is good design, not cheating.
Go back to a simple version
Run the simplest analysis that could address the question: descriptive statistics, a model with only the main variable, a plot, ten transcripts coded by hand, a timeline of events. If that works, add complexity one step at a time until something breaks. That's where the problem is.
Check the data before the method
Many stuck analyses turn out to be data problems: missing values coded as zeros, duplicated rows, a merge that silently dropped cases, mismatched units, a variable that means something different in one wave or site. Look at raw distributions, count your cases at every step, and read a few records in full. In qualitative work, check that codes are being applied consistently across the material.
Keep an analysis log
Each time you run something, note the date, the data file, what you changed and what happened. It's the analysis version of a lab notebook, and it stops you going in circles. Commented scripts saved next to their outputs do part of this for you.
Talk it through with someone
Explain the problem out loud to a peer, lab mate or your supervisor, starting from the question rather than the error message. Saying it to another person often shows you the gap. Many universities also have a statistics consulting service or research computing help desk for graduate students; check what yours offers, and bring your question, a description of the data and the simplest version you've tried.
Write while you analyse
Draft rough methods and first results paragraphs now. Writing makes you say what each step is for, and the steps you can't justify in a sentence are often the ones causing trouble. When the analysis does settle, you'll already be partway into the chapter.
Words you can use
Is this you? Signs to look for
- Your model won't converge, your code throws errors you don't understand, or your themes won't settle into anything coherent.
- You've been adjusting the same analysis for weeks, and every fix creates a new problem.
- You can't say which version of the data or script produced which result.
- You're no longer sure what the analysis is meant to tell you.
Why this happens
Analysis tends to get stuck for one of a few reasons. It has grown more complicated than the question needs. There's a problem in the data nobody has found yet. Or the question has drifted, and you're asking the analysis to answer something it wasn't set up for. From the inside these feel identical: circular, frustrating work that never resolves.
It happens in statistical, computational and qualitative work alike, and it can be isolating, because nobody else knows your data as well as you do. What usually helps is going backwards, to a simpler version and a plainer statement of what you need, with someone else looking at it alongside you.
Worth a look (2)
- Sage Research MethodsBooks, case studies, videos and practice data sets on research methods, from interviews to regression. Your library usually provides access.
- PRISMAThe reporting guideline for systematic reviews and meta-analyses, with the checklist, flow diagram templates and extensions for scoping and other reviews. Many journals expect you to follow it.
When to bring someone in
If you've been stuck on the same technical problem for more than a week or two, bring in your statistics consulting service, research computing team or a methods specialist in your department. A committee member with methods expertise is a good person to ask, especially if the problem might change your design. If you suspect the analysis can't answer your question at all, that's a conversation for your supervisor.
Find the right person, or talk to Ify