Chi Square Expected Count Less Than 5 SPSS: Can You Still Use the Test?
Chi square expected count less than 5 SPSS is probably the exact phrase you typed into Google five minutes ago, right after your crosstabs output showed that dreaded footnote. I have reviewed this exact SPSS output hundreds of times across twelve years of dissertation guidance work, and I can tell you straightaway: it is not the end of your analysis. But it is also not something you should ignore and hope nobody notices.
- One cell mildly below 5, in a larger table, with none below 1: report the chi square result as usual, add a one line caveat.
- 20% or more of your cells below 5, or any single cell below 1: switch to Fisher’s Exact Test, Monte Carlo, or collapse categories with conceptual justification.
- Total sample under roughly 20 cases: the real fix is a redesign or more data, not a statistical workaround.
The Short Answer First
Yes, you can often still use the chi square test, but it depends on how many cells are affected and how badly. SPSS flags this as a warning, not an automatic disqualification. The real question is whether your violation falls under a minor caveat or a full method switch.
Yes, But It Depends on How Many Cells and By How Much
If only one or two cells out of a large table are mildly below 5, most examiners will accept the result with a short note. If a third or more of your cells are affected, or if any single cell has an expected count below 1, you have a genuine problem. I will show you exactly where that line sits in a moment.
What “Chi Square Expected Count Less Than 5 SPSS” Actually Means
Before fixing anything, understand what SPSS is actually telling you. This single idea trips up more postgraduate students than any other part of the chi square output.
Expected vs Observed Frequency, Explained Simply
Observed frequency is the actual number of cases you counted in each cell of your contingency table. Expected frequency is the number SPSS calculates mathematically, assuming there is no relationship between your two variables at all. The calculation uses your row totals, column totals, and the degrees of freedom in your table. When your sample is small or your categories are unevenly split, some of those calculated expected numbers drop below 5, even though your real, observed numbers might look fine.
Expected cell count, in a chi square test, is the frequency SPSS would predict for a cell if the two variables being tested were completely unrelated to each other, calculated from the row and column totals of your contingency table.
Why SPSS Flags This in the Crosstabs Footnote
Run Analyze, then Descriptive Statistics, then Crosstabs, and tick the Chi-square box under Statistics. Below your output table, SPSS prints a footnote such as “2 cells (25.0%) have expected count less than 5.” This is SPSS doing its job honestly. It is warning you that the maths behind the chi square test starts to behave unreliably once too many expected counts get this small.
Cochran’s Rule: The Actual Statistical Standard
This is the chi square expected frequency rule everyone references but almost nobody explains properly. Chi square itself is a non-parametric test, meaning it does not assume your data follows a normal distribution, but it does assume your expected counts are large enough to behave reliably. The rule comes from William Cochran’s 1954 work on contingency tables, and it has two parts.
The 20% Rule (No More Than 1 in 5 Cells)
No more than 20% of your cells should have an expected count below 5. So in a 2×3 table with six cells, you can tolerate one cell dipping below 5 without major concern. Cross that 20% threshold and your p-value becomes genuinely unreliable, inflating your chance of a false positive result.
The Absolute Rule (No Cell Below 1)
Even if you stay under the 20% limit, no single cell should have an expected count below 1. This is the stricter companion rule, and SPSS minimum expected count less than 5 warnings often hide this second, harder problem underneath the first. One cell at 0.6 can distort your whole chi square statistic regardless of how the other cells look.
This Is One of Three Assumptions, Not the Only One
Expected count is the assumption everyone panics about, but chi square assumption violated SPSS problems can also come from two other places: your observations not being independent of each other, and your sample not being randomly drawn. Neither of those shows up as a neat SPSS footnote, so they get missed more often, even though committees do ask about them.
Is This Rule Still Considered Valid Today?
In my experience reviewing dissertations across UK and US universities, examiners still apply Cochran’s rule almost without exception, even in 2026. Some statisticians argue it is conservative for larger tables, but no supervisor I have worked with has ever accepted “the rule is outdated” as a defence in a viva or defence panel. Treat it as the standard until your specific methodology advisor tells you otherwise in writing.
How Serious Is Your Violation? (Severity Check)
Not every expected count less than 5 chi square warning carries the same weight. I grade these into two categories for every client I work with.
Yellow Flag: Minor Violation, Report With a Caveat
One cell, mildly below 5, in a table with six or more cells, and none of your cells sitting below 1. Report the chi square result as usual, add a one line footnote acknowledging the minor violation, and move on. This is genuinely common and genuinely fine.
Red Flag: You Need to Switch Methods
Too many cells expected count below 5, meaning you have crossed the 20% threshold, or any cell is sitting below 1, or your total sample size is under about 20 cases. At this point, reporting a standard chi square result without a fix is a real methodological weakness. This is where the three fixes below come in.
Fix 1: Combine or Collapse Categories
When Collapsing Makes Sense (and When It Doesn’t)
Combine categories chi square SPSS only when the merged groups still make logical sense to your research question. For example, “Strongly Disagree” and “Disagree” merging into “Disagree” on a Likert scale is defensible. Merging “Male” and “Female” into one category to fix your numbers is not defensible, because you destroy the variable you are actually testing.
How to Recode in SPSS Without Losing Meaning
- Go to Transform, then Recode into Different Variables
- Select your categorical variable and create a new variable name
- Define the old and new values, grouping sparse categories together
- Re-run your Crosstabs and Chi-square test on the new variable
- Check the footnote again to confirm the expected count problem is resolved
I have seen students collapse five response categories into three purely to pass the 20% rule, with no theoretical justification. Examiners notice this immediately, and it reads as data manipulation rather than genuine analysis. Only collapse categories you can defend on conceptual grounds, not statistical convenience alone.
Fix 2: Fisher’s Exact Test
Is It Only for 2×2 Tables? (Clearing the Myth)
Fisher exact test instead of chi square SPSS is widely believed to work only on 2×2 tables, and that belief is outdated. SPSS’s standard output does restrict the automatic Fisher’s Exact statistic to 2×2 tables, but the Exact Tests add-on module extends Fisher’s Exact to larger tables as well. IBM’s own SPSS documentation confirms the Exact Tests module is a separate licensed add-on, not part of the base package, which is why some students see the Exact button and others don’t. If your university licence includes Exact Tests, you are not limited to 2×2 at all.
For 2×2 tables specifically, some researchers also apply Yates’ continuity correction as a milder adjustment before jumping to Fisher’s Exact, though Fisher’s remains the more precise fix when expected counts are genuinely too low.
Step-by-Step: Running Fisher’s Exact in SPSS
- Go to Analyze, then Descriptive Statistics, then Crosstabs
- Add your two categorical variables into Row and Column boxes
- Click Statistics and tick Chi-square
- Click Exact (visible if the Exact Tests module is installed) and select Exact
- Run the analysis and read the Fisher’s Exact Test row in your output, not the Pearson Chi-Square row
Fix 3: Monte Carlo Exact Test
When to Use This Over Fisher’s
Monte Carlo exact test SPSS becomes your option when your table is too large for Fisher’s Exact to compute in reasonable time, which happens with bigger tables and larger samples. Monte Carlo gives you a simulated, highly accurate estimate of the exact p-value instead of the asymptotic chi square p-value that assumes large samples.
Step-by-Step: Enabling Monte Carlo in SPSS Crosstabs
- Follow the same Crosstabs path as above
- Click the Exact button in the Crosstabs dialog
- Select Monte Carlo instead of Exact
- Set your confidence interval, 99% is the conventional default
- Run the test and report the Monte Carlo significance value alongside your standard chi square statistic
Reading the Monte Carlo Output
Read the Monte Carlo significance figure exactly the way you read any p-value: below 0.05, your result is statistically significant, above 0.05, it is not. The difference is only in how that p-value was obtained. It is simulated across thousands of resamples rather than assumed from large-sample theory, which is precisely why it holds up better when your expected counts are small.
Other Options Worth Knowing About
Effect Size for Small Samples: Cramér’s V
Once you have a usable p-value, report Cramér’s V alongside it to show the strength of the relationship, not just whether one exists. One caution from experience: Cramér’s V can run slightly inflated in sparse, small-sample tables, so mention this as a limitation rather than presenting the figure as fully precise.
The G-Test: A Less Common Alternative
A small number of researchers use the G-test, also called the log-likelihood ratio test, instead of Pearson chi square when samples are small. It behaves similarly to chi square in most cases, but dissertation committees are far less familiar with it, so I would not default to it without checking with your supervisor first.
How to Report This in Your Dissertation
Examiners are not looking for perfection here, they are looking for evidence that you noticed the issue and handled it deliberately. The templates below follow APA 7th edition conventions for reporting statistical tests, which is the format most UK and US universities expect by default.
Methods Section Wording Template
“Prior to conducting the chi square test, expected cell frequencies were examined. As [X]% of cells had expected counts below 5, [categories were collapsed / Fisher’s Exact Test was applied / Monte Carlo simulation was used] in accordance with Cochran’s (1954) guidance.”
Results Section Wording Template
“A chi square test of independence was conducted. Given that [proportion] of cells had expected counts below 5, the Fisher’s Exact Test result is reported: p = [value]. This indicates [significant/non-significant association] between [variable 1] and [variable 2].”
Limitations Section Wording Template
“A limitation of this analysis is the small expected cell frequencies in [X] of the [Y] cells, which may affect the reliability of the chi square statistic. [Corrective method] was applied to address this, though results should be interpreted with appropriate caution given the sample constraints.”
A Short Case Study
I worked with a public health masters student in Delhi whose survey had a 4×3 table comparing age group against preferred vaccination information source. Three of her twelve cells showed expected counts below 5, and one sat at 0.8. Her first instinct was to drop the “60 plus” age group entirely because it had the fewest respondents.
I advised against that, since dropping an entire group changes her research question. Instead, we combined “60 plus” with “45 to 59” into a single “45 plus” category, which was conceptually defensible since both groups shared similar information seeking behaviour in her literature review. The revised table passed Cochran’s rule cleanly, and her examiner’s only comment was to ask her to explain the merge rationale in one sentence, which she had already written in her methodology chapter.
Can I Report Chi Square With Low Expected Counts? My Honest Take
After reviewing this exact issue across more dissertations than I can count, here is where I land. If your violation is minor, yellow flag territory, report the standard chi square result with a one line caveat and move on, no drama needed. If it is a red flag, pick the fix that fits your table size and sample constraints, not the fix that is easiest to Google.
A pattern I keep seeing across SPSS help sites: they explain Cochran’s rule correctly, then jump straight to “just use Fisher’s Exact Test” without mentioning that standard SPSS only runs Fisher’s automatically on 2×2 tables. Students without the Exact Tests module end up stuck, assuming they have done something wrong, when the real issue is a licensing gap nobody warned them about. A few guides also recommend dropping entire categories to solve the problem statistically, without considering whether that destroys your research question conceptually. Statistics is not just about passing a threshold, it is about keeping your analysis honest to what you actually set out to measure.
If you are still unsure which path applies to your specific table, a SPSS tutor session can walk through your actual output in fifteen minutes rather than guessing from a blog post. Before you even reach the analysis stage, running your numbers through a sample size calculator at the design phase can prevent this entire problem. The same logic of checking assumptions before trusting a p-value applies elsewhere too. It applies when you are running a Shapiro-Wilk normality test, checking VIF for multicollinearity, or deciding between Pearson and Spearman correlation. And if your chi square result feeds into a broader comparison involving post-hoc tests, the same care applies when choosing between Tukey HSD and Games-Howell.
FAQ
How many cells can have an expected count below 5 before it’s a real problem?
No more than 20% of your total cells should fall below 5, and no single cell should sit below 1. Cross either line and your chi square p-value becomes unreliable.
Is Fisher’s Exact Test only valid for 2×2 tables?
No. SPSS’s default output limits automatic Fisher’s Exact to 2×2 tables, but the Exact Tests add-on module extends it to larger tables as well.
What’s the minimum sample size needed for a chi square test?
There is no single fixed number, but most methodologists recommend at least 20 total cases, with enough spread that your smallest expected cell count clears 5. Check this at the design stage with a sample size calculator rather than after data collection.
Can I still report my chi square p-value if the assumption is technically violated?
Yes, if the violation is minor, one cell slightly below 5 in a larger table, report it with a brief caveat. For larger violations, switch to Fisher’s Exact or Monte Carlo instead.
What’s the difference between Fisher’s Exact and Monte Carlo methods in SPSS?
Fisher’s Exact calculates the precise exact p-value directly, suited to smaller tables. Monte Carlo simulates thousands of random samples to estimate the exact p-value, useful when the table is too large for direct exact calculation.
Does this rule apply to goodness-of-fit tests too, or just tests of independence?
Yes, Cochran’s rule applies to both chi square goodness-of-fit and independence tests, and most chi square assumptions dissertation chapters should mention this check regardless of which type you are running.
Written by Siddharth Gupta, 12-year dissertation and statistics consultant. Connect on LinkedIn.