To handle missing data in research, first diagnose why the data is missing (MCAR, MAR, or MNAR), then choose a method that matches that pattern — multiple imputation is the current best-practice default, while deleting cases (listwise deletion) should be a last resort. If you've ever opened your dataset and found scattered blank cells right before a deadline, you already know the panic: delete too much and you lose statistical power, guess wrong and you bias your entire result.
Key Takeaways
- Missing data falls into three types: MCAR (missing completely at random), MAR (missing at random, related to observed variables), and MNAR (missing not at random, related to the unobserved value itself).
- Multiple imputation is the recommended default for MCAR and MAR data because it preserves sample size and accounts for uncertainty, unlike single-value replacement.
- Listwise deletion is only safe when data is MCAR and the loss is small — typically under 5% of cases — otherwise it wastes power and can bias estimates.
- Never replace missing values with the mean as a default; mean imputation shrinks variance and distorts correlations.
- Always report your missing-data handling in the Methods section, including the percentage missing and the method used, per APA 7 transparency standards.
Why does missing data matter so much?
Missing data matters because it can quietly bias your results and shrink your statistical power without any error message warning you. If 20% of your participants skipped a question and those skippers differ systematically from responders, any conclusion you draw is built on a distorted subsample. Research in psychology and medicine routinely loses 10–30% of data points to dropout, skipped items, or equipment failure — so how you handle those gaps often matters as much as the analysis itself.
The mistake most non-statisticians make is treating missing data as a formatting nuisance rather than a threat to validity. Deleting every row with a blank feels tidy, but it can throw away half your sample and introduce bias you'll never detect from the output alone.
What are the three types of missing data (MCAR, MAR, MNAR)?
The three types of missing data are MCAR, MAR, and MNAR, and identifying which one you have determines every choice that follows.
- MCAR (Missing Completely At Random): The missingness is unrelated to any variable, observed or not. Example: a lab tablet crashed randomly and lost some responses. This is the "best" kind of missing — deletion won't bias results, only reduce power.
- MAR (Missing At Random): The missingness is related to other observed variables, but not to the missing value itself. Example: older participants are more likely to skip an income question, but among people of the same age, skipping is random. Most real-world data is MAR.
- MNAR (Missing Not At Random): The missingness depends on the unobserved value itself. Example: people with the highest incomes are the ones most likely to leave income blank. This is the hardest case and cannot be fixed by standard methods alone.
You can formally test whether data is MCAR using Little's MCAR test — a non-significant result (p > .05) supports the assumption that data is missing completely at random.
How much missing data is too much?
There is no universal cutoff, but a common rule of thumb is that under 5% missing on a variable is usually trivial, while over 10% requires careful handling and reporting. Above 40% missing on a single variable, many researchers consider dropping the variable entirely rather than imputing it, because you'd be inventing more data than you observed. The critical factor is not just the amount but the pattern: 3% missing that is MNAR can bias results more than 20% that is MCAR.
What are the main methods for handling missing data?
The main methods for handling missing data, from least to most defensible, are listwise deletion, single imputation (mean/median), regression imputation, and multiple imputation.
Method comparison table
| Method | How it works | Best when | Main risk |
|---|---|---|---|
| Listwise deletion | Drops any case with a missing value | Data is MCAR and <5% missing | Loses power; biases results if not MCAR |
| Pairwise deletion | Uses all available data per analysis | MCAR, correlation-based analyses | Inconsistent sample sizes across tests |
| Mean/median imputation | Fills blanks with the variable average | Almost never recommended | Shrinks variance, distorts correlations |
| Regression imputation | Predicts missing values from other variables | Quick single-dataset estimates | Overstates precision (no error term) |
| Multiple imputation (MI) | Creates several complete datasets and pools results | MCAR or MAR data | Computationally heavier; needs software |
For most graduate research, multiple imputation is the recommended default. It generates several plausible complete datasets (usually 5–20), runs your analysis on each, and pools the estimates — crucially preserving the uncertainty about what the missing values might have been. Tools like StatRyx can run multiple imputation and pool the results automatically, so you don't need to hand-code the pooling formulas.
A worked example: comparing deletion vs. imputation
Suppose you survey 200 participants on a stress scale and a sleep-quality scale, and 30 participants (15%) skipped the sleep question. You want the correlation between stress and sleep.
Option A — Listwise deletion: You drop the 30 incomplete cases, leaving n = 170. You find r = .28, p = .001. But if the skippers were the poorest sleepers (an MAR or MNAR pattern), you've removed exactly the people who would strengthen the relationship — biasing r toward zero.
Option B — Multiple imputation: You generate 10 imputed datasets, compute the correlation in each, and pool them. The pooled result is r = .34, 95% CI [.21, .46], p < .001. This estimate uses all 200 participants and reflects the real uncertainty in the imputed values — which is why the confidence interval is honestly a touch wider than a naive single-imputation CI would be.
The difference between r = .28 and r = .34 might change whether your effect clears a meaningful threshold — that's the practical cost of getting missing data wrong. If you're unsure which correlation to run in the first place, see our guide on choosing between Pearson and Spearman.
How do I report missing data in APA 7 format?
APA 7 requires you to report the amount of missing data, the likely mechanism, and the method used to handle it, all in the Methods section. A model sentence:
"Of the 200 participants, 30 (15%) had missing data on the sleep-quality measure. Little's MCAR test was non-significant, χ²(4) = 3.87, p = .424, supporting a missing-completely-at-random assumption. Missing values were addressed using multiple imputation (m = 10 imputed datasets), and estimates were pooled using Rubin's rules."
Notice the APA conventions: test statistics and p are italicised, the leading zero is dropped on p values below one (p = .424), and you state both the mechanism test and the handling method. Transparency here is what separates a defensible thesis chapter from one a reviewer flags.
Can StatRyx handle missing data automatically?
StatRyx detects missing data in your uploaded dataset, flags the percentage missing per variable, runs Little's MCAR test, and can apply multiple imputation before running your chosen analysis — then writes the whole thing up in APA 7 for you. Instead of wrestling with SPSS's Missing Value Analysis module or scripting mice in R, you get the mechanism test, the imputation, the pooled result, and the reportable sentence in one pass. That matters most for non-statisticians who know their data has gaps but aren't sure which method the gaps actually call for.
**Stop calculating this by