Written and maintained by CASRAI Editorial Board
Published
On this page
Complete case analysis (CCA) is a way of handling missing data in which you analyse only the participants who have a recorded value for every variable the analysis needs, and drop everyone else. It is also called listwise deletion. It is valid in some situations and biased in others, and it always costs sample size, so you have to decide whether the reasons for the missing data make it safe and then report how many cases you lost.
Definition: what is complete case analysis?
In the words of Sterne and colleagues in the BMJ, complete cases are “those individuals who have no missing data in any of the variables required for that analysis.” Hughes and colleagues define CCA as an analysis restricted to individuals with complete information on all variables of the main analysis. The set of complete cases therefore changes with the model: a correlation between two variables uses everyone who has both, while a regression with eight predictors uses only people who have all nine variables (eight predictors plus the outcome).
The terms complete-case analysis, listwise deletion and casewise deletion refer to the same procedure. The alternative that keeps partial cases is pairwise (available-case) deletion, covered in the listwise vs pairwise deletion comparison. Where the data should be filled in rather than dropped, see the multiple imputation protocol.
Formula: retention and the complete-case estimate
CCA has no single test statistic. Two quantities describe it. The first is the share of the sample you keep:
retention = ncc / n
where n is the number of participants in the full sample and ncc is the number with no missing value on any of the p variables in the model. The second is the estimate computed on those cases only. For a mean of variable y:
ȳcc = (1 / ncc) × Σi∈C yi
where C is the set of complete cases, i indexes participants, yi is the value for participant i, and Σi∈C sums over complete cases only. Regression coefficients, odds ratios and other estimates are handled the same way: fit the model to the rows in C.
When is complete case analysis unbiased?
Validity depends on why the data are missing, not simply on how much is missing. The standard classification is missing completely at random (MCAR: no systematic differences between observed and missing values), missing at random (MAR: any systematic differences can be explained by the observed data) and missing not at random (MNAR: neither). For diagnosing which applies, see MCAR, MAR and MNAR: diagnosing your missing data.
The two sources read for this page give more specific rules than “CCA needs MCAR”:
- Sterne and colleagues state that when data are missing only in an outcome measured once per individual, CCA is not biased, provided every variable associated with the outcome being missing is included as a covariate.
- For missing predictors, the same paper states CCA does not cause bias if the reasons for the missing data are unrelated to the outcome.
- Hughes and colleagues report that for most regression models CCA gives unbiased results when the chance of being a complete case does not depend on the outcome after taking the covariates into consideration, which includes some situations where data are missing not at random.
The practical consequence, noted in the Hughes abstract, is that there are situations in which CCA is unbiased while multiple imputation assuming MAR is biased. CCA is not automatically the weaker choice. In the other direction, STROBE guidance says CCA can be biased and is always inefficient, because bias arises when people with missing data are not typical of the whole sample and inefficiency arises from the reduced sample size.
Worked example 1: how quickly cases are lost
This is an illustrative composite, not real study data. Suppose a study enrols n = 200 participants and the analysis needs five variables with these proportions missing: age 5%, BMI 8%, smoking status 10%, income 15% and the outcome 4%. If the gaps fall independently across variables (an assumption made only to show the arithmetic), the expected share with no gap at all is the product of the five complete proportions:
| Variable added to the model | Proportion observed | Cumulative retention | Expected ncc (of 200) |
|---|---|---|---|
| Age | 0.95 | 0.95 | 190.0 |
| + BMI | 0.92 | 0.95 × 0.92 = 0.8740 | 174.8 |
| + Smoking status | 0.90 | 0.8740 × 0.90 = 0.7866 | 157.3 |
| + Income | 0.85 | 0.7866 × 0.85 = 0.6686 | 133.7 |
| + Outcome | 0.96 | 0.6686 × 0.96 = 0.6419 | 128.4 |
No single variable is more than 15% missing, yet the model keeps about 64% of participants. This is the cumulative effect Sterne and colleagues describe: exclusion of a substantial proportion of the original sample and a substantial loss of precision and power.
Worked example 2: bias when missingness depends on the outcome
Also illustrative. Ten participants have outcome scores 12, 14, 15, 17, 18, 20, 22, 25, 28 and 30. The full-sample mean is 201 / 10 = 20.1. Suppose the three highest scores (25, 28, 30) are never recorded, because high scorers drop out. The complete cases are the other seven, with sum 118 and mean 118 / 7 = 16.86, which is 3.24 below the true mean. Here the chance of being a complete case depends on the outcome itself, the condition Hughes and colleagues identify as the one that CCA does not tolerate. If missingness had instead followed a variable such as age that you adjusted for, the rule above would apply and the regression estimate could remain unbiased.
A published case: 115 patients, 82 complete
Sterne and colleagues cite a cost-effectiveness analysis in which costs were obtained for 115 patients but were complete for only 82, a retention of 82 / 115 = 71.3%. The complete case analysis implied chemotherapy was not cost effective (mean net monetary benefit −£3346), while the multiple imputation analysis implied it was (£1186), although confidence intervals were wide. Sterne and colleagues use it to show how the choice of method can change the conclusion.
Interpretation: reading a complete case result
| What you see | How to read it | Source |
|---|---|---|
| ncc is much smaller than n | Precision and power fall; wide confidence intervals are expected even if the estimate is unbiased | Sterne et al. 2009 |
| Only the outcome (measured once) is missing | CCA is not biased if the predictors of missingness are covariates in the model | Sterne et al. 2009 |
| Missing predictor values, reasons unrelated to the outcome | No bias from CCA; specialist methods may recover precision | Sterne et al. 2009 |
| Complete-case and multiple-imputation results differ | Compare them and discuss the difference in the paper | STROBE item 12c |
| Concern that data are MNAR | Cannot be settled from observed data; use sensitivity analyses | Sterne et al. 2009 |
Complete case analysis versus multiple imputation
Multiple imputation is not a default upgrade. Hughes and colleagues report that its efficiency gains over CCA are greatest when there are small amounts of missing data on many variables, or when auxiliary variables (not in the main analysis) carry information about the missing values. That matches the pattern in worked example 1: many variables with modest gaps each, where CCA loses a third of the sample. When the gaps are concentrated in the outcome and the covariates explain the missingness, the Sterne and Hughes rules above say CCA can already be unbiased. Pooling imputed results follows Rubin’s rules. Hughes and colleagues also state that transparent reporting of how missing data could affect the main results is crucial, whichever method you choose.
Assumptions and common mistakes
- Treating “less than X% missing” as a safe rule. The sources read here tie validity to the missingness mechanism, not to a percentage. A small share of outcome-dependent loss can still bias an estimate.
- Letting software decide silently. Many modelling functions drop incomplete rows by default, so the n in your model can be lower than your sample n. Always print the analysed n and compare it with the enrolled n.
- Reporting one n for the whole paper. Each model uses its own complete cases, so n varies by analysis. STROBE item 12c asks authors to explain how missing data were addressed.
- Assuming MAR can be tested. Sterne and colleagues note that it is not possible to distinguish MAR from MNAR using observed data, so the choice rests on subject-matter reasoning plus sensitivity analysis.
- Ignoring attrition. In longitudinal designs, dropout that relates to the outcome is the usual route to bias; see attrition bias.
How to report complete case analysis (APA style example)
State the enrolled n, the analysed n and percentage, the method, and the missingness handling. A model sentence, using the illustrative numbers from example 1:
Of the 200 participants enrolled, 128 (64.0%) had complete data on all five variables and were included in the regression analysis; 72 (36.0%) were excluded listwise. Missing values were most frequent for income (15%). Estimates from multiple imputation are reported in the Supplement and did not change the conclusions.
The last sentence is only appropriate if you ran that comparison. The STROBE checklist item 12c asks authors to explain how missing data were addressed, and its guidance says results from other approaches should be compared with complete case analyses and important differences discussed. Format N, n and percentages according to your journal’s APA edition.
How to calculate complete case analysis in Excel, R or Python
Excel
Add a helper column beside the data in A2:E201. In F2 enter =IF(COUNTBLANK(A2:E2)=0,1,0) and fill down. Then =SUM(F2:F201) gives ncc and =SUM(F2:F201)/ROWS(F2:F201) gives retention. Filter column F to 1 to analyse complete cases.
R
The base R function complete.cases() returns a logical vector indicating which cases have no missing values. The tidyr function drop_na() drops rows where any specified column is missing.
cc = df[complete.cases(df), ]nrow(cc) / nrow(df) # retention# tidyverse alternativecc = tidyr::drop_na(df, age, bmi, smoker, income, outcome)colSums(is.na(df)) # missing values per variable
Python
The pandas method DataFrame.dropna() removes rows with missing values; restrict it to the model’s variables with subset.
vars_ = ["age", "bmi", "smoker", "income", "outcome"]cc = df.dropna(subset=vars_)print(len(cc) / len(df)) # retentionprint(df[vars_].isna().sum()) # missing per variable
Whichever tool you use, record the analysed n for every model and keep the code that built the complete-case subset, so the exclusion count in the manuscript can be reproduced exactly.
Sources
- Sterne et al., BMJ 2009;338:b2393: Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls, checked 2026-11-19
- Hughes et al., Int J Epidemiol 2019;48(4):1294-1304: Accounting for missing data in statistical analyses: multiple imputation is not always the answer, checked 2026-11-19
- EQUATOR Network: STROBE item 12c, Statistical methods, missing data, checked 2026-11-19
- R documentation: complete.cases, Find Complete Cases, checked 2026-11-19
- tidyr documentation: drop_na, Drop rows containing missing values, checked 2026-11-19
- pandas documentation: DataFrame.dropna, checked 2026-11-19
Data: When complete case analysis is biased or unbiased, by missingness scenario
Compiled from Sterne et al. (BMJ 2009) and Hughes et al. (Int J Epidemiol 2019): each row states a missing-data scenario and what those authors report about the bias of complete case analysis (CCA) in it.
| Missingness scenario | CCA result as reported | Source |
|---|---|---|
| Missing only in an outcome measured once per person, all variables linked to missingness included as covariates | Not biased | Sterne et al. 2009 |
| Missing in predictor variables, reasons for missingness unrelated to the outcome | No bias; precision and power may be lost | Sterne et al. 2009 |
| Chance of being a complete case does not depend on the outcome given the covariates (most regression models) | Unbiased, including some MNAR situations | Hughes et al. 2019 |
| Missingness depends on no variable (MCAR), exposure coefficient | Unbiased in linear and logistic regression | Hughes et al. 2019 |
| Missingness depends on the outcome, exposure coefficient in linear regression | Biased | Hughes et al. 2019 |
Sources
- Sterne et al., BMJ 2009;338:b2393 (accessed 2026-11-19)
- Hughes et al., Int J Epidemiol 2019;48(4):1294-1304 (accessed 2026-11-19)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







