What Is Outlier Analysis? How to Read and Handle Outliers in SPSS
What Is Outlier Analysis
Outlier analysis is the process of identifying observations whose values differ clearly from most of the remaining data. In English, you will see the terms outlier, anomaly, and extreme value. A row of data should be checked when it is unusually high, unusually low, or contains a combination of responses that is very different from the general pattern in your sample.
Suppose you survey satisfaction using a Likert scale from 1 to 5. A respondent who selects 5 for every item is not automatically a statistical outlier. They may genuinely be very satisfied. By contrast, a respondent who enters an age of 222, or enters 99 for an item that only allows values from 1 to 5, needs to be checked immediately because the value may have been coded incorrectly.
The key point is that an outlier is not automatically incorrect data. Outlier analysis helps you answer three questions: is this observation unusual, what caused the unusual value, and do you have enough evidence to keep or remove it? You can also read outliers when you need to distinguish detection methods for different types of data.
In an SPSS file, checks usually rely on frequency tables, charts, boxplots, z-scores, multivariate distances, or model residuals. There is no single number that applies to every research topic. The appropriate threshold also depends on the variable being measured, sample size, data-collection method, and analysis objective.
Why Outlier Analysis Matters in Quantitative Research
An outlier can change the mean, standard deviation, correlation, and regression coefficient. A few extreme observations can also pull the regression line toward them. Your conclusion about the relationship between variables may therefore change substantially depending on whether you keep or remove those observations.
In scale analysis, unusual data can change Cronbach's Alpha or the corrected item-total correlation. In EFA, several highly different observations can affect the correlation matrix and factor structure. In regression, you also need to check whether an observation has excessive influence on the model, rather than looking only at the value of each individual variable.
Treat this as a data-quality check, not a search for reasons to delete rows that make the results look bad. If you remove a respondent only because the p-value is no longer attractive, that decision will be difficult to defend. If you remove the row because the respondent entered an incorrect code, answered outside the questionnaire range, or there is clear evidence of a collection error, your justification is stronger.
An outlier is also different from a moderator or mediator. A moderator changes the strength or direction of a relationship, while a mediator explains the mechanism linking an independent variable to a dependent variable. You can read moderator and mediator if you are confusing these concepts with an unusual observation.
Before checking for outliers, review the scale, data type, and coding. The article what is a scale can help you check how the items were measured and entered into SPSS.
What Outlier Value Is Acceptable
The question “what outlier value is acceptable” can lead to a mechanical approach. A large or small value alone is not enough to conclude that the data are wrong. For example, respondent income is often right-skewed in many surveys. A person whose income is higher than that of most respondents may be a valid observation and should not be deleted automatically.
Assess the observation according to the check you are using. The table below gives a practical decision framework, with each row identifying the relevant methodological or software source. The sources support the checking procedure, rather than claiming that every study must use the same threshold.
| Check | Sign to examine | Initial interpretation and treatment | Source |
|---|---|---|---|
| Value outside the valid range | Codes such as 0, 6, or 99 in a Likert scale from 1 to 5 | Check the original questionnaire and codebook, then correct the code or remove the row if there is evidence of an error | (Nguyễn Đình Thọ, 2011) |
| Boxplot | One or several points separated from the box and whiskers | Compare the point with the response form, and do not delete it only because it falls outside the chart | (Field, 2013) |
| Z-score | A standardized score that differs greatly from most of the sample | Treat this as a signal for further checking, and do not use one number alone to make the decision | (Field, 2013) |
| Multivariate distance | One row differs across several variables at the same time | Check the response pattern, measurement model, and effect on the next analysis | (Field, 2013) |
| Regression influence | One observation causes the coefficient or residual to change substantially | Compare models with and without the observation, then record the decision clearly | (Field, 2013) |
The rows in the table are checking guidance based on SPSS functions, not academic thresholds that require citation. When the methodological literature for your topic requires a specific level for an index, use the exact source agreed on by your supervisor or specified in your research design. When you have no clear basis, describing the observation as “requiring further checking” is safer than writing “acceptable” or “unacceptable.”
If you find an out-of-range code such as 99 in a Likert variable, this is usually a data error rather than a natural outlier. If the value remains within the valid range but is very different, check the original response and the observation's influence before deciding what to do.
How to Read Outliers in SPSS Output
Open your SPSS file and start with Analyze > Descriptive Statistics > Frequencies. Select the variables you need to check, click Statistics to view descriptive statistics, and select a chart if needed. For quantitative variables, Analyze > Descriptive Statistics > Explore allows you to view boxplots, a descriptive table, and checks for extreme values.
To examine outliers in regression, go to Analyze > Regression > Linear. Place the dependent variable and independent variables in the correct fields, click Save, and select suitable diagnostic values such as standardized residuals or Cook's distance if your analysis design requires them. Pay attention to the variable names SPSS creates in Data View so that you can match them to the correct data row.
The table below is illustrative output. The numbers only simulate the shape of the output and are not results from a real study.
| Case Number | Minimum | Maximum | Mean | Std. Deviation | Value to check |
|---|---|---|---|---|---|
| 1 to 118 | 1.00 | 5.00 | 3.62 | 0.71 | Most observations fall within the questionnaire range |
| 119 | 1.00 | 5.00 | 4.18 | 0.46 | The mean is high, but this is not enough to call it an outlier |
| 120 | 1.00 | 99.00 | 12.40 | 8.91 | The code may have been entered incorrectly, so check the original form |
| 121 | 1.00 | 5.00 | 1.20 | 0.44 | The responses are low, so check whether the case is valid |
In Case Processing Summary, check how many cases are valid and how many were excluded because of missing data. A row that disappears from an analysis because of missing data is not automatically an outlier. Record these as separate issues.
In Descriptives, review Minimum, Maximum, Mean, and Std. Deviation. If Minimum or Maximum falls outside the range allowed by the questionnaire, return to Variable View and the codebook first. If the range is valid but the deviation is large, open the boxplot and find the corresponding case number.
In a boxplot, a circle or marked point far from the box is a visual warning. You need to click it or match its case number, then open Data View at the relevant row. Check the complete response pattern for that row, not only the variable with the highest value.
If an outlier appears in regression, also review Coefficients, Residuals Statistics, and the diagnostic measures you saved. An observation with a large residual does not automatically invalidate the model. The question is whether it changes the conclusion, coefficient, or model fit to a meaningful extent.
How to Handle an Unacceptable Outlier
Work from the clearest cause to the most sensitive decision. This order helps you avoid deleting data simply because the output does not look good yet.
Step 1: Check the range and coding
Open Variable View and check Values, Missing, and Measure. Compare them with the original questionnaire. If a Likert variable should contain only values from 1 to 5 but Data View contains 0, 6, or 99, determine whether 99 is a missing code or an entry error before you recode it.
Keep an original copy of the data and work on a duplicate. Record the variable name, case number, old value, new value, and reason for the change. This record matters when your supervisor asks why the sample size after cleaning differs from the initial sample size.
Step 2: Compare the record with the questionnaire
If the data were entered from paper forms, review the relevant form. If you used an online form, check the response time, missing answers, and screening conditions you set. A respondent who selects 1 for every item may genuinely hold that view, or may have answered carelessly. You need to consider several signs together.
Signs such as an extremely short completion time, many skipped items, or the same repeated response pattern may support removing a record under a data-cleaning criterion defined in advance. If this criterion is not yet stated in the methods section, discuss it with your supervisor before applying it to many cases.
Step 3: Run the analysis with and without the observation
When you do not yet have a basis for deletion, create a temporary analysis version for comparison. Record the sample size, Mean, Std. Deviation, correlation, or regression coefficient in both versions. The aim is to learn whether the result is sensitive to the observation, not to select the version with the more attractive p-value.
If you keep the observation, you can state that it falls within the measurement range and that there is no evidence that the data are incorrect. If you remove it, state the reason, criterion, and number of rows removed. Do not write only “outliers were removed” without explaining the condition used.
Step 4: Transform data only with a methodological basis
Some studies may require a log transformation or treatment for skewness, but this is an analysis-design choice, not an automatic error-correction button. You need to understand how the transformation changes the interpretation of the results. For Likert data used in scale analysis, transforming individual items without a clear basis can make the measurement procedure difficult to explain.
Distinguishing Outliers from Missing Data and Miscoding
Missing data are cells with no response or a missing-data marker. An outlier is an existing value that differs from most observations. Miscoding occurs when a value is entered contrary to the coding rules, such as entering 9 instead of 4 or using 99 to mean “no response” while SPSS still treats 99 as a valid number.
All three situations can occur in the same file, but they require different treatment. For missing data, examine the missing-data mechanism and how SPSS excludes cases in each analysis. For miscoding, correct the value or declare it as missing after checking the codebook. For a valid outlier, you will usually keep it and explain the decision.
Do not confuse an unusual point with a moderator variable. An unusual point is a characteristic of one observation in the data. A moderator is a variable in the research model that is tested to determine whether the relationship between two variables changes at different levels of that moderator. These concepts belong to different stages of the analysis.
A univariate outlier is unusual on one variable. A multivariate outlier may have individual values that are not especially strange, while the combination of several values is unusual. For example, age and income may both be within valid ranges, but their combination may be very different from the sample. Looking at each column separately is therefore not enough when you run regression or a model with several variables.
Common Errors
The first error is deleting every point outside the boxplot. A boxplot is a warning tool. You still need to check the valid range, response form, and influence of the observation.
The second error is replacing an extreme value with the mean without recording the change. This alters the original data and may create an artificial observation. If the method requires this treatment, describe the rule applied, the number of affected rows, and the reason for choosing it.
The third error is entering a missing code but forgetting to declare it in SPSS. In that situation, the value 99 may enter Mean, Maximum, regression, or EFA. Check both the Missing column in Variable View and the frequency table after declaring the code.
The fourth error is reporting outliers without giving the case numbers. In your appendix or working file, keep the list of rows you checked. In the results chapter, you can briefly report the number of observations examined, the number of records removed, and the basis for removal.
You can adapt the following wording for your thesis: “The data were checked using minimum and maximum values, boxplots, and appropriate diagnostic measures. Unusual observations were compared with the scale range and collection information before a decision was made to retain or remove them, in order to avoid excluding data solely because the test results changed.”
Frequently asked questions
What is an outlier in SPSS
An outlier is an observation that differs clearly from most of the data. In SPSS, you can identify signs of one through Frequencies, Explore, boxplots, standardized scores, or diagnostic measures in Regression. The detection result is a basis for checking, not an instruction to delete the data.
Is an outlier necessarily incorrect data
Not necessarily. An outlier may be an incorrectly entered value, an out-of-range code, an invalid respondent, or a genuinely unusual observation in the population. Remove it only when you have evidence of a data error or a clearly defined cleaning criterion.
Do I need to remove outliers before running Cronbach's Alpha and EFA
Check data quality before running the analysis, but there is no rule that every outlier must be removed. Determine whether the observation results from a coding error, an invalid response, or a valid but rare value. If you remove it, rerun the analysis and record the change at each stage of the cleaning process.
SPSS reports an outlier, but I want to keep the data. What should I do
You can keep it if the observation is within the measurement range and there is no evidence of an error. Record the case number and reason for keeping it, and consider comparing the results before and after the check. This explanation is more defensible than keeping the data without discussing the unusual sign.
Can I replace an outlier with the mean
Some research designs allow extreme data to be treated through transformation or replacement, but you should not do this automatically for every case. Replacing a value with the Mean removes information from the observation and must be reported in the methods section. Before choosing this approach, check the cause, proportion of affected data, and requirements of the model.
Open your .sav file again, check Variable View, run Analyze > Descriptive Statistics > Explore, save the list of cases that need review, and record a decision for each row before running the next analysis. If you need to run and interpret the analysis directly on your own data, M4 SPSS and SmartPLS data analysis is the module for this step.