What Is an Outlier? How to Identify and Handle Outliers

What Is an Outlier

An outlier, also called an unusual observation or an extreme value, is an observation whose value lies far from most of the data for the same variable or for a combination of several variables. For example, most respondents may rate a service from 2 to 5, but one record contains the value 99. Another example is a respondent whose income is much higher than that of the rest of the group. That record may result from a data-entry error, an invalid response, or a genuinely different case in the population.

You need to distinguish an outlier from a value that is simply high or low. A value of 5 on a 1 to 5 Likert scale is not an outlier just because it is at the upper end of the scale. Whether a value is an outlier depends on its position in the data distribution and on the purpose of your analysis. Common checks include a Boxplot, standardized Z-score, the interquartile range IQR, and Mahalanobis distance when you examine several variables at the same time.

For a single variable, IQR is calculated as the distance between the third quartile and the first quartile:

IQR = Q3 - Q1

An observation is commonly flagged as a suspected outlier if it is lower than Q1 - 1,5 × IQR or higher than Q3 + 1,5 × IQR. This is a screening rule that tells you what to investigate next. It is not an automatic instruction to delete the data row. When you read the result, also examine the median, standard deviation, and the respondent's context.

Why Outliers Matter in Quantitative Research

An outlier can change the mean, standard deviation, correlation coefficient, and regression coefficient. A few extreme observations can pull the regression line toward them, making a coefficient larger, smaller, or even change its sign. As a result, the p-value and your conclusion about a hypothesis may also change.

The presence of an outlier does not prove that the data are wrong. A customer with genuinely high spending is still part of the population if your study covers all customers. Deleting that respondent simply to make the output look cleaner changes the target population and may leave you with a difficult question during the defence.

Outliers usually lead to four practical decisions:

1. Keep the observation because it is valid and reflects the population.
2. Correct the data if you can identify an entry error, such as entering 55 instead of 5.
3. Remove the record if there is clear evidence that the respondent is invalid.
4. Run a sensitivity analysis by comparing results before and after handling the observation, so you can see whether the conclusion depends on a few extreme cases.

Save a copy of the original data before making any changes. Create additional variables or a separate cleaned data file, record the case ID, write down the reason for each decision, and save the result after every run. This gives you a clear explanation for any decrease in the number of observations shown in the Descriptive Statistics or Regression output.

Outliers can also point to questionnaire-quality problems. A respondent who selects the same option for every item, completes the questionnaire unusually quickly, or leaves many items blank may produce a questionable record. These signs still need to be assessed together with the screening rules stated in your research method. Do not classify a record based on only one sign.

What Outlier Level Is Acceptable

The question “how many outliers is good” can easily lead you in the wrong direction. There is no fixed outlier percentage that is automatically better when it is lower. Your goal is to identify observations that are statistically unusual and then decide whether each one is valid in the real-world context of the study.

The thresholds below are common screening points. Each row indicates a sign that needs checking. None of them automatically decides that you should delete the data.

The IQR rule is a quartile-based screening tool. For regression, also read about variance and the variance formula so you understand why an extreme value can increase dispersion sharply.

Do not set a target of “zero outliers.” In a real sample, some unusual observations may be completely normal. A defensible target is valid data, criteria stated in advance, and a reason for every handling decision.

How to Read Outliers in the Output

In SPSS, start with Analyze > Descriptive Statistics > Explore. Move the variable you want to check into Dependent List, select Plots, and tick Boxplots and Normality plots with tests if needed. The Descriptives table reports the mean, median, standard deviation, and quartiles. The Boxplot shows points that fall outside the whiskers.

SPSS commonly labels distant points with the case number. Trace that number back to the original data, identify the respondent, check the variable code, and confirm the entered value. A point outside the whisker is a signal for investigation. It does not prove that the record is wrong.

Illustrative output

The table below is illustrative output, not the result of a specific study. Suppose you check the variable TN1, which represents income in millions of VND, using Analyze > Descriptive Statistics > Explore.

In this example, the value 85 is an outlier under the IQR rule. Open Data View, find the correct record, and check the original questionnaire or data-entry source. If the questionnaire confirms that the respondent really had an income of 85 million VND, you can retain the observation and explain how you checked it. If the original file shows 8,5 but the value was entered as 85, correct it based on that evidence and record the change history.

With several independent variables, a separate Boxplot for each variable may miss a case that is not extreme on any single variable but is unusual when the variables are considered together. Mahalanobis distance is more suitable for that situation. State which variables you used, which criterion you applied, and how many records you handled. Do not take one distance value from the software and delete the record immediately without checking the original data.

How to Handle an Unacceptable Outlier

You can handle outliers in an order that moves from the least intervention to the most intervention. This preserves more information and gives you a clearer explanation for your decisions.

Check Entry and Coding Errors

First, compare the suspected value with the questionnaire, the Excel file, or the original collection source. Check the variable labels, coding range, and reverse-coded variables as well. A value of 99 may represent “no response,” but if you treat it as a valid number, the mean and regression results will be wrong. In SPSS, use Analyze > Descriptive Statistics > Frequencies to inspect the minimum, maximum, and frequency of each value.

Retain a Valid Observation

If the data are confirmed as genuine and the respondent belongs to the target population, retaining the observation is usually well supported. You can report the number of detected outliers, the checks you performed, and the result of the sensitivity analysis. The committee may ask why you did not remove the case. Your answer should refer to the scope of the population, not to whether the coefficient looked better after deletion.

Remove an Invalid Record

Remove a record only when you have a clear reason, such as a case outside the sampling criteria, an incorrect value that cannot be recovered, a response outside the scale, or too much missing data under the rule defined in advance. Record the case ID, the relevant variable, the reason for removal, and the remaining number of observations. Do not delete the case directly from the original file.

Transform or Limit Values

Log transformation, winsorization, or value replacement may be considered in some models. State the formula, the reason for using it, and its effect on the results. Do not reduce every large value to the same level simply to reduce skewness. If your undergraduate thesis does not require this technique, retaining the data and presenting a sensitivity analysis is often easier to defend.

After handling the data, rerun the main tables and compare the coefficients, p-value, R², or other relevant indicators. If the conclusion changes, report that change. If the conclusion does not change, still document the checking procedure rather than claiming that no outlier existed.

Distinguishing Outliers from Similar Data Points

An outlier describes how unusual an observation is. It differs from the following concepts:

• Extreme value is the largest or smallest value in the sample. An extreme value may also be an outlier, but not every maximum or minimum is an outlier.
• Standard deviation measures dispersion around the mean. A large standard deviation does not identify which record is an outlier.
• Variance is the squared standard deviation and also describes the dispersion of the entire variable. It does not replace checking individual observations.
• Skewness describes the direction and degree of asymmetry in the overall distribution. A right-skewed distribution may contain outliers, but skewness does not directly identify which record should be deleted.
• Leverage indicates that an observation has an independent-variable value far from the centre of the predictors in a regression model. High leverage does not necessarily mean that the model is strongly affected.
• Influential point is an observation that has a substantial effect on the model result. A point can be an outlier with little influence, or it can be less extreme on each individual variable but highly influential in regression.

Therefore, when your lecturer asks whether an outlier invalidates the model, connect your answer to specific diagnostics such as Cook's Distance, leverage, residuals, and the results from rerunning the analysis. Do not use the word outlier for every number that differs from the mean.

Common Errors

The most common error is deleting every point outside the Boxplot. A Boxplot helps you find points that need investigation, but the handling decision must be based on the original data and the scope of the sample.

The second error is checking for outliers only after deleting the “bad” rows. At that point, you no longer know what problems existed in the original data. Save the original file, create a cleaned copy, and export a list of the records you reviewed.

The third error is confusing missing codes with genuine values. Codes such as 0, 9, 99, or 999 must be interpreted using the codebook. If a Likert variable from 1 to 5 contains the value 99, that usually indicates a coding issue that must be handled before calculating the mean or running EFA.

The fourth error is looking only at the mean. The mean, median, and standard deviation give you an initial picture, but you need to combine them with a Boxplot, frequencies, and the questionnaire context. You should also check the Statistics topic so that you place data diagnosis correctly within the analysis process.

The final error is failing to report data removal. If N changes between the descriptive, correlation, and regression tables, the reader needs to know why. A short explanation with the before-and-after numbers is more credible than ignoring the change.

Frequently asked questions

What is an outlier in SPSS?

In SPSS, an outlier is an observation whose value differs substantially from most of the data. You can usually detect it through a Boxplot in Analyze > Descriptive Statistics > Explore, through descriptive statistics, or through model-diagnostic indicators.

SPSS may label the case number on the chart. Use that number to find the record in Data View, then check the original data before deciding whether to retain, correct, or remove it.

How many outliers are acceptable?

There is no fixed outlier percentage that is acceptable for every study. An outlier is a signal to investigate, not a measure that you must reduce to zero.

State the rule you used, the number of observations flagged, the reason for handling them, and the result of comparing the analysis before and after handling. A valid outlier can be retained if it belongs to the target population.

Should I delete outliers before running regression?

You should not delete them automatically before identifying the cause. Check entry errors, missing codes, sampling criteria, and the observation's influence on the model. If the record is valid, you may retain it and report a sensitivity analysis.

If the record is incorrect or does not meet the data criteria, you can remove it according to the rule stated in your method. Keep the original file so the decision can be checked later.

Why does the Boxplot show an outlier when the data are still correct?

A Boxplot uses the interquartile-range rule to flag points that lie far from most observations. The rule cannot tell whether a respondent genuinely belongs to an unusual group.

For example, a customer with high spending may be a valid case. The point on the Boxplot only tells you to check the record, the original questionnaire, and the population scope.

Can an outlier distort Cronbach's Alpha and EFA?

An outlier can affect the mean, dispersion, and relationships among variables, which may influence some scale-analysis results. The size of the effect depends on the data and the number of records, so it cannot be predicted for every file.

Check the data before running Cronbach's Alpha and EFA, record the suspicious cases, and compare the results when you retain them and when you handle them for a defensible reason. Do not remove an item merely because one respondent has an outlier.

Open a copy of your data file now, run Analyze > Descriptive Statistics > Explore, save the Boxplot, and create a list of records to verify before editing. If you want to run the complete analysis on your own .sav or .csv file, you can use DoThesis M4 data analysis.