What Are Outliers? How to Identify and Handle Them

What Are Outliers?

Outliers are observations whose values differ clearly from most of the rest of the data. You may also see them described as unusual observations or abnormal values. For example, most respondents rate an item on a Likert scale from 1 to 5, but one row contains 55 because of an input error. That case needs to be checked immediately.

An outlier is not automatically incorrect data. A respondent whose income is much higher than the rest of the sample may still be a genuine respondent. Conversely, a questionnaire entered incorrectly, a wrongly reverse-coded item, or a respondent who selected the same option throughout the questionnaire can also produce an unusual observation.

Technically, you detect outliers by examining how far a value lies from the overall distribution. For a single variable, you usually examine the z-score, boxplot, or interquartile range. With multiple variables, an observation may not look unusual on any one variable but may become unusual when the variables are considered together, for example through the Mahalanobis distance.

The z-score is commonly written as follows:

z = (observed value - mean) / standard deviation

The standard deviation shows how dispersed the data are. You can also read what std means in SPSS if you have the Descriptive Statistics table open and are unsure what the Std. Deviation column means.

When you find an outlier, first return to the original data row, questionnaire, and coding rules. The software only flags an observation that deserves attention. Deciding whether it is an input error, a valid case, or data that should be removed requires the context of your study.

Why Outliers Matter in Quantitative Research

Outliers matter because some analyses are sensitive to very large or very small values. In linear regression, a few observations can pull the regression line toward themselves and change the coefficient, standard error, and p-value. Your conclusion about whether an independent variable affects the dependent variable may then change substantially.

In descriptive statistics, outliers can pull the mean upward or downward. Suppose you are analysing time spent using an application. Most respondents use it for 1 to 3 hours per day, but one record shows 300 hours. The mean no longer represents the usual group in your sample. The median and boxplot make this situation easier to see.

In assumption testing, an outlier may be related to a skewed distribution, unequal variance, or unusual residuals. For regression, examine the input values, residuals, and influence of each observation. Looking only at the Frequencies table and deleting rows with large numbers is difficult to defend when the committee asks how you cleaned the data.

Outliers can also result from questionnaire design errors. For example, you may use a scale from 1 to 5 but code 99 as “no answer” for one item, then import 99 as a valid value. In the output, 99 looks like an outlier, but it is actually a missing value that has not been declared.

In the analysis workflow, check outliers after reviewing coding ranges and missing values, and before finalising the regression model or other main inferential analyses. This is part of data cleaning, together with checking incomplete questionnaires, straight-lining responses, and reverse-coded items. If you are still designing your scale, read what a scale is to distinguish items, coding, and composite scores.

How Many Outliers Are Acceptable?

There is no single outlier threshold that applies to every type of data. A threshold can be used for initial screening, but it cannot automatically become a reason to remove an observation. You should state the detection method, the number of observations flagged, the reason for checking them, and the final decision.

The table below contains reference points commonly used in quantitative analysis. Each point is a supporting criterion, not an automatic data-deletion rule.

If you use a z-score in your thesis, report the absolute value you used to flag cases. Avoid writing that every observation above the threshold was deleted, because a point in the tail of the distribution may still be a valid value in the population.

The question “how many outliers are acceptable?” often focuses on the wrong issue. The goal is not to make the data contain no unusual points. The goal is to determine whether those points are valid, whether they have an inappropriate influence, and whether they change the conclusion.

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 they fit your checking objective. The Case Processing Summary table shows the number of valid observations and missing values, while the boxplot lets you see cases marked with circles or asterisks.

A circle on a boxplot usually represents a case outside the box area, while an asterisk represents a case farther away according to the rule used by SPSS. The number beside the point is the case number in the file, not the value of the variable. Return to Data View to see which respondent that row belongs to and whether other variables on the same row look unusual.

You can also check Descriptives for Minimum, Maximum, Mean, and Std. Deviation. If a scale only permits values from 1 to 5 but Maximum is 55, the cause is often data entry or the way missing values were declared. If the minimum and maximum are valid, the outlier may still be a genuine response.

The following is an illustrative output table, not the result of a specific study. The numbers are constructed to demonstrate how to read SPSS output.

In the illustrative table, case 24 is flagged because its z-score is large. You cannot yet conclude that it must be deleted. If the questionnaire records 95 million VND and the variable is measured in millions, the value may be valid. If the respondent entered 95 while the instructions required values in thousands, check the entry rules and the possibility of a unit error.

For regression, go to Analyze > Regression > Linear, select Statistics, and tick the relevant diagnostic options. Then go to Save to save standardized residuals or predicted values if needed. Tables such as Coefficients, Residuals Statistics, and residual plots help you determine whether one observation is changing the model substantially.

If you use SmartPLS, you generally will not handle outliers through a table with the same name as an SPSS table. Check the .csv file before importing it, the coding range, descriptive statistics for the variables, and the Outer Loadings and Path Coefficients results after running the model. SmartPLS is designed for PLS-SEM, but cleaning the data rows still needs to happen in the original file.

How to Handle Outliers That Do Not Meet the Criteria

Handle them in order, moving from the most certain cause to the decision with the greatest potential effect.

Correct an Input or Coding Error

If checking the questionnaire shows that a value was entered incorrectly, correct it to the original value and keep a record of the change. If 99, 999, or another special code represents a missing value, declare it correctly in SPSS under Variable View, in the Missing column, or convert it to missing before running the analysis. This is the easiest option to explain because you have evidence of the error.

Check the Measurement Unit and Reverse-Coded Items

An outlier may appear because one variable uses a different unit from the others. For a Likert scale, check whether the reverse-coded item has been recoded. An item that should range from 1 to 5 but contains a negative value or a value above 5 needs to be reviewed before calculating the mean score for the construct.

Keep It if It Is a Valid Observation

If the respondent exists, the response is complete, the value is realistic, and the case follows the collection rules, you have grounds to keep it. You can report that the data contained an unusual observation but that its validity was checked. You do not need to make every boxplot look neat by deleting respondents whose characteristics differ from the majority.

Run a Sensitivity Analysis

When a valid outlier has a strong influence, run the model with and without that observation to compare the direction of the effect, coefficient size, and statistical significance. This is a sensitivity analysis. It tells you whether the conclusion depends too heavily on one data row. State clearly which version is the main analysis and why.

Remove an Observation Only with a Methodological Basis

Remove a case only when you have a verifiable reason, such as an input error that cannot be recovered, a response outside the coding range, an invalid questionnaire under criteria set in advance, or an observation that does not belong to the target population. Record the case number, relevant variable, reason, and sample size before and after removal. Do not delete multiple rows simply because Cronbach's Alpha increases.

If removing one observation changes a result from significant to non-significant, or the reverse, present the situation honestly to your supervisor. A transparent procedure is more defensible than a polished table that you cannot explain.

Distinguishing Outliers from Unusual Points and Leverage

“Outlier” is commonly used for an observation that differs from the rest, but regression requires you to distinguish three related concepts. An unusual point on the outcome variable may have a large residual. Leverage concerns an independent-variable value that lies far from the centre of the data. Influence shows how much that observation actually changes the model result.

A case with an X value far from the rest may have high leverage but still lie close to the regression line, so it may not have strong influence. Conversely, a case with a large residual may affect the model when it is combined with high leverage. Looking at one z-score column is therefore insufficient when you are assessing regression.

Outliers are also different from missing values. A missing value is data that is absent or coded as no response. An outlier is an existing value that differs from the overall distribution. A code of 99 may look like an outlier in the output, but after reading the codebook you may determine that it is a missing value.

You should also distinguish outliers from moderators and mediators. A moderator changes the strength or direction of a relationship, while a mediator explains the transmission mechanism between variables. These are research-model concepts, not names for an unusual data point. You can read more about moderators, moderating, and mediators if you are confusing the terms while writing your hypotheses.

Common Errors When Handling Outliers

The first error is using one number as a deletion command. A high z-score only indicates that the observation needs review. You still have to check the value range, questionnaire, coding, and research objective.

The second error is deleting an outlier before saving the original file. Create a copy, name the file according to the processing round, and record the cases that changed. If your supervisor asks why the sample size decreased, you can open the tracking table instead of relying on memory.

The third error is looking only at the mean. Also check the minimum, maximum, median, boxplot, residuals, and related variables. A large value may be valid in the context of income, revenue, or number of uses even if it differs from most of the sample.

The fourth error is combining several causes into the sentence “the data were cleaned.” That sentence is too vague. In the methods chapter, state the checking criteria, number of rows removed, reason for removal, and final sample size.

The fifth error is reporting illustrative numbers as the result of your study. The output table in a guide only helps you understand the columns. When writing your thesis, replace it with the actual numbers from your .sav or .csv file, and save the syntax or processing steps so the analysis can be reproduced.

Frequently asked questions

What are outliers in SPSS?

In SPSS, outliers are cases flagged because their values lie far from most of the data under the checking method being used. You usually identify them through the boxplot in Explore, z-scores, or influence diagnostics in regression.

The number beside a boxplot point is usually the case number. Use it to find the corresponding row in Data View, then check the original data before deciding what to do.

How many outliers are acceptable?

There is no fixed outlier percentage that every thesis must meet. One study may contain valid unusual points, while another may contain an input code that needs to be corrected.

Report how you detected the cases and why you kept or removed them. Signals such as a large z-score, a position outside the boxplot, or Cook's Distance are prompts for checking, not automatic deletion criteria.

Do outliers always need to be removed?

No. If the observation belongs to the target population, contains a valid response, and represents a realistic value, you may keep it. Removing it only because it improves the mean, Alpha, or p-value is difficult to defend.

If the valid case has a strong influence, consider a sensitivity analysis and present the differences between the models. The final decision should be consistent with the research design and your supervisor's guidance.

What should you do about outliers in SPSS?

Check the coding, missing values, measurement units, reverse-coded items, and original questionnaire in sequence. Then rerun the descriptive statistics or regression on a copy of the data instead of editing the original file directly.

If you cannot find an error, keeping or removing the case should depend on its validity and influence. Record the case number, reason, and sample size before and after processing in your analysis log.

Can outliers reduce Cronbach's Alpha?

They can, if one or more unusual observations change the correlations between items. However, a low Alpha does not prove that an outlier is the cause.

Review Reliability Statistics, Item-Total Statistics, the data in each row, and the scale content together. Removing an item or case requires evidence from measurement and data quality, not just the Cronbach's Alpha if Item Deleted column.

When your file is open, create a copy, run Analyze > Descriptive Statistics > Explore for the main variables, record the flagged cases, and check each row before making changes. If you need to run the complete analysis on your own .sav or .csv file, M4 analysis from DoThesis supports SPSS and SmartPLS workflows, with results you can check and use when writing the next section.