Outlier Removal Guide: Conservative vs. Aggressive Choices

Outlier removal is not one universal rule. The real decision is how much evidence you want before deleting a row from your dataset. Conservative outlier removal means you delete only rows that look extreme under multiple checks. Aggressive outlier removal means you delete rows as soon as one reasonable method flags them.

That distinction matters because thesis data cleaning usually removes rows, not isolated cells. One participant can create several flagged cells across different variables, so a row-based review is clearer than counting one outlier cell after another.

If you want to be conservative, delete only where both methods agree. When IQR and z-score both flag the same row, the case is very likely to be a real outlier rather than a harmless edge case.

This approach is safer for thesis projects with small or medium sample sizes because every deleted row reduces power. It is also the better default when losing true observations would be more damaging than keeping a few noisy ones.

If you want cleaner data fast, you can delete rows whenever either IQR or z-score flags them. This aggressive outlier removal rule catches more unusual cases and can be justified when bad data points would distort the analysis more than a few lost rows.

Use this rule carefully. A row flagged by only one method may still be a legitimate extreme observation, especially in skewed data, small samples, or variables where high values are plausible in the real world.

The safest practical workflow is to start with rows flagged by both methods and then review the rest manually. That gives you a defensible first pass without pretending every flagged row is automatically wrong.

• Use this sequence in your thesis data cleaning:
• Step 1: Review rows flagged by both IQR and z-score first - these are your highest-priority cases.
• Step 2: Check rows flagged by only one method against the raw dataset and study context.
• Step 3: Ask whether the value is a typo, an impossible value, or a rare but plausible observation.
• Step 4: Record your decision rule before rerunning the analysis.

SPSS does not give you one magic outlier button, so run at least two checks before deleting rows. For z-scores, go to Analyze → Descriptive Statistics → Descriptives and save standardized values. Rows with |z| > 3.29 are extreme enough to inspect closely.

For an IQR-style visual check, use Graphs → Chart Builder → Boxplot. Points outside the whiskers are flagged as outliers and help you spot rows that need manual review. Use both screens together: z-scores quantify the deviation, while the box plot shows whether the value is isolated or part of a skewed distribution.

Always document three things: the rule you used, how many rows were deleted, and whether your main results changed after deletion. A clean methods sentence is: 'Rows flagged by both IQR and z-score were reviewed as potential outliers; 4 rows were removed before the main analysis.'

Do not be surprised if the number of detected outliers changes after deletion. Outlier thresholds are recalculated on the new dataset. When you remove one extreme row, the mean, standard deviation, quartiles, and IQR can change enough to reveal new rows that were not flagged before. That is why a dynamic outlier count is normal, not automatically a software bug.