You ran your Cronbach’s alpha in SPSS, and the number staring back at you is 0.62. Not great. You notice the “Alpha if Item Deleted” column shows that removing one or two items could push that number to 0.75 or higher. Now comes the question every researcher faces at some point: do you drop those items, or do you keep them?
Learning how to decide whether to drop an item from a scale after reliability analysis is one of the most practical skills in psychometric research. The decision affects your measurement quality, your validity evidence, and whether reviewers will accept your scale during peer review. Yet most methodology textbooks give only a brief treatment of this topic, leaving graduate students and researchers to figure it out from forum posts and scattered guides.
In this guide, I walk you through the decision-making process from start to finish. You will learn what Cronbach’s alpha really tells you, how to interpret the item-total statistics table, what thresholds to use for item-total correlation, and when removing an item actually helps versus when it creates new problems. I also cover the situations where you should never remove items at all, even when the statistics seem to support it.
If you are developing a new instrument, this information will help you build a tighter, more reliable measure. If you are adapting an existing scale, it will help you make defensible decisions you can report transparently in your manuscript. For real-world examples, you can explore a full scale development methodology from published research to see how these decisions play out in practice.
Table of Contents
Understanding Reliability Analysis and Internal Consistency
Reliability analysis answers a simple question: do the items on your scale measure the same underlying construct consistently? When you have a five-item Likert scale measuring job satisfaction, you expect all five items to move together. Someone who agrees with “I enjoy my work” should also tend to agree with “I look forward to coming to work.” Internal consistency reliability captures this expectation statistically.
Cronbach’s alpha is the most widely reported measure of internal consistency. It ranges from 0 to 1, with higher values indicating that your items hang together more tightly. The coefficient reflects the average inter-item correlation and the number of items on your scale. Adding more items generally increases alpha, while removing items generally decreases it, unless the removed item is poorly correlated with the rest.
Cronbach’s Alpha Interpretation: What the Numbers Mean
Most researchers use a common set of benchmarks to interpret Cronbach’s alpha. Nunnally’s classic guidelines, proposed in 1978, remain the most cited reference. Here is how to read your alpha values:
Below 0.60: Unacceptable for most research purposes. Your scale needs significant revision.
0.60 to 0.70: Acceptable for early-stage scale development or exploratory research. Questionable for established measures.
0.70 to 0.80: Good. This is the most common target range for published scales.
0.80 to 0.90: Very good. Most well-developed psychological and educational instruments fall here.
Above 0.90: Excellent, but potentially indicates redundancy. Items may be too similar, essentially asking the same question twice.
These thresholds are guidelines, not absolute rules. A short three-item scale may struggle to reach 0.70 even with well-written items, simply because alpha is sensitive to the number of items. A 30-item scale may hit 0.95 with items that are near-duplicates. Context matters, and you should always consider alpha alongside the nature and length of your instrument.
Why Alpha Alone Is Not Enough
Cronbach’s alpha gives you a single number summarizing your entire scale, but it cannot tell you which specific items are helping or hurting. Two scales with identical alpha values can have very different item-level performance. One might have five strong items all pulling their weight. The other might have four excellent items plus one terrible item dragging the average down.
This is where the item-total statistics table becomes essential. When you run reliability analysis in SPSS and request item-level diagnostics, the output includes several columns that break down each item’s contribution. These statistics are the real tools for deciding whether to drop an item from a scale after reliability analysis. Without them, you are working in the dark.
For a deeper look at how published researchers handle this process, the Mathematics Teaching Anxiety Scale validation study demonstrates the full workflow from initial item pool through reliability testing and final scale construction.
Key Statistics for Deciding Whether to Drop an Item
The item-total statistics table in SPSS produces four critical columns that you need to understand. Each one gives you a different angle on whether an item belongs in your scale or should be removed. Let me walk through them one at a time.
1. Corrected Item-Total Correlation
The corrected item-total correlation is arguably the single most useful diagnostic for item-level decisions. It measures the correlation between an individual item and the sum of all other items on the scale, excluding itself. The correction prevents artificial inflation that would occur if an item were correlated with a total that includes itself.
Here is how to interpret the values you will see in this column:
Below 0.20: Poor. The item is not measuring the same construct as the rest of the scale. Strong candidate for removal.
0.20 to 0.30: Marginal. The item has a weak relationship with the overall scale. Consider the content carefully before deciding.
0.30 to 0.40: Acceptable. The item contributes meaningfully to the construct. Generally safe to keep.
0.40 to 0.50: Good. The item is well-aligned with the overall scale.
Above 0.50: Very good. The item is strongly connected to the underlying construct.
Researchers on statistics forums frequently debate whether 0.20, 0.30, or some other value should serve as the cutoff for removal. The honest answer is that context drives the decision. For a new scale in early development, a stricter threshold of 0.30 makes sense. For an established scale where you want to preserve comparability with previous studies, you might only remove items below 0.20.
2. Alpha if Item Deleted
This column tells you what Cronbach’s alpha would become if you removed each individual item while keeping all others. It is one of the most commonly misunderstood statistics in reliability output, so let me be precise about what it means and how to use it.
The logic is straightforward. If the current alpha for your 10-item scale is 0.78, and the “Alpha if Item Deleted” value for Item 5 is 0.81, that means removing Item 5 would increase your overall alpha from 0.78 to 0.81. Item 5 is dragging the scale down. Conversely, if removing Item 5 would drop alpha to 0.74, then Item 5 is contributing positively and should be kept.
Here is the key interpretation rule: if the “Alpha if Item Deleted” value for an item is higher than your current overall alpha, that item is weakening the scale. The bigger the gap between the two numbers, the stronger the case for removal.
But there is a trap here that catches many researchers. A very small increase, like 0.78 to 0.79, is rarely worth removing an item for. You are gaining a trivial improvement in alpha while losing an item that may have content validity value. I recommend only considering removal when the increase is meaningful, typically 0.02 or more, and when the item also shows a low corrected item-total correlation.
3. Scale Mean if Item Deleted
This column shows what the mean total score would be if a particular item were removed. While less directly useful for the removal decision, it helps you understand the contribution of each item to the overall score. If removing an item would dramatically shift the scale mean, that item carries a lot of weight in your scoring system, and you should think carefully about the consequences of dropping it.
4. Scale Variance if Item Deleted
This column shows what the variance of total scores would be without each item. Large changes in variance after item removal can signal that the item is measuring something different from the rest of the scale. It is a secondary diagnostic that supports or contradicts what you see in the item-total correlation and Alpha if Item Deleted columns.
When all three signals align, when low item-total correlation, high Alpha if Item Deleted, and unusual variance shifts all point to the same item, you have a strong, defensible case for removal.
Decision Framework: How to Decide Whether to Drop an Item from a Scale After Reliability Analysis
Now I will give you a structured decision process. Follow these steps in order. Each step builds on the previous one, and you should never make a removal decision based on a single statistic alone.
Step 1: Check the Corrected Item-Total Correlation First
Start by scanning the corrected item-total correlation column. Any item below 0.20 is an immediate candidate for removal. These items are not measuring the same construct as the rest of your scale, and they are almost certainly pulling your alpha down.
Items between 0.20 and 0.30 require judgment. Look at the content of the item. Does it capture something conceptually different from the other items? Is the wording confusing or ambiguous? Could a reverse-coded item have been improperly scored? Answer these questions before deciding.
Step 2: Cross-Check with Alpha if Item Deleted
For every item you flagged in Step 1, check the “Alpha if Item Deleted” column. If removing the item would increase alpha, the evidence for removal grows stronger. If removing it would actually decrease alpha, then despite the low item-total correlation, the item is still contributing something to overall reliability.
This happens more often than you might think. Sometimes an item with a modest correlation of 0.25 still adds enough reliable variance to improve the overall coefficient. When alpha drops without the item, keep it, even if the correlation looks borderline.
Step 3: Consider the Content and Construct Definition
Statistics tell you part of the story, but construct validity matters just as much. Ask yourself whether the item taps into an important facet of your construct that no other item covers. If your scale measures “teaching effectiveness” and the item in question is the only one addressing classroom management, removing it might improve alpha while narrowing your construct coverage.
This is why content validity should always run alongside statistical analysis. A panel of subject-matter experts should have reviewed your items before you reached the reliability stage. If an expert flagged an item as essential during content validation, think twice before removing it based on a marginal statistical gain.
Step 4: Remove Items One at a Time
This is a critical rule that many researchers break. Never remove multiple items simultaneously based on a single reliability run. The item-total statistics are interdependent. Removing Item 3 changes the correlations and alpha-deletion values for Items 7 and 9. An item that looked marginal in the first run might look fine once a truly problematic item is removed.
The correct process is iterative. Remove the worst-performing item first. Re-run the reliability analysis. Re-examine the updated item-total statistics. Then decide whether to remove another item. Repeat this cycle until every remaining item meets your retention criteria or until further removal stops improving alpha.
Step 5: Know When to Stop
Knowing when to stop is just as important as knowing when to start. Here are clear stopping rules:
Stop when your alpha reaches an acceptable level (0.70 or higher for most applications).
Stop when removing additional items no longer increases alpha.
Stop when you have fewer than three items remaining on any subscale.
Stop when further removal would compromise the content coverage of your construct.
Stop when items being considered for removal have corrected item-total correlations above 0.30.
I have seen researchers strip their scales down to three items chasing an alpha of 0.90. By that point, the scale measures such a narrow slice of the construct that it loses practical and theoretical utility. Reliability is important, but it is not the only quality that matters.
When You Should NOT Remove Items
This topic gets almost no attention in most guides, yet it is critical. There are situations where removing items is the wrong decision regardless of what the statistics say.
If you are using a published, validated scale, you should generally not remove items without a very strong justification. Published scales have established norms, cut-off scores, and factor structures. Removing items makes your results incomparable to other studies using the same instrument. If your alpha is low on a published scale, the problem may be your translation, your sample, or your administration method rather than the items themselves.
If peer reviewers ask why you modified a validated instrument, “the alpha improved from 0.72 to 0.76” is not a compelling answer. You would need to demonstrate that the item is genuinely defective in your context, not just statistically suboptimal.
For researchers working with advanced methods beyond classical test theory, a Classical Test Theory and Item Response Theory comparison can provide richer diagnostic information that may change your item-level decisions.
Reporting Original and Revised Alpha Values
One of the biggest gaps in the published literature is how to report reliability when items have been removed. The answer is simple: report both values transparently. State the original alpha with all items included, describe which items were removed and why, then report the revised alpha after removal.
For example: “The initial 12-item scale showed a Cronbach’s alpha of 0.68. Examination of item-total statistics revealed that two items had corrected item-total correlations below 0.20 (Item 4: r = 0.14; Item 9: r = 0.11). After removing these items, the 10-item scale achieved an alpha of 0.79.”
This level of transparency lets readers evaluate your decisions and compare your scale performance with other studies. It also protects you against criticism during peer review.
Step-by-Step Guide: Running Reliability Analysis in SPSS
For readers who want a practical reference for running the analysis itself, here is the SPSS procedure that produces the item-total statistics you need for making removal decisions.
Running the Analysis
In SPSS, navigate to Analyze, then Scale, then Reliability Analysis. Move all items belonging to your scale into the Items box. Click the Statistics button and select Item, Scale, and Scale if item deleted. These three options give you the full item-total statistics table. Click Continue, then OK.
The output will include two key tables. The Reliability Statistics table shows your overall Cronbach’s alpha at the top. The Item-Total Statistics table shows one row per item with all four diagnostic columns discussed earlier.
Checking for Reverse-Coded Items
Before running reliability analysis, make sure all items are scored in the same direction. If your scale includes reverse-coded items and you have not reversed them, the item-total correlations will be negative, and alpha will be artificially low. This is the single most common reason for alarmingly poor reliability values.
To reverse-score an item in SPSS, use Transform, then Recode into Same Variables or Recode into Different Variables. For a 5-point Likert scale, recode 1 to 5, 2 to 4, 3 to 3, 4 to 2, and 5 to 1. After reverse-scoring, re-run the reliability analysis and check whether the problem persists.
I have seen cases where alpha jumped from 0.45 to 0.82 simply because one reverse-coded item was properly scored. Always verify your scoring direction before making any item removal decisions.
Running Subscale Reliability Separately
If your instrument has multiple subscales, run reliability analysis separately for each one. Combining subscales into a single reliability run will produce artificially low alpha values because items from different dimensions are not expected to correlate with each other. A 20-item scale with four five-item subscales should have four separate alpha calculations, not one combined analysis.
Using R, JASP, or Jamovi Instead of SPSS
The same principles apply if you are using open-source software. In R, the psych package provides the alpha() function, which produces item-total statistics equivalent to SPSS output, including the “Alpha if Item Deleted” column (labeled as “raw.alpha” in the drop output). JASP and Jamovi both offer point-and-click reliability analysis with the same diagnostic options. The thresholds and decision rules are identical regardless of which software you use.
Common Pitfalls and Warnings When Removing Scale Items
Now let me address the most common mistakes researchers make during this process. These insights come from years of forum discussions, peer review experiences, and real research scenarios shared by graduate students and practitioners.
Pitfall 1: Removing Items to Chase a Higher Alpha
The temptation to keep pruning items until alpha reaches an impressive number is strong, especially when reviewers or advisors emphasize reliability benchmarks. But alpha inflation through item removal can mask deeper problems with your scale. If you remove five items from a 12-item scale to get alpha from 0.65 to 0.85, you may have stripped away construct breadth and narrowed your measure to a single narrow dimension.
Always ask yourself: does the shorter scale still measure what I intended to measure? If the answer is uncertain, you have gone too far.
Pitfall 2: Ignoring Sample Size Effects
Cronbach’s alpha is sensitive to sample size, and so are item-total correlations. With a very small sample of 30 participants, your correlations will be unstable and your alpha estimates will have wide confidence intervals. An item that looks problematic with n = 30 might look fine with n = 300.
As a general rule, you need at least 100 participants for stable reliability estimates, and ideally 300 or more for scale development work. If your sample is below 50, interpret all item statistics cautiously and avoid removing items unless the evidence is overwhelming.
Pitfall 3: Removing Items Without Re-Examining Validity
Every time you remove an item, you change the scale. The factor structure may shift. The content validity may weaken. The relationship between your scale and external criteria may change. After finalizing your item set, you should re-run any validity analyses, including exploratory or confirmatory factor analysis, to confirm that your shortened scale still holds together conceptually.
For advanced psychometric approaches that go beyond alpha, advanced Rasch analysis procedures can provide item-level fit statistics that complement or improve upon classical item-total correlation diagnostics.
Pitfall 4: Not Reporting Changes Transparently
Some researchers remove items and only report the final alpha value, as if the original scale never existed. This is poor research practice and can be flagged during peer review. Always document what you started with, what you removed, and what you ended with. Reviewers and future researchers need this information to evaluate and replicate your work.
Pitfall 5: Treating Alpha as the Only Quality Indicator
Reliability is one pillar of measurement quality. Validity is the other. A scale can be highly reliable and still not measure what it claims to measure. Always pair your reliability analysis with construct validity evidence, including factor structure, convergent validity, discriminant validity, and criterion-related validity. Researchers interested in how validity evidence supports scale development can explore authentic assessment research methodology for applied examples.
Summary of Decision Criteria
Let me consolidate the key thresholds and rules into a single reference you can use when evaluating your own reliability output.
For corrected item-total correlation: items below 0.20 should almost always be removed. Items between 0.20 and 0.30 warrant a closer look at content and wording. Items above 0.30 generally should be kept.
For Alpha if Item Deleted: if removing an item would increase alpha by more than 0.02 and the item also has a low corrected item-total correlation, remove it. If the increase is less than 0.01, the gain is too small to justify the content loss.
For iteration: always remove items one at a time and re-run the analysis after each removal. Never batch-remove items based on a single run.
For published scales: think twice before removing any item. Low alpha on a validated instrument may indicate a problem with your sample or administration rather than the scale itself.
For reporting: always document the original alpha, the items removed, the reasons for removal, and the revised alpha. Transparency protects you during review and helps future researchers.
FAQs
How do you determine the reliability of a scale?
You determine scale reliability by calculating Cronbach’s alpha using reliability analysis in SPSS, R, or similar software. An alpha of 0.70 or higher is generally considered acceptable for research purposes, 0.80 or higher is good, and values above 0.90 indicate excellent internal consistency but may suggest item redundancy.
How do you interpret Cronbach’s alpha if item deleted?
The Alpha if Item Deleted column shows what your overall Cronbach’s alpha would become if you removed that specific item. If the value is higher than your current alpha, the item is weakening the scale and is a candidate for removal. If it is lower, the item is contributing positively and should be kept.
What is a good item-total correlation?
A corrected item-total correlation of 0.30 or higher is generally considered good, indicating the item measures the same construct as the overall scale. Values between 0.20 and 0.30 are marginal and require judgment. Values below 0.20 suggest the item is poorly related to the rest of the scale and should be considered for removal.
What is the acceptable range for corrected item-total correlation?
The acceptable range for corrected item-total correlation is typically 0.30 to 0.50 or higher. Values below 0.20 are considered unacceptable and usually warrant item removal. Values between 0.20 and 0.30 are borderline and should be evaluated alongside the Alpha if Item Deleted statistic and content considerations before deciding.
How can you improve the reliability of a scale?
You can improve scale reliability by removing poorly performing items with low item-total correlations, ensuring reverse-coded items are properly scored, increasing the number of well-written items, refining ambiguous wording, and collecting data from a larger and more representative sample to stabilize your reliability estimates.
When should you NOT remove items from a validated scale?
You should avoid removing items from published, validated scales because doing so makes your results incomparable to other studies and disrupts established factor structures and norms. Only remove items if there is strong evidence the item is defective in your specific context, such as a translation error or cultural mismatch, and report the change transparently.
Should you report both original and revised Cronbach’s alpha?
Yes, always report both values. Document the original alpha with all items, identify which items were removed and why, then report the revised alpha after removal. This transparency allows reviewers and future researchers to evaluate your decisions and compare your results with other studies using the same instrument.
Conclusion
Deciding whether to drop an item from a scale after reliability analysis comes down to examining multiple statistics together rather than relying on any single number. The corrected item-total correlation tells you whether each item measures the same construct as the rest of the scale. The Alpha if Item Deleted column tells you whether removing a specific item would improve overall reliability. Content validity tells you whether the item captures something irreplaceable in your construct definition.
The strongest item-removal decisions happen when all three signals align: the item-total correlation is below 0.20, removing the item increases alpha meaningfully, and the content is not essential to construct coverage. When the signals conflict, lean toward keeping the item and investigate the root cause of the weak performance before making permanent changes to your instrument.
Remember that reliability is a means to an end, not the goal itself. A reliable scale that no longer measures the full construct is not an improvement. Use the thresholds and decision process in this guide as a framework, but always apply your own judgment as a researcher who understands what your scale is supposed to capture and why each item was written in the first place.