What to Do When Your Data Is Suspected of Being Fabricated
The situation you are facing
You submit your results tables and your supervisor asks, “Did you actually collect this data yourself?” The question may arise because too many responses follow exactly the same pattern, completion times are unusually short, the response rate looks suspiciously clean, or you cannot open the original file when asked to check it.
Your data may still be genuine even if you no longer have a collection history, cannot tell where each row came from, or edited the file directly in Excel without saving an earlier version. In that situation, the issue is traceability and your ability to explain the process. You cannot conclude that you committed fraud from that problem alone.
First, stop rerunning SPSS to search for a “better-looking” table. Create a separate folder and copy the .sav file, .csv file, questionnaire, form link, screenshots of the collection process, and all syntax files you used. Keep the original unchanged, and name each copy with the date and time so that every later change can be traced.
If the committee’s question concerns sample size or sampling, you can also review what to do when the committee asks about sample size and SPSS sample-size question from the committee. These are related issues, but they do not mean that the data are fabricated.
Common causes
A suspicion usually comes from a specific sign. Write down that sign before you explain anything. Do not simply answer, “I really collected it,” if you cannot identify the corresponding file, dates, sample source, and cleaning process.
| Observable sign | Possible cause | How to check it |
|---|---|---|
| Many rows have the same response pattern from beginning to end | Respondents selected one level throughout, data were duplicated, or you entered the same data repeatedly | Use Excel or SPSS to find rows with identical values across all items, then check completion times if the form recorded them |
| Completion times are almost identical | Data were entered manually, the form recorded time incorrectly, or data were combined from several sources | Open the timestamp column, inspect the time distribution, and compare it with the collection schedule |
| Every Cronbach's Alpha, EFA result, or model looks unusually good | The sample is too homogeneous, items were edited, or difficult rows were removed unintentionally | Open the original data file, review the row-removal history, and rerun the analysis from the uncleaned version |
| Results tables exist but the original data file does not | You saved only the output, lost the file, or someone else ran the analysis | Search for .sav, .csv, syntax, log files, SmartPLS files, and backups on your computer or cloud storage |
| The number of observations changes between tables | Missing values or outliers were removed differently at different stages | Record N for every analysis, check SPSS filters, and list the cases that were removed |
| The results differ from a reference article | The context, sample, scale, or processing method is different | Compare the research designs instead of changing your data to match the original article |
A commonly overlooked cause is the data-entry process. If you enter 250 questionnaires into Excel and copy a data range into another file, shifted columns or repeated rows can produce unusual output. Check the variable codes, the direction of reverse-coded items, missing values, and column order before discussing data integrity.
You should also distinguish “fabricated data” from “data that do not meet the expected criteria.” A low Alpha, low KMO, nonsignificant model, or small R² can occur with genuine data. An unattractive result is not evidence of fabricated data, just as an attractive result is not proof that the data are genuine.
How to handle the issue in priority order
Preserve the original data
Create three versions: an untouched original, a cleaned version with a log, and an analysis version for SPSS or SmartPLS. Avoid vague names such as final_final2.sav. You can use the structure 01_raw, 02_cleaning, and 03_analysis, together with the update date.
In the log, record the date, action taken, number of rows before and after, reason for removing rows, variables processed, and person who performed the action. If you have never kept a log, start now and state accurately that it was created retrospectively for traceability. Do not present it as though it had existed from the beginning.
Check duplicates and unusual response patterns
In SPSS, open Data > Identify Duplicate Cases to identify rows that are duplicates across a selected group of variables. To inspect each respondent’s response pattern, select the main items and review rows with the same response sequence. A matching sequence is not enough to conclude that someone cheated, because a respondent may genuinely choose the same level for several statements.
Combine this check with completion time, screening information, respondent codes if available, and the distribution channel. Flag suspicious rows for separate checking. Do not delete them immediately just because they make the results less attractive.
Compare the collection file with the analysis file
If you used Google Forms, download the response spreadsheet again at the latest available time and compare the row count, column names, timestamps, and responses. If you used fillform.info, DoThesis’s form system, check the exported response file and the corresponding collection history. When data came from paper forms or messages, keep the questionnaire codes and the data-entry rules.
For .csv files, open the file in a text editor and inspect the delimiter, header row, and number of fields in each row. For .sav files, check Variable View, Value Labels, Missing, and Measure. A file can have the correct variable names but incorrect value labels, which can lead you to explain the results incorrectly.
Rerun the analysis from a controlled clean file
Open the clean file in SPSS and check Analyze > Descriptive Statistics > Frequencies for categorical variables and scale items. Then check missing values, minimum, maximum, mean, and standard deviation. If a Likert variable coded from 1 to 5 contains 0 or 9, determine whether those values represent missing codes or entry errors before continuing.
In SmartPLS, check that the .csv file has one header row, unique variable codes, and numeric columns recognized correctly. Rerun the model from the measurement model through the structural model. If the results change, save both the old and new output and record that the change resulted from data cleaning or correction of a data error. Do not choose the table with the more convenient p-value.
Compare against output that can be reopened
Save the output file, syntax, or screenshots of each step. The example below is illustrative output, not the result of a real study. The table names and columns follow the way SPSS commonly displays them, so you can see which evidence to preserve.
| SPSS table | Column or statistic to save | Illustrative output | What to compare |
|---|---|---|---|
| Case Processing Summary | Valid, Excluded, Total | Valid 198, Excluded 12, Total 210 | The list of 12 excluded cases and the reason for exclusion |
| Reliability Statistics | Cronbach's Alpha, N of Items | 0.812, 5 | Whether the data file contains exactly 5 items |
| Item-Total Statistics | Corrected Item-Total Correlation | 0.421 to 0.691 | Which item was removed and at which round |
| KMO and Bartlett's Test | KMO, Sig. | 0.784, 0.000 | The variables entered into EFA and the corresponding matrix file |
| Coefficients | B, Beta, t, Sig., VIF | Beta 0.286, Sig. 0.012, VIF 1.84 | The model, independent variables, and cases used |
This table does not prove that the data are genuine when viewed alone. Its value is that you can reopen the file and reproduce the same result from the same data version. If your supervisor changes a variable or asks you to remove a group of cases, create a new version instead of overwriting the old output.
Assess the criteria you actually used
When explaining the analysis, state the exact criteria and their sources. For example, Cronbach's Alpha from 0.7 upward is commonly accepted according to (Nunnally, 1978), while corrected item-total correlation from 0.3 upward follows (Nunnally et al., 1994). For EFA, KMO from 0.5 upward and a statistically significant Bartlett’s test can be referenced to (Kaiser, 1974). Factor loading from 0.5 upward and total variance explained from 50% upward can be supported by (Hair et al., 2010).
| Point to explain | Reference level | Canonical source |
|---|---|---|
| Cronbach's Alpha | From 0.70 | (Nunnally, 1978) |
| Corrected Item-Total Correlation | From 0.30 | (Nunnally et al., 1994) |
| KMO | From 0.50 | (Kaiser, 1974) |
| Factor loading | From 0.50 | (Hair et al., 2010) |
| Total variance explained | From 50% | (Hair et al., 2010) |
| Outer loading in PLS-SEM | From 0.70 | (Chin, 1998) or (Hair et al., 2022) |
| VIF in PLS-SEM | Below 5 | (Hair et al., 2019) |
These are reference points, not tools for forcing your data to pass. If you use SmartPLS, keep the correct PLS-SEM indicators, including outer loading, CR, AVE, HTMT, VIF, R², and Q². CFI, TLI, and RMSEA belong to the CB-SEM context. They should not be added to make a SmartPLS report look better.
When you need to collect more data
Consider collecting more data when the actual number of valid cases is lower than the original design, an important target group is almost absent, many questionnaires are incomplete in the same section, or you cannot verify the source of part of the dataset. Collecting more responses does not resolve the problem of suspicious rows. Keep those rows separate, check them, and record a clear decision.
If you only want more responses because Alpha is low or the p-value is large, check the scale, coding, and hypotheses first. A nonsignificant result can still be valid. Continuing to collect responses until a small p-value appears creates a greater risk when someone asks about your process.
Sample size also needs to be explained in relation to the research design. For the rule of 5 to 10 observations per item, the usual source is (Hair et al., 2010). For regression, the rule of n at or above 50 + 8m is stated in (Tabachnick and Fidell, 2013). Neither rule replaces a report of the number of questionnaires distributed, the number of valid questionnaires, and the sampling method.
If the concern is that your results differ from the original article, read what to do when results differ from the original article. Differences in context, timing, sample characteristics, and measurement can produce different results without implying fabricated data.
How to explain the issue to your supervisor
Follow this order: where the data came from, when they were collected, the sampling criteria, how many questionnaires were excluded, the cleaning rules, and the file used to generate the output. Open the file directly when asked instead of showing only screenshots of results tables.
How to write this in your thesis: “Data were collected from [survey population] during [period], producing [N0] responses. After checking [missing values, completion time, duplicate rows, or validity criteria], [N1] responses were retained for analysis. The cleaning steps are documented in [appendix or data log], and the tables in Chapter 4 were generated from [data file version].”
If you have used a service to run SPSS, request the data file, syntax, item-removal decisions, and software version. You can perform the analysis yourself in SPSS, SmartPLS, JASP, or R, depending on your skills and your department’s requirements. A service may save time on technical operations, but it increases the risk that you cannot explain each output line. Doing the work yourself takes more time but gives you control over the evidence. The essential point is to never accept results that you cannot trace.
You can also send a one-page summary showing the initial N, valid N, number of excluded rows, exclusion reasons, file version, and main tables. This is more specific than arguing about whether the data “look real.”
How to prevent the problem during study design
When you create the questionnaire, use stable variable codes such as DV1, DV2, and DV3, and mark reverse-coded items clearly. Do not change item wording or item order between collection rounds without saving a new version. Each time you export the data, keep the timestamp and a separate file name.
In the questionnaire, use screening questions that fit the target population, limit each person to one response where the platform allows it, and record the distribution channel. Do not collect unnecessary identifying information, but maintain an ethical and approved way to check duplicate responses within the boundaries of your study.
Before official collection begins, enter several test responses into SPSS or SmartPLS. Check whether the Likert codes, reverse-coded items, missing values, and column names work correctly. An error found before collecting 300 responses is much easier to fix than an error discovered after several files have been combined.
You can follow related posts in the Scientific research category or on the scientific research page so that your methods, sample size, and analysis remain consistent from the start.
Common mistakes when correcting the problem
The first mistake is deleting every row with similar responses. Some respondents genuinely choose the same level for several items, so partial similarity is not enough to justify exclusion. Combine multiple signs and record the reason for each group of cases.
The second mistake is changing values directly so that Alpha, KMO, or the p-value reaches a threshold. This removes your ability to explain the analysis. If you discover that you changed a value by mistake, restore the original version, record the incident, and rerun the analysis under a controlled process.
The third mistake is saving only screenshots of the output. A screenshot shows one table at one point in time, but it does not show which data produced that table. Save the .sav or .csv file, output file, syntax, cleaning log, and scale version.
The fourth mistake is presenting an exact number without a source. Every threshold in the article and in your thesis should be connected to an appropriate reference. If you have no basis for a particular cutoff, describe the criterion required by your department or supervisor instead of inventing a threshold.
Frequently asked questions
If I am asked whether the data are genuine, do I have to redo the entire study?
Not necessarily. First identify whether the question concerns the sample source, data file, data entry, cleaning process, or analysis result. If you still have the original file and can reproduce the output, explain each part before deciding to collect the data again.
Does a high Alpha prove that the data are genuine?
No. Alpha reflects the internal consistency of the items in the scale under the analysis you performed. A high Alpha can occur in genuine data, overly homogeneous data, or duplicated data. To explain data integrity, you need additional evidence about the collection source, timestamps, original file, and cleaning process.
Should I delete questionnaires completed too quickly?
You can define an exclusion criterion if it is reasonable and applied consistently. However, “too quickly” must be considered together with the response content, screening questions, and questionnaire context. Record the number of excluded questionnaires, the threshold or basis used, and the list of excluded cases.
Should I be concerned if my results look better than the original article?
A difference from the original article does not by itself prove that the data are fabricated. The two studies may differ in sample, context, timing, scale translation, and cleaning process. Present those differences and retain the files so your supervisor can check them when needed.
What should I do if I lost the original data file but still have the output file?
Search your downloads folder, recycle bin, cloud history, email, the computer used for data entry, and all backups. If you cannot recover the file, explain that limitation to your supervisor. Do not build a new file and present it as the original. You can use the output to describe the analysis that was run, but the possibility of rechecking it will be limited.
Should I hire an SPSS service to deal with this concern?
A service may make the technical work faster, but it cannot replace evidence about where the data came from. If another person helps you, still obtain the data file, syntax, item-removal log, and explanation for each decision. If you do not understand the output, the risk of being questioned again during the defense remains.
Your next action is to back up the original file, create a trace table for every cleaning step, rerun the analysis from the clean version, and prepare an explanation using your actual data. If you need to run the analysis on your own .sav or .csv file, you can use M4 DoThesis data analysis to work through each step without changing the data.