Assumption or Hypothesis: Differences and Use in a Thesis

What Are Research Assumptions and Hypotheses?

If you are writing your theoretical framework, it is easy to use the words “assumption” and “hypothesis” interchangeably. The short distinction is this: a hypothesis is a research proposition put forward for testing with data, while an assumption is usually a condition or premise accepted so that the analysis can be performed. In a quantitative thesis, statements H1, H2, and H3 are usually research hypotheses, not research assumptions.

For example, “Service quality has a positive effect on customer satisfaction” is a hypothesis. You can measure service quality with items, collect data through a questionnaire, and then test the path coefficient or regression coefficient. By contrast, “The observations were collected independently” is an assumption or analytical condition. It is not a relationship between constructs in the model.

You can also read the separate explanations of what a hypothesis is, what a model is, and what a theory is before revising Chapter 2. These three concepts are related, but they do not replace one another.

In quantitative research, a useful hypothesis must contain measurable constructs, a sufficiently clear direction of association, and a relationship that can be tested with .sav or .csv data. A hypothesis is not a broad question such as “What do customers think about the service,” and it is not a conclusion already known from the data.

Components of a Research Model and Its Hypotheses

When you read a research model, separate it into the following components. This makes it easier to identify which statement is a construct, which is a hypothesized relationship, and which is only an analytical condition.

In the table above, “Service quality” and “Satisfaction” are constructs or latent variables. Their relationship is stated as hypothesis H1. A condition such as the absence of serious multicollinearity belongs to the assumptions of the method, not to a hypothesis about customer behavior.

Some sources or lecturers may use “research assumption” broadly to mean predictions that need to be tested. When writing your thesis, however, check your department’s convention and use one term consistently. If Chapter 2 uses “hypotheses H1 and H2,” keep the same wording in Chapters 3 and 4.

A research model can be presented as a diagram with arrows. Each arrow should correspond to at least one theoretical argument and one hypothesis. If the diagram contains an arrow but the explanation has no matching H statement, the committee may ask why that relationship appears and where you tested it.

The Original Model and Common Extensions

A model usually starts from a theory or an established model. For example, the Technology Acceptance Model, commonly called TAM, explains technology acceptance through constructs such as Perceived Usefulness and Perceived Ease of Use. Davis's (1989) original TAM study is a source you can use when presenting this foundation.

For research on intentional behavior, Ajzen's (1991) Theory of Planned Behavior is often used with the constructs of attitude, subjective norm, perceived behavioral control, and behavioral intention. In technology acceptance contexts, UTAUT by Venkatesh et al. (2003) extends the explanation through several factors that influence intention and usage behavior.

An extension needs a reason. Do not add a variable simply because you found a paper with a model containing more variables. You need to answer three questions: What gap does the new variable address, which theory supports the relationship, and can your data measure the variable through a questionnaire?

For example, you might use TAM as the original model and add “trust” in a digital banking context. Trust then needs a definition, a referenced scale, and a connection to a specific hypothesis. If trust appears only in the diagram without an argument, it is not yet a well-supported extension.

You also need to distinguish a conceptual model from a statistical model. The conceptual diagram shows theoretical relationships. A regression model or PLS-SEM model uses data to estimate those relationships. A polished diagram does not guarantee that the hypotheses will be supported.

To place the model in the right context, you can consult the article on research models and the Research Models topic page. If your topic concerns competitive strategy, Porter's Five Forces is an example of an analytical framework, but that framework does not automatically become a quantitative model with H1 to H4.

Common Scales for Each Construct

You can test a hypothesis only after converting its constructs into items. You cannot directly measure “satisfaction” with one general impression if your aim is to build a multi-item scale. You need to define the construct, select a scale source, adapt the wording to the Vietnamese context, and test it again with your own data.

The sources in the table are starting points, not validated translations for every topic. Check the original construct, the number of items, the sample context, and the scale coding. For a Likert scale, state the number of response levels and the meaning of each level. Likert (1932) is the classic source on attitude measurement.

After collecting data, you can check reliability with Cronbach's Alpha. An alpha of 0.7 or above is generally considered acceptable according to (Nunnally, 1978), while exploratory research may consider a level from 0.6 according to (Hair et al., 2010). Corrected Item-Total Correlation from 0.3 or above can be referenced according to (Nunnally and Bernstein, 1994). These thresholds do not justify deleting items mechanically. Review the item wording, the theory, and how the results change after deletion.

If you use SmartPLS, assess outer loading, Composite Reliability, AVE, and HTMT according to PLS-SEM logic. An outer loading of 0.7 or above can be referenced to (Chin, 1998), CR from 0.7 and AVE from 0.5 or above can be referenced to (Fornell and Larcker, 1981), and HTMT below 0.85 or 0.90, depending on how closely related the constructs are, can be referenced to (Henseler et al., 2015).

Applying This to Your Topic

Suppose your topic is “Factors affecting students’ intention to use digital banking applications.” You select Perceived Usefulness, Perceived Ease of Use, and Trust as three independent variables, with Behavioral Intention as the dependent variable. The model diagram may contain three arrows pointing to Behavioral Intention.

The four hypotheses could be written as follows:

• H1: Perceived Usefulness has a positive effect on students’ Behavioral Intention to use digital banking applications.
• H2: Perceived Ease of Use has a positive effect on students’ Behavioral Intention to use digital banking applications.
• H3: Trust has a positive effect on students’ Behavioral Intention to use digital banking applications.
• H4: Perceived Ease of Use has a positive effect on Perceived Usefulness.

H4 adds another relationship between two independent variables. Keep H4 only if you have an argument from theory and previous research. When you state “positive,” you are specifying the expected direction of the relationship. During testing, the sign of the estimated coefficient and the statistical significance level will show whether the data are consistent with that expectation.

A hypothesis needs three layers of information. The first is the two constructs or variables being connected. The second is the direction of the relationship, such as positive or negative. The third is the context and population, such as students who have used a digital banking application. Without the context, the hypothesis can become too general to match the questionnaire.

You also need to distinguish a research question from a hypothesis. The question “Does Perceived Usefulness affect Behavioral Intention?” opens the issue to be examined. The hypothesis turns that question into a directional prediction: “Perceived Usefulness has a positive effect on Behavioral Intention.” The research question raises the issue, while the hypothesis creates a proposition for testing.

How to write this in your thesis

You can adapt the following paragraph: “Based on [name of theory or model], this study proposes that [independent variable] has a [positive/negative] effect on [dependent variable] in the context of [research population and scope]. Therefore, hypothesis H[ number ] is stated as follows: H[ number ]: [independent variable] has a [positive/negative] effect on [dependent variable].”

In the results section, do not write “H1 is correct” simply because the p-value is small. A more precise sentence is: “The estimation results show that [independent variable] has a [positive/negative] effect on [dependent variable], with [coefficient] = [value] and p-value = [value]. Therefore, H[ number ] is supported/not supported in the research sample.” Replace every bracketed position with the actual numbers from your file.

Testing the Model with SPSS or SmartPLS

If your model contains only composite variables and regression relationships, SPSS may be suitable. The usual process includes cleaning the data, checking Cronbach's Alpha, running EFA if required by the research design, creating representative variables, and running regression. In SPSS, you will work with tables such as Reliability Statistics, Item-Total Statistics, KMO and Bartlett's Test, Total Variance Explained, Rotated Component Matrix, and Coefficients.

If the model contains latent variables, multiple items, mediation, or moderation, SmartPLS may be more convenient for PLS-SEM. Separate the process into measurement model assessment and structural model assessment, following the two-step approach presented by Anderson and Gerbing (1988). In SmartPLS, review Outer Loadings, Construct Reliability and Validity, Discriminant Validity, Path Coefficients, R Square, f Square, VIF, and the Bootstrapping results.

The table below is illustrative output, not the result of a real study. It shows how you might read a table of hypothesis results.

When using bootstrapping in SmartPLS, the number of bootstrap subsamples is commonly set at 5,000 according to (Hair et al., 2022). Read the sign of the Path coefficient to identify the direction of the effect, use P values or the confidence interval to assess significance, and compare the result with the direction stated in the hypothesis. R² describes how much of the dependent variable is explained by the predictor variables. The levels 0.75, 0.50, and 0.25 can be referenced according to (Hair et al., 2011), but do not turn them into a promise about the quality of your model.

In SPSS regression, the Coefficients table usually contains Unstandardized Coefficients, Standardized Coefficients, t, and Sig.. Check VIF, for which a value below 5 can be referenced according to (Hair et al., 2019). A Durbin-Watson value between 1 and 3 is commonly used when considering residual autocorrelation according to (Field, 2013). These are model-checking criteria, not hypotheses H1 to H4.

Common Errors When Using This Model

Using “assumption” for every H1 and H2 statement. This makes it difficult for the reader to know whether you mean a theoretical prediction or an analytical condition. Use “research hypothesis” consistently for relationships and “assumption” for conditions that need to be satisfied.

Writing a hypothesis without measurable variables. The statement “Good technology makes customers more satisfied” sounds reasonable, but it is not ready for analysis. Define what “good technology” means as a construct, identify the items used to measure it, and show where it appears in the model.

Adding hypotheses after seeing the p-value. Do not test many relationships and keep only those that are significant. Build the hypotheses before testing. If you conduct an additional analysis, identify it as exploratory analysis and explain the reason for it.

Treating an unsupported hypothesis as proof that the entire model is wrong. A hypothesis may be unsupported because of the sample context, measurement approach, sample size, or the fit of the relationship within a particular population. Report the output as it is. Do not reverse the hypothesis direction simply to make the table look better.

Mixing SPSS, SmartPLS, and AMOS criteria. CFI, TLI, and RMSEA are fit indices commonly used in CB-SEM, with (Hu and Bentler, 1999) as a reference. Do not include these indices in a SmartPLS report if you did not run CB-SEM. Conversely, outer loading, HTMT, and bootstrapping must be interpreted within the correct PLS-SEM procedure.

Failing to save data versions. When deleting an item or adjusting the model, save the file by analysis round, such as efa_v1.sav, efa_v2.sav, and pls_final.csv. You need to know which item was removed, why it was removed, and how the results changed when your lecturer asks.

Frequently asked questions

Which is correct in a thesis, hypothesis or assumption?

Both words have valid meanings, but they refer to different roles. For H1 and H2 statements describing relationships to be tested, “hypothesis” is the more suitable term. For a condition or premise of the method, you can use “assumption.”

If your department or lecturer has a required template that uses “research assumption,” follow that requirement and use it consistently throughout the thesis. Do not alternate between the two words for the same content.

What is a research hypothesis, and must I use H1 and H2?

A research hypothesis is a proposition predicting a relationship between measurable constructs that can be tested with data. The labels H1 and H2 help you number the relationships and cross-check them across the model, methods chapter, and results chapter.

Not every quantitative topic needs the same number of hypotheses. The number depends on how many relationships you have a theoretical basis to test.

Can one hypothesis have two dependent variables?

A hypothesis should clearly show the relationship between one predictor and one outcome so that testing, interpretation, and reporting remain transparent. If there are two dependent variables, separate them into two hypotheses or construct two suitable equations.

You also need to check whether the two variables are truly different constructs or simply two dimensions of one construct. If they belong to the same higher-order construct, the modeling approach will be different.

Does an unsupported hypothesis mean that I should change the data?

No. Keep the original data and check the coding, reverse-coded items, missing data, analysis procedure, and analytical assumptions. If the output contains no error, report the unsupported result and explain it using the context and suitable literature.

Deleting observations or changing the model simply to obtain a significant p-value creates substantial risk at the defense. Every adjustment needs a methodological reason and must be recorded.

Can a model contain both mediation and moderation hypotheses?

It can, provided you have a theoretical basis and the sample size, scales, and analytical design meet the requirements. A mediator explains the mechanism through which an effect is transmitted, while a moderator changes the strength or direction of a relationship.

Distinguish the two types clearly in the diagram, hypotheses, and output interpretation. For an indirect effect, a bootstrap confidence interval that does not contain 0 is a criterion that can be referenced according to (Preacher and Hayes, 2008).

Open your data file again, match every arrow in the model with one hypothesis, one group of items, and one specific testing method, then use M4 data analysis to run the analysis on your own .sav or .csv file.