The five stages of this chapter

1. Who the system is forStakeholder analysis. The profile of the 40 organisers and the 64 travellers.
2. Whether the measurements are trustworthyReliability, purification, dimensionality. This comes before any result.
3. What the numbers sayDescriptives, correlation, regression, hypothesis testing.
4. What the words sayThematic analysis of the free-text answers, which independently confirms the numbers.
5. What the system must therefore doThe design mandate, the diagrams, the architecture, and 23 numbered requirements.

Stage 1. Who the system is for

The 40 organisers

Who answered: 45.0 percent were event coordinators or events managers, 27.5 percent owners or general managers, 17.5 percent front desk or guest relations, 10.0 percent food and beverage managers.

What kind of property: 35.0 percent boutique hotels and guesthouses under 20 rooms, 30.0 percent mid-size hotels of 20 to 60 rooms, 27.5 percent beach or holiday villas, and only 7.5 percent large hotels and resorts.

The 64 travellers

Stage 2. Are the measurements trustworthy?

This stage comes before any finding, and the order is deliberate. If the questions do not hang together, nothing after them means anything.

What went wrong, and what was done about it

Two questions failed. Both of them were the deliberately reversed ones, item 7 and item 20, and they failed in both samples.

After flipping, both reversed items correlated negatively with their own group. In plain terms: people answered them in line with the other questions rather than against them, which means the reversal did not register.

Table 4.3 in your thesis. Notice that the two groups with nothing removed did not need removing. Only the two reversed items failed.

Why it is a known artifact, not bad data

This is the strongest part of the defence. Three pieces of evidence:

• It hit exactly the two reverse-worded items on each instrument and nothing else. Every forward-worded item behaved normally.
• It happened the same way in two independently recruited samples.
• There is published research documenting this. Negatively worded items often load onto a separate "method factor" rather than the thing they were written to measure, especially where people answer in a second language or at speed.

One more check: does each group measure one thing?

Principal component analysis was run on every purified group. Every single one returned exactly one strong result, which means each construct can legitimately be averaged into a single score.

Stage 3. What the numbers say

The averages

Everything sits above the 3.00 midpoint in both samples, which means both sides of the market recognise the problem and want the fix.

Two patterns worth noticing. IV3 is the highest of the three causes in both samples, so information fragmentation is the most strongly felt problem. And the outcome is the highest of all four in both samples, which means willingness to adopt is strong on both sides.

The correlations, where all three look strong

Every one is positive and statistically significant. And the rank order is identical in both samples, which on its own is a striking result, because these were two independently recruited groups answering parallel questionnaires.

The regression, where the truth comes out

All three causes were entered into the model together, and the analysis was run separately on each sample. That gives two independent tests of the same idea.

Both models are significant overall, and both pass their assumption checks. VIF under 5 and Durbin-Watson inside 1.5 to 2.5.

Why the correlations and the regression disagree

The verdict on the three hypotheses

Stage 4. What the words say

Items 24 and 25 on each survey were free text. They were analysed with the six-phase thematic analysis method, and the result independently confirms the statistics.

The adoption conditions, which became your requirements

Question 25 on the organiser survey asked what would make them adopt a platform, or stop them. The answers turned directly into your non-functional requirements. This is how a feeling becomes something testable.

Stage 5. What the system must therefore do

The design mandate

The free-text answers sharpen it in three ways:

• Because staleness rather than absence is the primary failure, the requirement is for live-derived availability, not a status note the organiser types in and forgets.
• Because the information failure happens during events and not only while planning, the organiser view must be operational and live, not a report that runs overnight.
• Because adoption is conditional on reliability, simplicity, security, proof and cost, those become specified and testable requirements rather than generic quality talk.

The diagrams

The requirements

14 functional and 9 non-functional, each traced to its evidence. Four functional ones are marked critical, and they are the four that implement the finding.

Numbers from this chapter

If they ask you

Your correlations support all three hypotheses. Why did you report the regression instead?

Because correlation asks whether two things move together, and regression asks what each one contributes once the others are accounted for. My three predictors overlap heavily, with IV2 and IV3 correlating at 0.738 in the tourist sample, so at the bivariate level each predictor gets credit for variance it shares with the others.

The regression is the test that distributes that credit correctly. Reporting only the correlations would have supported all three hypotheses handsomely and would have been misleading, and Section 7.3 says explicitly that the temptation to do that was resisted.

Your organiser alpha was 0.159. How is that acceptable?

It is not, and it was not accepted. That figure is before instrument purification. One reverse-coded item, item 20, returned a corrected item-total correlation of minus 0.623 after reverse-scoring. Removing it by the standard 0.30 threshold lifted the scale to 0.690.

Table 4.3 reports both the initial and the final value for every construct, because showing only the final figure would conceal the diagnostic step that made the scale trustworthy.

Removing items until your alpha looks good sounds like fishing. How is that legitimate?

The rule was fixed in advance rather than chosen afterwards. Items below the conventional 0.30 corrected item-total threshold are removed, worst first, recomputing after each removal, and no scale may drop below 3 items.

Only 2 items in the whole instrument crossed that threshold, and they were exactly the 2 reverse-coded items, in both independent samples. No forward-worded item was touched, and no further removals were made once those two were out.

If I had been fishing, the removals would have been scattered across the instrument. Instead they were confined to the two items with a documented failure mode.

Is 64 responses enough for a regression?

It is enough for what the study claims and not enough for what it explicitly does not claim. With 64 cases and 3 predictors there are roughly 21 cases per predictor, which exceeds the commonly applied minimum of 10 to 15 for stable estimates but falls short of the 50 to 100 per predictor recommended for optimal stability.

So the model is adequately powered to detect a large effect and underpowered for small ones. That is precisely why the non-significant coefficients are reported as a failure to detect a unique effect rather than as evidence that no effect exists.

You defined the dependent variable as having three sub-dimensions but only scored one. Why?

Because the evidence decided. The principal component analysis showed the 4 surviving items forming a single dimension rather than three separable factors.

So it is scored as one unidimensional composite and the three conceptual sub-dimensions are retained as facets of the construct rather than as separately scored subscales. Establishing them as distinct measurable subscales would require a longer instrument with enough items per sub-dimension, and that is a named future recommendation in Chapter 7.