DataSuite 2 Documentation
English Русский
On this page

Writing it up

The cards are on the screen; the last step is turning them into a paragraph a reader can check. This page says which numbers go into it, in what order, and what has to be said about how they were made – walked through on the four café findings of Choosing an analysis. The rule underneath is the same for a café owner’s memo and a journal article: a report gives the reader enough to judge the finding and enough to reproduce it, and nothing that is only there because the app printed it.

Two paragraphs

A quantitative report has two parts that answer two different questions. The Method says what was done – the data, the fixes, the design, the test and the settings – so that someone with the same file could get the same card. The Results says what was found – the estimate, its uncertainty, the test and its verdict – so that the reader can judge it without seeing the card. Every module’s manual ends with a reporting checklist in exactly those two halves; the walk below is the checklist filled in.

Method

The Method paragraph has five things in it, and the first three are the same whatever the analysis.

The data. Where it came from, how many cases, what the variables were and how they were measured: “200 loyalty-card members of a two-branch café, surveyed by e-mail eight weeks after a menu change; per-visit spend and visit counts for the eight weeks before and after from the till; six 1–5 agreement items, a 0–10 satisfaction score, a yes/no recommendation and the wait at the last visit from the survey.” A reader who does not know what overall is cannot judge anything said about it, so define the variables in words, not by column name.

What was done to it. Every step of the codebook page is a sentence here, because each one changed a number: “Branch was recorded as 1 and 2 and recoded to its names. Age was entered as 999 for five members who declined, and those cells were treated as missing. One member’s after-window spend was a single catering order; that cell was excluded from the spend analyses (n = 199), the member’s other data retained.” State the rule, not just the outcome – an after-window average above 50 is not a per-visit spend – so that the reader can disagree with it. Then how missing values were handled: “analyses use the cases complete on their own variables (pairwise deletion), so n varies by analysis and is reported with each result.” The outliers and missing data page has the full list of what to disclose.

The design and the roles. Which cases were compared and how they relate: “Branches were compared as independent groups; before-and-after measures were compared within members, each member serving as their own control.” The reader needs the design to know which test is appropriate, and it is the sentence most often left out.

The tests and the settings. The test by name, the direction, and the app-wide settings that shaped the p: “Welch’s t-test for the branch comparisons, the paired t-test for the before–after change, Pearson’s r for the correlations; all tests two-tailed at α = 0.05, with confidence intervals at 95%. The six survey items were compared between branches as one family, with Holm adjustment.” A setting left at its default still goes in when it affects the numbers; the settings page marks the ones that do.

The tools. The software and the packages behind the tests, which the app’s R console and each manual’s Reproducibility section name, and the citation the app collects in its citation card. One sentence: “Analyses were run in DataSuite 2, which computes its statistics with R (the packages are listed in the appendix).”

Results

A result is one sentence per finding, and the sentence has a fixed skeleton: the estimate in the outcome’s own units, its confidence interval, the test with its statistic and degrees of freedom, the exact p, the effect size, the sample size. In that order, because that is the order of importance – the reader wants to know how big first, how sure second, and the test statistic is the receipt. The hypothesis testing, effect sizes and confidence intervals pages each say what their part of the skeleton needs; here the four café findings assembled from them.

A difference between groups. “Station members spent more per visit than Riverside members before the menu change: 9.52 (SD 4.05) against 7.91 (SD 2.97), a difference of 1.61 (95% CI 0.62 to 2.60), Welch’s t(179.5) = −3.20, p = 0.002, d = −0.45, n = 99 and 101.” Everything the card showed is here, and the direction has been put into words – Station more than Riverside – so that no reader has to work out which group the app subtracted from which; the statistic and d keep the sign the card printed, which follows the app’s group order. The group SDs are there because the reader can rebuild d from them; the group intervals are not, because the interval on the difference is the one that answers the question. A non-significant difference is reported in the same shape: “The branches did not differ in whether members would recommend the café – 72% at Riverside, 63% at Station, χ²(1) = 1.71, p = 0.19, V = 0.10 (95% CI 0 to 0.24), n = 200 – the interval on V allows anything from no association to a modest one.” Did not differ means the test found no evidence; write “no significant difference” rather than “no difference” and let the interval say how much room is left.

A change within the same cases. “Spend per visit rose after the change by 1.09 on average (SD of the change 2.86, 95% CI 0.69 to 1.49), paired t(198) = 5.37, p < 0.001, dz = 0.38, n = 199. Visits per member rose by 0.64 (95% CI −0.003 to 1.28), t(199) = 1.96, p = 0.051, dz = 0.14, n = 200.” The second sentence is the one to get right: it reports the boundary result exactly and lets the interval show that the data are compatible with no change and with a rise of a visit and a quarter. What it does not do is say “visits did not change” or drop the finding – a p of 0.051 rounds to nothing only in a report that is hiding it. The boundary section says why.

A relationship. “Overall satisfaction correlated with the taste rating (r = .55, 95% CI .44 to .64, p < .001, n = 193) and, negatively, with the wait at the last visit (r = −.30, p < .001, n = 178). In a regression on both, each point of taste was worth 0.81 points of satisfaction (95% CI 0.60 to 1.01) and each minute of waiting cost 0.13 (0.05 to 0.21), R² = .32, F(2, 175) = 41.5, p < .001, n = 178.” A correlation is its own effect size, so the sentence needs no second one; a regression coefficient’s unit is the sentence – points of satisfaction per point of taste – and the reader should be told the units rather than left to read B. Where a family of correlations was tested, say how many and which adjustment; where predictors were compared, β is the number that puts them on one scale and belongs in the sentence that compares them.

A structure. “The three food items formed a scale of acceptable consistency (Cronbach’s α = .71, 95% CI .63 to .77; ω = .72) and the three service items a somewhat weaker one (α = .67, .58 to .74; ω = .67), with the reverse-keyed item recoded. An exploratory factor analysis (minimum residual, oblimin) recovered the two intended factors – parallel analysis retained two, loadings .59 to .76 on Food and .58 to .72 in absolute value on Service, the reverse-keyed item loading negatively, no cross-loading above .20, factor correlation .39 – and a two-factor confirmatory model fit the data, χ²(8) = 6.41, p = .60, CFI = 1.00, RMSEA = .000 (90% CI .000 to .071), SRMR = .034, where a one-factor model did not, Δχ²(1) = 57.7, p < .001.” The reporting sections of the latent variables page and the factor and SEM and CFA manuals say which fit indices and which decisions – the extraction, the rotation, the rule for the number of factors – a reader needs to see.

What the reader needs and what they do not

Three habits separate a report that can be checked from one that can only be believed.

Estimate, interval, p – all three, every time. A p on its own says whether; an estimate says what; an interval says how sure. The café’s branch gap of 1.61 with an interval of 0.62 to 2.60 tells a reader more than p = 0.002 ever could, and a report that gives only the p has thrown away the two numbers that mattered. The same goes for a non-significant result: its interval is the finding.

Exact numbers, sensibly rounded. p to two or three decimals, never “p < 0.05” when the value is known, and “< 0.001” only when it is. Estimates to the precision the measurement had – a spend to the cent, a 0–10 score to one decimal – and effect sizes to two decimals. The app prints more decimals than a sentence should carry; rounding is the writer’s job, and rounding a p of 0.051 to 0.05 is not rounding.

The receipts. Degrees of freedom with every statistic, the sample size with every result, and the notes the app printed under a table – a member excluded, a correction applied, an exact test used, a replicate count – in the Method or in a footnote. A reader who wants to reproduce the analysis needs the file, the fixes, the design and the settings; give those and they can. The app’s reproducibility sections say which R functions ran, and a project file saved from the app records every choice made.

And three things a report does not need: the group-by-group intervals when the interval on the difference is given; every effect size the app offers when one was chosen in advance; and the word proven, which no sample supports. The finding is what the data showed, stated with its uncertainty; the reader decides what it means.

Where this appears in DataSuite