ImpactMojoStatistics Without Code 101www.impactmojo.in

ImpactMojo 101 Series · Free Forever

Statistics
Without
Code 101

Real Analysis in jamovi and JASP — from a Survey CSV to a Reported Result, with Every Click Written Down

Data & TechnologySouth Asia Focus100 SlidesFree Access

ImpactMojoStatistics Without Code 101www.impactmojo.in

Agenda

What We Cover

Why Point-and-Click, and Its Limits

Slides 3–10

Getting Data In and Clean

Slides 11–19

Describing Data

Slides 20–27

Comparing Groups

Slides 28–37

Correlation and Regression

Slides 38–45

Categorical Outcomes

Slides 46–54

Scales and Factor Analysis

Slides 55–62

The Modules: Beyond the Basics

Slides 63–70

Bayesian Analysis in JASP

Slides 71–78

Effect Sizes, Power and Reporting

Slides 79–90

Choosing Tools, Teaching and Practice

Slides 91–99

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section One

Why Point-and-Click, and Its Limits

ImpactMojoStatistics Without Code 101www.impactmojo.in

The Tools

jamovi and JASP: free, open, built on R, and menu-driven

jamovi

Free and open source, from a small team that came out of the JASP project in 2017. A spreadsheet on the left, results on the right, every analysis a form with tick-boxes, and every result an R computation underneath. Saves data and analyses together in one file. A library of add-on modules covers mediation, mixed models, meta-analysis, SEM, item response theory, power and more. Runs on Windows, Mac, Linux and in a browser (jamovi Cloud). jamovi.org.

JASP

Free and open source, from the University of Amsterdam, led by Eric-Jan Wagenmakers, since 2013. The same idea with a stronger emphasis on Bayesian analysis alongside the classical tests, APA-formatted tables by default, and modules for SEM, meta-analysis, network analysis, machine learning and audit. jasp-stats.org.

Both replace SPSS for most of what a social-science or NGO analyst does, at no cost, on any machine, with output that is correct and formatted for a report. Both show the R code behind an analysis if asked, which makes them the gentlest route into R there is. Neither is a toy; the numbers are R's numbers.

This course teaches jamovi as the main tool, because its data editor, filters and computed variables make it the better place to clean a survey, and JASP for the Bayesian section, where it leads.
Menu paths are given for jamovi 2.x; JASP's are close and named where they differ.
Every analysis is shown as: the click path, the boxes to tick, the output, and the sentence to write.

Navarro and Foxcroft, Learning Statistics with jamovi (2019), is free online and is the textbook this course pairs with.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Who

Who this is for, and what it assumes

Reader	Situation	What this course gives
Master's student in a social-science department	SPSS licence expired or never existed; a thesis with a survey to analyse	A complete free toolkit and the analyses a thesis needs, done correctly
NGO monitoring and evaluation officer	A baseline in Excel, a funder who wants 'significance', no programmer on staff	The right test for each question, the effect size to report, and the plot for the report
Teacher of research methods	A classroom with mixed laptops and no budget	A tool that installs in minutes and runs the same everywhere, plus a teaching sequence
Practitioner moving toward R	Wants to understand the output before learning to write it	Syntax mode shows the R behind every click
Field researcher	Needs to check data quality in a district office with no internet	Descriptives, filters and plots offline, in an afternoon

Assumed: the statistics in Data Analysis 101 and Bivariate Analysis 101 at the level of knowing what a mean, a p-value and a regression coefficient are. Not assumed: any code. Worked examples use two illustrative datasets, a self-help-group survey of 640 women and a learning assessment of 1,200 children in 40 schools, described in section 02.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Limits

What a menu cannot do, so you know before you start

Complex survey design. jamovi accepts a case-weight variable, but only integer frequency weights, and neither jamovi nor JASP handles strata and clusters, so standard errors from a menu ignore the design. For NFHS, PLFS, DHS or any weighted survey, R's survey package or Stata's svy is required.
Large or repeated pipelines. Cleaning 40 monthly files the same way is a script's job. A menu does it once.
Advanced econometrics. Instrumental variables, panel fixed effects with clustering, difference-in-differences with staggered adoption, time series: partly available through modules, mostly not.
Reproducibility across people. The saved file records the analyses, but a reviewer cannot re-run a menu sequence the way they re-run a script.

The honest position: jamovi and JASP cover the analyses in a typical master's thesis, an NGO evaluation report and an undergraduate methods course, and cover them well. Where a question needs survey weights or a scripted pipeline, the course says so at that point and names the tool that does it. Knowing the edge is part of using the tool competently.

The commonest error this course can prevent: an NFHS analysis run unweighted in a menu and reported as national estimates. Section 02 returns to it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

The Interface

The jamovi window: data, analyses, results, and the three tabs

The left pane is the data: a spreadsheet where each column has a name, a measurement type (nominal, ordinal, continuous, ID) and a data type (integer, decimal, text), set by double-clicking the header. The right pane is the results: every analysis you run appends its tables and plots there, and clicking one reopens its options. The tabs across the top are Data (Setup, Compute, Transform, Filters, Add/Delete rows and columns), Analyses (the menus), and Edit (annotations).

The measurement type controls what analyses will accept the variable. A 1/2 gender code set as continuous will be averaged; set it nominal and label it.
Results update live when the data or an option changes; there is no 'run' button.
The three-dots menu (top right) holds Syntax mode, the number format, and the plot theme.
Save as .omv: data, analyses, options and annotations in one file; anyone with jamovi reopens it exactly.

JASP's equivalents

A data view on the left (editable since version 0.18), analyses grouped as Descriptives, T-Tests, ANOVA, Mixed Models, Regression, Frequencies and Factor, each with a classical and a Bayesian entry; results on the right, APA-formatted; a .jasp file that stores everything. A '+' button adds modules. The logic is identical; the menus are labelled slightly differently and this course notes the difference where it matters.

Ten minutes of clicking through the interface with the bundled Tooth Growth or Big Five dataset is worth more than any slide here. Do it before section 02.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Syntax Mode

Syntax mode: the R behind the click, and the bridge to code

The three-dots menu's Syntax mode shows, above every result, the R command that produced it, using jamovi's own jmv package: jmv::ttestIS(data = df, vars = score, group = treatment, effectSize = TRUE). Copy it into R with jmv installed and it runs, producing the same tables. The Rj module goes the other way: an editor inside jamovi where you write R against the loaded data and see the output in the results pane.

Use syntax mode to learn what the options do: tick a box, watch the argument appear.
Use it to document: paste the syntax into the methods appendix. It is the reproducible record of the analysis.
Use Rj for the one thing the menus lack: a weighted estimate with the survey package, a custom plot, a loop over districts.
JASP has no syntax mode as such, but its analyses map onto named R packages (BayesFactor, lavaan, metafor) that it documents.

The teaching use

A student who has run twenty t-tests by menu and seen the syntax each time has learned the grammar of an R function without noticing. The step to writing t.test(score ~ treatment, data = df) is small. That is the argument for jamovi over SPSS in a methods course: the same menus, plus the door to code, at no cost.

R for Development 101 in this series picks up where syntax mode leaves off.

ImpactMojoStatistics Without Code 101www.impactmojo.in

SPSS

Coming from SPSS: what maps, what differs, what you gain

SPSS	jamovi	Note
Variable View: type, label, values, measure	Double-click the column header: name, description, type, levels	Levels are labelled in place; no separate value-labels dialog
Analyze → Descriptive Statistics → Frequencies / Descriptives / Explore	Analyses → Exploration → Descriptives, with Split by	One dialog covers all three
Compare Means → Independent-Samples T Test	T-Tests → Independent Samples T-Test	Effect size and assumption checks are tick-boxes
General Linear Model → Univariate	ANOVA → ANOVA (or ANCOVA)	Post hoc, estimated marginal means and plots in one form
Regression → Linear / Binary Logistic	Regression → Linear Regression / Logistic Regression (2 Outcomes)	Model builder with blocks; collinearity and residual plots as options
Crosstabs with chi-square	Frequencies → Independent Samples (χ² test of association)	Row, column and expected counts as tick-boxes
Scale → Reliability Analysis; Dimension Reduction → Factor	Factor → Reliability Analysis; PCA; Exploratory Factor Analysis; CFA	ω alongside α; parallel analysis built in; CFA in the core
Syntax window	Syntax mode; Rj module	R, not SPSS syntax
Split File, Select Cases	Split by in each analysis; Data → Filters	Filters are live and stack
.sav files	Open directly; also .csv, .xlsx, .dta, .rds	Value labels are preserved on import

What you lose: nothing a thesis needs. What you gain: effect sizes and assumption checks in every dialog, omega, CFA and parallel analysis without add-ons, live results, and a file a co-author can open without a licence.

ImpactMojoStatistics Without Code 101www.impactmojo.in

This Course

How the course is arranged

DATA IN: import, types, labels, filters, computed variables, checks (02)

→

DESCRIBE: descriptives and plots that answer questions (03)

→

COMPARE: two groups, several groups, before and after (04)

→

RELATE: correlation and regression, with diagnostics (05)

→

CATEGORICAL: proportions, chi-square, logistic regression (06)

→

SCALES: reliability, EFA, CFA (07)

→

MODULES: mediation, mixed models, meta-analysis, SEM, IRT (08)

→

BAYES: the JASP way of answering the same questions (09)

→

REPORT: effect sizes, power, reproducibility (10)

→

PRACTICE: choosing tools, teaching, a full worked study (11)

Each analysis slide follows one pattern: the question in words; the menu path; the options to tick and why; the output table with illustrative numbers; the assumptions and what to do when they fail; and the sentence that goes in the report. The same pattern makes the course usable as a reference: find the question, follow the slide.

The statistics are the same as in every other course in this series. What is new is that every one of them is reachable without a line of code, and the course says exactly where.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Two

Getting Data In and Clean

ImpactMojoStatistics Without Code 101www.impactmojo.in

Import

Importing: CSV, Excel, SPSS, Stata, and what to check on arrival

Menu → Open → browse. jamovi reads .csv, .txt, .xlsx, .sav (SPSS), .dta (Stata), .sas7bdat, .rds and .RData, and its own .omv. Value labels from SPSS and Stata become levels. CSVs arrive with every column typed by inspection: a column of 1s and 2s is treated as continuous integer, a column with any text as nominal text. The first job after import is to walk the columns and set each one's type and levels deliberately.

KoboToolbox and ODK exports: download as XLS or CSV with labels, not codes, or with both; jamovi will show the labels.
Excel files with merged headers, notes above the data, or totals rows must be cleaned in Excel first; jamovi wants one header row and one row per case.
Dates arrive as text; jamovi has no date type. Compute the pieces you need (year, month, days since baseline) in the source or with a computed variable.
Indian number formats with commas (1,23,456) are text on import; strip them in the source.

The arrival checklist

Number of rows equals the number of interviews. Number of columns equals the questionnaire's variables. Every ID is unique (Data → Descriptives on the ID with Frequency tables shows duplicates). Missing values are recognised as missing, not as 999 or blank strings (set the missing codes in the column setup). Every categorical variable has the levels you expect and no stray ones ('Female', 'female', 'F'). Ten minutes; it prevents every later analysis being wrong.

Keep the raw file untouched and save the jamovi work as a separate .omv. The raw file is the audit trail.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Datasets

The two worked datasets used in this course

shg_survey.csv (n = 640 women, 32 villages)	Type	Levels or range
id	ID	1–640
village	Nominal	32 codes
treatment	Nominal	SHG member / non-member (village-randomised)
age	Continuous	18–60
educ	Ordinal	None / Primary / Secondary / Higher
hh_size	Continuous	1–14
income_own	Nominal	Yes / No
d1–d5	Ordinal	Decision-making items, 0 / 1 / 2
savings_rs	Continuous	Monthly savings, rupees, right-skewed
decide_score	Continuous	Sum of d1–d5, 0–10
lang	Nominal	Hindi / Bangla

learning.csv (n = 1,200 children, 40 schools)	Type	Levels or range
child_id, school_id	ID / Nominal
arm	Nominal	Remedial / Control (school-randomised)
grade	Ordinal	3 / 4 / 5
sex	Nominal	Girl / Boy
baseline	Continuous	Reading score, 0–100
endline	Continuous	Reading score, 0–100
attend_pct	Continuous	Attendance, 0–100
reads_para	Nominal	Can read a paragraph at endline: Yes / No

Both are illustrative and constructed for teaching. Every number in the output tables that follow is illustrative too, with magnitudes chosen to be realistic.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Types and Levels

Setting measurement types and labelling levels

Double-click a column header, or Data → Setup. Set the measure type: Nominal for categories with no order (village, sex), Ordinal for ordered categories (education, Likert items), Continuous for numbers with meaningful intervals (age, score, rupees), ID for identifiers that should never be analysed. For nominal and ordinal variables, the Levels box lists the values found; type a label beside each code (1 = Female, 2 = Male) and drag to reorder ordinal levels.

Likert items: ordinal for descriptives and frequencies; many analyses (means, t-tests, factor analysis) will treat them as continuous when asked, and the course says when that is defensible.
A binary 0/1 variable can be continuous (its mean is a proportion) or nominal (for chi-square); set it as the analysis needs, or keep two copies.
Add a description to every variable. It appears in output and saves the codebook lookup.

Why this step decides everything

A t-test dialog will not accept a nominal variable as the outcome; a chi-square will not accept a continuous one as a factor; a regression will dummy-code a nominal predictor and treat an ordinal one as ordered polynomial contrasts unless told otherwise. The types are not decoration; they are the analysis's first assumption, and they are set here.

For education (None, Primary, Secondary, Higher): ordinal, levels in that order, so every table and plot reads left to right.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Computed Variables

Computed and transformed variables: scores, recodes, logs

Need	Data → Compute (new variable) formula	Note
Scale score	`d1 + d2 + d3 + d4 + d5` or `MEAN(d1, d2, d3, d4, d5)`	MEAN handles missing items; set a minimum with `MEAN(..., ignore_missing = 1)` and check the count
Log of a skewed variable	`LN(savings_rs + 1)`	The +1 for zeros; say so
Change score	`endline - baseline`	Section 04 discusses when a change score is the wrong analysis
Standardised score	`Z(baseline)`	Mean 0, SD 1 across the sample
Per-capita	`savings_rs / hh_size`	Division by zero if hh_size can be 0; filter first
Condition	`IF(attend_pct >= 75, "Regular", "Irregular")`	Text result becomes nominal
Age band	`IF(age < 25, "18-24", IF(age < 40, "25-39", "40+"))`	Nested IF; or use Transform for many bands

Transform (Data → Transform) recodes an existing variable through a saved transformation that can be applied to many columns at once: reverse a 1–5 item (6 - $source), collapse categories, convert codes to labels. Both computed and transformed variables update live if the source data change, and both show their formula in the column header, which is the documentation.

Every computed variable is a decision (which items, how missing was handled, why the log). Write the formula into the methods; jamovi shows it, so copy it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Filters

Filters: analysing a subset without deleting anything

Data → Filters adds a filter row: a formula that is true for the cases to keep. age >= 18; treatment == "SHG member"; NOT(village == 17); attend_pct > 0 and baseline != NA. Filtered-out rows are greyed, kept in the file, and excluded from every analysis while the filter is active. Filters stack; each can be switched off; the active set is shown in every result's footnote.

The right way to exclude implausible values (a savings figure of 9,999,999), pilot cases, or a district with a known data problem: filter, note it, keep the rows.
The wrong way: deleting rows. The raw n and the analysed n must both be reportable.
Split by (in most analysis dialogs) runs the same analysis per group without filtering; use it for descriptives by arm, filters for exclusions.
Report every filter in the methods, with the n before and after.

Data quality checks by filter

A filter is also a query. endline < baseline - 30 shows children whose score fell by more than 30 points: possible ID mismatches. savings_rs > 20000 shows outliers to verify against the paper form. d1 == NA or d2 == NA shows partial refusals of the decision module, which section 05 of SEM 101 says are not random. Each is thirty seconds and each belongs in the data-quality log.

Survey Design 101's high-frequency checks can all be run in jamovi with filters and descriptives, in the field, on the day.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Missing Data

Missing values: recognising them, counting them, handling them

Blank cells and the codes you declare as missing (Setup → Missing values: 999, -99, 'Refused') are treated as missing. Descriptives reports N and Missing per variable; a frequency table shows missing as a row if asked. Most analyses drop cases listwise: a regression with six predictors loses every case missing any one of them, and the results footnote gives the n used.

Report the missingness per variable in the descriptives table. It is the first thing a careful reader looks for.
For a scale score, MEAN() with a minimum-items rule is a defensible partial-response treatment; state the rule.
Multiple imputation is not in the jamovi core. The Rj module with the mice package does it; or move to R for that analysis.
Do not replace missing with the mean. It is one click in some tools and it is wrong.

Whether it matters

Descriptives → split by a 'missing on the outcome' indicator (a computed variable: IF(endline == NA, 1, 0)) shows whether the children with no endline score differ at baseline from those with one. If they were weaker, attrition biases the endline mean upward and the report must say so. This is the one missing-data analysis every evaluation needs and it takes two clicks.

Attrition by arm is reported in every trial. A crosstab of arm by missing-endline, section 06, is the table.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Weights

The survey-weights problem, stated plainly

NFHS, PLFS, DHS, NSS and most large surveys sample some households with higher probability than others and supply a weight per case to correct for it, and they sample in clusters within strata, which changes every standard error. jamovi can weight cases, but only by an integer frequency weight, and it has nowhere to enter strata or clusters; fractional design weights must be rounded, which the survey's documentation forbids. An NFHS analysis run in the menus is an analysis of the sample, not of the population, and its confidence intervals are too narrow. JASP has no design support either.

For your own survey with equal-probability sampling and a simple design, the menus are correct and clustering can be acknowledged as a limitation or handled in R.
For a weighted national survey, use jamovi for exploration and data checks only. Estimate in R (survey, via the Rj module or RStudio) or Stata (svy).
InsightStack's data starters for DHS and PLFS load the files with the weights and design already set, in R, Stata and Python.
Say in the report which estimates are weighted and which are not. Never present an unweighted menu estimate as a population figure.

What Rj makes possible

Install the Rj module (Modules → jamovi library). In its editor: library(survey); des <- svydesign(ids = ~psu, strata = ~stratum, weights = ~wt, data = data); svymean(~stunted, des). Four lines, inside jamovi, on the loaded data, with correct standard errors. It is code, but it is four lines, and Survey Design 101 and the FieldStack notes give the design variables for each survey.

This is the one limit in the course that produces wrong published numbers when ignored. The rest produce inconvenience.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Data Checklist

Before any analysis: the data checklist

Check	How in jamovi	Done
Row count matches the interview count	Bottom of the data pane; Descriptives N on the ID
IDs unique	Descriptives → Frequency tables on ID; any count above 1
Every variable typed and labelled	Walk the headers; Setup
Missing codes declared	Setup → Missing values
Missingness per variable known	Descriptives → N, Missing
Ranges plausible	Descriptives → Min, Max; box plots
Categorical levels clean	Frequency tables; fix in source or Transform
Outliers verified or filtered with a note	Filters; box plots
Scale scores computed with a stated missing rule	Compute with MEAN()
Weights and design understood	If weighted: not in the menus
Raw file preserved; work saved as .omv	Save As

Eleven rows, forty minutes for a survey of this size. Every one of them has, in some published report, been skipped and produced a number that was wrong.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Three

Describing Data

ImpactMojoStatistics Without Code 101www.impactmojo.in

Descriptives

Exploration → Descriptives: the one dialog to learn first

Drag variables into the box, optionally a grouping variable into Split by, and tick what you want: N, missing, mean, median, mode, sum, SD, variance, range, minimum, maximum, standard error, confidence interval of the mean, IQR, skewness, kurtosis, Shapiro-Wilk, percentiles. Tick Frequency tables for nominal and ordinal variables. Under Plots: histogram, density, box plot, violin, dot plot, bar plot, Q-Q plot. Everything updates as you tick. Transpose the table if there are many variables and few statistics.

For a report's Table 1: N, mean, SD, median, IQR for continuous; frequency tables for categorical; split by arm.
For a data check: min, max, missing, and a box plot per variable.
For a distribution decision: histogram with density, skewness, and the Q-Q plot, before choosing a test in section 04.

JASP

Descriptives → Descriptive Statistics, the same options plus a correlation heat-map, a Pareto plot, and Likert plots for ordinal items. JASP's table is APA-formatted by default and can be copied as LaTeX or HTML.

Mean and SD describe a symmetric variable; median and IQR describe a skewed one. Savings in rupees is skewed: report the median, and the mean only alongside it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Table 1

A worked Table 1: the SHG survey by arm

	SHG member (n = 322)	Non-member (n = 318)	Missing
Age, mean (SD)	34.1 (8.2)	34.6 (8.5)	0
Education, n (%): None / Primary / Secondary / Higher	118 (37) / 96 (30) / 84 (26) / 24 (7)	124 (39) / 92 (29) / 80 (25) / 22 (7)	0
Household size, median (IQR)	5 (4–7)	5 (4–7)	0
Own income, n (%) yes	141 (44)	112 (35)	3
Monthly savings (Rs), median (IQR)	300 (100–600)	150 (0–400)	11
Decision-making score (0–10), mean (SD)	5.8 (2.1)	4.9 (2.2)	8
Language, n (%) Bangla	162 (50)	158 (50)	0

Illustrative. Built from one Descriptives run split by treatment, with frequency tables ticked, and the percentages computed from the counts. The Missing column is there because a reader asks.

What it shows

The arms are balanced on the pre-treatment characteristics (age, education, household size, language), as randomisation should make them; the outcomes (own income, savings, decision-making) differ in the direction the programme intends. Whether the difference is more than chance is section 04's question; whether it is large enough to matter is section 10's.

Do not put p-values on a balance table for a randomised design. Differences are chance by construction, and one in twenty will be 'significant'.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Plots

Plots that answer questions: which one, when

Question	Plot	Where in jamovi	Note
What does this variable look like?	Histogram with density	Descriptives → Plots	Skew, bimodality, floor and ceiling effects
Are there outliers, and do groups differ?	Box plot, or violin with box, split by group	Descriptives → Plots → Box plot; Split by	Tick Data points to see every case
Is it roughly normal?	Q-Q plot	Descriptives → Plots	Points on the line: yes; a curve: skew
How are categories distributed?	Bar plot	Descriptives → Plots → Bar plot	For nominal and ordinal variables
Do two continuous variables relate?	Scatter plot with fit line	Regression → Correlation Matrix → Plot; or the scatr module	Add group colour with Split by
Group means with uncertainty	Descriptives plot: mean with CI	T-Tests and ANOVA → Descriptives plots	The plot for a results section
Before and after, per person	Paired plot	Flexplot or esci module	Shows individual change, not only means
Interaction	Estimated marginal means plot	ANOVA → Estimated Marginal Means → Plots	Lines that cross or diverge

Every plot exports as PNG, PDF, SVG or EPS from its right-click menu, at the resolution a journal needs. The plot theme (three-dots menu) sets a consistent look; the default is fine and the greyscale option prints well.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Distributions

Reading a distribution before choosing a test

Monthly savings in the SHG survey: histogram piled at zero and to the left, a long right tail to Rs 5,000, skewness 2.4, Shapiro-Wilk p < .001, Q-Q plot curving away above the line. Reading score at endline: a symmetric hump with a small floor at zero, skewness 0.1, Shapiro-Wilk p = .08. The first calls for a median, a log transform or a non-parametric test; the second is fine for means and t-tests.

Shapiro-Wilk rejects normality for almost any large sample; read it with the plot, not instead of it. With n = 640 a p of .001 can accompany a distribution that is fine for a t-test.
What matters for t-tests and ANOVA is the sampling distribution of the mean, which is close to normal for n above about 30 per group unless the skew is extreme.
Extreme skew, many zeros, or a bounded 0–100 score with a floor: consider the median, a transform, a non-parametric test, or a model built for the shape (section 06 and the modules).

Likert items and scale scores

A single five-point item is ordinal, and its histogram is five bars. A score summing five such items is close enough to continuous for means and t-tests in most applied work, and Norman (Advances in Health Sciences Education 2010, 15:625) reviews the evidence that parametric tests are robust here. Report the item-level frequencies as well, because a mean of 3.2 hides whether the sample is split or clustered in the middle.

Look at every outcome variable's histogram once, split by group. It is the cheapest insurance in statistics.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Ordinal Items

Describing Likert and ordinal items: frequencies, not only means

A three-point decision-making item (no say, joint, alone) or a five-point agreement item is described by its frequency table: the share at each level, split by group. Descriptives with Frequency tables ticked gives it; a bar plot draws it; JASP's Likert plot (Descriptives → Plots) stacks the levels horizontally for several items at once, which is the standard figure for a battery. The mean of the codes (1.32) is a summary a reader cannot picture; '41% decide alone, 39% jointly, 20% have no say' is one they can.

d1: who decides on food purchases	SHG member	Non-member
No say	12%	24%
Jointly	38%	44%
Alone	50%	32%
n	320	316

Illustrative. The shift is from 'no say' and 'jointly' toward 'alone'; the mean of the codes rises from 1.08 to 1.38 and hides the shape.
For a battery of items, one stacked bar per item, ordered by the share at the top level, is the report figure; the surveymv module and JASP's Likert plot draw it.
Test the shift with the ordinal logistic model (section 06) or, descriptively, the chi-square on the 3×2 table.
Sum the items into a scale only after section 07 has shown they belong together.

Means of Likert items are defensible for a scale score and misleading for a single item. Show the distribution.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Crosstabs

Two categorical variables: the contingency table

Frequencies → Contingency Tables → Independent Samples. Rows: education; columns: own income. Under Cells tick Row percentages (or Column, depending on which variable is the 'given'). The table shows counts and percentages; the χ² test of association appears above it, with a p-value, and section 06 covers reading it. As a descriptive tool this is how every 'by' table in a report is made.

Education	Own income: Yes	No	Total
None	62 (26%)	180 (74%)	242
Primary	68 (36%)	120 (64%)	188
Secondary	84 (51%)	80 (49%)	164
Higher	28 (61%)	18 (39%)	46
Total	242 (38%)	398 (62%)	640

Row percentages when the row variable is the explanatory one (education) and you want to compare across its levels.
Layers (a third variable) split the table: education by income within each arm.
Small cells (below 5 expected) make the χ² unreliable; jamovi offers Fisher's exact test as a tick-box.
Export the table to Word via copy; it keeps its structure.

Illustrative. The gradient is the story: own income rises with education from a quarter to three-fifths. The p-value on the χ² (section 06) says the gradient is not chance; the percentages say what it is.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Describing Change

Describing before and after: the paired view

Baseline and endline reading scores for the same children. Descriptives on both gives two means; a computed change score (endline minus baseline) gives the mean change and its distribution, which is what the paired t-test in section 04 tests. But the mean change hides who changed: a paired plot (the esci module's paired-samples estimation, or Flexplot) draws a line per child from baseline to endline, split by arm, and shows whether the gain was general or concentrated among those who started low.

41.2

mean baseline reading score, both arms

Illustrative

+9.8

mean change in the remedial arm (SD of change 12.4)

Illustrative

+3.1

mean change in the control arm (SD 11.9)

Illustrative

Report the baseline mean, the endline mean and the change with its SD, per arm. Three numbers per arm; the fourth (the difference in changes) is the effect and section 04 tests it.
A change score is the right description. It is not always the right analysis: regressing endline on baseline and arm (section 05) is usually more precise, and the course says why there.
Ceiling: a child at 95 cannot gain 20. Check the baseline distribution for children near the top before interpreting small gains among them.
Regression to the mean: children selected for remedial classes because they scored low will score higher next time whatever happens. The control arm is the only protection.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Four

Comparing Groups

ImpactMojoStatistics Without Code 101www.impactmojo.in

Which Test

Choosing the comparison: a decision table

Question	Outcome	Groups	Test (jamovi menu)	Non-parametric alternative
Do two independent groups differ?	Continuous	2, different people	T-Tests → Independent Samples T-Test (Student or Welch)	Mann-Whitney U (same dialog)
Did the same people change?	Continuous	2, same people	T-Tests → Paired Samples T-Test	Wilcoxon signed-rank (same dialog)
Does a mean differ from a known value?	Continuous	1	T-Tests → One Sample T-Test	Wilcoxon (same dialog)
Do three or more groups differ?	Continuous	3+, different people	ANOVA → One-Way ANOVA (Welch or Fisher), or ANOVA for factorial designs	ANOVA → Non-parametric → Kruskal-Wallis
Do groups differ after adjusting for a covariate?	Continuous	2+	ANOVA → ANCOVA	—
Repeated measures over time?	Continuous	3+ occasions, same people	ANOVA → Repeated Measures ANOVA	Non-parametric → Friedman
Two factors at once (arm × sex)?	Continuous	Crossed	ANOVA → ANOVA with two factors	—
Do groups differ on a proportion?	Binary	2+	Section 06: contingency tables, logistic regression	Fisher's exact

The rows are the questions an evaluation asks. Note that 'do the arms differ at endline after accounting for baseline' is ANCOVA (or its regression twin in section 05), and it is usually the right analysis for a before-after-control design, ahead of the paired t-test on change scores.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Independent t

Two groups: the independent-samples t-test, done properly

Question: do SHG members and non-members differ in decision-making score? Path: T-Tests → Independent Samples T-Test; decide_score into Dependent Variables, treatment into Grouping Variable. Tick: Welch's (not Student's, because it does not assume equal variances and costs nothing); Mean difference with Confidence interval; Effect size (Cohen's d); Descriptives; Descriptives plot; Assumption checks: Normality (Shapiro-Wilk), Q-Q plot, Homogeneity (Levene's).

	Statistic	df	p	Mean difference	95% CI	Cohen's d
Welch's t	5.21	631	< .001	0.90	0.56 to 1.24	0.42

Group	N	Mean	SD	SE
SHG member	318	5.80	2.10	0.118
Non-member	314	4.90	2.20	0.124

The sentence

'SHG members scored higher on decision-making than non-members (5.8 vs 4.9 on a 0–10 scale; mean difference 0.90, 95% CI 0.56 to 1.24; Welch's t(631) = 5.21, p < .001; d = 0.42).' The difference and its interval carry the finding; the p-value confirms it is not chance; d says it is a moderate effect.

Illustrative. Note the clustering: women in the same village are not independent, and the CI is somewhat too narrow. Section 08's mixed models fix it; for a report, say it in the limitations.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Assumptions

Assumption checks in the dialog, and what to do when they fail

Check	Where	Reads as	If it fails
Normality (Shapiro-Wilk, Q-Q)	T-Tests and ANOVA → Assumption Checks	p < .05 rejects normality; the Q-Q plot shows how	With n above ~30 per group and moderate skew, proceed; with severe skew or small n, Mann-Whitney or a transform
Equal variances (Levene's)	Same	p < .05: variances differ	Use Welch's t or Welch's ANOVA (both are tick-boxes; use them by default anyway)
Independence	Not a check; a design fact	Clustered by village, school, enumerator?	Mixed model (section 08) or cluster-robust SEs in R; at least say so
Outliers	Box plot	Points beyond the whiskers	Verify the data; report with and without; do not delete silently
Sphericity (repeated measures)	RM ANOVA → Assumption Checks (Mauchly)	p < .05: sphericity violated	Tick Greenhouse-Geisser correction
Homogeneity of regression slopes (ANCOVA)	Add the interaction term and test it	Interaction significant: slopes differ	Report the interaction; ANCOVA's adjusted means are then misleading

Every check is one tick-box away, which is why there is no excuse for a report that does not mention them. The sentence: 'Variances were unequal (Levene's p = .02), so Welch's test was used; distributions were approximately normal by Q-Q plot.'

ImpactMojoStatistics Without Code 101www.impactmojo.in

Paired t

Before and after: the paired t-test, and why it is often not the analysis you want

T-Tests → Paired Samples T-Test; baseline and endline as the pair. Tick mean difference with CI, effect size, descriptives, the Wilcoxon alternative. Illustrative output for the remedial arm: mean change +9.8, 95% CI 8.8 to 10.8, t(599) = 19.4, p < .001, d = 0.79. For the control arm: +3.1, CI 2.2 to 4.0. Each is a within-group change, and each is 'significant', and neither is the programme effect.

The programme effect is the difference between the two changes (9.8 − 3.1 = 6.7 points), which the paired test does not estimate.
Two ways to get it in jamovi: an independent t-test on the change score by arm (simple; correct for a randomised design), or ANCOVA of endline on arm with baseline as covariate (more precise, section 05).
A paired test on the treatment arm alone is the single most common wrong analysis in NGO evaluation reports: it attributes secular growth, practice effects and regression to the mean to the programme.

When the paired test is right

When there is no control group and the question is descriptive: 'did scores change?' Then report it as change, not as effect. When the design is a crossover or a within-person comparison of two conditions. When the 'pair' is a matched pair of villages. In each case the sentence is about the pair, and the causal claim is the design's, not the test's.

The difference-in-differences in a menu: compute change, independent t-test by arm. Illustrative result: difference in changes 6.7 points, 95% CI 5.3 to 8.1, d = 0.55.

ImpactMojoStatistics Without Code 101www.impactmojo.in

ANOVA

Three or more groups: one-way ANOVA and post hoc comparisons

Question: does decision-making differ by education level? Path: ANOVA → ANOVA; decide_score as Dependent, educ as Fixed Factor. Tick: Effect size (ω² is less biased than η²); Assumption Checks; Post Hoc Tests with educ, Tukey correction, Effect size; Estimated Marginal Means with plot and table.

Source	SS	df	MS	F	p	ω²
educ	186.4	3	62.1	13.8	< .001	0.057
Residuals	2,862.0	636	4.5

Post hoc (Tukey)	Mean difference	SE	p_tukey	d
Primary − None	0.52	0.21	0.06	0.24
Secondary − None	1.14	0.22	< .001	0.53
Higher − None	1.61	0.35	< .001	0.75
Secondary − Primary	0.62	0.23	0.03	0.29

The sentence

'Decision-making differed by education, F(3, 636) = 13.8, p < .001, ω² = 0.06. Tukey-corrected comparisons showed higher scores for women with secondary (+1.1, p < .001) and higher (+1.6, p < .001) education than for those with none; the primary-none difference was small and not significant (+0.5, p = .06).' The estimated marginal means plot is the figure.

Illustrative. The omnibus F says something differs; the post hoc table says what. Report both, and the effect sizes, not the F alone.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Factorial

Two factors: does the programme work differently for girls and boys?

ANOVA → ANOVA; endline as Dependent; arm and sex as Fixed Factors; jamovi builds the model with both main effects and the interaction. Tick Estimated Marginal Means for arm × sex with a plot. The interaction term is the question: if it is significant, the programme's effect differs by sex, and the marginal-means plot shows how (lines not parallel). If not, report the main effects.

Source	F	df	p	ω²
arm	48.2	1, 1196	< .001	0.038
sex	3.1	1, 1196	0.08	0.002
arm × sex	0.6	1, 1196	0.44	0.000

Illustrative: a clear arm effect, no evidence that it differs by sex (interaction p = .44, ω² near zero). The sentence says so, and does not go on to test the effect 'separately for girls and boys', which invites finding a difference by chance.
Add baseline as a Covariate in the same dialog and the model becomes ANCOVA with factors, the standard evaluation model.
Type III sums of squares (the default) are right for unbalanced designs; leave them.
Simple effects (the effect of arm within each sex) are an option under Estimated Marginal Means; report them only when the interaction is significant.

Testing for heterogeneity by sex, caste, grade and district is eight interactions. One will be 'significant'. Pre-specify the one that matters.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Repeated Measures

Repeated measures: the same children at three points

ANOVA → Repeated Measures ANOVA. Define the within-subjects factor (time: baseline, midline, endline), drag the three score columns into its cells, add arm as a Between-Subjects Factor. Tick Assumption Checks (Mauchly's sphericity) and apply Greenhouse-Geisser if it fails; Estimated Marginal Means for time × arm with a plot. The time × arm interaction is the programme effect over time.

Data must be wide: one row per child, one column per occasion. Long data (one row per child-occasion) needs reshaping first, which jamovi cannot do; do it in the source or in Rj.
Any child missing an occasion is dropped entirely. With attrition, the mixed model (section 08) keeps them and is the better tool.
The sphericity assumption (equal variances of all pairwise differences) fails routinely; the correction is one tick.

Illustrative reading

Time × arm: F(1.8, 2,100) = 21.4, p < .001 after Greenhouse-Geisser (ε = 0.91). The plot: both arms rise from baseline to midline; the remedial arm keeps rising to endline while the control flattens. The effect appears in the second half of the year. That is a finding the two-wave design could not have produced, and it is the reason to collect a midline.

Report the marginal means at each occasion by arm with CIs, in a table and a plot. The F statistic is the footnote.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Non-parametric

Non-parametric tests: when, and what they actually test

Test	Replaces	Tests	Effect size
Mann-Whitney U	Independent t	Whether one group tends to have higher values (stochastic dominance), not the medians unless shapes match	Rank-biserial correlation (tick-box)
Wilcoxon signed-rank	Paired t	Whether differences tend to be positive	Rank-biserial
Kruskal-Wallis	One-way ANOVA	Whether any group tends higher	ε² (tick-box); Dwass-Steel-Critchlow-Fligner pairwise comparisons
Friedman	Repeated-measures ANOVA	Whether occasions differ in rank	Durbin-Conover pairwise
Spearman's ρ	Pearson's r	Monotonic association	ρ itself

Use them when the outcome is ordinal with few levels, when the sample is small and the distribution clearly non-normal, or when outliers dominate and cannot be resolved. Do not use them because Shapiro-Wilk was significant in a sample of 600. They lose a little power when the t-test would have been valid and they answer a slightly different question, which the report should state in words: 'savings tended to be higher among members (Mann-Whitney U = 38,240, p < .001; rank-biserial r = 0.24)', with the medians and IQRs alongside.

For skewed money variables, the alternative is often better than either: a t-test on the log (a ratio of geometric means), or a median regression in R.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Comparison Report

Reporting a group comparison: the table and the figure

Outcome	Remedial, mean (SD), n	Control, mean (SD), n	Difference (95% CI)	Test	Effect size
Endline reading (0–100)	51.0 (16.2), 600	44.3 (16.8), 600	6.7 (4.8 to 8.6)	Welch t(1,196) = 7.0, p < .001	d = 0.41
Change from baseline	9.8 (12.4), 600	3.1 (11.9), 600	6.7 (5.3 to 8.1)	Welch t(1,196) = 9.5, p < .001	d = 0.55
Attendance (%)	82.1 (11.0), 600	80.9 (11.6), 600	1.2 (−0.1 to 2.5)	Welch t(1,196) = 1.8, p = .07	d = 0.11
Reads a paragraph, n (%)	372 (62)	294 (49)	13 points (7 to 18)	χ²(1) = 20.4, p < .001	RR 1.27

Illustrative. One row per outcome, the same columns for each, the difference and its interval before the test, the effect size last. The last row is section 06's.

The figure

The descriptives plot from the t-test dialog: two means with 95% CIs, one per arm, for the primary outcome. Or, better, the esci module's estimation plot, which shows the two distributions, the means, and the difference with its interval on a separate axis. Export as SVG.

Clustering by school makes every interval here too narrow. Section 08 shows the mixed model that corrects it, and the report's limitations must mention it if the mixed model was not run.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Five

Correlation and Regression

ImpactMojoStatistics Without Code 101www.impactmojo.in

Correlation

The correlation matrix: Pearson, Spearman, and the plot

Regression → Correlation Matrix; drag in the continuous variables. Tick Pearson (for linear relationships between roughly normal variables), Spearman (for monotonic relationships, ordinal variables or outliers), Report significance, Confidence intervals, Flag significant, and Plot (a scatter matrix with densities and fit lines). Kendall's tau-b is there for small samples with many ties.

	Baseline	Endline	Attendance
Baseline	—
Endline	0.71 [0.68, 0.74]	—
Attendance	0.18 [0.12, 0.23]	0.24 [0.19, 0.29]	—

Illustrative. Baseline predicts endline strongly (r = 0.71), which is why ANCOVA is precise. Attendance relates weakly to both.
Report r with its CI; the p-value adds nothing when n = 1,200 and everything is 'significant'. r = 0.18 is a small association whatever its p.
The scatter plot is not optional. Anscombe's quartet has four datasets with identical r and entirely different shapes.
Correlation is symmetric and says nothing about direction or cause. Every phrase in the report should be 'associated with'.

A correlation matrix with twenty variables and stars on the significant ones is a fishing net. Pick the correlations the question needs.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Linear Regression

Linear regression: the dialog, the output, and the sentence

Question: what is the programme effect on endline reading, adjusting for baseline and grade? Path: Regression → Linear Regression; endline as Dependent; baseline into Covariates; arm and grade into Factors. Tick: Model Fit (R², adjusted, F); Model Coefficients: Confidence interval, Standardized estimate; Assumption Checks: Q-Q plot of residuals, Residual plots, Collinearity statistics, Durbin-Watson if ordered; Estimated Marginal Means for arm.

Predictor	Estimate	SE	95% CI	t	p	Std. β
Intercept	14.6	1.4	11.8 to 17.4	10.3	< .001
baseline	0.72	0.02	0.68 to 0.76	33.5	< .001	0.71
arm: Remedial − Control	6.6	0.68	5.3 to 7.9	9.7	< .001	0.20
grade: 4 − 3	1.9	0.83	0.3 to 3.5	2.3	0.02
grade: 5 − 3	3.4	0.84	1.8 to 5.1	4.1	< .001

R² = 0.55, adjusted 0.55, F(4, 1,195) = 362, p < .001. Illustrative.

The sentence

'Adjusting for baseline score and grade, children in the remedial arm scored 6.6 points higher at endline (95% CI 5.3 to 7.9, p < .001), about 0.4 SD of the baseline distribution. Baseline score was the strongest predictor (0.72 points per baseline point). The model explained 55% of the variance in endline scores.'

This is ANCOVA in regression form and the standard analysis for a randomised two-wave design. Its CI (5.3 to 7.9) is narrower than the change-score t-test's (5.3 to 8.1) because baseline is used as a covariate rather than subtracted with coefficient 1.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Diagnostics

Regression diagnostics: the plots and what they say

Check	Where	Good	Bad, and the fix
Residuals vs fitted	Assumption Checks → Residual plots	A flat band of points	A curve: non-linearity, add a squared term or transform; a funnel: heteroskedasticity, use robust SEs (Rj: sandwich) or transform the outcome
Q-Q plot of residuals	Same	Points on the line	Heavy tails or skew: check outliers; for large n the estimates are fine, the intervals slightly off
Collinearity (VIF, tolerance)	Assumption Checks → Collinearity statistics	VIF below 5, ideally below 3	Two predictors measure the same thing; drop one or combine
Influential cases (Cook's distance)	Assumption Checks → Cook's distance	All small	A few cases drive the fit; verify them, report with and without
Autocorrelation (Durbin-Watson)	Assumption Checks	Near 2	Only for ordered data; for clustered data the fix is a mixed model
Linearity of each covariate	Scatter plots; add a squared term	No improvement	Curvilinear: keep the square, interpret with a plot

Three ticks, thirty seconds. The two most common problems in evaluation data are a bounded outcome with floor effects (the residual plot shows it) and clustering (no plot shows it; you know it from the design). The first argues for a plot of the outcome's distribution in the report; the second for section 08.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Categorical Predictors

Factors in regression: dummy coding, reference levels and contrasts

A nominal predictor dragged into Factors is dummy-coded automatically: one coefficient per level against a reference level, which is the first level in the variable's order. Reference Levels in the dialog changes it; for education, 'None' is the natural reference and the coefficients read as 'Primary versus None', 'Secondary versus None'. For an ordinal factor jamovi offers polynomial contrasts (linear, quadratic trends); for most reporting the simple (dummy) contrasts are clearer.

The estimated marginal means (tick it) give the adjusted mean outcome at each level, which is what a report table wants, with CIs.
An ordinal predictor with many levels (age bands) can be entered as a factor or as a numeric covariate; the factor is safer, the covariate is one number.
Interactions: Model Builder → select two terms → add as interaction. Section 04's factorial logic applies, with the marginal-means plot as the reading.

The Model Builder

Regression → Linear Regression → Model Builder lets you add predictors in blocks and compares the blocks: R² change and its F-test from adding the programme arm to a model that already has baseline and grade. That is the 'what does the programme add' question in one table, and it is where hierarchical regression in a thesis lives.

Reference level: choose the one the comparison is about, and say which in the table note. 'Female' as reference means the coefficient is 'Male minus Female'.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Interpreting

Reading coefficients: units, standardisation, and what 'controlling for' means

Unstandardised (Estimate): the change in the outcome, in its units, per unit of the predictor, other predictors held constant. 6.6 reading points for the programme; 0.72 endline points per baseline point.
Standardised (Std. β): the change in SDs of the outcome per SD of the predictor. Useful to compare predictors within a model; meaningless for a binary predictor's 'SD'.
Holding constant means comparing children with the same baseline and grade. It does not mean the programme effect is causal unless the design made it so; here it did.
R² is how much variance the model explains; a small R² with a precise, important coefficient is a good model of an outcome with many causes.
Non-significant is not 'no effect'; the CI says what effects are compatible. Attendance: 1.2 points, CI −0.1 to 2.5. Small effects are not ruled out.

Observational data

The same regression on the SHG survey, decide_score on membership, education, age and own income, gives a membership coefficient adjusted for those three. Membership was randomised by village, so it is a causal effect. Own income was not, so its coefficient is an association: women with income decide more, or women who decide more earn income, or both come from something unmeasured. The dialog does not know the difference. The analyst does, and the report's verbs must show it.

Econometrics 101 and Impact Evaluation 101 in this series are the courses on when a coefficient is an effect. This course is on how to get the coefficient.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Curves and Logs

Non-linear relationships: logs, squares, and reading them

Savings in rupees on age: the scatter plot shows savings rising to the forties and falling after. A straight line misses it. Compute age_sq (age^2) and add both as covariates; the coefficient on the square is negative and significant; the estimated marginal means plot (or the Flexplot module) draws the curve. Alternatively compute the log of savings as the outcome, in which case coefficients read as proportional changes: 0.05 means about 5% more savings per year of age.

Log outcome: coefficients are approximately percentage changes for small values; exactly (exp(b) − 1) × 100.
Log predictor: the coefficient is the change in the outcome per 1% change in the predictor, divided by 100.
Zeros: log(x + 1) is common and changes the interpretation for small x; say so. Half the sample saving zero is a two-part question (do they save; how much), and section 06's logistic model answers the first part.
A square term's peak is at −b₁/(2b₂); compute it and report it in the predictor's units.

The plot decides. If the scatter with a smooth line (Correlation Matrix → Plot, or Flexplot) is straight, the linear model is fine; if it bends, model the bend; if it does something no polynomial fits, the relationship may be a threshold, and a factor with cut points is more honest than a curve.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Regression Report

Reporting a regression: the table a thesis examiner expects

	Model 1: baseline only	Model 2: + arm	Model 3: + grade, sex, attendance
Baseline score	0.74 (0.71, 0.77)	0.73 (0.70, 0.76)	0.72 (0.68, 0.76)
Remedial arm		6.7 (5.4, 8.0)	6.6 (5.3, 7.9)
Grade 4 (vs 3)			1.9 (0.3, 3.5)
Grade 5 (vs 3)			3.4 (1.8, 5.1)
Boy (vs girl)			−0.8 (−2.1, 0.5)
Attendance (per 10 points)			0.9 (0.3, 1.5)
R²	0.50	0.54	0.55
ΔR² (p)		0.04 (< .001)	0.01 (< .001)
n	1,200	1,200	1,196

Illustrative. Unstandardised coefficients with 95% CIs; the three models from the Model Builder in three columns; the note states the reference levels, the outcome's units and the clustering caveat.

Why three columns

The arm effect barely moves as covariates are added (6.7 to 6.6), which is what randomisation predicts and what a reader checks. In an observational study, a coefficient that halves when a control is added was confounded by it, and the columns show that too. The ΔR² row says what each block added.

Stars are optional; intervals are not. A coefficient with a CI is readable without the p-value; the reverse is not true.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Six

Categorical Outcomes

ImpactMojoStatistics Without Code 101www.impactmojo.in

Proportions

One proportion and a goodness-of-fit: the Frequencies menu

Frequencies → Proportion Test (2 Outcomes) tests whether a binary variable's proportion differs from a value: is the share of children who can read a paragraph different from the state's 50% figure? Tick the confidence interval. Frequencies → N Outcomes (χ² Goodness of Fit) tests whether the distribution across categories matches expected proportions: does the sample's caste distribution match the district's census shares? Enter the expected proportions in the dialog.

Illustrative: 666 of 1,200 children read a paragraph, 55.5%, 95% CI 52.7 to 58.3; against 50%, p < .001.
The CI is the result; the test against an arbitrary value rarely is.
Goodness-of-fit against census shares is a representativeness check for a survey, and belongs in the methods.

Binary variables as continuous

A 0/1 variable set as continuous has a mean equal to the proportion, and Descriptives gives it with a CI, split by group, in one step. A t-test on a 0/1 outcome by arm estimates the difference in proportions with a CI and is close to the two-proportion z-test; it is a defensible shortcut for a report, and the contingency table (next slide) is the formal version.

Report proportions with n and CI: '55.5% (666 of 1,200; 95% CI 52.7 to 58.3)'. Never a percentage without its n.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Chi-square

Two categorical variables: the chi-square test of association

Question: does the share of children reading a paragraph differ by arm? Path: Frequencies → Contingency Tables → Independent Samples; arm in Rows, reads_para in Columns. Tick: χ², χ² continuity correction (for 2×2), Fisher's exact; Comparative Measures: Odds ratio, Relative risk, with CIs; Nominal: Phi and Cramer's V; Cells: Row percentages, Expected counts.

Arm	Reads: Yes	No	Total	Row %
Remedial	372	228	600	62.0%
Control	294	306	600	49.0%
Total	666	534	1,200	55.5%

Statistic	Value	df	p
χ²	20.6	1	< .001
Relative risk (Remedial vs Control)	1.27		95% CI 1.15 to 1.39
Odds ratio	1.70		95% CI 1.35 to 2.14
Cramer's V	0.13

The sentence

'Children in the remedial arm were more likely to read a paragraph at endline (62% vs 49%; difference 13 percentage points; risk ratio 1.27, 95% CI 1.15 to 1.39; χ²(1) = 20.6, p < .001).' The percentage-point difference and the risk ratio are what a programme reader can use; the odds ratio overstates the effect when the outcome is common, as here.

Illustrative. Cramer's V of 0.13 is a small association by convention and a 13-point difference in reading is not small in policy. Effect-size conventions are not the last word.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Small Cells

Small samples, small cells, and paired categorical data

The χ² approximation fails when expected counts are small (below 5 in more than 20% of cells is the old rule). Tick Expected counts to see them; use Fisher's exact test (a tick-box) for 2×2 tables with small cells, and collapse categories or use Fisher's generalisation for larger tables.
Larger tables (education × income, 4×2): the χ² says the variables are associated; the row percentages say how; Cramer's V says how strongly. Post hoc, compare the standardised residuals (available as a tick-box) to see which cells drive it.
Ordered categories (education × a Likert item): the χ² ignores the order; Spearman's correlation or an ordinal regression uses it and has more power.
Paired categorical data (the same child at baseline and endline, reads or not): Frequencies → Paired Samples (McNemar test). The question is whether more children moved from No to Yes than from Yes to No.

Layers

Add a third variable to Layers and the table repeats within each level: arm × reads within each grade. jamovi reports a χ² per layer. This is how a report shows that the programme's effect on reading held in every grade, or did not, without a logistic model. It is also where Simpson's paradox lives: an association in the pooled table that reverses within every layer, because the layer variable confounds it.

Illustrative McNemar: of 600 remedial children, 190 moved from not reading to reading and 22 the other way; p < .001. The 22 are worth a look.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Logistic

Logistic regression: a binary outcome with several predictors

Path: Regression → Logistic Regression → 2 Outcomes (Binomial); reads_para as Dependent; baseline into Covariates; arm, grade, sex into Factors. Tick: Model Fit (deviance, AIC, McFadden's R²); Model Coefficients: Odds ratio with CI; Estimated Marginal Means for arm (these are predicted probabilities, and the useful output); Prediction: Classification table, Accuracy, AUC and ROC curve; Assumption Checks: Collinearity.

Predictor	Estimate (log odds)	SE	Odds ratio	95% CI	p
Intercept	−2.84	0.23			< .001
baseline (per point)	0.064	0.005	1.07	1.06 to 1.08	< .001
arm: Remedial	0.61	0.13	1.84	1.42 to 2.38	< .001
grade: 4 vs 3	0.28	0.16	1.32	0.97 to 1.81	0.08
grade: 5 vs 3	0.55	0.16	1.73	1.26 to 2.38	< .001
sex: Boy	−0.09	0.13	0.91	0.71 to 1.17	0.47

Estimated marginal means: predicted probability of reading, remedial 0.63 (0.59 to 0.67), control 0.48 (0.44 to 0.52), at mean baseline. AUC 0.79. Illustrative.

The sentence

'Adjusting for baseline score, grade and sex, children in the remedial arm had 1.8 times the odds of reading a paragraph (95% CI 1.4 to 2.4, p < .001); the predicted probability was 63% versus 48% at the mean baseline score, a difference of 15 percentage points.' Give the probabilities; nobody thinks in odds.

The odds ratio (1.84) is larger than the risk ratio (1.27) because the outcome is common. Reporting the odds ratio as if it were a risk ratio ('84% more likely') is a routine misreading.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Reading Logistic

Reading logistic output: odds, probabilities, and fit

The coefficient is a change in log odds; exponentiated, an odds ratio; neither is a change in probability. The change in probability depends on where you start, which is why the marginal means (predicted probabilities at set values) are the reportable result.
McFadden's R² is not variance explained and its values are low (0.2 is good); do not compare it with a linear R².
AUC is the probability that a random reader has a higher predicted probability than a random non-reader; 0.5 is chance, above 0.8 is strong. It measures discrimination, not calibration.
The classification table at a 0.5 cut-off is often misleading when the outcome is unbalanced; report it, but do not choose the model on it.
Separation (a predictor that perfectly predicts the outcome) produces enormous coefficients and SEs; the fix is a penalised model in R, or collapsing the predictor.

Three or more categories

Regression → Logistic Regression → N Outcomes (Multinomial) for unordered categories (main occupation: farm, non-farm, none); Ordinal Outcomes for ordered ones (a five-point satisfaction item, or a decision-making item scored 0/1/2). The ordinal model (proportional odds) gives one odds ratio per predictor for being in a higher category and assumes it is the same across the thresholds; that assumption is not tested in jamovi's dialog and should be checked in R for a paper.

The ordinal model is the right analysis of a single Likert item as an outcome. A t-test on it is the common one.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Rates and Counts

Counts and rates: when the outcome is 'how many'

Number of clinic visits, number of days absent, number of loans: counts, bounded at zero, skewed, often with many zeros. A linear regression on a count gives negative predictions and wrong intervals. The jamovi core has no Poisson dialog; the GAMLj module's Generalized Linear Model does Poisson and negative binomial, with an offset for exposure (days at risk) and coefficients that exponentiate to rate ratios.

Poisson assumes variance equals the mean; counts almost always have more variance (overdispersion), and negative binomial handles it. GAMLj reports the dispersion.
Many zeros beyond what the model expects: a zero-inflated or hurdle model, in R. Or split the question: any visits (logistic) and how many among those with any (count).
A rate ratio of 1.3 reads as '30% more visits per child-year in the treatment arm'.

The bounded-score problem

A reading score from 0 to 100 with a pile at 0 and a pile at 100 is neither a count nor a continuous normal variable. Linear regression is usually tolerable for the middle and wrong at the ends; a beta regression or a Tobit is the formal answer, in R. For a report, the linear model with a plot of the distribution by arm and a sentence about the floor is honest.

GAMLj, by Marcello Gallucci, is the single most useful module to install. Section 08.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Attrition Table

A worked attrition table: who was lost, by arm

	Remedial	Control	Test
Enrolled at baseline	640	640
Endline score available	600 (93.8%)	600 (93.8%)	χ²(1) = 0.0, p = 1.00
Baseline score of those lost, mean (SD)	34.1 (15.0), n = 40	36.2 (15.8), n = 40	Welch t(78) = 0.6, p = .54
Baseline score of those retained	41.7 (16.0)	41.5 (16.3)
Lost vs retained, pooled	35.2 vs 41.6		Welch t(94) = 3.5, p < .001

Illustrative. Built from one contingency table (arm × missing-endline indicator) and two t-tests using a computed indicator and filters. Attrition is equal across arms and lost children were weaker at baseline in both arms equally.

The sentence

'Attrition was 6.2% in both arms. Children lost to follow-up had lower baseline scores than those retained (35.2 vs 41.6, p < .001) but did not differ by arm (p = .54), so the comparison between arms is unlikely to be biased by attrition; the endline means overstate the level for the full cohort.' Four numbers, one paragraph, and the evaluation's main threat addressed.

Every evaluation report needs this table and most do not have it. It costs three dialogs.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Categorical Report

Reporting categorical results: the checklist

Result	Report	Not
A proportion	% with n and 95% CI	% alone
Two proportions	Both %, the difference in points with CI, the risk ratio with CI, the test	The odds ratio alone when the outcome is common
A table	Counts and row or column % (say which), the χ² with df and p, Cramer's V	Percentages with no counts
A logistic model	Odds ratios with CIs, predicted probabilities at meaningful values, n, AUC	'Significant predictors' with stars
Small cells	Fisher's exact; the cells collapsed and why	A χ² with expected counts of 2
Attrition	By arm, with baseline comparison of lost and retained	'Some children could not be followed up'
Ordinal outcome	The ordinal model's odds ratio, or the item distribution by group	A mean of the codes as if continuous, without saying so

The pattern is the same as section 04's: the estimate and its interval first, the test second, the effect size in units a reader can use. For categorical outcomes the usable unit is the percentage point.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Seven

Scales and Factor Analysis

ImpactMojoStatistics Without Code 101www.impactmojo.in

Reliability

Reliability analysis: alpha, omega, and the items that hurt

Path: Factor → Reliability Analysis; d1 to d5 into Items. Tick: Cronbach's α, McDonald's ω; Item statistics: item-rest correlation, α if item dropped, ω if item dropped; Correlation heatmap. Reverse-scaled items go in their own box and are flipped automatically.

Item	Mean	SD	Item-rest r	α if dropped	ω if dropped
d1 food purchases	1.32	0.71	0.58	0.74	0.75
d2 large purchases	1.04	0.78	0.64	0.72	0.73
d3 visiting family	1.21	0.74	0.60	0.73	0.74
d4 own health care	1.18	0.76	0.55	0.75	0.76
d5 children's schooling	1.55	0.62	0.31	0.81	0.82
Scale				α = 0.79	ω = 0.80

Reading it

Alpha 0.79 and omega 0.80 are acceptable. Item d5 correlates weakly with the rest (0.31) and the scale would be more reliable without it (0.81). That is a reason to look at d5, not a reason to delete it: SEM 101's discussion of that item applies. Report ω over α when they differ, because α assumes equal loadings.

Illustrative. Ordinal items with three categories: the reliability from the raw scores understates the ordinal reliability. jamovi's dialog uses the raw scores; say so, or compute the ordinal version in Rj.

ImpactMojoStatistics Without Code 101www.impactmojo.in

EFA

Exploratory factor analysis: how many factors, and which items go together

Factor → Exploratory Factor Analysis; the twelve empowerment items into Variables. Extraction: minimum residuals (default) or maximum likelihood; Rotation: oblimin (the factors may correlate); Number of factors: based on parallel analysis (the recommended default), or eigenvalue above 1, or a fixed number to compare. Tick Assumption Checks (Bartlett's, KMO), Factor summary, Factor correlations, Model fit measures, and Hide loadings below 0.3 for readability. Scree plot under Additional Output.

KMO above 0.7 and Bartlett's p < .001 say the items correlate enough to factor. KMO below 0.6 says stop.
Parallel analysis compares the eigenvalues to those from random data of the same size; it is the one factor-count rule that survives scrutiny.
Principal Component Analysis is a separate dialog (Factor → PCA) and is not factor analysis; use it for data reduction, not for scale structure.

Illustrative result

Parallel analysis: three factors. After oblimin rotation, d1 to d5 load on factor 1 (0.45 to 0.79, d5 lowest), m1 to m3 on factor 2 (0.68 to 0.83), a1 to a4 on factor 3 (0.52 to 0.74); one cross-loading (a1 on factor 1, 0.32). Factor correlations 0.31 to 0.46. The scale has the structure it was designed with; d5 and a1 are the items to discuss. Uniqueness above 0.8 on an item means the factors explain little of it.

EFA on the pilot or on half the sample; CFA (next slide) on the rest. Both in one dialog each.

ImpactMojoStatistics Without Code 101www.impactmojo.in

CFA

Confirmatory factor analysis in the jamovi core

Factor → Confirmatory Factor Analysis. Create three factors, name them, drag the items into each. Estimates: Standardized estimate, Factor covariances; Model Fit: Test for exact fit (χ²), CFI, TLI, RMSEA with CI, SRMR; Additional Output: Residual covariances, Modification indices, Path diagram. The estimator is maximum likelihood; missing data by full-information ML (a tick) or exclude cases listwise.

Fit measure	Value	Reads as
χ² (df), p	118.6 (51), < .001	Exact fit rejected, as expected at n = 640
CFI	0.955	Above 0.95
TLI	0.942	Just under 0.95
RMSEA [90% CI]	0.046 [0.035, 0.057]	Good
SRMR	0.045	Good

Illustrative. The structure holds by every approximate-fit measure; the residual covariances show where the small misfit is.
jamovi's core CFA uses ML only, which is not ideal for three-category items; the SEMLj module (section 08) adds the ordinal estimator (WLSMV) and full SEM.
Modification indices are shown; the reason to look at them and not to act on them is in SEM 101, section 06.
The path diagram is a report figure; export it as SVG.

Reliability, EFA and CFA in three core dialogs: jamovi covers scale validation to the level a master's thesis needs, without SPSS or AMOS.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Scale Scores

From factor to score: what to compute and use downstream

Sum or mean of items: Compute with MEAN(); transparent, comparable across studies, weights every item equally. The right choice for most reports.
Factor scores: EFA and CFA dialogs can save them (a tick under Additional Output writes them to the data as new columns). Weighted by loadings; standardised; not comparable across samples; fine for a regression within this study.
Standardised sum: Z() of the mean score, so effects read in SDs.
Categorised ('high' vs 'low' empowerment at the median): loses information, invites arbitrary cut points, and should be avoided unless a threshold has meaning.

Sub-scales

Three factors give three scores (decision-making, mobility, assets) and a total. Report the programme's effect on each: a programme that moves mobility and not decision-making is a finding the total hides. Three t-tests or one ANCOVA per sub-scale, with a note that three tests were run. The MANOVA dialog (ANOVA → MANCOVA) tests them jointly, and is rarely what a reader wants.

Whatever the score, its reliability and its construction go in the methods, with the formula. A reviewer will ask.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Translation Check

Comparing languages: the invariance check jamovi can and cannot do

The survey ran in Hindi and Bangla. Before comparing scores across languages, the items should work the same way in each. In the jamovi core: run the CFA with Split by language and compare the loadings and fit by eye; run reliability split by language. That is the configural check and a rough metric check. The formal invariance sequence (metric, scalar, with ΔCFI) is in the SEMLj module's multigroup option, or in R with lavaan and semTools.

Illustrative: loadings within 0.1 of each other in both languages for ten items; d3 loads 0.72 in Hindi and 0.51 in Bangla; α 0.80 and 0.76.
A loading that differs by 0.2 is an item to take back to the translators and the cognitive-interview notes.
Report the split CFA in the methods and the formal test if the comparison across languages is a result of the paper.

SEM 101, section 08, is the full treatment. The point here is that the first, informative look costs one tick (Split by) and most multilingual surveys never take it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

IRT

Item response theory in a module: when items should not be summed

A reading test where items range from letters to paragraphs is not a set of interchangeable questions; a child who reads the paragraph has answered the easier items too. Item response theory models each item's difficulty and discrimination and places children on a common ability scale. The snowIRT and psychoPDA modules bring Rasch and 2PL/graded response models into jamovi: item parameters, item characteristic curves, person ability estimates, fit statistics, and differential item functioning by group.

Use it for ability and knowledge tests, and for scales where items differ sharply in severity (food insecurity: worrying about food versus going a whole day without).
The ability estimate is a better outcome than the raw score when items differ in difficulty and when tests differ across waves.
DIF by language or sex is the IRT version of measurement invariance, item by item.

Item Response Theory 101 in this series is the course; the module is where a jamovi user can run its examples without code. Install: Modules → jamovi library → search 'IRT'.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Scale Report

Reporting a scale: the paragraph and the table

The paragraph

'Decision-making was measured with five items (Appendix A) scored 0 to 2 and summed (range 0–10; mean of available items used where one item was missing, n = 14). Internal consistency was acceptable (McDonald's ω = 0.80, Cronbach's α = 0.79). Exploratory factor analysis on the pilot (n = 180) and confirmatory factor analysis on the main sample supported a three-factor structure across the twelve empowerment items (CFI = 0.955, RMSEA = 0.046, SRMR = 0.045); item d5 loaded weakly (0.41) and was retained for comparability with the source scale. Loadings were similar in the Hindi and Bangla versions except for item d3 (0.72 vs 0.51).'

Scale	Items	Range	ω	α	Mean (SD)	Missing
Decision-making	5	0–10	0.80	0.79	5.4 (2.2)	8
Mobility	3	0–6	0.81	0.80	3.1 (1.6)	5
Control over assets	4	0–8	0.76	0.75	3.8 (1.9)	11

Illustrative. The paragraph and the table between them answer every question a methods reviewer asks about a scale, and every number in them came from the three Factor dialogs.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Eight

The Modules: Beyond the Basics

ImpactMojoStatistics Without Code 101www.impactmojo.in

The Library

The jamovi library: modules worth installing, and what each adds

Module	Adds	Use it for
GAMLj (Gallucci)	General linear model, generalized linear model (logistic, Poisson, negative binomial, ordinal, multinomial), mixed models, generalized mixed models; simple effects, plots, contrasts	Clustered data (children in schools), counts, and any regression with interactions plotted properly
jAMM (Gallucci)	Mediation and moderation with bootstrapped indirect effects, conditional effects, path diagrams	Mechanism questions
SEMLj (Gallucci and Jentschke)	Full SEM via lavaan: syntax or diagram interface, ordinal estimators, multigroup, invariance, fit	Everything in SEM 101
MAJOR (Hamilton)	Meta-analysis via metafor: effect sizes from various inputs, random effects, forest and funnel plots, moderators	Everything in Systematic Reviews 101, section 07
esci (Cumming and Calin-Jageman)	Estimation statistics: effect sizes with CIs and the plots that show them	Reports that lead with the difference, not the p
Flexplot (Fife)	Graphics-first modelling: scatter, paired, interaction plots with fitted models	Seeing the model before fitting it
jpower	Power analysis for t-tests and proportions	Section 10
snowIRT, psychoPDA	Item response theory	Section 07
Rj	An R editor inside jamovi	The one thing the menus lack
TOSTER	Equivalence tests	Showing two things are the same, not merely not different
walrus	Robust statistics (trimmed means, bootstrapped tests)	Outlier-heavy data
Distraction	Distributions: plots and probabilities	Teaching

Modules → jamovi library → Install. Each appears as a new menu under Analyses. Module versions matter for reproducibility; note them in the methods.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Mixed Models

Clustered data: the mixed model in GAMLj

Children in 40 schools; the programme was assigned by school; children in one school resemble each other. Every t-test and regression so far has treated the 1,200 children as independent and produced intervals that are too narrow. Path: Linear Models (GAMLj) → Mixed Model; endline as Dependent; baseline as Covariate; arm and grade as Factors; school_id as Cluster variable; in Random Effects, add the intercept by school. Tick Fixed effects estimates with CI, Random components (variance, ICC), Estimated marginal means.

Effect	Estimate	SE	95% CI	p	Compare: OLS CI
baseline	0.72	0.02	0.68 to 0.76	< .001	0.68 to 0.76
arm: Remedial	6.5	1.42	3.6 to 9.4	< .001	5.3 to 7.9
Random intercept SD (school)	4.1
Residual SD	10.6
ICC	0.13

Reading it

The programme estimate barely changes (6.6 to 6.5); its interval doubles in width (5.3–7.9 to 3.6–9.4), because with 40 schools the effective sample for a school-level treatment is 40, not 1,200. The ICC of 0.13 says 13% of the variance in endline scores is between schools. This is the analysis the design requires, and the interval it gives is the one to report.

Illustrative. A cluster-randomised trial analysed without accounting for clusters is the most common statistical error in education evaluations, and the module fixes it in one dialog.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Mediation

Mediation in jAMM: membership, savings, decision-making

Medmod (core) does a simple mediation with a bootstrap; jAMM does the general case. Path: jAMM → GLM Mediation Model; decide_score as Dependent; savings_rs as Mediator; treatment as Factor; age, educ as Covariates. Tick Bootstrap (percentile, 5,000), confidence intervals, the path diagram, and the individual regressions. Output: the a, b, c′ paths, the indirect effect with its bootstrap CI, the total effect.

Effect	Estimate	95% CI (bootstrap)
a: treatment → savings	142 Rs	98 to 186
b: savings → decide (per 100 Rs)	0.21	0.13 to 0.29
Indirect (a × b)	0.30	0.18 to 0.44
Direct (c′)	0.60	0.25 to 0.95
Total	0.90	0.56 to 1.24

Illustrative. A third of the total effect runs through the savings path; two-thirds does not. The sentence: 'the association is consistent with partial mediation through savings; the mediator was not randomised.'
Moderation in the same module: a moderator variable and its interaction with the predictor, with simple slopes at ±1 SD and a plot.
Moderated mediation: both, with the conditional indirect effect at levels of the moderator.
SEM 101, sections 07 and 10, cover what these estimates assume. The module makes them easy to get; it does not make them causal.

Clustering by village is not handled in jAMM's bootstrap. For a paper, the SEM in SEMLj or lavaan with clustered errors is the version to report.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Meta-analysis

Meta-analysis in MAJOR: pooling published effects

A dataset with one row per study: effect size and its standard error (or means, SDs and ns per arm, or a correlation and n, or a log odds ratio and CI; the module accepts each). Path: MAJOR → the entry matching the input type. Tick the model (random effects, REML), heterogeneity statistics (Q, I², τ²), the forest plot, the funnel plot, and publication bias tests (Egger's, trim-and-fill, only with ten or more studies). Moderators for a meta-regression.

Illustrative: seven SHG studies, pooled SMD on decision-making 0.18, 95% CI 0.09 to 0.27, I² = 52%, τ² = 0.01, prediction interval −0.05 to 0.41.
The forest plot is the result; the prediction interval is the honest one.
Everything in Systematic Reviews 101 section 07 applies, including what the module does not decide: which studies to include and how to extract them.

The module is the metafor package with a form on it. A meta-analysis that a master's student can run in an afternoon is the right scale of ambition for a thesis chapter; the search and screening are the weeks of work.

ImpactMojoStatistics Without Code 101www.impactmojo.in

SEM Module

SEMLj: structural equation models with a form or a syntax box

SEMLj offers two entries: a syntax interface where you type lavaan model syntax into a box, and an interactive interface where you define latent variables and paths by dragging. Both run lavaan and return the same tables: loadings, paths, fit, modification indices, the path diagram, standardised estimates, indirect effects if defined, and multigroup analysis with constraints for invariance testing. Estimators include ML, MLR and WLSMV for ordinal items; missing data by FIML.

The worked SEM in SEM 101, section 07, runs in the syntax box as written there.
Invariance: the multigroup panel with 'loadings', 'intercepts' or 'thresholds' constrained, compared by fit in the output.
Bootstrap for indirect effects is a tick; clustering is not supported in the dialog and needs lavaan in Rj.

For a student who has done sections 02 to 07, this module is the route into SEM without AMOS, without a licence, and with better estimators than AMOS offers.

ImpactMojoStatistics Without Code 101www.impactmojo.in

GLM Plots

Interactions, simple effects and plots in GAMLj

GAMLj's General Linear Model is the core regression dialog with better defaults for interpretation: centred covariates, simple effects and simple slopes on request, estimated marginal means at chosen values, and plots of any predictor with any moderator on the horizontal axis and separate lines, with CIs or raw data. The question 'does the programme work better for children who started lower?' is an arm × baseline interaction, and the plot of endline against baseline with one line per arm is its answer.

Illustrative: interaction coefficient −0.06 (95% CI −0.10 to −0.02): the programme's effect is larger for lower baseline scores; simple effects at baseline −1 SD: 9.1 points; at +1 SD: 4.1 points.
Covariates are centred by default so main effects read at the mean; the option to change this is there.
Effect sizes (partial η², ω²) per term; contrasts of your choosing.

If you install one module, install GAMLj and use its GLM instead of the core regression for anything with an interaction or a cluster. The core dialogs are for learning; GAMLj is for reports.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Rj: the escape hatch, with four lines of R

Need	Rj code (data is the loaded dataset)
Weighted estimate with survey design	`library(survey); d <- svydesign(ids=~psu, strata=~stratum, weights=~wt, data=data); svymean(~decide_score, d)`
Cluster-robust standard errors	`library(sandwich); library(lmtest); m <- lm(endline ~ baseline + arm, data); coeftest(m, vcov = vcovCL(m, cluster = data$school_id))`
Multiple imputation	`library(mice); imp <- mice(data, m = 20); pool(with(imp, lm(endline ~ baseline + arm)))`
Reshape long to wide	`library(tidyr); pivot_wider(data, names_from = wave, values_from = score)`
A custom plot	`library(ggplot2); ggplot(data, aes(baseline, endline, colour = arm)) + geom_point(alpha = .3) + geom_smooth(method = "lm")`

Each snippet runs inside jamovi on the data already loaded, with the output in the results pane beside the menu analyses. Packages must be installed in the R that jamovi uses (the module's documentation says how). This is not a way to avoid learning R; it is a way to learn it four lines at a time, for exactly the thing the menu cannot do.

The first snippet is the answer to section 02's weights problem. It is short. It is still code, and it is still required.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Nine

Bayesian Analysis in JASP

ImpactMojoStatistics Without Code 101www.impactmojo.in

Why Bayes

The same questions, a different kind of answer

A p-value is the probability of data at least this extreme if there were no effect. It cannot say how likely the effect is, cannot support the null, and does not accumulate as data arrive. A Bayesian analysis starts from a prior distribution over the effect, updates it with the data, and reports the posterior: the range of effects now credible, and the Bayes factor, which says how much more the data support one hypothesis than another. JASP was built to make this as easy as a t-test, and in JASP every classical dialog has a Bayesian twin.

A 95% credible interval means what most people think a confidence interval means: 95% probability the effect lies in it, given the data and the prior.
A Bayes factor of 8 for the alternative means the data are 8 times more likely under an effect than under none. A Bayes factor of 8 for the null is evidence for no effect, which a p-value can never give.
Evidence accumulates: analyse as data arrive without the multiple-testing problem of repeated p-values.
The price: a prior must be chosen and stated, and the result depends on it, usually mildly.

When it earns its place in development work

Small pilots where 'not significant' would be read as 'does not work': the Bayes factor says whether the data favour no effect or are simply uninformative. Sequential monitoring of a programme's outcomes. Combining a new study with a prior estimate from the literature. Reporting to a decision-maker who wants 'how likely is it that the effect is at least 5 points', which the posterior answers directly and the p-value cannot.

Van Doorn and colleagues, 'The JASP guidelines for conducting and reporting a Bayesian analysis' (Psychonomic Bulletin & Review 2021, 28:813), is the ten-page guide to read first.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Bayesian t-test

The Bayesian independent-samples t-test in JASP

Path: T-Tests → Bayesian Independent Samples T-Test; decide_score as Dependent, treatment as Grouping. Tick: Bayes factor BF₁₀; Descriptives; Plots: Prior and posterior with additional info, Bayes factor robustness check, Sequential analysis. Prior: the default Cauchy with scale 0.707 on the standardised effect size (Rouder and colleagues 2009), or a narrower one if you have reason to expect a small effect.

	BF₁₀	Error %	Median δ	95% credible interval
SHG member vs non-member	4.8 × 10⁴	< 0.001	0.41	0.25 to 0.57

Illustrative, matching section 04's classical result (d = 0.42). The data are tens of thousands of times more likely under an effect than under none; the effect is credibly between a quarter and a half of a standard deviation.

Reading the plots

The prior-and-posterior plot shows the prior (wide, centred on zero) and the posterior (narrow, centred on 0.41); the two dots at zero are the Savage-Dickey ratio that is the Bayes factor. The robustness plot shows BF₁₀ across prior widths: if it stays large from a narrow to a wide prior, the conclusion does not depend on the prior. The sequential plot shows the Bayes factor as cases accumulate; here it crosses 100 by about n = 200.

The sentence: 'The data provided extreme evidence for a difference (BF₁₀ = 4.8 × 10⁴); the posterior median effect size was 0.41 (95% CI 0.25 to 0.57), robust across prior widths.'

ImpactMojoStatistics Without Code 101www.impactmojo.in

Bayes Factors

Reading Bayes factors: the scale, the direction, and the null

BF₁₀	Evidence for H₁ (an effect)	BF₀₁ = 1/BF₁₀	Evidence for H₀ (no effect)
1	None; the data do not distinguish	1	None
1 to 3	Anecdotal	1 to 3	Anecdotal
3 to 10	Moderate	3 to 10	Moderate
10 to 30	Strong	10 to 30	Strong
30 to 100	Very strong	30 to 100	Very strong
Above 100	Extreme	Above 100	Extreme

The Lee and Wagenmakers (2013) adaptation of Jeffreys's labels. They are conventions for description, not thresholds for decision; a BF of 2.9 and one of 3.1 are the same evidence.

BF₁₀ is H₁ over H₀; BF₀₁ the reverse. JASP lets you choose which to display; state which you report.
The Bayes factor compares two specific hypotheses under specific priors. A BF for 'an effect of any size' differs from one for 'a positive effect' (one-sided, a tick-box) and from one under a narrow prior.
Evidence for the null is the distinctive gift: illustrative, attendance by arm, BF₀₁ = 4.2, moderate evidence that attendance did not differ. A p of .07 could not have said that.
Bayes factors and credible intervals answer different questions (is there an effect; how big is it) and both should be reported.

'BF₁₀ = 1.4' means the study was too small to tell. Say that, rather than 'no significant effect'.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Bayesian Models

Bayesian ANOVA, regression and correlation in JASP

ANOVA → Bayesian ANOVA compares all models built from the factors (null, arm, sex, arm + sex, arm + sex + interaction) and reports each model's Bayes factor against the null and the inclusion Bayes factor for each effect, averaging over models. Regression → Bayesian Linear Regression does the same over all subsets of predictors and gives posterior inclusion probabilities and model-averaged coefficients with credible intervals. Regression → Bayesian Correlation Matrix gives Bayes factors and posteriors for each correlation.

Illustrative Bayesian ANOVA: inclusion BF for arm above 10⁶; for sex 0.6; for the interaction 0.2. The data favour leaving the interaction out, which is more informative than 'p = .44'.
Bayesian regression's model averaging is a principled answer to 'which predictors matter' that stepwise regression pretends to give.
Frequencies → Bayesian Contingency Tables gives a Bayes factor for association and a posterior for the odds ratio.

What is not in the menus

Custom priors from a previous study (informed priors) are possible in the t-test and regression dialogs but limited; hierarchical models with random effects for schools are in the Bayesian Mixed Models entry in recent versions and need long sampling; anything beyond that is brms in R. The menus cover the standard analyses in Bayesian form and that is most of what a thesis needs.

Run the classical and Bayesian versions side by side for the primary outcome. When they agree, say so; when they differ, the Bayes factor is usually telling you the evidence was weaker than the p-value implied.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Priors

Priors: what the defaults are, and when to change them

JASP's default priors are designed to be reasonable for psychology's effect sizes: a Cauchy on the standardised effect with scale 0.707 for t-tests, a multivariate Cauchy for ANOVA fixed effects, a JZS prior for regression. They put substantial weight on medium and large effects. Field programmes' effects are typically small (0.1 to 0.3 SD), and a default prior that expects larger effects can produce a Bayes factor mildly favouring the null when the true effect is small and the sample modest.

State the prior in the methods every time: 'Cauchy prior on effect size, scale 0.707 (JASP default)'.
For a field evaluation, consider scale 0.5 or a one-sided prior if only a positive effect is plausible, and report the robustness plot.
An informed prior from a meta-analysis (a normal centred on the pooled SMD with its SD) is legitimate and powerful, and must be stated and justified.
Never choose the prior after seeing which one gives the answer you want. The robustness check exists to show that you did not.

Estimation without Bayes factors

If the question is 'how big', not 'whether', the posterior distribution and its credible interval are the answer and the Bayes factor can be left aside. With a wide prior and a decent sample, the credible interval is nearly the confidence interval with the interpretation people wanted all along. That is the gentlest use of Bayesian methods and often the most useful.

Kruschke's Doing Bayesian Data Analysis (2nd ed., Academic Press, 2015) is the estimation-first text; JASP's Learn Bayes module teaches the ideas interactively.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Bayes Report

Reporting a Bayesian analysis: the van Doorn checklist, condensed

Element	Report
Hypotheses	What H₀ and H₁ are, and whether one- or two-sided
Prior	The distribution and its parameters for every tested effect, and why (default, informed from a source, or chosen for the expected effect size)
Software	JASP version and the R package underneath (BayesFactor)
Bayes factor	BF₁₀ or BF₀₁, stated; the numerical error percentage if sampling was used
Posterior	Median or mean of the effect size, the 95% credible interval, and the plot
Robustness	The Bayes factor across prior widths (the plot or its range)
Sequential	If data were analysed as they arrived, the sequential plot
Interpretation	Evidence described with the conventional labels, without treating them as thresholds; the null supported where the BF supports it
Classical companion	The frequentist result alongside, where readers expect it

Van Doorn and colleagues (2021) give the full list and worked examples. A Bayesian analysis reported without its prior is unreviewable; with it, it is more transparent than the classical version, because the assumption is on the page.

ImpactMojoStatistics Without Code 101www.impactmojo.in

jamovi's Bayes

The same analyses in jamovi: the jsq module

The jsq module ports JASP's Bayesian t-tests, ANOVA, regression, correlation and contingency tables into jamovi, with the same priors, Bayes factors, posterior plots and robustness checks. Install it from the library and the Bayesian entries appear beside the classical ones. The results are the same BayesFactor package's; the interface is jamovi's, so the data-cleaning tools of section 02 and the modules of section 08 are in the same window.

Use JASP when the Bayesian analysis is the main event and you want the fuller dialogs, the Learn Bayes module and the newest methods.
Use jamovi with jsq when the Bayesian result is a companion to a classical workflow already in jamovi.
Either way, the reporting checklist is the same.

A student who can run a t-test, read a Bayes factor and a credible interval, and state a prior has more statistical literacy than most published applied papers show. It takes an afternoon in JASP.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Ten

Effect Sizes, Power and Reporting

ImpactMojoStatistics Without Code 101www.impactmojo.in

Effect Sizes

Effect sizes: which one, and what it means in your units

Analysis	Effect size (tick-box)	Reads as	Convention, and its limits
Independent t	Cohen's d (also Hedges' g, Glass's Δ)	Difference in SD units	0.2 / 0.5 / 0.8 small / medium / large (Cohen 1988); in education Kraft (2020) argues 0.05 / 0.20 as the meaningful boundaries
Paired t	Cohen's d (paired)	Mean change over the SD of changes	Not comparable to the independent d; say which
ANOVA	ω² (preferred), η², partial η²	Share of variance explained by the factor	0.01 / 0.06 / 0.14; partial η² is inflated in multi-factor designs
Correlation	r	Strength of linear association	0.1 / 0.3 / 0.5
Regression	Unstandardised B in outcome units; standardised β; R², ΔR²	B is the one a reader can use	β for comparison within a model only
Contingency table	Percentage-point difference; risk ratio; odds ratio; Cramer's V	Points for policy; RR for comparison; V for strength	V: 0.1 / 0.3 / 0.5; OR overstates RR when outcomes are common
Non-parametric	Rank-biserial r; ε²	Probability-based difference	As for r
Logistic	Odds ratio; predicted probability difference	Points, again	Give the probabilities

Every dialog in this course has an effect size tick-box and every result in a report should carry one. The unit that matters is the outcome's own: 6.6 reading points, 13 percentage points, Rs 150 a month. The standardised version is for comparison across studies and goes second.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Interpreting Size

How big is big? Benchmarks against the setting, not Cohen

Cohen's benchmarks came from psychology experiments in the 1960s and 1970s and he warned against using them outside that context. In education, Kraft (Educational Researcher 2020, 49:241) reviewed hundreds of interventions and found the median effect on achievement was about 0.1 SD; a 0.2 SD effect from a cheap, scalable programme is large. In development economics, cash-transfer effects on school enrolment of a few percentage points are policy-relevant. The benchmark is what comparable programmes achieve, at what cost, for whom.

Translate the SD effect into the outcome's units and into something concrete: 0.4 SD on reading is 6.6 points, about a third of a year's typical growth in this sample.
Compare with the literature's effects for similar programmes; Systematic Reviews 101 is how to find them.
Divide by cost where cost is known: points per Rs 1,000 per child.
Report the CI's lower bound as the conservative reading: 'at least 5 points'.

'A medium effect (d = 0.42)' tells a programme officer nothing. '0.9 points on a 10-point decision-making scale, about a fifth of the gap between women with no schooling and those with secondary schooling' tells them something.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Power

Power analysis before the study: jpower and the cluster problem

The jpower module (or G*Power, free, Faul and colleagues, Behavior Research Methods 2007, 39:175) answers: how many respondents to detect an effect of size d with 80% power at α = 0.05? For an independent t-test with d = 0.3, about 176 per group; d = 0.2, about 394 per group; d = 0.5, 64 per group. Enter the smallest effect that would matter, not the effect you hope for. jpower also gives the power curve and the effect detectable with the sample you can afford.

For proportions: jpower's two-proportion entry; a 10-point difference from 50% needs about 390 per group.
For regression and ANOVA: G*Power's f² and f entries; Impact Evaluation 101 covers the formulas.
After the study, do not compute 'observed power'; it is a function of the p-value and adds nothing. Report the CI.

Clustered designs

When schools or villages are the unit of assignment, the effective sample is far smaller than the number of children. The design effect is 1 + (m − 1) × ICC: with 30 children per school and ICC 0.13, it is 4.8, so 1,200 children carry the information of 250. Neither jpower nor G*Power handles this in its basic dialogs; multiply the required n by the design effect, or use a cluster-RCT calculator (the PowerUpR package, or Optimal Design). Forty schools of thirty is a modest trial for an effect of 0.2 SD.

Power is decided before data collection. The number to put in the proposal is the one that comes out of this slide.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Intervals

Confidence intervals: reading them, drawing them, leading with them

A 95% confidence interval is the range of effects compatible with the data at the conventional level; across many studies, 95% of such intervals contain the true value. Its width is the precision; its position is the estimate; whether it crosses zero is the significance test, read at a glance. It answers 'how big, and how sure' where the p-value answers only 'whether'. Every dialog in jamovi and JASP offers it as a tick-box and every reported estimate should carry it.

'6.5 points (95% CI 3.6 to 9.4)': the effect is very probably between a modest and a large one; the study cannot pin it closer.
'1.2 points (95% CI −0.1 to 2.5)': a small positive effect is plausible and so is none; the study was not large enough to decide.
Two intervals that overlap can still differ significantly; test the difference, do not eyeball the overlap.
A wide interval is a finding about the study's size, and the report says so.

Drawing them

The descriptives plot in the t-test and ANOVA dialogs draws group means with CIs; the estimated marginal means plot draws adjusted means with CIs; the esci module draws the difference itself with its interval on a separate axis, which is the figure Cumming and Calin-Jageman recommend. A forest-style plot of all secondary outcomes' effects with CIs, one row each, is the single most informative figure an evaluation report can contain, and esci or the MAJOR module's forest plot will draw it from a small table of estimates.

Lead with the interval. The p-value is derivable from it; the reverse is not true.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Multiple Testing

Many outcomes, many subgroups: keeping the false-positive rate honest

Twenty tests at α = 0.05 on outcomes with no true effect produce one significant result on average. An evaluation with eight outcomes and five subgroups has forty tests. The fixes: declare a primary outcome and treat the rest as secondary; correct the p-values within a family (Bonferroni is conservative, Holm is better and equally simple, Benjamini-Hochberg controls the false discovery rate); or pre-register the tests so a reader can count them.

jamovi's post hoc dialogs offer Tukey, Scheffé, Bonferroni and Holm corrections as tick-boxes; use Holm.
For a set of outcomes, compute the Holm-adjusted p-values by hand (rank the p-values, multiply the smallest by k, the next by k − 1, ...) or in Rj with p.adjust.
Subgroup findings not pre-specified are exploratory; say so in the sentence that reports them.
A 'significant' effect in one subgroup and not another is not evidence of a difference between them; test the interaction.

The honest evaluation report has one primary result with an interval, a table of secondary results with adjusted p-values or an explicit count, and an exploratory section labelled as such. The dishonest one has twelve bullet points, each significant, and no count.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Misreadings

The five misreadings that a menu makes easy

Misreading	Why it happens	The correction
p < .001 means a big effect	The dialog shows p first and largest	p depends on n; the effect size is separate and must be reported
p = .07 means no effect	'Not significant' reads as 'nothing'	The CI shows what effects remain plausible; a Bayes factor may say the data were uninformative
Significant in the treatment arm, not in control, so the programme worked	Two paired tests instead of one comparison	Test the difference between arms; section 04
Controlled for it, so it is not confounding	Regression dialogs make adjustment feel like proof	Adjustment handles measured confounders in the form entered; unmeasured ones remain; causal claims need a design
The model fits (R² = 0.55), so the coefficients are causal	Fit and causation look alike in a table	R² is prediction; the arm coefficient is causal only because arm was randomised

None of these is a software problem, and all of them are more common in menu-driven work because the menu removes the moment of choice where a person writing code would have to think. The cure is the sentence: write the result in words before reading the p-value.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Reproducibility

Making menu work reproducible: the .omv file, annotations and syntax

The .omv file holds the data, every analysis, every option and every result, and reopens identically on any machine with jamovi. Add annotations (Edit tab, or click above any result) saying what each analysis is for and what you concluded. Turn on syntax mode and the R call sits above each result. Together these make a menu analysis as reproducible as a script, provided the file is shared.

Keep one .omv per project, analyses in the order of the report, annotated. Delete exploratory dead ends or move them to a second file.
Record the jamovi version and every module's version in the first annotation; module updates change defaults.
Export the syntax of every reported analysis to a text file as an appendix.
Deposit the .omv with the data on OSF or Zenodo when the report is published; a reader with jamovi can check every table.

JASP

The .jasp file does the same; annotations are added by clicking the result titles; the file can be exported as HTML or PDF with all tables and plots; and the JASP team's own guidelines ask for the .jasp file to be shared with a paper. The Open Science Framework has a JASP integration that stores it directly.

A reviewer who can open your file and see the analysis is a reviewer who trusts the table. That is the whole case for saving and sharing it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Export

Getting results out: tables, plots and the report

Output	How	Note
A table to Word	Right-click the table → Copy; paste into Word	Pastes as a real table; the number format (three-dots menu) is applied
A table to LaTeX	Right-click → Export → LaTeX (JASP); copy as HTML and convert (jamovi)	JASP's is cleaner
A plot	Right-click the plot → Export → PDF, SVG, EPS or PNG	SVG or PDF for print; PNG at 300 dpi for Word
All results	Menu → Export → HTML or PDF	The whole results pane as a document; useful for an appendix or a supervisor
The data with computed variables	Menu → Export → CSV	For use in R or Stata; filters are applied if 'filtered rows' is chosen
The syntax	Syntax mode on; select and copy	Into the methods appendix
APA formatting	JASP tables are APA by default; jamovi's are close	Round to two decimals in the report; p to three, or '< .001'

The report is written in a word processor; the numbers come from the software by copy, never by retyping. Retyped numbers are the source of the mismatches between tables and text that referees find.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Report Structure

The results section of an evaluation report, from jamovi output

SAMPLE: Table 1 by arm (Descriptives); attrition table (Contingency, t-tests)

→

PRIMARY OUTCOME: ANCOVA or mixed model estimate with CI and effect size; the descriptives plot

→

SECONDARY OUTCOMES: one table, same columns, adjusted p-values or a count

→

MECHANISM AND HETEROGENEITY: pre-specified interaction or mediation, labelled; exploratory ones labelled

→

SCALES: reliability and structure, in the methods or an appendix

→

ROBUSTNESS: with and without outliers; non-parametric companion; clustered version

→

LIMITATIONS: clustering if not modelled; attrition; measurement; the verbs used

Every step maps to a dialog in this course, and the whole sequence for an evaluation of this size takes a competent analyst two days in jamovi, of which the first is data checking. Academic Writing 101 covers the prose around it.

Lead with the primary outcome's estimate and interval, in units. Everything else supports that sentence.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Worked Report

A worked results paragraph, assembled from the course's outputs

Illustrative

'Of 1,280 children enrolled at baseline in 40 schools, 1,200 (93.8%) were assessed at endline, with equal attrition by arm; children lost to follow-up had lower baseline scores in both arms (Table 2). Arms were balanced at baseline (Table 1). Adjusting for baseline score and grade, and accounting for clustering by school, children in the remedial arm scored 6.5 points higher on the 100-point reading assessment at endline (95% CI 3.6 to 9.4, p < .001; ICC = 0.13), about 0.4 SD of the baseline distribution and a third of the typical annual gain in this sample. The share reading a full paragraph rose from 49% to 62% (difference 13 points, 95% CI 7 to 18). The effect was larger for children with lower baseline scores (interaction p = .004; 9.1 points at one SD below the baseline mean, 4.1 at one SD above) and did not differ by sex (interaction p = .44). Attendance did not differ by arm (1.2 points, 95% CI −0.1 to 2.5; Bayesian BF₀₁ = 4.2, moderate evidence of no difference). Results were similar in the unadjusted analysis and with the ten influential cases removed (Appendix Table A3). All analyses were conducted in jamovi 2.6 with the GAMLj module; the analysis file is deposited at [repository].'

Every number came from a dialog in sections 02 to 09. The paragraph reports the estimate before the test, the unit before the SD, the cluster-adjusted interval, the pre-specified interaction and its plot, the null finding as a finding, and the file location.

Read it against the misreadings slide. None of the five is in it.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Reporting Checklist

Before you submit: the statistics checklist

Every result	Done
Estimate in the outcome's units, with a 95% CI
The test statistic, df and exact p (or '< .001')
An effect size, named
n for that analysis (after missing and filters)
Assumption checks mentioned, and the choice they led to
Direction and magnitude in words a non-statistician can use

The report	Done
Primary outcome declared; secondary outcomes counted or adjusted
Clustering handled or acknowledged
Attrition table by arm
Scales' reliability and structure reported
Exploratory analyses labelled
Causal verbs only where the design supports them
Software and module versions; the .omv or .jasp file deposited
Numbers copied, never retyped; text matches tables

Fourteen lines, and the difference between an NGO report a funder can act on and one they cannot.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Section Eleven

Choosing Tools, Teaching and Practice

ImpactMojoStatistics Without Code 101www.impactmojo.in

The Tools

The point-and-click field: what each tool is for

Tool	Cost	Best at	Weak at
jamovi	Free, open	Data cleaning in place, the standard analyses with effect sizes, modules for mixed models, SEM, meta-analysis, IRT; syntax mode	Survey weights; long-to-wide reshaping; dates
JASP	Free, open	Bayesian analyses, APA tables, the Learn Bayes module, SEM and meta-analysis modules, a plot builder and an AI assistant in recent versions	Data editing (improving); survey weights
gretl	Free, open	Econometrics: time series, panels, instrumental variables, with menus	Psychometrics; plots are dated
PSPP	Free, open	An SPSS clone for those who must match SPSS output	Few effect sizes; little development
BlueSky Statistics; R Commander	Free (BlueSky has a paid tier)	R with menus; BlueSky is broad and generates R code	Smaller communities
Excel and Google Sheets	Free or licensed	Descriptives, pivot tables, quick plots, the Analysis ToolPak's t-tests and regression	Everything else; no assumption checks; errors travel silently
SPSS	Licence, expensive	The institutional default; broad; good documentation	Cost; no omega, CFA or parallel analysis without add-ons; syntax is its own language
Stata	Licence	Survey weights, panels, econometrics; menus plus a command line worth learning	Cost; weaker psychometrics and plots than R
R with RStudio	Free, open	Everything, reproducibly	Not point-and-click; the learning curve is the cost

For a department: jamovi as the teaching tool and R as the destination. For an NGO: jamovi for evaluation analysis, with Rj or a colleague in R for weighted surveys. For an econometrics course: gretl beside jamovi. SPSS is not needed for anything in this course.

ImpactMojoStatistics Without Code 101www.impactmojo.in

When to Move On

When the menu is not enough: the signs, and the next tool

Sign	Why the menu fails	Next tool
The survey has weights, strata and clusters	No design support	R survey (via Rj or RStudio); Stata svy
The same cleaning must be done to twenty files	Menus do it once	R or Python scripts; Data Analysis 101
Panel data with fixed effects and clustered errors	GAMLj can approximate; the standard tools are elsewhere	Stata xtreg; R fixest; gretl
Difference-in-differences with staggered adoption	Not in any module	R did, fixest; Impact Evaluation 101
Time series	Not in jamovi or JASP	gretl; R forecast; Time Series 101
Multiple imputation	Rj only	R mice
Custom or publication-quality figures	Fixed plot options	R ggplot2; Data Visualization 101
A reviewer asks for the code	A menu sequence is not code	Syntax mode exported; or redo in R
Text, images, networks	No dialogs	R or Python

None of these is a reason not to start in jamovi. Each is a reason the next course exists. The syntax mode is the bridge, and the analyst who has run every analysis in this course and read the R behind it has done the first month of an R course already.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Teaching

Teaching with jamovi: a twelve-week sequence for a methods course

Week	Topic	jamovi	Assessment
1	Data: types, levels, cleaning, the checklist	Setup, Compute, Filters	A cleaned file with a data log
2	Describing: distributions and plots	Descriptives, all plots	Table 1 and three plots from a supplied survey
3	Comparing two groups; effect sizes	Independent and paired t-tests	A comparison with CI, d and a sentence
4	Several groups; assumptions	One-way ANOVA, post hoc, non-parametric	An ANOVA with post hoc and a plot
5	Factorial designs and interactions	ANOVA with two factors; marginal means	An interaction plot and its reading
6	Correlation and simple regression	Correlation matrix, linear regression	A regression table with diagnostics
7	Multiple regression and ANCOVA	Model builder, factors, marginal means	The three-column regression table
8	Categorical outcomes	Contingency tables, logistic regression	A crosstab and a logistic model with probabilities
9	Scales	Reliability, EFA, CFA	A scale validation paragraph
10	Clustered data and mediation	GAMLj mixed model; jAMM	A mixed model with ICC
11	Bayesian inference	JASP t-test and ANOVA	A Bayes factor with prior stated
12	Reporting, reproducibility, and the edge	Annotations, syntax, export, Rj	A full results section from a supplied dataset, with the .omv file

Every week's assessment is a piece of a report; week 12's is the report. The sequence matches Navarro and Foxcroft's chapters and the sections of this course, and it installs on a classroom of mixed laptops in the first ten minutes of week 1.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Institutions

Adopting jamovi in a department or an organisation

Install: one download per machine, no licence server, no activation, works offline. The 'solid' release for teaching; 'current' for new features.
Lab machines: install once; modules install per user profile from the library, which needs internet the first time. Pre-install the five modules used in the course.
jamovi Cloud: a browser version for students without laptops; check the current free tier and data-privacy terms before putting participant data on it.
Data protection: files stay on the machine; nothing is uploaded unless the cloud version is used. For sensitive survey data this is the argument for the desktop version.
Version control: fix the version for a course or a project and note it; a mid-semester update that changes a default confuses everyone.

The SPSS conversation

An institution paying for SPSS licences it uses for t-tests and regressions is paying for nothing this course does not provide. The transition costs are the teaching materials (Navarro and Foxcroft is free and complete) and the habit. The gains are the licence, the effect sizes and assumption checks that SPSS hides behind extra dialogs, omega and CFA and parallel analysis without add-ons, and students who leave with a tool they can keep.

The strongest argument is the last one. A graduate who learned SPSS at a university that paid for it loses the tool on graduation. One who learned jamovi does not.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Worked Study

A worked study, dialog by dialog: the SHG evaluation from CSV to report

Step	Dialog	Output	Section
Import and set up	Open; Setup for 15 variables; missing codes 999	Typed, labelled data	02
Check	Descriptives with min/max/missing; Filters for implausible savings; frequency tables	Data log; 6 cases filtered with reasons	02
Scores	Compute: decide_score = MEAN(d1..d5); log_savings	Two new variables with formulas	02
Scale	Reliability; CFA split by language	ω = 0.80; structure holds; d3 differs by language	07
Table 1	Descriptives split by treatment; frequency tables	Balance table	03
Primary outcome	GAMLj mixed model: decide_score ~ treatment + age + educ + (1 \| village)	Effect 0.85, 95% CI 0.42 to 1.28; ICC 0.09	08
Secondary	Mixed model on log_savings; logistic on income_own	Two more rows, same columns	05, 06, 08
Mechanism	jAMM mediation through savings	Indirect 0.30, CI 0.18 to 0.44, labelled consistent-with	08
Heterogeneity	GAMLj interaction treatment × income_own, pre-specified	Interaction p = .21; no evidence	08
Bayesian companion	JASP Bayesian t-test	BF₁₀ extreme; posterior d 0.41	09
Report	Copy tables; export plots; annotate; syntax to appendix; deposit .omv	Results section and file	10

Illustrative numbers. Eleven steps, no code except the optional weighted check in Rj, and every output a reviewer can reopen. The cluster-adjusted primary estimate (0.85, CI 0.42 to 1.28) is smaller and less precise than the naive t-test's (0.90, CI 0.56 to 1.24), and it is the honest one.

Two days for an analyst who knows the course; a week for one learning it. Both produce a report that a funder's methods reviewer will accept.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Pitfalls

The dozen errors that menu analyses make easy

Error	Section	Fix in one line
Analysing a weighted survey unweighted	02	R survey via Rj; say what is unweighted
Wrong variable types (a code averaged, a number as text)	02	Set every type on import
Deleting rows instead of filtering	02	Filters, with notes
Paired t-test on one arm as the programme effect	04	Compare arms: change-score t or ANCOVA
p-values on a balance table	03	Descriptives only
Ignoring clustering	04, 08	GAMLj mixed model
Odds ratio reported as relative risk	06	Give probabilities and the RR
Item deletion by 'alpha if dropped'	07	Look at the item; keep it unless there is a reason
Twelve outcomes, no correction, no primary	10	Declare a primary; Holm; count the tests
Non-significant read as no effect	10	The CI; a Bayes factor
Retyped numbers	10	Copy tables
No file shared	10	Deposit the .omv

ImpactMojoStatistics Without Code 101www.impactmojo.in

Reading List

Where to go next

Resource	What it covers	Notes
Navarro and Foxcroft, Learning Statistics with jamovi (2019)	The whole of sections 02 to 07, as a textbook with the dialogs	Free at learnstatswithjamovi.com; the companion to this course
jamovi user guide and the module documentation (jamovi.org)	Every dialog; GAMLj, jAMM and SEMLj have their own sites with worked examples	Free
Van Doorn et al., Psychon Bull Rev 2021, 28:813	Conducting and reporting a Bayesian analysis in JASP	Section 09; ten pages
Cumming and Calin-Jageman, Introduction to the New Statistics (2nd ed., Routledge, 2024)	Estimation, effect sizes and CIs, with the esci module	Sections 04 and 10
Lakens, Improving Your Statistical Inferences (free online course and book)	p-values, power, effect sizes, equivalence tests, pre-registration	The clearest treatment of section 10's material
Kraft, Educational Researcher 2020, 49:241	Interpreting effect sizes in education	Section 10
Kruschke, Doing Bayesian Data Analysis (2nd ed., Academic Press, 2015)	Bayesian estimation from first principles	For those who want more than section 09
Datalab.cc's jamovi video series (Barton Poulson)	Screen-by-screen walkthroughs	Free; good for a first week
ImpactMojo: Data Analysis 101, Bivariate Analysis 101, Multivariate Analysis 101, Econometrics 101, Impact Evaluation 101, SEM 101	The statistics behind each dialog, and the designs that make coefficients causal	impactmojo.in/101-courses/

ImpactMojoStatistics Without Code 101www.impactmojo.in

Summary

What to remember

jamovi and JASP are free, open, correct, and enough for a thesis or an evaluation report. SPSS is not required.
Set every variable's type and labels on import; filter, never delete; compute scores with a stated missing rule.
Plot every outcome before testing it. The plot chooses the test.
The estimate and its interval come first, in the outcome's units; the p-value second; the effect size named.
Compare arms, not before-and-after within one arm.

Children in schools and women in villages are clustered; the GAMLj mixed model is one dialog and it is the honest analysis.
Reliability, EFA and CFA are three core dialogs; use them before summing a scale.
Weighted surveys are the one thing the menus cannot do. Four lines in Rj can.
A Bayes factor can support the null; a p-value cannot. Report the prior.
Annotate, keep the syntax, deposit the file. A menu analysis is reproducible if the file is shared.

The menu removes the moment where a coder has to think. Put it back by writing the sentence before reading the p-value.

ImpactMojoStatistics Without Code 101www.impactmojo.in

Statistics Without Code 101 · Complete

Now go run it,
and write the sentence first.

Every analysis a thesis or an evaluation needs, in a free tool, with the estimate before the p-value and the file shared afterwards. Explore the rest of the ImpactMojo 101 Series, free forever.

More 101 Courses Explore ImpactMojo Dataverse

CC BY-NC-ND 4.0·Free Forever·ImpactMojo 101 Series