01Confirmatory or exploratory?
Confirmatory analysis tests hypotheses, and often a measurement or structural model, that you specified from theory before looking at the data, ideally in a written or pre-registered analysis plan. Exploratory analysis looks for patterns and generates hypotheses for later testing. Both are legitimate, but they should be labelled honestly: presenting exploratory findings as if they were predicted (sometimes called HARKing, hypothesising after the results are known) inflates false positives. We help you keep the two apart in your chapters.
02Data compiling and recoding
We set up a data entry format in Excel, import it into SPSS or other software, label variables and values, recode and compute scale scores, and check for data entry errors, outliers and missing data. Missing data are examined for their pattern (missing completely at random, at random, or not at random) and handled with suitable methods such as multiple imputation or full information maximum likelihood, rather than simply deleting cases.
03Descriptive statistics
Analysis usually starts with descriptive statistics for the sample, such as age, sex, occupation, income and marital status: frequencies and percentages for categorical variables, and the mean, standard deviation, median, minimum, maximum and range for continuous ones, with tables and graphs. The reliability of composite scores is then checked before hypothesis testing.
04Assumptions and diagnostics
We check the assumptions your tests rely on: normality of residuals (with plots, since tests such as Shapiro-Wilk flag trivial departures in large samples), linearity, homogeneity of variance (Levene's test, with Welch corrections where needed), independence, sphericity for repeated measures (with Greenhouse-Geisser correction where it fails) and multicollinearity (variance inflation factors). Where assumptions fail, we discuss transformations, robust or non-parametric methods, or a different model.
05Hypothesis testing
Each hypothesis is matched to a suitable test, such as t-tests, chi-square tests, ANOVA and ANCOVA, correlation, linear or logistic regression, mixed models or non-parametric alternatives. A p-value tells you how compatible the data are with the null hypothesis; it is not the probability that the hypothesis is true, and a non-significant result is not proof of no effect.
06Effect sizes and confidence intervals
We report effect sizes with confidence intervals: for example Cohen's d or Hedges' g for mean differences, partial eta squared or omega squared for ANOVA, R-squared and standardised coefficients for regression, and odds or risk ratios for binary outcomes. Cohen's small, medium and large benchmarks are a rough guide only; your field's earlier findings are a better yardstick.
07Multiple comparisons
Running many tests raises the chance of a false positive. We plan corrections that suit your design: Tukey's HSD or planned contrasts after ANOVA, Bonferroni or the more powerful Holm method to control the family-wise error rate, or the Benjamini-Hochberg procedure to control the false discovery rate in larger sets of tests.
08Confirmatory factor analysis and SEM
Confirmatory factor analysis (CFA) tests whether your items measure the constructs you intended; structural equation modelling (SEM) then tests the hypothesised paths between constructs. We help you choose an estimator suited to your data (for example robust maximum likelihood, or WLSMV for ordinal Likert items), check identification, and report the chi-square test with fit indices such as CFI, TLI, RMSEA (with its confidence interval) and SRMR.
09Reading fit indices with care
Hu and Bentler (1999) suggested cut-offs of about 0.95 or above for CFI and TLI, 0.06 or below for RMSEA and 0.08 or below for SRMR. These came from simulations of particular models with continuous data and are widely treated as stricter rules than their authors intended; they behave differently with ordinal data, small samples and complex models. We use them as guides, look at local fit (residuals and modification indices), change a model only where theory supports it, and report every change openly.
10Covariance-based or PLS-SEM?
Covariance-based SEM (in AMOS, Mplus, LISREL, lavaan in R or Stata) tests how well a theoretical model reproduces the data. Partial least squares SEM (for example in SmartPLS) is a variance-based method aimed at prediction, with its own quality criteria, such as composite reliability, average variance extracted and the HTMT ratio; CB-SEM fit indices do not carry over to it. We help you choose and justify the approach that suits your aims.
11Validity of your instruments
Content validity is judged by experts on whether the items cover the construct, for example with a content validity index. Face validity is whether the instrument appears, to participants and experts, to measure what it claims; it is the weakest form of evidence. Construct validity is supported by CFA, convergent validity (for example average variance extracted of 0.50 or more) and discriminant validity (for example the Fornell-Larcker criterion or HTMT).
12Reliability
After validity, we assess reliability: internal consistency with Cronbach's alpha or McDonald's omega, composite reliability in SEM, test-retest reliability for stable traits, and inter-rater reliability where observers score data.
13Study design drives the analysis
Descriptive studies use frequencies, percentages, cross-tabulations and measures of central tendency and variability. Cohort studies compare exposed and unexposed groups over time with relative risks, incidence rate ratios, Kaplan-Meier curves and Cox regression; case-control studies use odds ratios. Randomised experiments compare intervention and control groups with t-tests, chi-square tests, ANCOVA, regression or repeated-measures models, analysed by intention to treat. Quasi-experimental designs, where randomisation is not feasible, use interrupted time series, difference-in-differences or propensity score methods; a simple pre-post paired t-test cannot separate the intervention from other changes over time. Longitudinal studies use linear mixed-effects models, generalised estimating equations or repeated-measures ANOVA. Diagnostic accuracy studies report sensitivity, specificity, predictive values (which depend on prevalence), likelihood ratios and ROC curves with the AUC.
14Causal claims
Randomised trials and well-designed quasi-experiments support causal conclusions; observational analyses need careful design and theory to do so. Structural equation models test whether data are consistent with a causal model but cannot prove it on their own, and Granger causality shows that one time series helps predict another, not that it causes it. We help you word your conclusions to match your design.
15Software
We work in SPSS, R (including lavaan for CFA and SEM), Stata, AMOS, Mplus, SAS, LISREL, SmartPLS, JASP, jamovi, Minitab and JMP. SPSS is easy to learn and widely used in psychology, health and social sciences. R is free, flexible and has packages for almost every method, but needs coding. Stata is strong for panel data, regression and epidemiology. SAS is common in pharmaceutical and clinical trial work. JASP and jamovi are free point-and-click packages built on R. Minitab suits quality control and Six Sigma, and JMP interactive, visual exploration. Where you can, we supply syntax or scripts so your analysis is reproducible.