Overview
The type of dependent variable decides the family of regression you need. A continuous outcome such as test scores suits ordinary least squares; a yes/no outcome needs logistic regression; unordered categories call for multinomial logistic; ordered categories such as Likert-type ratings may suit an ordinal (proportional odds) model; and counts such as hospital visits need Poisson or negative binomial regression. Using the wrong family produces coefficients that are hard to interpret or simply wrong.
Our work covers the whole modeling cycle: defining predictors from your theory, preparing variables (coding categorical predictors, centering, handling missing data), estimating the model, checking diagnostics, and translating coefficients into quantities your reader understands — predicted probabilities, marginal effects or incidence rate ratios rather than raw log-odds.
Every analysis comes with commented code and a record of modeling decisions, so the results can be reproduced and defended. We also help you avoid common over-claims, such as treating an observational regression coefficient as a causal effect.
Who it’s for and the problems it solves
This is for you if you are…
- Master's and doctoral students testing hypotheses about predictors of an outcome
- Health and social science researchers modeling binary, ordinal or count outcomes
- Authors responding to reviewer comments about multicollinearity, model fit or interaction terms
- Analysts in public, nonprofit and business settings who need a defensible explanatory model
Common problems we help with
- You are unsure whether your outcome needs linear, logistic, ordinal or Poisson regression
- Your logistic regression output shows odds ratios you cannot explain in words
- VIF values are high and your predictors change sign when you add controls
- The proportional odds assumption fails in your ordinal model
- Count data show overdispersion or a large number of zeros
- Your supervisor wants a hierarchical (blockwise) regression showing R² change and you do not know how to set it up
What’s included — and what isn’t
Included
- Choosing the model family from your outcome's measurement level and distribution
- Multiple linear regression with standardized and unstandardized coefficients, confidence intervals and R² / adjusted R²
- Binary logistic regression with odds ratios, predicted probabilities, classification tables and ROC/AUC where useful
- Multinomial logistic regression with a justified reference category
- Ordinal logistic regression with a test of the proportional odds assumption and alternatives when it fails
- Poisson, negative binomial and zero-inflated or hurdle models for count outcomes, with offsets for exposure
- Hierarchical (blockwise) regression reporting ΔR² or likelihood-ratio tests between blocks
- Interaction terms and non-linear effects, with simple slopes or marginal-effects plots
- Diagnostics: residual plots, influential cases (Cook's distance, leverage), VIF, heteroskedasticity tests and robust standard errors
- Publication-style tables and annotated code in SPSS, R, Stata or Python
Not included
- Stepwise predictor selection as a way to find significant results without theoretical justification
- Describing correlational coefficients as causal effects in your write-up
- Writing your results or discussion sections for you — you receive explanations to write from
- Removing observations or recoding variables purely to change significance
What you receive
- Model specification note
- The outcome, model family, predictors, controls, coding decisions and missing-data approach, agreed before estimation.
- Diagnostics report
- Residual, influence and collinearity checks with plots and a short statement of what each means for your model.
- Results tables
- Coefficients or odds/incidence rate ratios with 95% confidence intervals, model fit statistics and block comparisons, formatted to your style guide.
- Interpretation guide
- Plain-language explanation of each key coefficient, with predicted values or marginal-effects plots where they aid understanding.
- Reproducible code
- SPSS syntax, R script, Stata do-file or Python notebook that reproduces the analysis from the raw data.
How it works
Clarify the question
We look at your hypotheses, outcome and candidate predictors and confirm which regression family fits.
Prepare variables
Recoding, dummy coding, centering, transformation and missing-data decisions are made and logged with you.
Estimate and diagnose
We fit the model, run diagnostics and discuss any problem — for example overdispersion or influential cases — before interpreting results.
Interpret and report
We produce tables and figures and translate coefficients into meaningful quantities.
Walkthrough
A session to go through the output and answer your questions so you can write it up yourself.
Models and diagnostics we use
- OLS linear regression
- For continuous outcomes, checked for linearity, independence, homoskedasticity (Breusch–Pagan, residual plots) and normality of residuals; HC3 robust standard errors when variance is unequal.
- Binary logistic
- Odds ratios with confidence intervals, likelihood-ratio tests, pseudo-R² (Nagelkerke, McFadden) reported with caution, Hosmer–Lemeshow or calibration plots, and average marginal effects.
- Multinomial and ordinal logistic
- Proportional odds checked with the Brant test or a likelihood-ratio comparison; partial proportional odds or multinomial models when the assumption fails.
- Count models
- Poisson with dispersion checks, negative binomial for overdispersion, zero-inflated or hurdle models when excess zeros have a substantive explanation.
- Hierarchical entry
- Blockwise entry of controls and focal predictors, reporting change in R² or deviance so the added explanatory value is visible.
- Collinearity and influence
- Variance inflation factors, condition indices, Cook's distance and DFBETAs, with sensitivity analyses rather than automatic deletion.
Interpreting coefficients people can understand
A logistic coefficient of 0.41 is a change in log-odds, which almost nobody can picture. We convert it into an odds ratio (1.51), and where helpful into predicted probabilities at meaningful values — for example, the probability of the outcome for a typical participant with and without the exposure. The same applies to incidence rate ratios in count models and to interactions, which are nearly always clearer as a plot than as a coefficient.
We also distinguish statistical from practical significance: a tiny coefficient can be significant in a large sample, while a sizeable one can be imprecise in a small sample. Confidence intervals and standardized or marginal effects make this visible.
Hierarchical regression is not multilevel modeling
The word 'hierarchical' is used for two different things. Hierarchical regression means entering predictors in blocks to see what each block adds. Hierarchical linear modeling means accounting for nested data such as students within schools. If your observations are clustered, a blockwise regression will understate standard errors and we will suggest multilevel modeling instead.
Limits of regression we will point out
Regression on observational data shows conditional associations. Omitted variables, reverse causality and measurement error can all bias coefficients, and adding more controls does not automatically fix this — controlling for a variable that lies on the causal path can introduce bias. We help you choose controls from theory and word conclusions carefully.
Models with many predictors relative to the sample size overfit: they describe your sample well and generalize poorly. We keep models parsimonious, report adjusted R² or information criteria, and suggest cross-validation or penalized regression when prediction is the actual goal.
- Events-per-variable in logistic models is checked, and small-sample alternatives such as Firth's penalized likelihood are considered
- Pre-specified hypotheses are separated from exploratory models in the reporting
- Robustness checks (alternative codings, excluding influential cases) are reported, not hidden
What the work can look like
Illustrative example — fictional data
Sample excerpt: logistic regression on vaccine uptake
A public health master's student modeled whether adults received a seasonal vaccine (yes/no) using age group, chronic illness and trust in health services (n = 1,020). The numbers are invented for illustration.
Model result
Each one-point increase in trust (1–7 scale) was associated with higher odds of vaccination, OR = 1.38, 95% CI [1.24, 1.54], controlling for age group and chronic illness. The model improved on the controls-only model, likelihood-ratio χ²(1) = 34.9, p < .001.
Predicted probabilities
At low trust (2) the predicted probability of vaccination for a 45–64-year-old without chronic illness was 0.31; at high trust (6) it was 0.62. This framing is included for the discussion because odds ratios are often misread as risk ratios.
Diagnostics
VIFs were below 1.6, no case had a Cook's distance above 0.5, and the calibration plot showed close agreement between predicted and observed rates across deciles.
Illustrative example with invented numbers, not a real client project. Your deliverables reflect your own data and reporting requirements.
What you’ll need to provide
- Research questions and hypotheses, including which variables are predictors and which are controls
- The anonymized dataset and codebook
- Any theoretical model, prior studies you are following, or required model specification
- Style guide or journal requirements for tables
- Supervisor or reviewer comments if you are revising an existing analysis
You don’t need everything ready to ask for a quote — tell us what you have and we’ll say what’s missing.
What affects price and turnaround
Price depends on
- Number of models and outcomes
- Model family and complexity (interactions, non-linear terms, zero-inflation)
- Amount of data preparation, recoding and missing-data handling
- Whether robustness and sensitivity analyses are required
- Delivered analysis, guided learning sessions or review of your own work
Turnaround depends on
- Clarity of the hypotheses and variable list at the start
- Data quality and whether several files need merging
- Time needed for you to decide on issues raised by diagnostics
Every project is quoted individually. How we price our work.
Questions people ask before starting
Should I use stepwise regression?
Usually not. Stepwise selection capitalizes on chance, inflates coefficients and gives misleading p-values. We prefer entering predictors based on theory, and for prediction problems we suggest penalized methods such as LASSO with cross-validation.
My Likert-scale outcome — is it linear or ordinal?
A single Likert item is ordinal and usually suits ordinal regression. A scale score averaged over several items is often treated as continuous. We look at the number of categories and the distribution and explain the trade-offs.
Can you handle missing data?
Yes. We examine the missingness pattern and use listwise deletion only when it is defensible; otherwise multiple imputation or full-information methods, with the approach documented for your methods section.
Which software will you use?
Whatever suits you and your institution: SPSS, R, Stata or Python. If your preferred package cannot fit a needed model (for example, zero-inflated models in base SPSS), we explain the alternative.
Will you interpret the results for my thesis?
We explain every result clearly and can check your draft for accuracy. The written interpretation in your thesis is yours, and you should acknowledge statistical assistance as your institution requires.
Can you respond to reviewer comments about my regression?
Yes. We map each comment to an analysis change or a reasoned response, run any additional models and help you draft the technical part of your reply for you to finalize.