Before any statistics, the single biggest determinant of whether an fMRI study works is its experimental design — how conditions are arranged in time. A perfectly preprocessed dataset from a poorly designed task cannot be rescued by analysis. This note covers the conceptual choices; the quantitative treatment of how those choices affect estimation precision (design efficiency and variance) is developed in 3.2 General Linear Model (Part 3).
1. Why timing is everything
The BOLD response to a brief event is not instantaneous: it is the neural time course convolved with the haemodynamic response function (1.1 BOLD Signal and HRF), a sluggish blur that peaks ~5–6 s later and lasts ~20 s. Two consequences follow, and they drive every design decision:
- Because the HRF is slow, responses to events closer together than ~20 s overlap and add up. Your design determines how much they overlap.
- Because we detect a condition by how its predicted BOLD differs from other conditions and from baseline, conditions that are too similar in timing become statistically indistinguishable (collinear regressors).
2. The three design families
| Design | What it is | Best for |
|---|---|---|
| Blocked | Conditions grouped into long (~15–30 s) blocks of repeated trials | Detecting where an effect is — maximum statistical power |
| Event-related | Individual trials spaced out, often in randomised order | Estimating the shape/timing of the response; separating trial types; avoiding predictability |
| Mixed | Sustained block state + embedded transient events | Dissociating sustained vs transient components |
Blocked designs stack many responses on top of each other, producing a large, easily detected signal difference between blocks — this is why the active-vs-passive checkerboard task used elsewhere in this tutorial is blocked. The cost is that you cannot see the shape of a single response, and blocks can be predictable (attention, anticipation confounds).
Event-related designs trade raw detection power for flexibility: you can present trials in random order, model trial types separately, and even estimate the HRF shape itself. The cost is smaller per-event signal, so you need more trials and careful timing.
3. The detection ↔ estimation trade-off
These two goals pull in opposite directions:
- Detection power — how reliably you can say a voxel is active at all. Maximised by making the design regressor vary a lot (long blocks; conditions maximally separated). Blocked designs win.
- Estimation efficiency — how precisely you can recover the response amplitude or shape for each condition. Maximised by rapid, jittered event-related designs that decorrelate the regressors.
No single design maximises both; you choose based on your question. The formal version of this trade-off — design efficiency — is worked through with simulations in 3.2 General Linear Model.
4. Jitter: varying the inter-stimulus interval
If every trial is separated by a fixed interval, the condition regressors become correlated with each other and with the baseline, and the model cannot tell their overlapping responses apart. Jittering the inter-stimulus interval (ISI) — drawing it from a distribution rather than fixing it — breaks that regularity, decorrelates the regressors, and lets the GLM separate overlapping responses. Rapid event-related designs are only viable because of jitter.
Rule of thumb
Fixed short ISIs are the classic way to accidentally build a design you cannot analyse. If trials are close together, randomise the spacing.
5. Beyond on/off: parametric designs
Not every manipulation is categorical. When a trial carries a graded quantity — reaction time, stimulus intensity, difficulty, subjective value, prediction error — you can model the response as scaling with that quantity using a parametric modulator rather than a plain on/off regressor. This is a design decision (you must record the per-trial values), and the modelling side is covered in 3.2 General Linear Model.
6. Design principles (Max-Min-Con)
A compact checklist for a robust task, adapted from the classic “Max-Min-Con” rule:
- Maximise the variance due to your effect of interest — separate your conditions well; use enough trials; make contrasts strong.
- Minimise error variance — keep runs short enough to limit motion/fatigue, control the environment, use a comfortable, learnable task.
- Control nuisance variance — counterbalance and randomise condition order so that time, arousal, and practice effects are not confounded with condition; balance low-level stimulus properties across conditions.
And two practical musts:
- Include a baseline/rest the conditions can be contrasted against.
- Record accurate onsets and durations in a BIDS
events.tsv(2.2 BIDS Structure) — the GLM is only as good as its timing file.
Where next
Once the design exists, the responses it predicts become the regressors of the general linear model: 3.2 General Linear Model.
References
- Amaro, E., Jr, & Barker, G. J. (2006). Study design in fMRI: basic principles. Brain and Cognition, 60(3), 220–232. doi:10.1016/j.bandc.2005.11.009
- Dale, A. M. (1999). Optimal experimental design for event-related fMRI. Human Brain Mapping, 8(2–3), 109–114.
- Liu, T. T. (2004). Efficiency, power, and entropy in event-related fMRI with multiple trial types. NeuroImage, 21(1), 401–413.
- MRC CBU — Design efficiency in fMRI.