THEME & GSEQ Setup Guide

1. What is sequential analysis — and why does data format matter?

Lag sequential analysis (GSEQ) tests whether behaviour B follows behaviour A significantly more (or less) often than chance, at lag +1, +2, +3 and so on. The key statistic is the adjusted residual (z > 1.96 = significant positive association; z < −1.96 = significant inhibition). It needs events coded as a time-ordered sequence. Reference: Bakeman & Quera (2011).

T-pattern detection (THEME) finds recurring temporal structures — chains of events that consistently occur in the same order within relatively invariant time windows. Its defining concept is the critical interval [t+d1, t+d2]: after event A at time t, this interval contains event B more often than chance (Magnusson 1996, 2000). Unlike GSEQ, THEME uses real timestamps, so the timing matters, not just the order.

Why your code-window design decides what analysis is possible

In Sportscode, only code-button presses generate timestamped instances with their own start time, duration and CSV row. Label-group selections do not — they are attributes of the enclosing button press and share its timestamp. This is the single most consequential decision: if you capture behavioural dimensions as label groups, all dimensions within one press share one timestamp and cannot be analysed as individually timed sequential events.

OpenTag works the same way: each event is a timestamped instance; attributes are descriptors that share the event's time. Coding each action as its own event gives you genuine inter-action timing; bundling actions as attributes on one event does not.

2. Understanding your data type

Before choosing a design, decide which of two published, valid paradigms fits your research question.

If you only want to know what kind of episode follows what, Paradigm 1 is enough and you can assign a constant duration (Section 8 / the converter's constant-duration option). If you want the temporal structure of action–reaction chains, you need Paradigm 2 — one timestamped event per action.

3. Three code-window designs

Design 1 — Label-based single button Paradigm 1

One code button (e.g. Actor) pressed once per episode. Label groups capture the dimensions: Attack Initiator, Key Technique, Opponent Reaction, Outcome. The converter joins the dimension values into one code, e.g. CutKick_NaeryeoAxe_GuardRaises_Counter3Pt.

Analysis: GSEQ tells you which episode types follow each other; THEME finds which episode types cluster in time. Individual action timing within an episode is not recoverable.

Design 2 — All code buttons Paradigm 2

Every action type has its own code button; boundary buttons (Start, End) delimit sequences; no label groups.

Start                       — referee reset / whistle
P1_CutKick, P1_NaeryeoAxe…  — Player 1 techniques
P2_CutKick, P2_NaeryeoAxe…  — Player 2 techniques
P1_GuardRaises, P2_Retreat… — reactions
P1_Win_1Pt … P2_Win_5Pt     — scoring
NoScore                     — reset, no point
End                         — whistle after resolution

The button name is the complete event code — no labels, no joining. The converter reads each row as a code and splits on Start/End into one GSEQ session per sequence. THEME receives genuine timed events. This is the recommended design for new studies about the sequential/temporal structure of individual actions.

Design 3 — Hybrid code buttons + labels

Code buttons for primary actions (P1Attack, P2Attack, P1Win, Start, End), with label groups on the attack buttons for technique and reaction. Each press gets one timestamp; technique/reaction labels share it, but between-action timing is real. Use when there are too many technique combinations to warrant individual buttons but you still need per-action timestamps.

4. Naming rules (all designs)

These apply to every button name and label value. Violations produce sanitised codes that are unreadable or that collide.

5. Coding protocol & timestamp precision

Designs 1 & 3 (label-based)

Set the Actor button's lead time to 8–10 s so pressing it rewinds to cover the full sequence. Pause, rewind and select labels in any order while the button is held; close the instance on release (or via a deactivation link from the Outcome label group).

Design 2 (all code buttons) — post-hoc frame-stepping

1. Watch the full sequence through without pressing anything.
2. Note what happened (techniques, players, outcome).
3. Rewind using frame-step keys.
4. Park the video on the exact frame each action begins.
5. Press the button at that frame; release when the action ends.
6. Move to the next action and repeat.

Slower (~3–4 min per sequence) but it produces precise timestamps — the precision that makes THEME genuinely temporal rather than merely sequential. In OpenTag, use the frame-step controls and tag each action as its own event at its true frame.

6. Combination techniques

Combinations (two techniques in rapid succession by the same player) are common in combat sports. Three options:

• A — Ignore structure: code the sequence once, label the primary scoring technique. Loses co-occurrence detail; fine when combinations are rare.
• B — Composite atomic labels (best for Design 1): add explicit combination labels (AxeDollyo, AxeSewo). Watch a sample first to find all attested combinations before building the window — adding labels mid-project means recoding.
• C — Two presses, identical window (Design 2 only): press the player's button twice over the same window, once per technique. Equal durations in %SS = lag-0 co-occurrence = correct simultaneity.

7. Sequence boundaries & GSEQ sessions

GSEQ analyses patterns within sessions. Without boundaries, the last event of sequence 17 and the first of sequence 18 look like a lag transition — a meaningless pairing.

• Design 2: the converter splits each Start→End block into one GSEQ session automatically — the cleanest approach.
• Design 1: use the converter's sequence-assignment panel. Auto-detect uses a gap threshold (default 4 s, suited to the taekwondo reset/bow). Review the editable table and override as needed.
• Post-hoc: if data was collected without boundaries, add a sequence-ID column in your spreadsheet; the converter reads it.

8. Data volume requirements

Adjusted residuals are unreliable when observed frequencies are very low. A widely used working floor is 20 instances per event code (criterion or target).

The converter's frequency audit panel surfaces this before you download — codes under 20 are flagged amber, under 5 red. If a code is too sparse: code more matches, combine rare codes with GSEQ's LUMP command, or redesign with fewer/broader codes.

lump P1Attack = P1_CutKick P1_FeintBody P1_StepIn P1_DirectAttack

Worked example: 3 matches × ~30 sequences × 40 unique codes ≈ 360 events ≈ 9 per code — below the floor. Lump or broaden.

9. Step-by-step: importing into THEME

1. Open THEME (PatternVision; free for academic use). File → New Project.
2. File → Import observation → select observation.txt from the converter's THEME ZIP.
3. File → Import category table → select vvt.vvt from the same ZIP.
4. Verify the event list shows all codes in lowercase.
5. Search → T-pattern detection.

Detection settings: minimum occurrences 3–5; significance p = .005 (THEME default, used across the published studies); minimum pattern length 2; disable the "fast" requirement (per Lapresa 2013, so valid slower patterns aren't rejected).

Read the dendrogram: each node = one event; a branch = a temporal relationship (critical interval); longer chains = more complex recurring structures. To generalise across matches, look for patterns recurring in multiple sessions — Diana et al. (2017) used ≥ 80% of matches as the filter.

10. Step-by-step: importing into GSEQ 5

1. Open GSEQ 5 (Bakeman & Quera — free download). File → Open → the .mds from the converter. (Commercial alternative: GSEQ via Mangold INTERACT.)
2. GSEQ compiles .mds → .mex automatically. Compile errors = illegal characters in code names — check the converter preview.
3. Type FREQ to see frequencies; flag anything below 20.
4. Type the table command:

   table CriterionCode given all codes lag 1 to 5

5. Type STATS to compute adjusted residuals.

Interpreting: z > +1.96 → target follows criterion significantly more than chance (p < .05); z < −1.96 → significant inhibition. Report criterion, target, lag, adjusted residual and direction. Use LUMP (Section 8) before TABLE when individual codes are too sparse.

11. Methods text & citations

A ready-to-adapt methods paragraph:

References

• Bakeman, R., & Quera, V. (2011). Sequential analysis and observational methods for the behavioral sciences. Cambridge University Press. https://doi.org/10.1017/CBO9781139017343
• Borrie, A., Jonsson, G. K., & Magnusson, M. S. (2002). Temporal pattern analysis and its applicability in sport. Journal of Sports Sciences, 20(10), 845–852. https://doi.org/10.1080/026404102320675675
• Camerino, O., Chaverri, J., Anguera, M. T., & Jonsson, G. K. (2012). Dynamics of the game in soccer: Detection of T-patterns. European Journal of Sport Science, 12(3), 216–224. https://doi.org/10.1080/17461391.2011.566362
• Diana, B., Zurloni, V., Elia, M., Cavalera, C. M., Jonsson, G. K., & Anguera, M. T. (2017). How game location affects soccer performance: T-pattern analysis of attack actions. Frontiers in Psychology, 8, 1415. https://doi.org/10.3389/fpsyg.2017.01415
• Lapresa, D., Arana, J., Anguera, M. T., & Garzón, B. (2013). Comparative analysis of sequentiality using SDIS-GSEQ and THEME. Journal of Sports Sciences, 31(15), 1687–1695. https://doi.org/10.1080/02640414.2013.796061
• Magnusson, M. S. (1996). Hidden real-time patterns in intra- and inter-individual behavior. European Journal of Psychological Assessment, 12(2), 112–123.
• Magnusson, M. S. (2000). Discovering hidden time patterns in behavior: T-patterns and their detection. Behavior Research Methods, Instruments, & Computers, 32(1), 93–110. https://doi.org/10.3758/BF03200792