Check How Missing TikTok Captions Change Your Research Sample

Tool Limits And Native Fields

TikTok Transcript current public build 0.1.36 accepts startUrls with one to 100 known public video URLs, short links, or numeric video IDs. It does not search profiles, keywords, or sound pages. preferredLanguages selects existing public tracks in preference order, while an empty list uses the first public track. It does not translate, transcribe uncaptured audio, collect comments, or infer content codes.

One dataset row represents each submitted input, including nonbillable status outcomes. Exact fields include inputIndex, inputUrl, videoId, videoUrl, authorHandle, durationSeconds, languageCode, text, segments with startSeconds, endSeconds, and text, status, statusReason, isBillable, and scrapedAt. Preserve video IDs as strings. Caption language may use a regional value such as eng-US. Null or empty transcript with a failure status is not spoken silence.

Statuses include succeeded, no_public_captions, language_unavailable, duplicate_input, not_tiktok, invalid_input, unsupported_url, video_unavailable, not_public, blocked_or_challenged, and extraction or caption failures. Success requires nonempty text. isBillable true means a successful native transcript result event, but other rows remain evidence. Platform usage is separate from a result event, so do not call the whole run free. OUTPUT fields inputsReceived, resultRows, billableTranscriptRows, nonBillableResultRows, duplicateVideosSkipped, and statusCounts are run accounting, not sample representativeness.

The current one-video English control succeeded with the exact requested video ID, nonempty text, and twelve segments of the checked timing shape. This proves a public caption path on that source, not all videos, perfect spoken word accuracy, or coding validity. Replace your whole sample list with one intended known video, set preferredLanguages to `["en"]` only for English study or leave it empty for the first public track.

Eight Steps For Bounding Missing Data

1. Start from the selected video ledger, including every intended known video. Deduplicate known video IDs without dropping input outcome evidence. Declare the target population and one binary content code per video before calculations.
2. Keep outcomes distinct: validated eligible caption, caption absent, requested language unavailable, temporary or source access failure, unsupported input, duplicate, and validation unresolved. A source failure is not evidence that the video lacks captions. Exclude genuinely invalid or out-of-scope inputs under the declared rule; do not mix them into missing in-scope outcomes.
3. Define N as the eligible selected unique videos, K as those with a verified code result, P as verified positive codes, and M as eligible unknown outcomes equal to N minus K. Reconcile these numbers. Treat a failed or empty caption as an unknown code, not a negative. Validation errors can also make outcomes unknown.
4. Report the observed success subset fraction P divided by K only when K is greater than zero. Label it as such. Report missing fraction M divided by N only when N is greater than zero. These are different denominators. If no outcomes are verified, do not divide by zero or make a content claim.
5. Calculate assumption-free binary bounds for the selected eligible sample: lower P/N assumes every unknown is code zero, and upper (P+M)/N assumes every unknown is code one. State that these are logical ranges for this selected sample, not a confidence interval, probability of truth, or all-TikTok estimate. A confidence interval describes sampling uncertainty under a model; this interval describes unknown outcomes.
6. For a planned group comparison, calculate the same ledger and bounds for each group. If ranges overlap, missing outcomes can reverse or erase the apparent ordering; overlap alone is not a statistical significance test. Record any claim that survives all allowed assignments versus one that depends on assumptions.
7. Investigate availability by declared language, creator, or period if those source attributes are known. Differences are descriptive signals, not proof of a mechanism. Fix invalid inputs and inspect source failures under a bounded process. Do not replace unavailable video content with translated or generated text without changing and documenting the measurement method.
8. State the study conclusion at the supported scope. If the bounds are too wide for the research decision, seek validated outcomes or narrow the question. Do not silently enlarge the dataset by selecting only easy caption successes.

Original Fictional Worked Example

All examples below use original fictional data and do not come from a real TikTok source.

A fictional study selects twenty eligible unique videos. Twelve have validated code outcomes, of which seven are positive. Eight outcomes remain unknown for distinct caption, access, or validation reasons.

The observed fraction is 7/12=58.3%. The missing fraction is 8/20=40%. Selected-sample bounds are 7/20=35% and (7+8)/20=75%. Thus 35 percent to 75 percent is the logical code proportion range under arbitrary unknown outcomes, not an estimated accuracy interval. Do not report 58.3 percent as the verified full selected-sample proportion.

Fictional Ledger And Group Bounds

The known-only order where Group A is greater than Group B is not robust. Assign Group A unknowns negative and Group B unknowns positive to get 50 percent versus 80 percent. The opposite assignment gives 70 percent versus 20 percent. Reconcile Group A plus Group B: twenty selected, twelve known, seven positive, eight unknown.

Missing-reason ledger

Invalid/out-of-scope inputs and duplicate observations are outside eligible unique N, not extra unknown videos.