1,284

Voice Personas

Who is doing the asking, and does it change the answer?
Why this screen exists

Three hundred field researchers are three hundred slightly different instruments. Each has a manner, an accent and a way of nudging, and the correlation that induces within one enumerator's workload inflates variance in a way nobody measures. At a modest intra-interviewer correlation of 0.02 and 120 interviews each, a nominal 3.5 lakh field sample carries the precision of roughly 1.03 lakh.

Personas here do the opposite. Assignment is randomised within stratum, not matched to maximise cooperation. Matching would raise the response rate and quietly confound the result. Randomising means the persona effect becomes a measured quantity that can be corrected for, which is the table at the bottom of this screen.

Language and register matching is different and uncontroversial: a Purvanchal register for eastern Uttar Pradesh raises comprehension without touching who is willing to answer.

Active. Persona effect is measurable and corrected for. Cooperation is slightly lower than matching would give.
Asha
Hindi · Neutral urban
Verified end to end16.2% assigned
Delhi, UP urban, Haryana, MP urban
Ramesh
Hindi · Purvanchal
Verified end to end14.8% assigned
Eastern UP, Bihar, Jharkhand
Devendra
Hindi · Malwa and Nimar
Verified end to end3.0% assigned
MP rural, Rajasthan east, Chhattisgarh
Nusrat
Urdu · Dakhini leaning
Verified end to end6.4% assigned
Urdu households nationally
Ananya
Bengali · Rarh standard
Verified end to end9.1% assigned
West Bengal, Tripura, Assam Barak valley
Sneha
Marathi · Pune standard
Verified end to end8.6% assigned
Maharashtra, Goa
Gopal
Telugu · Coastal Andhra
Verified end to end7.9% assigned
Andhra Pradesh, Telangana
Karthik
Tamil · Chennai standard
Verified end to end7.2% assigned
Tamil Nadu, Puducherry
Bhavana
Kannada · Old Mysore
Verified end to end6.1% assigned
Karnataka
Jignesh
Gujarati · Central Gujarat
Verified end to end5.8% assigned
Gujarat, DNH and DD
Sasmita
Odia · Coastal Odisha
Verified end to end4.4% assigned
Odisha
Meera
Malayalam · Central Kerala
Verified end to end4.2% assigned
Kerala, Lakshadweep
Harleen
Punjabi · Malwa
Verified end to end3.9% assigned
Punjab, Chandigarh, Haryana border belt
Kalpana
Assamese · Upper Assam
Verified end to end2.4% assigned
Assam, Arunachal foothills
Measured persona effectthis wave, n = 2,14,600
PersonaAssignedCooperationQ11 refusalEffect on estimateCorrection
Asha Hindi16.2%46.2%13.8%+0.3 ptsApplied
Ramesh Hindi14.8%49.1%11.2%−0.4 ptsApplied
Devendra Hindi3.0%47.8%12.4%+0.1 ptsApplied
Nusrat Urdu6.4%44.7%16.4%+0.9 ptsApplied
Ananya Bengali9.1%45.9%14.6%−0.3 ptsApplied
Sneha Marathi8.6%43.8%14.1%−0.2 ptsApplied
Gopal Telugu7.9%47.3%12.6%−0.5 ptsApplied
Karthik Tamil7.2%41.6%18.9%+0.6 ptsApplied
Bhavana Kannada6.1%44.2%15.3%+0.2 ptsApplied
Jignesh Gujarati5.8%48.4%11.9%−0.1 ptsApplied
Sasmita Odia4.4%46.7%13.1%+0.4 ptsApplied
Meera Malayalam4.2%42.9%17.2%+0.7 ptsApplied
Harleen Punjabi3.9%45.1%14.8%−0.2 ptsApplied
Kalpana Assamese2.4%44.8%13.9%+0.3 ptsApplied

Effect is the difference in the incumbent vote share estimate against the persona-pooled mean, after stratum controls. Randomisation is what makes this column interpretable at all.

The argument in one number
rho = 0

One interviewer, two lakh calls, identical wording and pacing every time. There is no interviewer variance term because there is one interviewer. That is arithmetic, and it is the hardest thing for a field-based competitor to answer.

Held back deliberately

No persona is assigned on inferred community, religion or caste. Voice, name and register are selected on language and region only. Inferring community from a name and then choosing a matching voice is a short walk to a product we will not build.