Snap Shot
- The question: How do researchers and evaluators code the questions and answers in a recorded child forensic interview, how reliable is that coding, and what can it tell a court about the accuracy of a child’s account?
- What the research shows: Coding classifies each interviewer utterance (invitations, facilitators, directive wh- questions, option-posing questions, and suggestive questions) and counts the details the child reports in response. Trained coders reach high agreement, and an automated model has matched trained coders. Studies with independent evidence of what happened find that details elicited by open invitations are more often confirmed, and that errors are more often elicited by focused and suggestive questions.
- What it does not show: Coding measures how a detail was obtained, not whether it is true. Many answers to closed questions are accurate, open recall can reflect earlier contamination, coding schemes differ across laboratories, and the research has not established a proportion of closed questions that marks an interview as unreliable.
- The law: An expert who codes an interview must meet Rule 702 (as amended in 2023), Daubert, Kumho, Mil. R. Evid. 702 and Houser, and in Arizona Rule 702 as amended in 2024; experts may explain how question form bears on reliability but may not vouch for or against a child’s truthfulness (Brooks, Lindsey, Moran).
- Issues for each side: Prosecutors can use coding to show which central details the child volunteered in free recall; defense counsel can use it to show which details first appeared in the interviewer’s questions. Both should expect written coding rules, raw counts, and a check on reliability, and neither should present a coding result as a measure of truthfulness.
Most child forensic interviews are now recorded, and the recording is often the central evidence in a child abuse prosecution. Because the whole exchange is preserved, the questioning can be examined systematically. In the 1990s researchers developed methods for classifying every interviewer utterance by type and counting what the child said in response, and those methods became the main measure of interview quality in field research and in evaluations of interviewing protocols.
The same methods now reach court, where both sides retain experts to analyze recorded interviews and judges hear argument about whether particular details were suggested. A coding analysis can bring discipline to those disputes by replacing an impression that an interview was “open” or “leading” with a description another analyst can check. It can also be misused, as when a percentage of closed questions is offered as though it measured whether a child is telling the truth.
This paper describes how coding works, its role in the NICHD Investigative Interview Protocol, its reliability, its use in tracing the source of details, and what studies with independent evidence of the events show about question type and accuracy, followed by the limits of the method and the governing law. The question hierarchy and the practical review of a single interview are covered in “Question Types and Evaluating a Forensic Interview from the Record” in the Child Abuse Allegations hub; this paper concentrates on the coding method and its research uses and limits.
What Interview Coding Is
Interview coding grew out of field studies of investigative interviews with children in Israel and the United States (Lamb et al., 1996; Sternberg et al., 1996). Coders work from verbatim transcripts of the substantive part of the interview and assign each interviewer turn to one category. The categories used in most later research distinguish broad invitations that ask for free recall without limiting the topic; facilitators, which are brief neutral prompts or repetitions that keep the child talking; directive questions, usually wh- questions, that follow up on content the child has introduced; option-posing questions, which require the child to accept, reject, or select between alternatives the child has not offered; and suggestive questions, which signal the expected answer or assume information the child has not provided (Lamb, Orbach, Warren, et al., 2007). The categories are ordered by how much of the content the interviewer supplies; the later ones rely more on recognition than on recall and are treated as carrying greater risk.
Coding has a second half. Coders also count what the child reports. A detail is typically each new piece of information that bears on the allegation, such as who was present, what was done, with what objects, and where or when, and coders link each one to the prompt that drew it out (Lamb, Orbach, Warren, et al., 2007). Combining the two produces measures of productivity: how many details each kind of prompt yields, and what share of the child’s account came from free recall rather than from focused questioning. In one early field study of 22 interviews of 5- to 11-year-olds, open invitations produced longer and more detailed answers than directive, leading, or suggestive utterances regardless of age, although younger children’s answers were shorter (Lamb et al., 1996).
Coding schemes have grown more refined. Wh- questions asking what happened or how an action occurred are more productive than other wh- questions (Ahern et al., 2018), and researchers have identified hybrid forms. In 434 interviews of 4- to 12-year-olds, about one in four questions beginning “tell me” was a “faux invitation” that actually asked a wh- or yes/no question, and such questions were associated with less productive and less responsive answers (Henderson et al., 2020). Such refinements make coding more informative, but two studies, or two experts, may not be counting the same thing.
Coding and the NICHD Protocol
Coding was the tool used to build and test the NICHD Investigative Interview Protocol. Unscripted field interviews relied heavily on focused questions; in them the average invitation yielded about five details, whereas in a later field study a scripted opening invitation, preceded by narrative practice during rapport building, yielded many times that number (Lamb, Orbach, Warren, et al., 2007). Because interviewers drifted back to focused questions even after productive open answers, the researchers scripted the entire interview, and coding became the measure of whether the protocol changed interviewer behavior (Lamb et al., 2018).
A meta-analysis of five comparison studies found that protocol interviews contained many more invitations (g = 1.60) and fewer option-posing (g = −0.95) and suggestive prompts (g = −0.63) than standard interviews, and that children in protocol interviews provided more central details in response to invitations (g = 0.90); the authors noted that the advantage was less consistent for preschool children (Benia et al., 2015). These outcomes are coding measures: they show that a protocol changes how interviewers ask and how much children say. Whether the additional information is accurate is a separate question, addressed below.
Reliability: Can Coders Agree?
Coding is useful in court only if another qualified coder applying the same rules would reach substantially the same result. Studies report inter-rater reliability as percentage agreement and as Cohen’s kappa, which corrects agreement for chance. Trained coders generally reach high agreement, but that depends on training, written rules, and practice.
Errors follow predictable patterns. When postgraduate students, online trainees, and police officers coded transcripts after studying a coding manual, overall error rates were low, but the trainees shared four misunderstandings: that any wh- question is a specific question, that any request beginning “tell me” is open-ended, that an open question cannot include specific content, and that a specific question cannot draw an elaborate answer; trainees who stopped practicing became less accurate (Powell et al., 2013). Researchers have also documented that terms such as “open,” “closed,” “leading,” and “specific” are defined differently across psychological and linguistic literatures and across agencies, so that the same question may fall in different categories in different schemes (Oxburgh et al., 2010).
Automation is a recent development. A machine-learning language model developed and tested with more than 350,000 questions from 1,435 forensic interviews and 416 trial transcripts of 3- to 17-year-olds distinguished invitations, wh- questions, option-posing questions, and non-questions with 95% agreement with trained human coders (κ = .93); when blind recoders examined the disagreements, the human codes were more often the ones in error, frequently because coders overlooked invitations (Szojka et al., 2025). The study also shows the limits. The model did not attempt to identify suggestive questions separately; the blind recoders who resolved the hard cases agreed with each other at κ = .73; and errors shared by the model and the original coders could not be detected.
For legal purposes, a coding analysis should state the scheme, the definitions of each category and of a “detail,” the unit coded (question, turn, or sentence), how ambiguous items were resolved, and whether a second coder checked any of the work. Without that information, a reported proportion cannot be evaluated or reproduced.
Tracing the Source of Details
The forensic value of coding often lies less in totals than in sequence. Tracing a detail to the utterance that first introduced it shows whether the child volunteered it, whether the child elaborated on it in response to an open prompt, or whether it first appeared in the interviewer’s question and was adopted by the child. In a close analysis of 24 interviews of seven children, suggestive questions drew a disproportionate share of the children’s self-contradictions and open invitations drew none (Lamb & Fauchier, 2001), so the question that preceded a contradiction can be informative, although the sample was small.
Sequential coding shows that the link between question and answer is more complicated than a tally suggests. In 80 unstructured field interviews of 40 children aged 3 to 7 about known events, a hierarchical coding system separated leading questions that suggested accurate information from those that suggested inaccurate information, and individual interviewers were consistent in their questioning style (Gilstrap, 2004). A lagged sequential analysis of 41 interviews by professional interviewers of 3- to 7-year-olds about a staged event found that leading questions as a group did not increase children’s acquiescence; one type, leading questions with inaccurate content, was followed by agreement, and children’s responses could be predicted from their own preceding responses, apart from the intervening question (Gilstrap & Ceci, 2005). Coding that records only question type cannot make the distinction between accurate and inaccurate suggestions, because that distinction requires knowing what actually happened.
Tracing sources also depends on the record. Measured against audio recordings, investigators’ contemporaneous notes of 20 interviews left out more than half of the interviewers’ utterances and about one in four incident-relevant details, and credited details to more open prompts than the ones that had produced them (Lamb et al., 2000). Coding is only as complete as the recording and transcript, and earlier conversations are usually beyond its reach.
What Coding Can and Cannot Show About Accuracy
Coding measures how information was obtained. Linking question type to accuracy requires independent evidence of what happened, which real cases rarely provide. The designs researchers have used each support the conclusion that free recall tends to be more accurate than answers to focused questions, and each shows how far that conclusion can be pressed.
In a case study in which an audio recording captured the last of a series of abusive incidents, more than half of the details the child reported in the investigative interview were corroborated by the recording, and nearly all of the corroborated details were central to the allegation (Orbach & Lamb, 1999). In a case study of four girls aged 8 to 15, photographic and audio records corroborated about 86% of the sexual acts the girls alleged but only about 43% of the coercive acts they reported, and omissions of documented acts were common (Bidrose & Goodman, 2000). An uncorroborated statement in such a study is not shown to be false; the evidence simply did not capture it.
A field study compared the accounts of 43 children, averaging about 10 years old, with those of 52 youths who admitted abusing them. About two thirds of the children’s details were not addressed by the perpetrators, but among those that were, perpetrators confirmed a larger share of the details children gave to invitations than of those given to focused prompts; question type, however, did not predict which details the perpetrators contradicted (Lamb, Orbach, Hershkowitz, et al., 2007). An analogue study allowed full verification by staging an event for 128 children aged 5 to 7 and interviewing them with the NICHD protocol four to six weeks later. Children were highly accurate regardless of prompt type; invitations produced more detail without loss of accuracy, and the errors that did occur were mostly about peripheral information and more often followed directive or yes/no questions (Brown et al., 2013). A recent comparison of 50 forensic interviews with case records found that details matched the records about 93% of the time for an older group of children and about 83% for a younger group, whose errors concerned mainly dates, times, and context; case records are an imperfect standard, and accuracy was not analyzed by question type (Ahn & Choi, 2026).
Taken together, these studies support two propositions both sides can rely on. In these studies, details from a child’s free recall, especially central details, were generally accurate, and children questioned well gave accurate accounts of abuse. Errors are more likely in response to focused, option-posing, and suggestive questions, and in peripheral, temporal, and contextual details. The studies do not support reading accuracy off a coding sheet. Many answers to focused questions are correct, a suggested detail may still be true, and a detail given in free recall may reflect earlier conversations the recording does not show.
Limits of the Evidence
Most coding research is field research on real interviews, where the question form is known but the truth usually is not. The studies that verified accuracy are few, small, or both: two case studies, a perpetrator-confirmation sample in which most details could not be checked, an analogue study of a staged, non-abusive event with young children, and a records comparison in which the records themselves were the standard. Most field data come from a few research groups and countries, some from interviews in Hebrew or Korean under national protocols. Coding categories and definitions of a “detail” vary across studies, so effect sizes are not always comparable. The coding studies reviewed here generally did not analyze results by the child’s sex or gender, and the verification samples were too small or too uniform to do so; Bidrose and Goodman (2000) studied four girls. Most important, all of these findings describe groups of answers. They describe the average risk attached to a form of question, not whether a particular child’s particular answer is accurate.
Questions to Ask About This Research
When an expert reports that a given percentage of an interviewer’s questions were option-posing or suggestive, counsel can ask which coding scheme was used, how each category and the unit of analysis were defined, whether a second coder checked the work and what agreement was obtained, and what benchmark the percentage is being compared against, since the research has not established a validated cutoff. When Lamb, Orbach, Hershkowitz, et al. (2007) is cited to show that free-recall details are accurate, counsel can ask how many details could actually be checked and whether the finding also held for contradicted details, which it did not. When Brown et al. (2013) is cited, counsel can ask whether results with 5- to 7-year-olds recalling a staged, benign event four to six weeks later apply to the child and the events at issue. When automated coding is offered (Szojka et al., 2025), counsel can ask which categories the model codes, whether suggestive questions were identified by a person, and who reviewed the output. And when a detail is said to have originated in the interviewer’s question, counsel can ask whether earlier, unrecorded conversations might account for it, in either direction. These questions are explained in “Reading Research Claims: A Guide for Attorneys” on this site.
Interview Coding and the Law
Federal and Military Standards
Expert testimony based on coding a recorded interview is governed by Federal Rule of Evidence 702, as amended December 1, 2023, under which the proponent must show that it is more likely than not that the testimony rests on sufficient facts or data and reliable methods reliably applied to the case. Daubert v. Merrell Dow Pharmaceuticals, Inc. (1993) and Kumho Tire Co. v. Carmichael (1999), which extended the gatekeeping obligation to all expert testimony, including testimony based on technical or other specialized knowledge, direct the court to the method’s reliability. For coding, that inquiry reaches the scheme, its definitions, and whether its application can be reproduced. In courts-martial, Military Rule of Evidence 702 tracks the federal rule; under Military Rule of Evidence 1102 the 2023 amendment took effect in the military rule in 2025 (Joint Service Committee on Military Justice, 2025), and practitioners should confirm current text. Military judges also apply the six factors of United States v. Houser (1993). An expert may describe the research on question form and may trace particular details to particular prompts in the recording, but may not tell the factfinder, directly or by statistical proxy, that the child is or is not truthful (United States v. Brooks, 2007).
Courts have treated the way a child was questioned as relevant to reliability. In Idaho v. Wright (1990), decided under a confrontation framework later displaced by Crawford v. Washington (2004), the Supreme Court held a young child’s statements to a physician insufficiently trustworthy where the interview was unrecorded and the questions were leading, but declined to require any particular interview procedure. In State v. Michaels (1994), the New Jersey Supreme Court required a pretrial taint hearing when a defendant shows some evidence of suggestive or coercive interviewing; where such hearings are available, coding and source-of-detail analysis are typically central to the evidence on both sides, and jurisdictions differ on whether they recognize them.
Arizona
Arizona Rule of Evidence 702 adopted the Daubert standard effective January 1, 2012, and was amended effective January 1, 2024, to conform to the 2023 federal amendment (Ariz. Sup. Ct. Order R-23-0004). State v. Lindsey (1986) and State v. Moran (1986) allow an expert to educate the jury about children’s statements and the effects of interviewing practices but bar opinions, direct or indirect, on whether a particular child is credible. State v. Salazar-Mercado (2014) permits “cold” expert testimony that educates the jury on general principles without applying them to the case; an expert who explains the coding method and the research in general terms may testify in that way, while an expert who codes the recording in the case applies principles to facts and must satisfy Rule 702(d).
Special Issues in Legal Practice
The principles described above apply with equal force to every party in a legal proceeding. The science operates the same way regardless of who is asking the questions, and the research summarized here does not by itself favor either side. What differs between the parties is not the science but the vantage point — the particular places where these issues tend to surface, and the particular errors each party is best positioned to guard against. The two summaries below identify those recurring issues. They are offered for orientation, not as litigation strategy; in every case the controlling question is what the evidence in that case will, and will not, support, and the expert’s role is to inform the trier of fact rather than to advocate for either side.
Special Issues for Prosecutors
The research supports several propositions prosecutors can legitimately present. Coding is a well-established research method, and the studies that could check accuracy against independent evidence found that details children volunteered in free recall, especially central details, were generally accurate, with errors more common in peripheral, temporal, and contextual details. Those studies were few and small and included a staged, non-abusive event, and closed questions carry more risk, especially for children with intellectual disability (see “Intellectual Disability and Memory”). Some focused questions are expected in nearly every interview and do not by themselves make it unreliable. A transparent coding analysis can document, detail by detail, that the central allegations first emerged in the child’s own narrative, that focused questions came after free recall, and that the child added to the account when invited to.
The research also counsels caution. Studies verifying accuracy are few and small, and none permits an expert to say that a particular child’s account is accurate because it was elicited by open questions; free recall can reflect earlier conversations. When a central detail surfaces only after an option-posing or suggestive question, it says less about the child’s own memory, and corroboration matters more. A coding analysis offered by the prosecution must meet the same standards as any other: a stated scheme, written definitions, raw counts, a reliability check, and candid treatment of the portions of the interview that do not support the prosecution’s position.
Special Issues for Defense
The research supports several propositions defense counsel can legitimately present. Coding can show, detail by detail, where key information first appeared, and details that first emerged in the interviewer’s option-posing or suggestive questions are more likely to be inaccurate as a group. It is legitimate to examine whether the record is complete, whether earlier conversations shaped the account, and whether an opposing analysis disclosed its coding rules and reliability, and to offer expert testimony on these matters.
The research also counsels caution about the limits that apply to the defense. A high proportion of focused questions is not evidence that the events did not occur, many answers to such questions are accurate, and in the sequential research leading questions as a group did not increase children’s agreement, although one type, those with inaccurate content, was followed by agreement. A tally that lacks a stated scheme or a reliability check, or that does not connect particular questions to particular details, can give a misleading picture. An expert cannot tell the factfinder that a particular child’s memory was altered or that the child is untruthful, and the most defensible analysis acknowledges the details the child volunteered.
This paper is for educational purposes only and is not legal advice. The author is a psychologist, not an attorney; summaries of law are general and must be verified against current authority or with a licensed attorney.
References
Ahern, E. C., Andrews, S. J., Stolzenberg, S. N., & Lyon, T. D. (2018). The productivity of wh- prompts in child forensic interviews. Journal of Interpersonal Violence, 33(13), 2007–2015. https://doi.org/10.1177/0886260515621084
Ahn, J., & Choi, S. (2026). Distinct patterns in the accuracy of child testimonies: Evidence from a cluster-based comparison with case records. Legal and Criminological Psychology, 31(1), 217–228. https://doi.org/10.1111/lcrp.70010
Benia, L. R., Hauck-Filho, N., Dillenburg, M., & Stein, L. M. (2015). The NICHD Investigative Interview Protocol: A meta-analytic review. Journal of Child Sexual Abuse, 24(3), 259–279. https://doi.org/10.1080/10538712.2015.1006749
Bidrose, S., & Goodman, G. S. (2000). Testimony and evidence: A scientific case study of memory for child sexual abuse. Applied Cognitive Psychology, 14(3), 197–213. https://doi.org/10.1002/(SICI)1099-0720(200005/06)14:3<197::AID-ACP647>3.0.CO;2-6
Brown, D. A., Lamb, M. E., Lewis, C., Pipe, M.-E., Orbach, Y., & Wolfman, M. (2013). The NICHD Investigative Interview Protocol: An analogue study. Journal of Experimental Psychology: Applied, 19(4), 367–382. https://doi.org/10.1037/a0035143
Gilstrap, L. L. (2004). A missing link in suggestibility research: What is known about the behavior of field interviewers in unstructured interviews with young children? Journal of Experimental Psychology: Applied, 10(1), 13–24. https://doi.org/10.1037/1076-898X.10.1.13
Gilstrap, L. L., & Ceci, S. J. (2005). Reconceptualizing children’s suggestibility: Bidirectional and temporal properties. Child Development, 76(1), 40–53. https://doi.org/10.1111/j.1467-8624.2005.00828.x
Henderson, H. M., Russo, N., & Lyon, T. D. (2020). Forensic interviewers’ difficulty with invitations: Faux invitations and negative recasting. Child Maltreatment, 25(3), 363–372. https://doi.org/10.1177/1077559519895593
Joint Service Committee on Military Justice. (2025). Amendments to the Military Rules of Evidence since 1984 (G. E. Maggs, Ed.; 2025 ed.). https://jsc.defense.gov/Portals/99/MRE%20Amendments%202025.pdf
Lamb, M. E., Brown, D. A., Hershkowitz, I., Orbach, Y., & Esplin, P. W. (2018). Tell me what happened: Questioning children about abuse (2nd ed.). Wiley. https://doi.org/10.1002/9781118881248
Lamb, M. E., & Fauchier, A. (2001). The effects of question type on self-contradictions by children in the course of forensic interviews. Applied Cognitive Psychology, 15(5), 483–491. https://doi.org/10.1002/acp.726
Lamb, M. E., Hershkowitz, I., Sternberg, K. J., Esplin, P. W., Hovav, M., Manor, T., & Yudilevitch, L. (1996). Effects of investigative utterance types on Israeli children’s responses. International Journal of Behavioral Development, 19(3), 627–637. https://doi.org/10.1177/016502549601900310
Lamb, M. E., Orbach, Y., Hershkowitz, I., Horowitz, D., & Abbott, C. B. (2007). Does the type of prompt affect the accuracy of information provided by alleged victims of abuse in forensic interviews? Applied Cognitive Psychology, 21(9), 1117–1130. https://doi.org/10.1002/acp.1318
Lamb, M. E., Orbach, Y., Sternberg, K. J., Hershkowitz, I., & Horowitz, D. (2000). Accuracy of investigators’ verbatim notes of their forensic interviews with alleged child abuse victims. Law and Human Behavior, 24(6), 699–708. https://doi.org/10.1023/A:1005556404636
Lamb, M. E., Orbach, Y., Warren, A. R., Esplin, P. W., & Hershkowitz, I. (2007). Enhancing performance: Factors affecting the informativeness of young witnesses. In M. P. Toglia, J. D. Read, D. F. Ross, & R. C. L. Lindsay (Eds.), The handbook of eyewitness psychology: Vol. 1. Memory for events (pp. 429–451). Lawrence Erlbaum Associates.
Orbach, Y., & Lamb, M. E. (1999). Assessing the accuracy of a child’s account of sexual abuse: A case study. Child Abuse & Neglect, 23(1), 91–98. https://doi.org/10.1016/S0145-2134(98)00114-8
Oxburgh, G. E., Myklebust, T., & Grant, T. (2010). The question of question types in police interviews: A review of the literature from a psychological and linguistic perspective. International Journal of Speech, Language and the Law, 17(1), 45–66. https://doi.org/10.1558/ijsll.v17i1.45
Powell, M. B., Benson, M., Sharman, S. J., Guadagno, B., & Steinberg, R. (2013). Errors in the identification of question types in investigative interviews of children. International Journal of Police Science & Management, 15(2), 144–156. https://doi.org/10.1350/ijps.2013.15.2.308
Sternberg, K. J., Lamb, M. E., Hershkowitz, I., Esplin, P. W., Redlich, A., & Sunshine, N. (1996). The relation between investigative utterance types and the informativeness of child witnesses. Journal of Applied Developmental Psychology, 17(3), 439–451. https://doi.org/10.1016/S0193-3973(96)90036-2
Szojka, Z. A., Yashraj, S., & Lyon, T. D. (2025). Automated question type coding of forensic interviews and trial testimony in child sexual abuse cases. Law and Human Behavior, 49(2), 163–172. https://doi.org/10.1037/lhb0000590
Legal Authorities
Crawford v. Washington, 541 U.S. 36 (2004).
Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993).
Idaho v. Wright, 497 U.S. 805 (1990).
Kumho Tire Co. v. Carmichael, 526 U.S. 137 (1999).
State v. Lindsey, 149 Ariz. 472, 720 P.2d 73 (1986).
State v. Michaels, 136 N.J. 299, 642 A.2d 1372 (1994).
State v. Moran, 151 Ariz. 378, 728 P.2d 248 (1986).
State v. Salazar-Mercado, 234 Ariz. 590, 325 P.3d 996 (2014).
United States v. Brooks, 64 M.J. 325 (C.A.A.F. 2007).
United States v. Houser, 36 M.J. 392 (C.M.A. 1993).
Fed. R. Evid. 702.
Mil. R. Evid. 702, 1102.
Ariz. R. Evid. 702.
Ariz. Sup. Ct. Order R-23-0004 (amending Ariz. R. Evid. 702, effective Jan. 1, 2024).
Related topic papers: Suggestibility of Children’s Memory · Memory Development · Source Monitoring · External Sources of Memory Distortion · Memory Expert Testimony
Related topics from Arizona Forensic Services: Child Abuse Allegations · Courts-Martial