In her Keynote Article, Johanne Paradis does a service for the research and the clinical communities with this comprehensive review of the literature encompassing bilingualism and specific language impairment (SLI). Her work and the work of her colleagues for more than a decade have been persistent in bringing these two threads together: using bilingual (BL) data to speak to theoretical issues and using research findings to inform clinical practice. I am not alone in my appreciation (and admiration) of her many contributions, which are nicely pulled together here and placed in better perspective than in the several articles that have presented much of the work in isolation.
My Commentary focuses on the clinical implications section of the Keynote, which I find both enlightening and disturbing. Theories of SLI, Paradis notes, were developed based on observations of monolingual (ML) children. To assess their generality, we must ask how consistent those perspectives are with data from BLs. Paradis' starting point for her Keynote is the similarity of test scores and acquisition patterns between typically developing BL (TD-BL) and ML children with SLI (SLI-ML). Throughout she emphasizes the very important issue of how different “BL” individuals and groups are from each other: BL is a very heterogeneous term; furthermore, BL profiles change over time. It is not surprising that conflicting results emerge from studies of BLs in different social circumstances and at different stages of bilingualism. We cannot, and should not, discuss BLs without specifying the language history and dominance of the sample. Paradis is emphatic in making this point and in following the principle.
DOUBLE DELAY?
Much is made of the fact that in many reports TD-BL and (S)LI-ML look alike to the uninformed eye; but if one waits a few years, this temporary overlap begins to disappear. Paradis' work shows that, after approximately 3 years, TD-BL will start to score closer to the average range for TD-ML, and about that time, their acquisition sequence and error patterns will also diverge from those of SLI-ML. Before that time, (S)LI-BL may score lower than SLI-ML, prompting the label “doubly disadvantaged.”
I am predisposed to Paradis' argument that one cannot reasonably compare language-impaired BL children to language-impaired ML children on a given set of structures to evaluate “double delay” until TD-BL children have shown us that children of their language background could reasonably be expected to have achieved the critical mass of input necessary to learn the structures under consideration. In Kohnert's (Reference Kohnert and Goldstein2004) terms, one must distinguish when the reasons for incomplete acquisition are internal to the child, such as a processing deficit, as opposed to when they are external to the child, which is attributable to factors beyond the child's control, as in the typical BL circumstance of reduced input. When we are trying to identify a child with poor language function, it is poor logic to confuse internal and external factors.
However, in most reports, observers have focused on evaluating outcomes, not causes, and so the internal and external distinction plays no role. There is a very straightforward sense in which what Paradis argues as not being double delay is in fact just that: BOTH internal AND external. Paradis reports Leonard's (Reference Leonard1998) proposal that SLI-ML children's poor processing abilities lead to difficulty representing vulnerable structures. Weak representations necessitate a greater amount of exposure for the same amount of learning and thus are “functionally equivalent to infrequent exposure.” Turning this around, one could say that the potentially reduced exposure to a given structure seen in BLs is “functionally equivalent” to poor processing abilities found in SLI. Thus, the BL child with SLI has both the actual processing deficit AND its functional equivalent.
Paradis' long discussion involving effect sizes argues that there is no double delay in the data Steenge (Reference Steenge2006) reports as an example of “additional disadvantage.” However, Paradis' figure 4, based on Steenge's data, presents a perfect illustration of double delay by adding a time dimension to the hypothetical illustrations Paradis creates in figure 3. The slope of the TD-BL line is the steepest of the four lines, and we might extrapolate that in 6 or 7 more years the two lines depicting typical development (ML and BL) will intersect. However, the line for language impaired BL learners is flat. If we extrapolate from its slope, those children appear unable to ever catch up to language impaired ML. (Luckily, we are not warranted to make the extrapolation, but the potential inference rather belies the optimistic possibility of “bilingualism as therapy.”)
WHAT IS THE CLINICIAN TO DO IN THE INTERIM?
Paradis' focus is on achieving parity with ML performance as embodied in ML norms. At least 10 times in her Keynote, she asks the question in some form: “How much exposure is sufficient to meaningfully compare sequential bilinguals to monolingual?” or “How and when [can] bilingual language development be referenced to monolingual development”; or she reports some instance of BLs catching up to MLs. Implicit in those questions and reports is the uncomfortable underlying assumption that BL is somehow defective until TD-BL achieves parity with TD-ML and diverges from language impaired ML. I think this is fundamentally the wrong perspective.
BYPASSING ML NORMS
Why aim at ML norms? The same important line of inquiry that Paradis pursues that attempts to reconcile BL performance and ML norms points the way to BL benchmarks, if not norms, that is, BL development referenced to the development of other BLs. The BL profile (Oller, Pearson, & Cobo-Lewis, Reference Oller, Pearson and Cobo-Lewis2007) appears asynchronous when referenced to ML development, but the BL profile has utility in its own right. Paradis knows this very well. She explicitly states, “[the profile effect] suggests that norms should be collected for monolingual and bilingual separately, and [they] need to be organized in function of both chronological age and target language exposure time.” However, in the rush to show parity with ML norms, that message gets lost.
Paradis' work brings us in 6-month steps to almost 3 years of exposure for a child introduced to English at around age 5. However, when does the “target state” arrive? Other estimates suggest a longer period of catching up to MLs on ML norms: notably, Cummins (Reference Cummins1979) at 5–7 years or Collier and Thomas (Reference Collier and Thomas2004) at 5 years, depending on the program. Kohnert and Bates (Reference Kohnert and Bates2002) do not report norms, but they show a full shift of dominance after 10.3 years. If there is no point of intersection between ML and BL development, one might find a point where the BL curve levels off and is no longer rising relative to the ML.
Rather than wait Paradis' 3 years (or Cummins’ 7 years) before attempting to identify dual language children with language impairment, one needs to describe expected patterns of simultaneous BL development and expected patterns of sequential BL development by age AND by exposure time (and perhaps condition of exposure; as, e.g., Kohnert, Reference Kohnert and Goldstein2004, does). Because one would need such benchmarks for about 3 years at least according to Paradis' estimates, why not just use them all the time?
Paradis points out in the Keynote that bilingualism is most often not a choice. I would counter that it is the only reasonable choice. The choice to be ML is out of step with global politics and economics and with children's lives in the 21st century. We can no longer afford an ML frame of reference. The typical ML may be a reasonable reference for other MLs, but it is a faulty guide for other groups. Similarly, BL norms would not be useful for MLs.
In institutional research at the University of Miami (Pearson, Reference Pearson1993), I found college students, many of them raised and schooled in the United States, still functioning below the ML expectation on the college entrance exam to the point of averaging below the cutoff point for admission. Had the admissions officials paid attention to those scores, they would not have admitted most of those BL students. They would never have learned that these low-scorers performed as well as their ML peers in the very domain the test was supposed to predict: the BLs’ average grade point average after four semesters was equivalent to, or slightly higher than, the ML average. In a regression analysis, a score on the standardized entrance test lower than the cutoff predicted competent performance. MLs needed a higher score for the same prediction. Granted, these high-achieving high school graduates are far above the level to define SLI, but this is an easy setting in which to see that the test is not the goal; it is just a pointer to a goal. BL norms should go straight to the BL behaviors and bypass the ML norm. We need to determine what score a BL child needs on a given test to predict good performance in the domain of interest, or we need a different test.
I realize that in the real world there is utility for ill-equipped speech–language therapists to know when they can trust a diagnosis on the basis of the only tests they have: those normed on MLs. Paradis' data and logic improve their practice. However, we must remember that these tests are not ends in themselves. They are indicators of satisfactory academic progress, narrative understanding, or school readiness, to name a few uses of standardized testing. Even assessments called “inventories” are not exhaustive beyond a child's first few words. After that, they are samples of the individual's behavior compared to an appropriate group's behavior on the same tasks.
BL norms will be no more applicable to MLs than the reverse. In the college admissions example above, if a college based admission for MLs on the BL prediction, they would admit unqualified MLs. Admissions officers would demand separate norms for MLs. We cannot cease our demand until we have them.
You have
Access
In her Keynote Article, Johanne Paradis does a service for the research and the clinical communities with this comprehensive review of the literature encompassing bilingualism and specific language impairment (SLI). Her work and the work of her colleagues for more than a decade have been persistent in bringing these two threads together: using bilingual (BL) data to speak to theoretical issues and using research findings to inform clinical practice. I am not alone in my appreciation (and admiration) of her many contributions, which are nicely pulled together here and placed in better perspective than in the several articles that have presented much of the work in isolation.
My Commentary focuses on the clinical implications section of the Keynote, which I find both enlightening and disturbing. Theories of SLI, Paradis notes, were developed based on observations of monolingual (ML) children. To assess their generality, we must ask how consistent those perspectives are with data from BLs. Paradis' starting point for her Keynote is the similarity of test scores and acquisition patterns between typically developing BL (TD-BL) and ML children with SLI (SLI-ML). Throughout she emphasizes the very important issue of how different “BL” individuals and groups are from each other: BL is a very heterogeneous term; furthermore, BL profiles change over time. It is not surprising that conflicting results emerge from studies of BLs in different social circumstances and at different stages of bilingualism. We cannot, and should not, discuss BLs without specifying the language history and dominance of the sample. Paradis is emphatic in making this point and in following the principle.
DOUBLE DELAY?
Much is made of the fact that in many reports TD-BL and (S)LI-ML look alike to the uninformed eye; but if one waits a few years, this temporary overlap begins to disappear. Paradis' work shows that, after approximately 3 years, TD-BL will start to score closer to the average range for TD-ML, and about that time, their acquisition sequence and error patterns will also diverge from those of SLI-ML. Before that time, (S)LI-BL may score lower than SLI-ML, prompting the label “doubly disadvantaged.”
I am predisposed to Paradis' argument that one cannot reasonably compare language-impaired BL children to language-impaired ML children on a given set of structures to evaluate “double delay” until TD-BL children have shown us that children of their language background could reasonably be expected to have achieved the critical mass of input necessary to learn the structures under consideration. In Kohnert's (Reference Kohnert and Goldstein2004) terms, one must distinguish when the reasons for incomplete acquisition are internal to the child, such as a processing deficit, as opposed to when they are external to the child, which is attributable to factors beyond the child's control, as in the typical BL circumstance of reduced input. When we are trying to identify a child with poor language function, it is poor logic to confuse internal and external factors.
However, in most reports, observers have focused on evaluating outcomes, not causes, and so the internal and external distinction plays no role. There is a very straightforward sense in which what Paradis argues as not being double delay is in fact just that: BOTH internal AND external. Paradis reports Leonard's (Reference Leonard1998) proposal that SLI-ML children's poor processing abilities lead to difficulty representing vulnerable structures. Weak representations necessitate a greater amount of exposure for the same amount of learning and thus are “functionally equivalent to infrequent exposure.” Turning this around, one could say that the potentially reduced exposure to a given structure seen in BLs is “functionally equivalent” to poor processing abilities found in SLI. Thus, the BL child with SLI has both the actual processing deficit AND its functional equivalent.
Paradis' long discussion involving effect sizes argues that there is no double delay in the data Steenge (Reference Steenge2006) reports as an example of “additional disadvantage.” However, Paradis' figure 4, based on Steenge's data, presents a perfect illustration of double delay by adding a time dimension to the hypothetical illustrations Paradis creates in figure 3. The slope of the TD-BL line is the steepest of the four lines, and we might extrapolate that in 6 or 7 more years the two lines depicting typical development (ML and BL) will intersect. However, the line for language impaired BL learners is flat. If we extrapolate from its slope, those children appear unable to ever catch up to language impaired ML. (Luckily, we are not warranted to make the extrapolation, but the potential inference rather belies the optimistic possibility of “bilingualism as therapy.”)
WHAT IS THE CLINICIAN TO DO IN THE INTERIM?
Paradis' focus is on achieving parity with ML performance as embodied in ML norms. At least 10 times in her Keynote, she asks the question in some form: “How much exposure is sufficient to meaningfully compare sequential bilinguals to monolingual?” or “How and when [can] bilingual language development be referenced to monolingual development”; or she reports some instance of BLs catching up to MLs. Implicit in those questions and reports is the uncomfortable underlying assumption that BL is somehow defective until TD-BL achieves parity with TD-ML and diverges from language impaired ML. I think this is fundamentally the wrong perspective.
BYPASSING ML NORMS
Why aim at ML norms? The same important line of inquiry that Paradis pursues that attempts to reconcile BL performance and ML norms points the way to BL benchmarks, if not norms, that is, BL development referenced to the development of other BLs. The BL profile (Oller, Pearson, & Cobo-Lewis, Reference Oller, Pearson and Cobo-Lewis2007) appears asynchronous when referenced to ML development, but the BL profile has utility in its own right. Paradis knows this very well. She explicitly states, “[the profile effect] suggests that norms should be collected for monolingual and bilingual separately, and [they] need to be organized in function of both chronological age and target language exposure time.” However, in the rush to show parity with ML norms, that message gets lost.
Paradis' work brings us in 6-month steps to almost 3 years of exposure for a child introduced to English at around age 5. However, when does the “target state” arrive? Other estimates suggest a longer period of catching up to MLs on ML norms: notably, Cummins (Reference Cummins1979) at 5–7 years or Collier and Thomas (Reference Collier and Thomas2004) at 5 years, depending on the program. Kohnert and Bates (Reference Kohnert and Bates2002) do not report norms, but they show a full shift of dominance after 10.3 years. If there is no point of intersection between ML and BL development, one might find a point where the BL curve levels off and is no longer rising relative to the ML.
Rather than wait Paradis' 3 years (or Cummins’ 7 years) before attempting to identify dual language children with language impairment, one needs to describe expected patterns of simultaneous BL development and expected patterns of sequential BL development by age AND by exposure time (and perhaps condition of exposure; as, e.g., Kohnert, Reference Kohnert and Goldstein2004, does). Because one would need such benchmarks for about 3 years at least according to Paradis' estimates, why not just use them all the time?
Paradis points out in the Keynote that bilingualism is most often not a choice. I would counter that it is the only reasonable choice. The choice to be ML is out of step with global politics and economics and with children's lives in the 21st century. We can no longer afford an ML frame of reference. The typical ML may be a reasonable reference for other MLs, but it is a faulty guide for other groups. Similarly, BL norms would not be useful for MLs.
In institutional research at the University of Miami (Pearson, Reference Pearson1993), I found college students, many of them raised and schooled in the United States, still functioning below the ML expectation on the college entrance exam to the point of averaging below the cutoff point for admission. Had the admissions officials paid attention to those scores, they would not have admitted most of those BL students. They would never have learned that these low-scorers performed as well as their ML peers in the very domain the test was supposed to predict: the BLs’ average grade point average after four semesters was equivalent to, or slightly higher than, the ML average. In a regression analysis, a score on the standardized entrance test lower than the cutoff predicted competent performance. MLs needed a higher score for the same prediction. Granted, these high-achieving high school graduates are far above the level to define SLI, but this is an easy setting in which to see that the test is not the goal; it is just a pointer to a goal. BL norms should go straight to the BL behaviors and bypass the ML norm. We need to determine what score a BL child needs on a given test to predict good performance in the domain of interest, or we need a different test.
I realize that in the real world there is utility for ill-equipped speech–language therapists to know when they can trust a diagnosis on the basis of the only tests they have: those normed on MLs. Paradis' data and logic improve their practice. However, we must remember that these tests are not ends in themselves. They are indicators of satisfactory academic progress, narrative understanding, or school readiness, to name a few uses of standardized testing. Even assessments called “inventories” are not exhaustive beyond a child's first few words. After that, they are samples of the individual's behavior compared to an appropriate group's behavior on the same tasks.
BL norms will be no more applicable to MLs than the reverse. In the college admissions example above, if a college based admission for MLs on the BL prediction, they would admit unqualified MLs. Admissions officers would demand separate norms for MLs. We cannot cease our demand until we have them.