Home Curriculum Vitae Research Publications Interesting links


Autobiographical summary of my research


bulletInitial work
bulletMemory modelling
bulletThe SAM model
bulletThe REM model: extensions to implicit and semantic memory paradigms
bulletStatistical analysis
bulletNeural network models and regression analyses
bulletApplied research
bulletDecision making
bulletRecent projects
bulletRetrieval inhibition



I have been very lucky to have been involved in collaborations with a large number of excellent researchers. This started during my studies at the University of Nijmegen where I worked as a Master's student with Charles de Weert on opponent process theories for colour vision. My supervisor at the University of Nijmegen, the late Eddy Roskam, shaped my attitude towards mathematical modelling in psychology. The most important influence on my research came from my long-lasting collaboration with Richard Shiffrin. I have also been fortunate to have had a number of excellent PhD students that have now themselves become important researchers. These include José Kerstholt, René Zeelenberg, Diane Pecher, Eric-Jan Wagenmakers, Pauline Spaan and Sander Daselaar.


Initial work
My first experience in research was as a research assistant in the lab of my older brother Wijnand Raaijmakers at the Department for Physiological and Comparative Psychology at the University of Nijmegen. I ran a number of aversive conditioning experiments, the aim of which was to test the then popular although controversial hypothesis that injecting brain matter from conditioned animals into the brains of naïve animals would lead to transfer of the conditioned response (some researchers had claimed to have obtained such results with aplysia). These experiments were not successful, i.e. no transfer could be shown.

During my studies for the MSc at the University of Nijmegen, I had enlisted for an advanced course in Perception. However, since the number of staff members involved in that course outnumbered the number of students, I came to work on a one-to-one basis with Charles de Weert on Colour Vision. As part of the "course" we ran an experiment to test some assumptions from the axiomatic formulation of opponent-colours theory that had then just been developed by David Krantz. This resulted in my first scientific publication (Raaijmakers & De Weert, 1975).

However, since it would be difficult for me to work as an experimenter in this field (being a protanope), I switched to the area of mathematical learning theory and became interested in mathematical models for concept learning. I wrote a Master's thesis on that topic in which I developed an extension of Chumbley's Dual-Process Markov model for concept identification (see Raaijmakers, 1981). This was also originally planned to be the topic for my PhD dissertation.

[back to top]

Memory Modelling

The SAM model for recall and recognition
Around that time I obtained a fellowship from the Netherlands Organization for Scientific Research for working for a year at Indiana University with the late Frank Restle. However, since Frank Restle wasn't around for the first part of the year, I began to work with Richard Shiffrin. This collaboration proved to be quite successful and we have continued the collaboration for more than 25 years. After some preliminary work on a model that Shiffrin had developed in the late 60s, we started around December 1977 to work on what was initially mainly a model for free recall. After the initial work showed much promise, we realized that the model that we had developed could be extended to a general framework for retrieval from long-term memory. After returning to the University of Nijmegen, I changed to topic for my PhD and after a couple of months the manuscript for the dissertation was ready (the PhD was officially granted on June 26, 1979). The basic contents of the thesis can be found in the 1980 chapter in Bower's series The Psychology of Learning and Motivation. This model subsequently became known as the SAM model (for Search of Associative Memory).

One of the major accomplishments of the SAM model in its initial years, was the explanation that it provided for the part-list cuing puzzle. This phenomenon was originally interpreted (and still is sometimes) as evidence for inhibitory processes in recall and as evidence for the hypothesis that interitem associations were not important or effective in recall. We showed that this conclusion was false since the phenomenon could be explained by the SAM model (that does rely heavily on interitem associations). After about six months of experimenting with the model we finally discovered the reason why the SAM model could explain this result. A major part of the 1981 paper in Psychological Review was devoted to this explanation. We showed that the model could not just handle the basic phenomenon but also a large number of related findings. In my opinion the account provided by SAM still stands as the most comprehensive explanation of the part-list cuing effect (see also Raaijmakers and Phaf, 1999).

Following this initial work, the SAM model was extended to cued recall (Raaijmakers & Shiffrin, 1981b) and recognition (Gillund & Shiffrin, 1984). Mensink and Raaijmakers (1988) extended the model to interference and forgetting using a contextual fluctuation model based on earlier work by Estes. This latter variant was later also used to explain spacing effects in recall (Raaijmakers, 2003).

Although the SAM model was (and still is) a quite successful model for recall, Shiffrin and his colleagues (Ratcliff, Clark & Shiffrin, 1990; Shiffrin, Ratcliff & Clark, 1990) discovered a phenomenon that could not be explained by the SAM model for recognition without making additional assumptions. They showed that in recognition there is no effect of the strength of the other items on the list although there is an effect of the number of other items. The combination of the presence of list-length effects and the simultaneous absence of list-strength effects was problematic for SAM (and not just for SAM; Shiffrin et al. showed that it created problems for most of the then existing models for recognition). In order to explain this result in SAM, it was necessary to make a differentiation assumption, i.e. the assumption that as items get better encoded they also become less confusable with the target item.

[back to top]

The REM model: extensions to implicit and semantic memory paradigms
In 1994/1995, when Richard Shiffrin spent a sabbatical year at the University of Amsterdam, he developed (together with Mark Steyvers, then a student in Amsterdam) a new model for recognition based on Bayesian considerations. In this new model, termed REM (for Retrieving Effectively from Memory), the differentiation assumption fell out as a consequence of the Bayesian approach. Shiffrin and Steyvers (1997) showed that REM provided a satisfactory solution to the problems with the SAM recognition model.

At about the same time, I had begun a research program on implicit memory and especially the storage of new associative information in semantic memory (at least that was the initial intent). In these experiments we gave subjects cycles of primed lexical decision or perceptual identification tests. During the study phase, the same prime item was always paired to the same target item. In perceptual identification both the prime item and the target item were presented very briefly (about 30 msec) in such a way that the subjects were usually unable to identify the prime item. After each trial in lexical decision or perceptual identification, a study trial was given for that pair. Using this methodology, we showed in a number of experiments (see Schrijnemakers & Raaijmakers, 1997; Pecher & Raaijmakers, 1999, 2004) that automatic associative priming could be reliably observed for newly learned associations. However, the priming effect was highly dependent on the task used during the initial study phase. That is, when during the study phase a lexical decision task was used, priming for new associations could be observed but no priming was obtained when during final testing a perceptual identification task was used (and vice versa). These results (as well as others, see Pecher et al., 1998) led to the hypothesis that the semantic or lexical system was highly sensitive to contextual aspects of recent exposures.

At around that time, Lael Schooler (a visiting postdoc), Rich Shiffrin and myself were considering ways in which the newly developed REM model could be extended to deal with lexical/semantic memory. We assumed that items as well as cues could be represented as feature vectors with elements representing contextual, perceptual (e.g. orthography), and semantic features. As a first task, we set out to model the bias results obtained Ratcliff and McKoon (1997) in a 2-alternative forced-choice perceptual identification task. After looking at a number of different possibilities, Lael Schooler came up with the simple idea that the cue vector is filled with a number of correctly perceived features (the number depending on stimulus duration). The remaining perceptual features are filled with noise. Priming effects are due to adding a small amount of context information to the lexical/semantic vectors (in his simulations, Lael added just one extra context element). This model successfully handled the results that Ratcliff and McKoon had obtained.

We are currently extending this model to deal with other results in semantic and implicit memory research (see Raaijmakers, 2005). In one project we have applied the basic REM framework to lexical decision (see Wagenmakers et al, 2004). This model is a direct extension of the REM recognition model. It assumes that lexical decisions are based on the same kind of global familiarity decisions as in recognition memory except that the feature vectors involved come from the lexical-semantic system rather than episodic memory. More generally, we believe that implicit memory effects are best understood as being due to the addition of features to the lexical-semantic trace as a result of a recent encounter with the target item.

[back to top]

Statistical analysis

Michaelis-Menten equation

I have always had a strong interest in mathematical and statistical issues. One of the first research projects that I did in this area was instigated by a question my brother Wijnand (a biochemist) asked me about some problems concerning the analysis of binding data that were fitted to the so-called Michaelis-Menten equation (an equation of the form Y=aX/(b+X)). Traditional solutions to this equation were based on rewriting this equation in a way that led to a linear regression equation (e.g. 1/Y vs 1/X). However, such equations differ from the standard linear regression situation (the error term will also be transformed in complicated ways), leading to bias in the estimates for the parameters of the original equation. Using the assumption that the error in the Michaelis-Menten equation is proportional to the true value, I derived a maximum likelihood solution, including estimates for the standard errors of the parameters (see Raaijmakers, 1987).

As a sideline of this work, I became interested in the problem of regression analyses in which not just the dependent but also the independent variable has measurement error. This problem is known in statistics as that of structural and functional relationships or the errors-in-variables problem and in econometrics as two-stage least squares. In such cases, the standard estimate for the slope parameter is biased. Based on prior work by Robertson (1974) I ran a number of simulations that showed that a correction for bias (using the bias factor derived by Robertson) led to an unbiased estimate for the slope with a standard error that was well approximated by Robertson's equation. This solution was subsequently used in an analysis of the effects of measurement error in the covariate in the analysis of covariance (see Raaijmakers & Pieters, 1985, 1987).

[back to top]

The Language-as-fixed-effect-fallacy

During my term as editor of Acta Psychologica I handled a number of papers on semantic or associative priming in (at that time usually) lexical decision. In such experiments there are usually two random factors, Subjects and Items. Although Clark (1973) had described a procedure for how to analyze such data, many of these papers used a procedure that paid lip service to Clark's recommendations but was in fact not in accordance with those recommendations. Clark had advocated a quasi-F ratio as the correct analysis and had described a procedure by which a minimum value for the quasi-F ratio could be computed. That procedure entailed the running of two separate ANOVA's (one averaging over Subjects, the other over Items) the results of which could be used to compute the min F'. The majority over papers did report the results of the two ANOVA's yet failed to combine the results into the min F' statistic. Since this occurred over and over again (even in cases where a standard ANOVA would be correct) Chris Schrijnemakers and I wrote a paper, essentially reiterating Clark's points. We also added a discussion of designs where the items are counterbalanced over the conditions and showed (with the help of a statistician friend, Frans Gremmen) that in such designs the standard analysis is indeed correct. This work was published in the Journal of Memory and Language (Raaijmakers, Schrijnemakers & Gremmen, 1999). I wrote a second paper for the Canadian Journal of Experimental Psychology that discussed some of the criticisms that had come up after the JML paper was published (Raaijmakers, 2003).

[back to top]

Neural network models and regression analyses

During the time that I was working in an applied research institute, I became interested in the possibilities of neural network models (in particular backpropagation networks) for predicting complex relations. I obtained a grant for investigating whether and how physical descriptive measures of faces (e.g. the distance between the eyes, the mouth to face ratio etc) could be predicted using neural network models from subjective descriptive measures (such as whether the face was narrow or broad, whether the forehead was narrow or broad, etc). The idea behind this project was that such a network might be used in police investigations to search a database based on a verbal description of the perpetrator or even in a future automated system for face composites.
In order to establish some baseline, I used linear regression techniques to compare against the network results. The results were disappointing. Although a large network could predict the features of the faces on which it had been trained very well, the network models were quite inferior when it came to generalization to new faces or in the presence of input noise. It turned out that this was mostly due to the very large number of to be estimated parameters in the network models. However, even when the number of parameters was about equal to that in the regression models, the network models were not better. These results also lead one to question the apparent success of other neural network models (e.g. the NetTalk model). The results were written up (the manuscript can be found here) but it was not accepted for publication in a journal issue devoted specifically to the use of network models in such situations, because the reviewers claimed that my results were specific to this area that just happened to involve linear relations rather than the nonlinear ones where a network model would have been superior. However, this still leaves one to wonder whether many of the other published successes of neural network models were not more apparent than real.

[back to top]

Applied research

Decision making

During the time that I worked at the TNO Human Factors Research Institute, I did a number of projects on decision making and related problems. A major part of that work had to do with problems in what is called "Command and Control" (C2), i.e., situations in which a group of people is responsible for devising a plan of action based on information gathered from many sources and then has to monitor the action and take appropriate steps if the situation changes. Research in this area was (and still is) a bit messy. One of the reasons may be the fact that quite different situations are treated as though they involve similar decision problems. Peter Essens and I developed a framework for analyzing decision problems in C2 environments in which we made a distinction between situations requiring a rapid response and situations in which there was time for a more extended decision process. We claimed that the former type required pre-defined decision rules in the form of production rules: IF a certain situation arises, THEN act in the following way (see the following lecture that I gave on a symposium, unfortunately in Dutch). We never thought of publishing our views since we did not think that our views were very different from those in the existing literature (this was in 1988). Later developments showed that we were mistaken since a more commercially oriented group around Klein & Associates later formulated a position that was quite similar but received a lot of attention in the decision community (the so-called Naturalistic Decision Making approach). Klein and his followers held that decisions in real-life often involved what they called "recognition-primed decision making" which is basically the same as the production rule approach that we had described but didn't think was novel enough to deserve publication.

Other projects included a review of the literarture on the effects of fatigue and loss of sleep on decision making (see Raaijmakers, 1990), and a observation study on the use of various information sources by wheather forecasters (Raaijmakers, 1988).

[back to top]


I have been involved in a number of research projects on dementia. In one of the initial projects (with Marja Abbenhuis and Wijnand Raaijmakers) the starting point was the differentiation between dementia en depression in older patients. The idea at that time was that the poor memory performance of depressed patients was due to a lack of controlled or elaborative processing while the poor performance of demented patients was due to a basic memory deficit. Hence, if a test would be used that mainly involved automatic retrieval processes, the demented patients would still show a deficit while the depressed patients should perform normally. As a measure that mainly relied on automatic processes we took the repetiton priming effect in perceptual identification (at that time an effect that was thought to reflect the contribution of episodic memory). It quickly became evident that that choice wasn't a very good one since evidence became available that amnesic patients were performing remarkably well on such implicit memory measures. In the remainder of that project we focused on the conditions in which demented patients and aged persons performed did or did not show a repetition priming effect. One result that we obtained was that performance of aged persons was much poorer when the feedback on the correct response was auditory rather than visual: older persons did not profit as much from auditory feedback as younger adults.

In a more recent project (with Pauline Spaan and Cees Jonker) we ran a large longitudinal study in older patients with and without memory complaints using a variety of memory measures (episodic, semantic and implicit) in order to determine which measures best predicted the occurrence of dementia two years after the intial memory assessment. The tests that best predicted dementia were a paired-associate learning test that included semantically related word pairs, priming in perceptual identification, and a visual association test (interassociate pictures of common objects). We also found that tests that differentiate demented from nondemented subjects at the time of the diagnosis (i.e., after dementia is already evident) are not necessarily the same as the tests that best predict dementia two years before the diagnosis can be made.

I have also been involved in a project with neuroscientists from the Free University of Amsterdam in which neuroimaging measures (fMRI) were used to determine whether there were differences between younger and older participants in the activity in specific memory areas during encoding and/or retrieval of information. This was the PhD project of Sander Daselaar. We obtained evidence that the Medial Temporal Lobe (MTL) was involved in both the encoding and the retrieval of information. In this project we used both episodic memory tasks as well as semantic, procedural and implicit memory tasks (see Daselaar et al., 2001, 2002, 2003a, 2003b, 2003c).

[back to top]

New projects

Retrieval inhibition

I have recently become interested in the inhibition effects observed in the so-called retrieval practice paradigm. These effects have been claimed to be evidence for inhibition or represssion in human memory. We are exploring the hypothesis that the REM model can explain such results without making use of such inhibition concepts. We have also ran a number of experiments showing that the retrieval inhibition effect is not larger for strong items asa claimed by the inhibition account (see Jakab & Raaijmakers, 2009).

[back to top]