Quasi-experimental designs came about because of: 1) difficulty of applying the classical natural science method to the social

sciences 2) overemphasis on theory testing and development 3) high cost of classic natural science methods 4) development of new statistical tools that allowed for statistical control Comparison Group Pre-test/Post-test Design In a quasi-experimental design, the research substitutes statistical "controls" for the absence of physical control of the experimental situation. The most common quasi-experimental design is the Comparison Group Pre-test/Post-test Design. This design is the same as the classic controlled experimental design except that the subjects cannot be randomly assigned to either the experimental or the control group, or the researcher cannot control which group will get the treatment. In other words, participants do not all have the same chance of being in the control or the experimental groups, or of receiving or not receiving the treatment. Quasi-Experimental Design A quasi-experimental design is one that looks a bit like an experimental design but lacks the key ingredient -- random assignment. My mentor, Don Campbell, often referred to them as "queasy" experiments because they give the experimental purists a queasy feeling. With respect to internal validity, they often appear to be inferior to randomized experiments. But there is something compelling about these designs; taken as a group, they are easily more frequently implemented than their randomized cousins. I'm not going to try to cover the quasi-experimental designs comprehensively. Instead, I'll present two of the classic quasi-experimental designs in some detail and show how we analyze them. Probably the most commonly used quasiexperimental design (and it may be the most commonly used of all designs) is the nonequivalent groups design. In its simplest form it requires a pretest and posttest for a treated and comparison group. It's identical to the Analysis of Covariance design except that the groups are not created through random assignment. You will see that the lack of random assignment, and the potential nonequivalence between the groups, complicates the statistical analysis of the nonequivalent groups design. The second design I'll focus on is the regression-discontinuity design. I'm not including it just because I did my dissertation on it and wrote a book about it (although those were certainly factors weighing in its favor!). I include it because I believe it is an important and often misunderstood alternative to randomized experiments because its distinguishing characteristic -- assignment to treatment using a cutoff score on a pretreatment variable -- allows us to assign to the program those who need or deserve it most. At first glance, the regression discontinuity design strikes most people as biased because of regression to the mean. After all, we're assigning low scorers to one group and high scorers to the other. In the discussion of the statistical analysis of the regression discontinuity design, I'll show you why this isn't the case. Finally, I'll briefly present an assortment of other quasi-experiments that have specific applicability or noteworthy features, including the Proxy Pretest Design, Double Pretest Design, Nonequivalent Dependent Variables Design, Pattern

Matching Design, and the Regression Point Displacement design. I had the distinct honor of co-authoring a paper with Donald T. Campbell that first described the Regression Point Displacement Design. At the time of his death in Spring 1996, we had gone through about five drafts each over a five year period. The paper (click here for the entire paper) includes numerous examples of this newest of quasi-experiments, and provides a detailed description of the statistical analysis of the regression point displacement design. Quasi-Experimental Designs An experimental design is one in which participants are randomly assigned to levels of the independent variable. As we saw in our discussion of random assignment, experimental designs are preferred when the goal is to make causeand-effect conclusions because they reduce the risk that the results could be due to a confounding variable. However, there are some circumstances in which it is not practical or ethical to randomly assign participants. We cannot randomly assign participants to be male or female, and we should not randomly assign pregnant women to consume different quantities of alcohol. Random assignment is not an option for these variables, but we still want to make cause-and-effect inferences about gender and about alcohol. Is there anything we can do? Yes. Quasi-experiments are designed to maximize internal validity (confidence in cause-and-effect conclusions) despite being unable to randomly assign. At the outset, it is important to know that quasi-experiments tend to have lower internal validity than true experiments. In this reading, well discuss four quasiexperimental approaches: 1) matching, 2) mixed designs, 3) singlesubject designs, and 4) developmental designs. The major threats to quasi-experimental designs are confounding variables: variables other than the independent variable that (a) tend to co-vary with the independent variable and b) are plausible causes of the dependent variable. Quasi-experiments are designed to reduce confounding variables as much as possible, given that random assignment is not available. For example, say that you are investigating the effects of bedtime stories on childrens reading performance. Your hypothesis is that children whose parents read to them will score higher on a reading test than children whose parents do not read to them. Since many parents feel strongly about reading to their children and are unlikely to give it up, random assignment will not be appropriate for this study. What confounding variables are likely in this situation? In other words, how might children who receive bedtime stories be different from children who do not receive bedtime stories? Perhaps their parents have a greater love of reading themselves, and either through genetics or environment the parents are influencing their children to read. Perhaps children who do not receive bedtime stories tend to have parents who work longer hours and have less time to spend with their children. Either of these factors parents love of reading or total contact time with parents could be a confounding variable in the relation between bedtime stories and reading performance. How do you rule those out? The first approach we will look at is matching.

Matching Lets say that our two biggest concerns with the bedtime story study mentio ned in the last paragraph are parent reading and parent contact time. Matching begins by measuring these potential confounding variables, along with the independent and dependent variables. Thus, all the participants in our study would be measured on four variables: 1) whether they receive bedtime stories (our independent variable), 2) how much their parents read (one of the potential confounding variables), 3) average number of parent contact hours per day with the child (the other potential confounding variable), and 4) reading performance (the dependent variable). Perhaps the most common matching procedure is pairwise matching, in which you identify a pair of participants one who gets a bedtime story and one who does not who are as similar as possible on the two potential confounding variables. Ideally, each pair would be exactly the same on each matching variable, but as the number of matched variables increases, this becomes less likely 1. Once you identify your matched pairs, you conduct a paired-samples 1 It turns out that pairwise matching is relatively inefficient because it throws out a lot of perfectly good data. Much better is a method called full matching that makes it possible for a person in one condition to be matched to several similar people in the other condition. See Hansen and Klopfer (2006) and the optmatch package in R.Quasi-Experimental ttest (or a within-subjects ANOVA) on the reading performance scores. If the reading scores differ significantly, you can be confident that the difference is not due to either parent reading or parent contact hours because each pair was matched on those variables. Those two potential confounding variables have been controlled through matching. Matching is superior to random assignment in one way: it guarantees that the participants in different conditions are not different on the matched variables. The problem with matching is that it does nothing to equalize the conditions on other variables. For example, the availability of childrens books in the home may be an important cause of childrens reading performance, but because it was not measured or matched, it will remain a potential confounding variable. Mixed Designs Mixed designs are designs with at least one between-subjects variable (such as treatment condition) and at least one within-subjects variable (such as time: before and after). Weve already discussed mixed designs as a way to combine the advantages of between- and within-subjects variables, but it turns out that they are especially useful when random assignment is not available. Consider a study on the effect of television exposure on childrens aggressiveness. Randomly assigning children to watch different amounts of television is problematic both because it is intrusive into family life and because it may be unethical if you suspect that television makes children aggressive. For these reasons, most studies on the relation between television exposure and aggression are correlational: you measure childrens level of television exposure as it is and do not manipulate it. But these studies suffer from a problem with inferring cause and effect. If you find that exposure to television is strongly correlated with aggressiveness, it could be because 1) television makes children aggressive, 2) aggressive children have a greater preference for watching television, or

3) some third factor, like amount of parent supervision, is actually causing both television exposure and child aggressiveness. To solve this problem, Joy, Kimball, and Zabrack (1986) took advantage of an unusual circumstance: television coming to a small Canadian town. Joy and colleagues used a mixed design in which the within-subjects variable was time (before and then two years after television arrived) and the main between-subjects variable was town (the town that had no television nicknamed Notel as well as two other towns that were chosen to be very similar to Notel). Thus, their design was a 2 (time: before and after) by 3 (town: Notel and the two other towns) mixed design with repeated measures on the first factor. The researchers measured aggressive acts per minute during free play on the playground in each town, and the results are plotted in Figure 1 below Types of Snowflakes Simple Prisms A hexagonal prism is the most basic snow crystal geometry (see the Snowflake Primer). Depending on how fast the different facets grow, snow crystal prisms can appear as thin hexagonal plates, slender hexagonal columns (shaped a lot like wooden pencils), or anything in between. Simple prisms are usually so small they can barely be seen with the naked eye. The examples at right show two stubby prisms and one thin plate. Snow crystal facets are rarely perfectly flat, being more typically decorated with various indents, ridges, or other features. Stellar Plates These common snowflakes are thin, plate-like crystals with six broad arms that form a star-like shape. Their faces are often decorated with amazingly elaborate and symmetrical markings. Plate-like snowflakes form when the temperature is near -2 C (28 F) or near -15 C (5 F), as dictated by the snow crystal morphology diagram. Sectored Plates Stellar plates often show distinctive ridges that point to the corners between adjacent prism facets. When these ridges are especially prominent, the crystals are called sectored plates. The simplest sectored plates are hexagonal crystals that are divided into six equal pieces, like the slices of a hexagonal pie. More complex specimens show prominent ridges on broad, flat branches.

Stellar Dendrites Dendritic means "tree-like", so stellar dendrites are plate-like snow crystals that

have branches and sidebranches. These are fairly large crystals, typically 2-4 mm in diameter, that are easily seen with the naked eye. Stellar dendrites are clearly the most popular snow crystal type, seen in holiday decorations everywhere. You can see these crystals for yourself quite well with just a simple magnifier. (See Snowflake Watching for more about observing snowflakes.) Fernlike Stellar Dendrites Sometimes the branches of stellar crystals have so many sidebranches they look a bit like ferns, so we call them fernlike stellar dendrites. These are the largest snow crystals, often falling to earth with diameters of 5 mm or more. In spite of their large size, these are single crystals of ice -- the water molecules are lined up from one end to the other. Some snowfalls contain almost nothing but stellar dendrites and fernlike stellar dendrites. It can make quite a sight when they collect in vast numbers, covering everything in sight. The best powder snow, where you sink to your knees while skiing, is made of stellar dendrites. These crystals can be extremely thin and light, so they make a low density snowpack. Hollow Columns Hexagonal columns often form with conical hollow regions in their ends, and such forms are called hollow columns. These crystals are small, so you need a good magnifier to see the hollow regions. Note how the two hollow regions are symmetrical in each column. Sometimes the ends grow over and enclose a pair of bubbles in the ice, as seen in the last picture on the right. Needles Needles are slender, columnar ice crystals that grow when the temperature is around -5 C (23 F). On your sleeve these snowflakes look like small bits of white hair. One of the amazing things about snow crystals is that their growth changes from thin, flat plates to long, slender needles when the temperature changes by just a few degrees. Why this happens remains something of a scientific mystery. Capped Columns These crystals first grow into stubby columns, and then they blow into a region of the clouds where the growth becomes plate-like. The result is two thin, plate-like crystals growing on the ends of an ice column. Capped columns don't appear in every snowfall, but you can find them if you look for them.

The first example at right shows three views of a capped column. The first view is from the side, showing the central column and the two plates edge-on. The other two views show the same crystal from one end, with the microscope focused separately on the two plates.

Double Plates A double plate is basically a capped column with an especially short central column. The plates are so close together that inevitably one grows out faster and shields the other from its source of water vapor. The result is one large plate connected to a much smaller one. These crystals are common -- many snowflakes that look like ordinary stellar plates are actually double plates if you look closely. The first picture at right shows a double plate from the side. The second picture shows a double plate with the microscope focused on the smaller plate. In the third picture, note the slightly out-of-focus hexagon that is about one-sixth as large as the main crystal. This hexagon is the second side of a double plate, connected to the main plate by a small axle. Split Plates and Stars These are forms of double plates, except that part of one plate grows large along with part of the other plate. The picture at right shows all eight ways to make a split star. Split plates and stars, like double plates, are common but often unnoticed. You may have to stare at these pictures a bit to see how the two distinct pieces fit together. Note how in each case the crystals are connected in the center with short axles. Triangular Crystals Plates sometimes grow as truncated triangles when the temperature is near -2 C (28 F). If the corners of the plates sprout arms, the result is an odd version of a stellar plate crystal. These crystals are relatively rare. Surprisingly, no one knows why snow crystals grow into these three-fold symmetrical shapes. (Note however that the molecular structure of triangular crystals is no different from ordinary six-sided crystals. The facet angles are all the same.) 12-Sided Snowflakes Sometimes capped columns form with a twist, a 30-degree twist to be specific. The two end-plates are both six-branched crystals, but one is rotated 30 degrees relative to the other. This is a form of crystal twinning, in which two crystals grow joined in a specific orientation. These crystals are quite rare, but sometimes a snowfall will bring quite a few. The

picture at the far right shows a 12-sider where the two halves are widely separated. Bullet Rosettes The nucleation of an ice grain sometimes yields multiple crystals all growing together at random orientations. When the different pieces grow into columns, the result is called a bullet rosette. These polycrystals often break up to leave isolated bullet-shaped crystals. Sometimes a bullet rosette can become a capped rosette, as shown in the example at the far right. Radiating Dendrites When the pieces of a polycrystal grow out into dendrites, the result is called a radiating dendrite (also called a spatial dendrite). The first example on the right shows radiating plates. The second example shows a fernlike stellar dendrite with two errant branches growing up out of the main plane of the crystal. Rimed Crystals Clouds are made of countless water droplets, and sometimes these droplets collide with and stick to snow crystals. The frozen droplets are called rime. All the different types of snow crystals can be found decorated with rime. When the coverage is especially heavy, so that the assembly looks like a tiny snowball, the result is called graupel. The first two pictures at right have relatively light rime coverage. The final example is completely covered with rime, but you can still see the six-fold symmetry of the underlying stellar crystal. Irregular Crystals The most common snow crystals by far are the irregular crystals. These are small, usually clumped together, and show little of the symmetry seen in stellar or columnar crystals. Artificial Snow Snow machines shoot a mixture of water and compressed air out of nozzles. The water comes out as fine droplets, and the air cools as it decompresses, causing the droplets to freeze. A fan blows the ice particles onto the slopes. You can see from the picture at right that artificial snow is made of frozen water droplets, with none of the elaborate structure found in real snow crystals. If you want to go outside and look at snow crystals for yourself, I recommend my new book -- Ken Libbrecht's Field Guide to Snowflakes. This book contains a much more complete list of the different snow crystal types, along with how to find them. Once you know what to look for, snowflake watching is a fascinating recreation!

Classifying Snowflakes How does one classify snowflakes? It's not so easy, because how you divide the different types is somewhat a matter of taste. There is a good analogy with breeds of dogs. The definition of different breeds is decided upon by a committee of people, and really one can make up as many breeds as one wants. And no matter how many different breeds you define, some dogs will be mixed, not belonging to any one breed. Snowflakes do come in different types, and you need to give them names if you want to talk about them. But there will never be a precise way to define the different types. I prefer the 35 types shown at the top of this page, but others have come up with alternate classifications schemes. Some of these are shown below.

International Classification System In 1951 the International Commission on Snow and Ice produced a fairly simple and widely used classification system for solid precipitation [1]. This system defines the seven principal snow crystal types as plates, stellar crystals, columns, needles, spatial dendrites, capped columns, and irregular forms. To these are added three additional types of frozen precipitation: graupel, ice pellets, and hail. This is a fine classification system, but in my opinion it's too simplistic to be very interesting.

Nakaya's Classification Physicist Ukichiro Nakaya (see Historic Snowflakes) created the first systematic classification scheme for snowflakes, in which he subdivided falling snow into 41 individual morphological types [2].